Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claims 1-4, 10, 12-15, 17, and 20 are pending. Claims 6-9, 16, and 19 have been canceled. This Office Action is responsive to the amendment filed on 02/20/2026, which has been entered into the above identified application.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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 1, 3-4, 10, 12-14, 17, and 20 are rejected under 35 U.S.C. 103 as being unpatentable Mormont et al. (“Multi-Task Pre-Training of Deep Neural Networks for Digital Pathology”, published 05/06/2020), hereinafter Mormont; in view of Finnie et al. (US 20200380359 A1, filed 05/27/2020), hereinafter Finnie; in further view of Moradi et al. (US 20190073770 A1, filed 09/06/2026). Mormont was cited in a previous Office Action.
Regarding Claim 1, Mormont teaches A method for disease classification in a medical system (Mormont: “we investigate multi-task learning as a way of pre-training models for classification tasks in digital pathology.” [Abstract]), the method comprising:
machine training a first machine-learned model to output an estimate of an anatomical or functional characteristic using a first training dataset having at least 1,000 samples (Mormont: “The core idea is to pre-train a model on a large dataset (the source task), and then somehow transfer the learned knowledge to facilitate training on a second dataset (the target task).” [Section I. Introduction]; “Datasets that Were Used for Multi-Task Pre-Training. CLF, DET and SEG Respectively Stand for classification, detection and segmentation. H&E, IHC and M3C Respectively Stand for Hematoxylin and Eosin, Immunohistochemistry and Masson's Trichrome. Images and Classes Columns Give the Number of Images and Classes of the Final (Possibly Transformed) Task for this Dataset” [Table I]; “The core idea is to pre-train a model on a large dataset (the source task), and then somehow transfer the learned knowledge to facilitate training on a second dataset (the target task). As the source task must be a large dataset, the most common choice is using ImageNet, a classification dataset containing more than 1 million natural images organized into 1000 classes, as a source.” [Section I. Introduction]);
generating a second machine-learned model to classify a disease by adapting the first machine-learned model by adding one or more fully connected layers and a SoftMax layer that receive an estimate of the anatomical or functional characteristic and output an estimation of a pathology (Mormont: “In this work, we study the two classical approaches of network transfer, namely feature extraction and fine-tuning. In both cases, θs is pre-trained on some source task(s), either ImageNet or several tasks simultaneously in the MTL setting. Feature extraction consists in using the pre-trained θs only to extract the feature vector it outputs for all images of the target task. The extracted features can then be used to learn a third-party classifier, a common choice being a linear SVM. Fine-tuning consists in further training θs on the target task. A fully connected layer and a softmax are attached to the shared network for generating the target task classes probabilities and the resulting network is trained using for instance stochastic gradient descent.” [Section IV.C. Transferring a Multi-Task Pre-Trained Network]);
machine training the second machine-learned model to output the classification of the disease (Mormont: “The core idea is to pre-train a model on a large dataset (the source task), and then somehow transfer the learned knowledge to facilitate training on a second dataset (the target task).” [Section I. Introduction]; “Fine-tuning consists in further training θs on the target task. A fully connected layer and a softmax are attached to the shared network for generating the target task classes probabilities and the resulting network is trained using for instance stochastic gradient descent.” [Section IV.C. Transferring a Multi-Task Pre-Trained Network]);
acquiring a medical scan of a patient (Mormont: “All the models trained using this hyperparameters combination (i.e. one per seed) are transferred to the target task tk using both transfer protocols (i.e. feature extraction or fine-tuning). The test set of tk is used solely to evaluate the resulting transfer performance whereas the training and validation sets can be used by the transfer protocol, feature extraction or fine-tuning, for training and hyperparameter tuning.” [Section IV.E. Final Performance Evaluation]); and
classifying, by the second machine-learned model, the disease of the patient from the medical scan (Mormont: “We have first created a pool of classification tasks from existing sources containing almost 900k digital pathology images. Using this pool, we have pre-trained a neural network in a multi-task setting in order to transfer the resulting model to unseen digital pathology tasks. Using a robust evaluation protocol, we have shown that transferring a model pre-trained in multi-task can be beneficial for the performance on the target task.” [Conclusion]).
