DETAILED ACTION
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 1-6, 9-15, and 19-20 are rejected under 35 U.S.C. 102 (a)(2) as being anticipated over LAD (US 20220351373 A1, Date Filed: 2022-03-18).
Re Claim 1, LAD discloses a method (see LAD: e.g., Fig. 1, and, -- a deep learning algorithm with the ability to provide a high-performance classifier to predict either the presence of geographic atrophy (GA), or the likelihood of progression from intermediate age-related macular degeneration to GA. The system can also be used for broader applications outside of eye disease.--, in abstract) comprising:
receiving an optical coherence tomography (OCT) volume image of a retina of a subject (see LAD: e.g., --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]);
generating, via a deep learning model, an output using the OCT volume image in which the output indicates whether nascent geographic atrophy is detected (see LAD: e.g., --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]); and
generating a map output for the deep learning model using a saliency mapping algorithm, wherein the map output indicates a level of contribution of a set of regions in the OCT volume image to the output generated by the deep learning model (see LAD: e.g., -- [0110] A representative example of GA is presented in FIG. 5. While the central scans have a high predictive value or probability (p) and are characterized by red dots concentrated in the lesional and perilesional areas, the peripheral retinal scans have very low p values and diffusely distributed red dots in the attention maps. In eyes with intermediate AMD that will convert to GA in one year, the red dots are mainly concentrated in large drusen or drusenoid PEDs and underlying choriocapillaris and choroid, hyperreflective foci, and areas of nascent GA or incomplete RPE and outer retinal atrophy (iRORA). To a lesser extent, the attention maps also mark the neurosensory retina overlying drusen (FIG. 4B). A representative case example of intermediate AMD that preceded GA by one year can be found in FIG. 5. In eyes with intermediate AMD that will not progress to GA, the attention maps are composed of red dots diffusely present in a large area of drusen and neurosensory retina overlying them; the probability value p is low (FIG. 4C).--, in [0110];
--trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]; and, -- Medical Imaging with Deep Learning. PMLR, pp. 721-732). There are four radiologists annotates each patch with both lesion label and lesion mask. A patch in the dataset is labeled as positive if more than two (i.e., ≥3) radiologists have annotated presence of a lesion, otherwise negative…. As all patches are centered on a lesion, we randomly shift masks used in the positive contrast to reduce overlaps between masks of positive and negative contrasts.--, in [0188]).
Re Claim 2, LAD further discloses wherein the saliency mapping algorithm comprises a gradient- weighted class activation mapping (GradCAM) algorithm and wherein the map output visually indicates the level of contribution of the set of regions in the OCT volume image to the output generated by the deep learning model (see LAD: e.g., -- saliency mapping schemes are considered as comparators for the proposed approach: (i) Gradient: standard gradient-based salience mapping; (ii) Grad-CAM (Selvaraju R, et al. (2017) Proceedings of the IEEE International Conference on Computer Vision, pp. 618-626): gradient-weighted class activation mapping; (iii) LRP (Bach S, et al. (2015) PloS One. 10(7):e0130140): layer-wise relevance propagation and its variants. CUB: Bird classification in the wild. In this task we want to qualitatively and quantitatively compare the causal relevance of saliency maps generated by WBP and its competitors.--, in [0154], and, -- [0172] Activation Based Methods. Methods under this category (such as CAM, Grad-CAM, guided Grad-CAM, Grad-CAM++) use a linear combination of class activation maps from convolutional layers to derive a saliency map--, in [0172]; and also see: --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]).
Re Claim 3, LAD further discloses the OCT volume image comprises a plurality of OCT slice images that are two-dimensional, and generating an evaluation recommendation based on at least one of the output or the map output, wherein the evaluation recommendation identifies a subset of the plurality of OCT slice images for further review (see LAD: e.g., Fig. 1, and, -- to find out the most important slices for GA diagnosis. The get final predicted probability of GA (GA score) for an image x at inference time--, in [0185], also see: --[0003] Disclosed herein is a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0003]);
Re Claim 4, LAD further discloses wherein the subset includes fewer than 5% of the plurality of OCT slice images (see LAD: e.g., Fig. 1, and, -- to find out the most important slices for GA diagnosis. The get final predicted probability of GA (GA score) for an image x at inference time--, in [0185]).