However, Mormont fails to expressly disclose machine training the second machine-learned model with few shot learning using a second training dataset having fewer than 200 samples to output the classification and an uncertainty of the classification, wherein at least some of the less than 200 samples of the second training dataset are synthetic examples, wherein the few shot learning comprises implementing a prototypical network that maps the data into an embedding space, forms a prototype representation for each pathology class as a mean of embeddings of support examples for a class, and classifies query examples by determining proximity of the query embedding to the prototype representations in the embedding space or wherein the few shot learning comprises implementing a matching network that uses an attention mechanism over a learned embedding of a support set of labeled examples to predict classes for a query set of unlabeled examples; determining an uncertainty of a classification; and displaying the classification and the uncertainty.
In the same field of endeavor, Finnie teaches machine training the second machine-learned model with few shot learning using a second training dataset to output the classification and an uncertainty of the classification, wherein at least some of samples of the second training dataset are synthetic examples (Finnie: “Advantageously, the example method comprises collecting a second set of digital images from the database, wherein the second set of digital images is sampled from digital images assigned to a few shot class; creating a second training set for a second stage of training comprising the second set of digital images, training the second artificial neural network using the second training set.” [0007]; “The device 100 comprises an output 224, configured to output a signal depending on the class and optionally on the confidence score. For example, for five classes the output 224 outputs a vector of scores. The length of the vector is equal to the number of classes, the class that has the highest score is the result of the prediction for the class, i.e. the class determined for the content of the digital image.” [0029]; “In a further aspect of the present invention, the digital image data may be augmented by applying one or more transformations to a digital image. A modified digital image may be created from the digital image sampled for one of the first training set or the second training set in particular by cropping, mirroring, rotating, smoothing, or contrast reduction to create the modified digital image.” [0086]),
wherein the few shot learning comprises implementing a prototypical network that maps the data into an embedding space, forms a prototype representation for each pathology class as a mean of embeddings of support examples for a class, and classifies query examples by determining proximity of the query embedding to the prototype representations in the embedding space or wherein the few shot learning comprises implementing a matching network that uses an attention mechanism over a learned embedding of a support set of labeled examples to predict classes for a query set of unlabeled examples (Finnie: “The device 100 comprises a second neural network 212 comprising a second feature extractor 212 and a second classifier 216. The second neural network 212 in the example is a prototypical model neural network. The second feature extractor 214 is configured according to the configuration of the first feature extractor 204. The second classifier 216 is for example a cosine distance classifier. The second feature extractor 214 is configured to determine features of a content of the digital image received at an input 218 in a feature space. The second classifier 216 is configured to classify the content of the digital image into a class. The second classifier 216 is configured to output the class at an output 222.” [0027]; “After transferring baseline's cosine classifier's weights to the prototypical neural network's feature space as prototypes, the distance from a test image's feature 506 is calculated in the same way for both, the many shot classes and the few shot classes. As depicted in FIG. 5, the test image's feature 506 is closer to a first star 508 representing a training image's feature, than the nearest centroid 510. When train images in the same class are very different from each other, this approach is very efficient, when the intra-class variance is high. This is in particular useful, when a test image is closer to a specific training image than the average feature of another cluster of training images.” [0081]);
determining an uncertainty of a classification (Finnie: “Advantageously, in accordance with the example embodiment of the present invention, a confidence score is determined for the class for the content. This provides a certainty of the result. For example, if it's not certain, that the prediction for a certain digital image can be trusted, because the confidence score is lower than a threshold, this digital image might not belong to any class the model has been trained with.” [0016]); and