Re Claim 5, LAD further discloses displaying the map output, wherein the map output comprises a saliency map overlaid on an individual OCT slice image of the OCT volume image and a bounding box around at least one region of the set of regions (see LAD: e.g., Fig. 1, “masking” read on overlay, -- in FIG. 1, to maximize the weighted binary cross-entropy loss, i.e., L(y, f (x)), the likelihood that scans from SD-OCT inputs, x, were correctly assigned (prognosticated) to either the GA or control groups, y, in the assessment of the upcoming year, while encouraging that i) regions masked-out by the saliency maps, x*, which we call negative contrasts, were not informative of GA, and ii) unmasked input scans, x, and regions masked-out by a randomly assigned saliency map, x′, did not affect the model's ability to predict GA.--, in [0101]; and, -- in FIG. 5. While the central scans have a high predictive value or probability (p) and are characterized by red dots concentrated in the lesional and perilesional areas, the peripheral retinal scans have very low p values and diffusely distributed red dots in the attention maps. In eyes with intermediate AMD that will convert to GA in one year, the red dots are mainly concentrated in large drusen or drusenoid PEDs and underlying choriocapillaris and choroid, hyperreflective foci, and areas of nascent GA or incomplete RPE and outer retinal atrophy (iRORA). To a lesser extent, the attention maps also mark the neurosensory retina overlying drusen (FIG. 4B). A representative case example of intermediate AMD that preceded GA by one year can be found in FIG. 5. In eyes with intermediate AMD that will not progress to GA, the attention maps are composed of red dots diffusely present in a large area of drusen and neurosensory retina overlying them; the probability value p is low (FIG. 4C).--, in [0110]
-- For WBP, we also tested the bounding box variant (see the Appendix for details). In Table 6, we see consistent performance gains in AUC score via incorporating PPI training (from 0.877 to 0.937), accompanied by the reductions in model variation evaluated by the standard deviations of AUC from the five-fold cross-validation. The gains are most significant when using our WBP for saliency mapping. We further compare the saliency maps generated by these different combinations. We see that without the additional supervision from PPI, competing solutions like Grad, GradCAM and LRP sometimes yield non-sensible saliency maps (attending to image corners). Overall, PPI encourages more concentrated and less noisy saliency maps. Also, different PPI-based saliency maps agree with each other to a larger extent. Our findings are also verified by experts (co-authors, who are ophthalmologists specializing in GA) confirming that the PPI-based saliency maps are clinically relevant by focusing on retinal layers likely to contain abnormalities or lesions.--, in [0142], and [0155], and, -- In FIG. 12, we compare saliency maps generated by Grad-CAM, WBP, WBP (box) to the ground truth lesion masks from expert annotations. Note that we have only supplied patch-label labels during training, not the pixel-level expert segmentation masks, which constitute a challenging task of weakly-supervised image segmentation. In line with the observations from the GA experiment, our PPI-training enhanced WBP saliency maps are mostly consistent with the expert segmentations.--, in [0161], and, -- [0165] Details on Causal Masking. In this work, we consider three types of causal masking: (i) the point-wise soft causal masking defined by Equation (2) in the main text, (ii) hard masking, and (iii) box masking. For the hard masking, for each image, we keep points with WBP weight larger than k times of the standard deviation of WBP weights of the whole image. We test k from 1 to 7 and achieve similar results. As the model performs slightly better when k=7, we set k as 7, for all experiments. For the box masking, we use the center of mass for these kept points as the center to draw a box. The height and width of this box is defined as center.sub.h/w±1.2 std.sub.h/w. In this way at least 90% of filtered points are contained in the box. For the soft masking, we set ω to 100 and σ to 0.25. We have also experimented with image-adaptive thresholds instead of a fixed σ for all inputs, i.e., set the threshold as mean value plus k times of the standard deviation of WBP weights of the whole image. The experiment comparison of these masking methods mention above is conducted on LIDC dataset. We repeat the experiments a few times and the results are consistent.--, in [0165]).
Re Claim 6, LAD further discloses wherein the identifying comprises:
identifying a potential biomarker region in association with a region of the set of regions as being associated with the nascent geographic atrophy (see LAD: e.g., --Patients with intermediate AMD (iAMD) are at increased risk for development of NVAMD or GA, and may be targeted in future clinical trials with the goal of preventing onset of advanced disease. Our long-term goal is to understand the mechanisms for progression to late stages of AMD and to develop predictive biomarkers that will facilitate clinical trials of iAMD and GA and timely standard of care treatment for NVAMD for improved clinical outcomes….Thirdly, AI can help identify candidate drug targets for future clinical trials if predictive features can be identified and linked to mechanisms of action. While a few DL models based on multimodal imaging have been developed with various degrees of success and performance characteristics, there is a need for very high-performance algorithms that can identify specific SD-OCT biomarkers of GA or NVAMD, essentially opening the “black box” of typical DL models.--, in [0086]-[0087], and, --Predictive models such as this may facilitate personalized prediction of AMD progression to inform standard of care treatments and clinical trial enrollment. There is a need for development of high performance classifiers based on DL algorithms predicting progression of GA independent of human graders based on longitudinal SD OCT datasets. Ideally, the DL model should help identify the specific SD-OCT features or biomarkers that can collectively increase the probability of new progression from intermediate AMD to the severe stage of nonexudative AMD, essentially opening the “black box” associated with such artificial intelligence algorithms.--, in [0117], [0191]);
generating a scoring metric for the potential biomarker region (see LAD: e.g., --Patients with intermediate AMD (iAMD) are at increased risk for development of NVAMD or GA, and may be targeted in future clinical trials with the goal of preventing onset of advanced disease. Our long-term goal is to understand the mechanisms for progression to late stages of AMD and to develop predictive biomarkers that will facilitate clinical trials of iAMD and GA and timely standard of care treatment for NVAMD for improved clinical outcomes….Thirdly, AI can help identify candidate drug targets for future clinical trials if predictive features can be identified and linked to mechanisms of action. While a few DL models based on multimodal imaging have been developed with various degrees of success and performance characteristics, there is a need for very high-performance algorithms that can identify specific SD-OCT biomarkers of GA or NVAMD, essentially opening the “black box” of typical DL models.--, in [0086]-[0087], and, --Predictive models such as this may facilitate personalized prediction of AMD progression to inform standard of care treatments and clinical trial enrollment. There is a need for development of high performance classifiers based on DL algorithms predicting progression of GA independent of human graders based on longitudinal SD OCT datasets. Ideally, the DL model should help identify the specific SD-OCT features or biomarkers that can collectively increase the probability of new progression from intermediate AMD to the severe stage of nonexudative AMD, essentially opening the “black box” associated with such artificial intelligence algorithms.--, in [0117], and, --[0183] Multi-view CNN Variation. We use a variant of the multi-view CNN model (Su H, et al. (2015) Proceedings of the IEEE International Conference on Computer Vision, pp. 945-953) to process the 3D OCT inputs, and use it as our baseline solution. The architecture of this model is outlined in FIG. 16. For each slice, the model feed it into a CNN network, and get the feature f.sub.i of slice i (f.sub.i=CNN(x.sub.i)), followed by a fully connected layer and a Sigmoid activation to get a probability score p.sub.i=sigmoid(FC.sub.1(f.sub.i))… [0185] Here δ is a trainable bias term parameter, initialized to a high value to stabilize the training, and gradually attenuated to a small number during training. τ is the temperature parameter, which is set to a small value to sharpen the attention weight, which helps us to find out the most important slices for GA diagnosis. The get final predicted probability of GA (GA score) for an image x at inference time, we compute the weighted summation of the probabilities w.sub.i of all 100 slices GA=Σ.sub.iw.sub.ip.sub.i.--, in [0183]-[0185], and [0191]; and, also see: --To quantitatively evaluate the causal relevance of competing saliency maps, we adopt the evaluation scheme proposed in (Hooker S, et al. (2019) Advances in Neural Information Processing Systems, pp. 9737-9748), consisting of masking out the contributing saliency pixels and then calculating the reduction in prediction score.)… We investigate how the different parings of PPI and saliency mapping schemes (i.e., Grad, GradCAM, LRP, WBP) affect performance. For WBP, we also tested the bounding box variant (see the Appendix for details). In Table 6, we see consistent performance gains in AUC score via incorporating PPI training (from 0.877 to 0.937), accompanied by the reductions in model variation evaluated by the standard deviations of AUC from the five-fold cross-validation.--, in [154]-[0155], [0160]); and
identifying the biomarker region as including at least one biomarker for a selected diagnosis of nascent geographic atrophy when the scoring metric meets a selected threshold (see LAD: e.g., -- we focus on the classification task of predicting the presence of a lesions, which is consistent with the setup of (Selvan R. et al. (2020) Medical Imaging with Deep Learning. PMLR, pp. 721-732). There are four radiologists annotates each patch with both lesion label and lesion mask. A patch in the dataset is labeled as positive if more than two (i.e., ≥3) radiologists have annotated presence of a lesion, otherwise negative. The ground-truth mask is the pixel-level union set of the four masks.--, in [0188], also see: --we consider three types of causal masking: (i) the point-wise soft causal masking defined by Equation (2) in the main text, (ii) hard masking, and (iii) box masking. For the hard masking, for each image, we keep points with WBP weight larger than k times of the standard deviation of WBP weights of the whole image. We test k from 1 to 7 and achieve similar results. As the model performs slightly better when k=7, we set k as 7, for all experiments. For the box masking, we use the center of mass for these kept points as the center to draw a box. The height and width of this box is defined as center.sub.h/w±1.2 std.sub.h/w. In this way at least 90% of filtered points are contained in the box--, in [0165]).
Re Claim 9, LAD further discloses generating an initial output for each OCT slice image of a plurality of OCT slice images that form the OCT volume image to form a plurality of initial outputs (see LAD: e.g., --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]); and
averaging the plurality of initial outputs to form a health indication output (see LAD: e.g., --In order to provide a mechanism to visually interpret model predictions, the model generates attention maps via weight backpropagation (WBP), which probabilistically masks out regions of the scan that do not contribute to the ability of the model for predicting GA, thus not contributing to the estimation of p.sub.i. Subsequently, pre-classification GA probabilities p.sub.i from different scans were aggregated, for which we considered different approaches, for instance, a simple average (i.e., mean pooling). However, as shown in FIG. 3, some image scans were more informative of GA than others, in terms of their p.sub.i values when examining the differences in distribution of the GA group relative to the controls, indicating that scan position may be leveraged for improved GA identification.--, in [0100]; and,
--[0104] FIG. 3 shows that the proposed model roughly identified scans from the range 35-75 of 100 as the most discriminative of GA prediction. Specifically, we present mean (solid lines) and SD (shaded areas) of cross-validated pre-classification p(GA) values (i.e.,) stratified into GA and non-GA groups, from which it is apparent that average p(GA) in the range 35-75 is substantially higher in the GA group than that of the controls. Moreover, that scans out of this range (1-34 and 76-100) are much less informative of GA status, as demonstrated by low predicted pre-classification p(GA) values generally lower than 0.25 (FIG. 3).--, in [0104];
-- For the box masking, we use the center of mass for these kept points as the center to draw a box. The height and width of this box is defined as center.sub.h/w±1.2 std.sub.h/w. In this way at least 90% of filtered points are contained in the box. For the soft masking, we set ω to 100 and σ to 0.25. We have also experimented with image-adaptive thresholds instead of a fixed σ for all inputs, i.e., set the threshold as mean value plus k times of the standard deviation of WBP weights of the whole image. The experiment comparison of these masking methods mention above is conducted on LIDC dataset. We repeat the experiments a few times and the results are consistent.--, in [0165])
Re Claims 10-15, claims 10-15 are the corresponding system claims to claims 1-6 respectively. Thus, claim 10-15 are rejected for the similar reasons as for claims 1-6. Furthermore, LAD further discloses a system comprising: a non-transitory memory; and a hardware processor coupled with the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform the method (see LAD: e.g., --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]).