displaying the classification and the uncertainty (Finnie: “The device 100 comprises an output 224, configured to output a signal depending on the class and optionally on the confidence score. For example, for five classes the output 224 outputs a vector of scores. The length of the vector is equal to the number of classes, the class that has the highest score is the result of the prediction for the class, i.e. the class determined for the content of the digital image.” [0029]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated machine training the second machine-learned model with few shot learning using a second training dataset to output the classification and an uncertainty of the classification, wherein at least some of the samples of the second training dataset are synthetic examples, wherein the few shot learning comprises implementing a prototypical network that maps the data into an embedding space, forms a prototype representation for each pathology class as a mean of embeddings of support examples for a class, and classifies query examples by determining proximity of the query embedding to the prototype representations in the embedding space or wherein the few shot learning comprises implementing a matching network that uses an attention mechanism over a learned embedding of a support set of labeled examples to predict classes for a query set of unlabeled examples; determining an uncertainty of a classification; and displaying the classification and the uncertainty, as taught by Finnie to the method of Mormont as both of these methods are directed towards transfer learning strategies for adapting a model to classify samples in a related task. In making this combination and training the second machine learning model as a prototypical network, utilizing synthetic examples in training, and determining an uncertainty of each classification, it would allow the method of Mormont to access “a pretrained feature space” without having to train the extractor from scratch (Finnie: [0011]), as well as determine whether a “prediction for a certain digital image can be trusted” so as to determine if a “digital image might not belong to any class the model has been trained with” (Finnie: [0016]).
Mormont and Finnie still fail to expressly disclose machine training the second machine-learned model with few shot learning using a second training dataset having fewer than 200 samples.
In the same field of endeavor, Moradi teaches machine training the second machine-learned model with few shot learning using a second training dataset having fewer than 200 samples (Moradi: “Using the feature maps of a trained segmentation network, deviations from normal anatomy can be learned by a binary convolutional network on an extremely unbalanced training dataset with as little as one positive for 17 negative samples.” [0017]; “In this example, one of the trained segmentation models (n=16) is used as the source of features, and the corresponding 100 testing images from 10 patients are used as the negative samples. For each test, the negative samples were randomly divided into two equal sized (50) sets for training and testing the classification network. For the 30 positive samples of each disease, they were randomly divided into 10 samples for training and 20 samples for testing.” [0041]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated machine training the second machine-learned model with few shot learning using a second training dataset having fewer than 200 samples, as taught by Moradi to the method of Mormont and Finnie because both of these methods involve transferring the knowledge derived from a segmentation model to train a classification model on a previously unseen target class. In making this combination and training the classification model with less than 200 samples, it addresses the limited number of samples that exists in fields such as medical imaging and provides evidence that the transference of knowledge from a model trained on anatomical features can allow for the classification of diseases with a very small number of positive samples (Moradi: [0017]-[0021]).
Regarding Claims 14 and 20, they are method and system claims that recite similar corresponding limitations to those of Claim 1. Therefore, they are rejected for the same reasons as Claim 1 above.
Regarding Claim 3, Mormont, Finnie, and Moradi teach the method of Claim 1, wherein the second machine-learned model comprises a multi-task model (Mormont: “The structure of our multi-task neural network is similar to those of [17] and [34] and is guided by the objective of pre-training a network for transfer.” [Section IV.A. Multi-Task Architecture]). Regarding Claim 4, Mormont, Finnie, and Moradi teach the method of Claim 1, wherein the first and second machine-learned models comprise neural networks (Mormont: “we have pre-trained a neural network in a multi-task setting in order to transfer the resulting model to unseen digital pathology tasks.” [Section VI. Conclusion]).
Regarding Claim 10, Mormont, Finnie, and Moradi teach the method of Claim 1, wherein at least some of the at least 1,000 samples of the first training dataset are synthetic examples (Mormont: “Classical data augmentation and normalization have been applied to the input images. We have used ImageNet statistics for normalizing the images as early experiments have shown no significant improvement by normalizing with per-task statistics. As data augmentation, we have applied simple random vertical and horizontal flips as well as extraction of a random square crop (if the image is not square already).” [Section IV.F. Hyperparameters Settings and Experiments]; In light of paragraph [0043] which states “the dataset to be used as training data is augmented with synthetic datasets”, BRI of introducing synthetic examples entails utilizing data augmentation).