Re Claim 19, LAD discloses system comprising:
a non-transitory memory; and
a hardware processor coupled with the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system (see LAD: e.g., --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]) to:
train a deep learning model using a training dataset that includes training OCT images that have been labeled as evidencing nascent geographic atrophy or not evidencing nascent geographic atrophy to form a trained deep learning model (see LAD: e.g., --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]; also see: --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]) );
receive an optical coherence tomography (OCT) volume image of a retina of a subject (see LAD: e.g., --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]);
generate, via the trained deep learning model, a classification score using the OCT volume image in which the classification score indicates whether nascent geographic atrophy is detected (see LAD: e.g., -- saliency mapping schemes are considered as comparators for the proposed approach: (i) Gradient: standard gradient-based salience mapping; (ii) Grad-CAM (Selvaraju R, et al. (2017) Proceedings of the IEEE International Conference on Computer Vision, pp. 618-626): gradient-weighted class activation mapping; (iii) LRP (Bach S, et al. (2015) PloS One. 10(7):e0130140): layer-wise relevance propagation and its variants. CUB: Bird classification in the wild. In this task we want to qualitatively and quantitatively compare the causal relevance of saliency maps generated by WBP and its competitors.--, in [0154], and, -- [0172] Activation Based Methods. Methods under this category (such as CAM, Grad-CAM, guided Grad-CAM, Grad-CAM++) use a linear combination of class activation maps from convolutional layers to derive a saliency map--, in [0172]; and also see: --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]);
generate a saliency volume map for the OCT volume image using a saliency mapping algorithm, wherein the saliency volume map indicates a level of contribution of a set of regions in the OCT volume image to a diagnosis of geographic atrophy generated by the deep learning model (see LAD: e.g., -- saliency mapping schemes are considered as comparators for the proposed approach: (i) Gradient: standard gradient-based salience mapping; (ii) Grad-CAM (Selvaraju R, et al. (2017) Proceedings of the IEEE International Conference on Computer Vision, pp. 618-626): gradient-weighted class activation mapping; (iii) LRP (Bach S, et al. (2015) PloS One. 10(7):e0130140): layer-wise relevance propagation and its variants. CUB: Bird classification in the wild. In this task we want to qualitatively and quantitatively compare the causal relevance of saliency maps generated by WBP and its competitors.--, in [0154], and, -- [0172] Activation Based Methods. Methods under this category (such as CAM, Grad-CAM, guided Grad-CAM, Grad-CAM++) use a linear combination of class activation maps from convolutional layers to derive a saliency map--, in [0172]; and also see: --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]);
detect a set of potential biomarker regions in the OCT volume image using the saliency volume map (see LAD: e.g., -- saliency mapping schemes are considered as comparators for the proposed approach: (i) Gradient: standard gradient-based salience mapping; (ii) Grad-CAM (Selvaraju R, et al. (2017) Proceedings of the IEEE International Conference on Computer Vision, pp. 618-626): gradient-weighted class activation mapping; (iii) LRP (Bach S, et al. (2015) PloS One. 10(7):e0130140): layer-wise relevance propagation and its variants. CUB: Bird classification in the wild. In this task we want to qualitatively and quantitatively compare the causal relevance of saliency maps generated by WBP and its competitors.--, in [0154], and, -- [0172] Activation Based Methods. Methods under this category (such as CAM, Grad-CAM, guided Grad-CAM, Grad-CAM++) use a linear combination of class activation maps from convolutional layers to derive a saliency map--, in [0172]; and also see: --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]); and
generate a report that confirms that nascent geographic atrophy is detected when at least one potential biomarker region of the set of potential biomarker regions meets a set of criteria and when the classification score meets a threshold (see LAD: e.g., -- saliency mapping schemes are considered as comparators for the proposed approach: (i) Gradient: standard gradient-based salience mapping; (ii) Grad-CAM (Selvaraju R, et al. (2017) Proceedings of the IEEE International Conference on Computer Vision, pp. 618-626): gradient-weighted class activation mapping; (iii) LRP (Bach S, et al. (2015) PloS One. 10(7):e0130140): layer-wise relevance propagation and its variants. CUB: Bird classification in the wild. In this task we want to qualitatively and quantitatively compare the causal relevance of saliency maps generated by WBP and its competitors.--, in [0154], and, -- [0172] Activation Based Methods. Methods under this category (such as CAM, Grad-CAM, guided Grad-CAM, Grad-CAM++) use a linear combination of class activation maps from convolutional layers to derive a saliency map--, in [0172]; and also see: --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]).
Re Claim 20, Lad further discloses wherein the saliency mapping algorithm comprises a gradient- weighted class activation mapping (GradCAM) algorithm and wherein the classification score is a probability that the OCT volume image evidences nascent geographic atrophy and wherein the threshold is a value selected between 0.5 and 0.8 (see LAD: e.g., --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]; also see: --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]);
filtering the saliency map to generate a modified saliency map (see LAD: e.g., -- we derived an architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction. Visual inspection of the saliency maps show that WBP, especially when coupled with PPI, is more robust to spurious features compared to competing approaches. Tested on natural image and medical image datasets, empirical results verify the combination of PPI and WBP consistently delivers performance gains across a wide range of tasks relative to competing solutions, and the gains are most significant where the application is complicated by small sample size, data heterogeneity, or confounded with spurious correlations.
[0163] Derivation of Convolutional Weight Back-propagation. Let's denote the input variable as I∈custom-character.sup.H×W, the convolutional filter weight as W∈custom-character.sup.(2S+1)×(2S+1), the output variable as Oϵcustom-character.sup.H×W, and the weight backpropagate to O as Ŵ∈custom-character.sup.H×W….If the convolutional layer is downsizing the input variable (i.e., strides), the {tilde over (W)}.sub.ijk.sup.l+1 is padded with zeros around the weights (left, right, up, and down) to display the input elements that the convolutional filter strides over. The number of padding zeros is equal to the number of strides minus 1.
[0165] Details on Causal Masking. In this work, we consider three types of causal masking: (i) the point-wise soft causal masking defined by Equation (2) in the main text, (ii) hard masking, and (iii) box masking. For the hard masking, for each image, we keep points with WBP weight larger than k times of the standard deviation of WBP weights of the whole image. We test k from 1 to 7 and achieve similar results. As the model performs slightly better when k=7, we set k as 7, for all experiments. For the box masking, we use the center of mass for these kept points as the center to draw a box. The height and width of this box is defined as center.sub.h/w±1.2 std.sub.h/w. In this way at least 90% of filtered points are contained in the box. For the soft masking, we set ω to 100 and σ to 0.25. We have also experimented with image-adaptive thresholds instead of a fixed σ for all inputs, i.e., set the threshold as mean value plus k times of the standard deviation of WBP weights of the whole image.--, in [0162]-[0165]; and, -- saliency mapping schemes are considered as comparators for the proposed approach: (i) Gradient: standard gradient-based salience mapping; (ii) Grad-CAM (Selvaraju R, et al. (2017) Proceedings of the IEEE International Conference on Computer Vision, pp. 618-626): gradient-weighted class activation mapping; (iii) LRP (Bach S, et al. (2015) PloS One. 10(7):e0130140): layer-wise relevance propagation and its variants. CUB: Bird classification in the wild. In this task we want to qualitatively and quantitatively compare the causal relevance of saliency maps generated by WBP and its competitors.--, in [0154], and, -- [0172] Activation Based Methods. Methods under this category (such as CAM, Grad-CAM, guided Grad-CAM, Grad-CAM++) use a linear combination of class activation maps from convolutional layers to derive a saliency map--, in [0172]).
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be /negated by the manner in which the invention was made.
Claims 7, 16 are rejected under 35 U.S.C. 103 as being unpatentable over LAD, and in view of Buckland (US 20210209758 A1).
Re Claim 7, LAD although discloses the OCT volume image comprises a plurality of OCT slice images that are two-dimensional, and output the results of geographic atrophy (GA) detection (see LAD: e.g., Fig. 1, and, -- to find out the most important slices for GA diagnosis. The get final predicted probability of GA (GA score) for an image x at inference time--, in [0185], also see: --[0003] Disclosed herein is a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0003]);
LAD however does not explicitly disclose generating an evaluation recommendation based on at least one of the output or the map output, wherein the evaluation recommendation identifies a subset of the plurality of OCT slice images for further review;
Buckland discloses the scoring metric comprises at least one of a size of the potential biomarker region or a confidence score for the potential biomarker region (see Buckland: e.g., -- A noticeable gap with respect to quantitative imaging of the retina is the lack of a calibration standard, and the strong dependence of imaging techniques on the length of the posterior chamber of the eye which is largely an unknown quantity, at least with respect to data that guides retinal imaging systems. Furthermore, the lack of standards has made the quantitative application of adult optical coherence tomograph (OCT) imaging systems to the under-developed eye of the child untenable and has further limited the clinical interoperability of imaging systems and their normative data to imaging systems of one model by one manufacturer, thereby dramatically increasing the cost for introducing new instruments to the market or qualifying new diagnostic imaging biomarkers that for broad use across a class of instruments.--, in [0057]; and, -- The characterization of the retina for use in automated or guided diagnostics and prognostics is quite challenging for the richness of the visual and neurological processes in sight. Certain measurements are established as standards of care and supported with normative data for the adult eye, yet the actual set of objective, quantitative biomarkers that are extracted from volumetric images of the retina is quite slim, and largely limited to pathology of the mature adult retina. Current normative datasets are device specific (constrained to a specific manufacturer and model), adult only, gender non-specific, and ethnicity non-specific. Useful normative values are also limited to total retinal thickness in the vicinity of the fovea, and nerve fiber layer thickness in a circumference about the optic nerve head. Yet the primary clinical application of OCT imaging of the fovea is an observational interpretation of edema associated with age-related macular degeneration that does not require measurement. Use of the nerve fiber layer thickness is intended to support the diagnosis of glaucoma, but the diagnostic value remains fraught. Given the exceptional progress in understanding the visual processes and the development of therapies, including gene therapies that are successfully treating blindness, there is clearly a need to develop and validate a more complete and robust set of imaging biomarkers to diagnose disease earlier, develop precision medical treatments, and screen patients as candidates for treatments. This problem is no less important for the pediatric population that has a lifetime of sight to consider. Some embodiments of the present inventive concept catalog measurements in three dimensions that are associated with retinal physiology, provide a methodology for searching for biomarkers associated with normal pathological development, and provide a system for verifying and validating such biomarkers with techniques of big data.--, in [0069]; and, -- A typical clinical OCT image of the retina may have an A-scan of length 1024 pixels, with 500-1000 A-scans per B-scan, and 100-500 B-scans per volume, therefore including a range of 1024×500×100 to 1024×1000×500 voxels, from 51.2 to 512 megapixels, respectively. As systems become faster and storage memory cheaper, the maximum image size may increase, though this range of values encompasses more information than is objectively clinically useable today. Considering only the boundary surface surfaces on a fully segmented image set, the number of voxels that may be analyzed for quantitative biomarkers is reduced by the ratio of number of layers (a-k) to the length of an A-scan, 11/1024, i.e. two orders of magnitude. The prospect of developing a finite discrete set of biomarkers from this reduced set of values remains daunting.--, in [0071]-[0073]; and,
-- [0155] With respect to segmentation, a conventional segmentation process is discussed above with respect to FIGS. 3A and 3B. However, some embodiments of the present inventive concept may use contouring techniques to connect/bridge sectors. Different techniques may be used depending on the quality of the image/sectors. In other words, one method or set of rules may be used for contiguous good sectors and a different set of rules may be used for a discontinuous sector. If the image quality is not good, a confidence level for the image may be low. For example, for a discontinuous sector, segmentation may be stopped after a plurality of discontinuous gaps in the image are encountered. Furthermore, it may be determined if the gap extends to an edge of the image. The discontinuous portions of the image or segment may be contoured but after a certain number of discontinuities, it may not be worth recovering the image as the image may not be useable. Various rules may be used to determine what images should be contoured and which images should be discarded.
[0156] In the process of developing and validating the local segmentation processes, attributes of the image classification and associated segmentation recipes, including methods and tuning parameters are coupled, as per FIG. 11. Among the quantification labels (QuantLabels 1113, FIG. 11) is the Image Quality Metric. The QuantLabels 1113 may be a group of quantitative labels designed to facilitate decision making on managing contours across neighboring segments.--, in [0155]-[0156]);
LAD and Buckland are combinable as they are in the same field of endeavor: deep learning neural network for eye diseases detection based on OCT images. Therefore it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify LAD’s method using Buckland’s teachings by including the scoring metric comprises at least one of a size of the potential biomarker region or a confidence score for the potential biomarker region to LAD’s calculating prediction probability in order to diagnose disease earlier, develop precision medical treatments, and screen patients as candidates for treatments based on interpretation of edema associated with age-related macular degeneration (see Buckland: e.g. in [0069], [0071]-[0073], [0155]-[0156]).
Re Claim 16, claim 16 is the corresponding system claim to claim 7 respectively. Thus, claim 16 is rejected for the similar reasons as for claim 7. Furthermore, LAD as modified by Buckland further disclose a system comprising: a non-transitory memory; and a hardware processor coupled with the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform the method (see LAD: e.g., --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]).
Claims 8, 17-18 are rejected under 35 U.S.C. 103 as being unpatentable over LAD, and in view of BOYD (US 20210279874 A1).
Re Claim 8, LAD further discloses generating a saliency map for an OCT slice image of the OCT volume image using thesaliency mapping algorithm, the saliency map indicating a degree of importance of each pixel in the OCT slice image for a diagnosis of nascent geographic atrophy (see LAD: e.g., --trained the model in a novel interpretable computer vision framework PPI that combines saliency mapping, causal reasoning, synthetic intervention and contrastive learning, to help the model perform robustly even on a small dataset with only 872 samples; 3) we leveraged the architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction, to communicate model insights and facilitate causal-informed reasoning. Empirical evidence confirms that our model yields encouraging results for high dimensional volume medical images with a small training dataset.
[0121] Limitations. Despite the promising results, our study has several limitations. First, the overall size of the dataset is relatively small for deep learning. In particular, although the AUC for GA events and GA predictions were encouraging, the standard deviations for 5-fold cross validations were understandably large. Though difficult in practice, a much larger dataset or a population with more GA events may enable more accurate deep-learning models to be trained and evaluated with high confidence. Another limitation is that some important features to GA were not used in the current model. In particular, we found features worked well for GA prediction in our other manuscript. It is possible that combine these features in our deep learning model may enable more accurate results. However, we focused on exploring the predictive power of OCT images without additional processing, feature engineering or expert annotation.--, in [0120]-[0121], and, --[0123] Deep neural networks have shown significant promise in comprehending complex visual signals, delivering performance on par or even superior to that of human experts. However, these models often lack a mechanism for interpreting their predictions, and in some cases, particularly when the sample size is small, existing deep learning solutions tend to capture spurious correlations that compromise model generalizability on unseen inputs. In this work, we propose a contrastive causal representation learning strategy that leverages proactive interventions to identify causally-relevant image features, called Proactive Pseudo-Intervention (PPI). This approach is complemented with a causal salience map visualization module, i.e., Weight Back Propagation (WBP), that identifies important pixels in the raw input image, which greatly facilitates the interpretability of predictions. To validate its utility, our model is bench-marked extensively on both standard natural images and challenging medical image datasets. We show this new contrastive causal representation learning model consistently improves model performance relative to competing solutions, particularly for out-of-domain predictions or when dealing with data integration from heterogeneous sources. Further, our causal saliency maps are more succinct and meaningful relative to their non-causal counterparts.--, in [0123]; and, --Specifically, SM encompasses techniques for post hoc visualizations on the input (image) space to facilitate interpretation of model predictions. This is done by projecting the key features used in prediction back to the input space, resulting in the commonly known saliency maps….[0127] In this work, we present a solution that accounts for the needs of causal representation identification and visual verification. Our key insight is the derivation of causally-informed saliency maps, which facilitate visual verification of model predictions and enable learning that is robust to nuisance (non-causal) invariances.--, in [0126]-[0127]; also see: --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]);
filtering the saliency map to generate a modified saliency map (see LAD: e.g., -- we derived an architecture-agnostic saliency mapping scheme called Weight Back Propagation (WBP), which faithfully captures the causally-relevant pixels/features for model prediction. Visual inspection of the saliency maps show that WBP, especially when coupled with PPI, is more robust to spurious features compared to competing approaches. Tested on natural image and medical image datasets, empirical results verify the combination of PPI and WBP consistently delivers performance gains across a wide range of tasks relative to competing solutions, and the gains are most significant where the application is complicated by small sample size, data heterogeneity, or confounded with spurious correlations.
[0163] Derivation of Convolutional Weight Back-propagation. Let's denote the input variable as I∈custom-character.sup.H×W, the convolutional filter weight as W∈custom-character.sup.(2S+1)×(2S+1), the output variable as Oϵcustom-character.sup.H×W, and the weight backpropagate to O as Ŵ∈custom-character.sup.H×W….If the convolutional layer is downsizing the input variable (i.e., strides), the {tilde over (W)}.sub.ijk.sup.l+1 is padded with zeros around the weights (left, right, up, and down) to display the input elements that the convolutional filter strides over. The number of padding zeros is equal to the number of strides minus 1.
[0165] Details on Causal Masking. In this work, we consider three types of causal masking: (i) the point-wise soft causal masking defined by Equation (2) in the main text, (ii) hard masking, and (iii) box masking. For the hard masking, for each image, we keep points with WBP weight larger than k times of the standard deviation of WBP weights of the whole image. We test k from 1 to 7 and achieve similar results. As the model performs slightly better when k=7, we set k as 7, for all experiments. For the box masking, we use the center of mass for these kept points as the center to draw a box. The height and width of this box is defined as center.sub.h/w±1.2 std.sub.h/w. In this way at least 90% of filtered points are contained in the box. For the soft masking, we set ω to 100 and σ to 0.25. We have also experimented with image-adaptive thresholds instead of a fixed σ for all inputs, i.e., set the threshold as mean value plus k times of the standard deviation of WBP weights of the whole image.--, in [0162]-[0165]);
LAD however does not explicitly disclose overlaying the modified saliency map on the OCT slice image to generate the map output,
BOYD discloses overlaying the modified saliency map on the OCT slice image to generate the map output (see BOYD: e.g., -- In some embodiments, the delta analysis can comprise displaying the subtracted ocular images. In some embodiments, the subtracted ocular images can be displayed as an overlay on preprocessed ocular images acquired at earlier clinical visits. In some embodiments, the plurality of analytical modules can comprise a broad image analytics module configured to perform a broad image analysis of the plurality of preprocessed ocular images, and the plurality of ocular images can be acquired from a plurality of subjects…. the aggregated quantitative features can be processed by a cognitive technology configured to perform one or more of: (i) cognitive analysis of phenotyping, genotyping, and/or epigenetics, (ii) biomarker identification, (iii) estimating a probability of an eye disease, (iv) estimating a probability of effectiveness of a treatment for an eye disease, (v) estimating a probability for recommending a treatment for an eye disease, and (vi) estimating a probability for recommending a clinical trial enrollment, (vii) estimating the likelihood of responding to a treatment for an eye disease, (viii) estimating the safety of a treatment of an eye disease.--, in [0006]);
LAD and BOYD are combinable as they are in the same field of endeavor: deep learning in eye image processing including (ii) biomarker identification, (iii) estimating a probability of an eye disease. Therefore it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify LAD’s method using BOYD’s teachings by including overlaying the modified saliency map on the OCT slice image to generate the map output to LAD’s output of saliency map and to generate the map output in order to output and display the result of OCT image analysis including (ii) biomarker identification, (iii) estimating a probability of an eye disease ..etc., (see BOYD: e.g. in [0006]).
Re Claim 17, claim 17 is the corresponding system claim to claim 8 respectively. Thus, claim 17 is rejected for the similar reasons as for claim 8. Furthermore, LAD as modified by BOYD further disclose a system comprising: a non-transitory memory; and a hardware processor coupled with the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform the method (see LAD: e.g., --[0002] Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Advanced disease results in either irreversible loss of vision in the dry form of geographic atrophy (GA), or in retinal exudation in neovascular AMD (NVAMD)…..a system using machine learning in detecting geographic atrophy (GA), the system comprising at least one processor; a memory; and a computing platform including the at least one processor and the memory, wherein the computing platform is configured for receiving, as input, a plurality of OCT volume scan images; generating, using a trained GA detection algorithm, a probabilistic likelihood that each scan is informative of GA and including high-probability scans in a dataset, and determining, using the included scans, whether GA is present or likely to occur, wherein the GA detection algorithm includes at least one machine learning algorithm and is trained using one or more data sets associated with related GA events; and outputting, by the GA detection algorithm, information indicating whether or not GA is present or likely to occur.--, in [0002]-[0003], and [0087]).
Re Claim 18, LAD as modified by BOYD further disclose wherein the map output comprises a saliency map overlaid on an individual OCT slice image of the OCT volume image (see LAD: e.g., Fig. 1, “masking” read on overlay, -- in FIG. 1, to maximize the weighted binary cross-entropy loss, i.e., L(y, f (x)), the likelihood that scans from SD-OCT inputs, x, were correctly assigned (prognosticated) to either the GA or control groups, y, in the assessment of the upcoming year, while encouraging that i) regions masked-out by the saliency maps, x*, which we call negative contrasts, were not informative of GA, and ii) unmasked input scans, x, and regions masked-out by a randomly assigned saliency map, x′, did not affect the model's ability to predict GA.--, in [0101]; and, -- in FIG. 5. While the central scans have a high predictive value or probability (p) and are characterized by red dots concentrated in the lesional and perilesional areas, the peripheral retinal scans have very low p values and diffusely distributed red dots in the attention maps. In eyes with intermediate AMD that will convert to GA in one year, the red dots are mainly concentrated in large drusen or drusenoid PEDs and underlying choriocapillaris and choroid, hyperreflective foci, and areas of nascent GA or incomplete RPE and outer retinal atrophy (iRORA). To a lesser extent, the attention maps also mark the neurosensory retina overlying drusen (FIG. 4B). A representative case example of intermediate AMD that preceded GA by one year can be found in FIG. 5. In eyes with intermediate AMD that will not progress to GA, the attention maps are composed of red dots diffusely present in a large area of drusen and neurosensory retina overlying them; the probability value p is low (FIG. 4C).--, in [0110]
-- For WBP, we also tested the bounding box variant (see the Appendix for details). In Table 6, we see consistent performance gains in AUC score via incorporating PPI training (from 0.877 to 0.937), accompanied by the reductions in model variation evaluated by the standard deviations of AUC from the five-fold cross-validation. The gains are most significant when using our WBP for saliency mapping. We further compare the saliency maps generated by these different combinations. We see that without the additional supervision from PPI, competing solutions like Grad, GradCAM and LRP sometimes yield non-sensible saliency maps (attending to image corners). Overall, PPI encourages more concentrated and less noisy saliency maps. Also, different PPI-based saliency maps agree with each other to a larger extent. Our findings are also verified by experts (co-authors, who are ophthalmologists specializing in GA) confirming that the PPI-based saliency maps are clinically relevant by focusing on retinal layers likely to contain abnormalities or lesions.--, in [0142], and [0155], and, -- In FIG. 12, we compare saliency maps generated by Grad-CAM, WBP, WBP (box) to the ground truth lesion masks from expert annotations. Note that we have only supplied patch-label labels during training, not the pixel-level expert segmentation masks, which constitute a challenging task of weakly-supervised image segmentation. In line with the observations from the GA experiment, our PPI-training enhanced WBP saliency maps are mostly consistent with the expert segmentations.--, in [0161], and, -- [0165] Details on Causal Masking. In this work, we consider three types of causal masking: (i) the point-wise soft causal masking defined by Equation (2) in the main text, (ii) hard masking, and (iii) box masking. For the hard masking, for each image, we keep points with WBP weight larger than k times of the standard deviation of WBP weights of the whole image. We test k from 1 to 7 and achieve similar results. As the model performs slightly better when k=7, we set k as 7, for all experiments. For the box masking, we use the center of mass for these kept points as the center to draw a box. The height and width of this box is defined as center.sub.h/w±1.2 std.sub.h/w. In this way at least 90% of filtered points are contained in the box. For the soft masking, we set ω to 100 and σ to 0.25. We have also experimented with image-adaptive thresholds instead of a fixed σ for all inputs, i.e., set the threshold as mean value plus k times of the standard deviation of WBP weights of the whole image. The experiment comparison of these masking methods mention above is conducted on LIDC dataset. We repeat the experiments a few times and the results are consistent.--, in [0165]; also see BOYD: e.g., -- In some embodiments, the delta analysis can comprise displaying the subtracted ocular images. In some embodiments, the subtracted ocular images can be displayed as an overlay on preprocessed ocular images acquired at earlier clinical visits. In some embodiments, the plurality of analytical modules can comprise a broad image analytics module configured to perform a broad image analysis of the plurality of preprocessed ocular images, and the plurality of ocular images can be acquired from a plurality of subjects…. the aggregated quantitative features can be processed by a cognitive technology configured to perform one or more of: (i) cognitive analysis of phenotyping, genotyping, and/or epigenetics, (ii) biomarker identification, (iii) estimating a probability of an eye disease, (iv) estimating a probability of effectiveness of a treatment for an eye disease, (v) estimating a probability for recommending a treatment for an eye disease, and (vi) estimating a probability for recommending a clinical trial enrollment, (vii) estimating the likelihood of responding to a treatment for an eye disease, (viii) estimating the safety of a treatment of an eye disease.--, in [0006]).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WEI WEN YANG whose telephone number is (571)270-5670. The examiner can normally be reached on 8:00 - 5:00 pm.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Amandeep Saini can be reached on 571-272-3382. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free).
If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/WEI WEN YANG/Primary Examiner, Art Unit 2662
Prosecution Insights