Regarding Claim 12, Mormont, Finnie, and Moradi teach the method of Claim 1, further comprising generating, by a processor, a clinical decision from the classification (Moradi: “A disease detection is received from the trained classification network. The disease detection indicating the presence or absence of a predetermined disease.” [0002]; “The components of computer system/server 12 may include, but are not limited to, one or more processors or processing units 16, a system memory 28, and a bus 18 that couples various system components including system memory 28 to processor 16.” [0050]).
Regarding Claim 13, Mormont, Finnie, and Moradi teach the method of Claim 12, further comprising estimating an uncertainty of the classification, and wherein generating comprises generating based on the uncertainty (Finnie: “Advantageously, in accordance with the example embodiment of the present invention, a confidence score is determined for the class for the content. This provides a certainty of the result. For example, if it's not certain, that the prediction for a certain digital image can be trusted, because the confidence score is lower than a threshold, this digital image might not belong to any class the model has been trained with.” [0016]).
Regarding Claim 17, Mormont, Finnie, and Moradi teach the method of Claim 14, wherein machine training the first classifier uses training data with a number of examples at least ten times a number of examples for machine training the second classifier (Mormont: “The core idea is to pre-train a model on a large dataset (the source task), and then somehow transfer the learned knowledge to facilitate training on a second dataset (the target task). As the source task must be a large dataset, the most common choice is using ImageNet, a classification dataset containing more than 1 million natural images organized into 1000 classes, as a source.” [Section I. Introduction]; Moradi: “In this example, one of the trained segmentation models (n=16) is used as the source of features, and the corresponding 100 testing images from 10 patients are used as the negative samples. For each test, the negative samples were randomly divided into two equal sized (50) sets for training and testing the classification network. For the 30 positive samples of each disease, they were randomly divided into 10 samples for training and 20 samples for testing.” [0041]).
Claims 2 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Mormont in view of Finnie and Moradi, as applied to Claims 1 and 14 above, in further view of Xue et al. (“Full left ventricle quantification via deep multitask relationships learning”, published 09/28/2017), hereinafter Xue. Xue was cited in a previous Office Action.
Regarding Claim 2, Mormont, Finnie, and Moradi teach the method of Claim 1, wherein the second machine-learned model is trained to classify cardiac disease (Moradi: “The segmentation network is trained on normal images only, but produces features for both normal and diseased cases. Significant gains are shown in positive detection rate using these feature maps on a classification network compared to using the original images with the same network. Examples provided herein include cases with two cardiac diseases that are each detected on a different binary classifier. It will be appreciated that the present disclosure is applicable to additional disease types by combining multiple binary classifiers, or by training a multiclass disease detector.” [0045].
However, they fail to expressly disclose wherein the first machine-learned model is trained for prediction of ejection fraction.
In the same field of endeavor, Xue teaches wherein the first machine-learned model is trained for prediction of ejection fraction (Xue: “Accurate quantification of left ventricle (LV) from cardiac imaging is among the most clinically important and most frequently demanded tasks for identification and diagnosis of cardiac diseases” [Introduction]; “Existing efforts in cardiac quantification have been limited to one index only, i.e, the LV cavity area,1 which relates to ejection fraction and is the easiest index to estimate among the above mentioned ones” [Introduction]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated wherein the first machine-learned model is trained for prediction of ejection fraction, as taught by Xue to the method of Mormont, Finnie, and Moradi because both of these methods are directed towards the diagnosis and management of cardiac diseases through deep learning and multi-task neural networks. As ejection fraction is a feature that is measured when diagnosing cardiac diseases, specifying the application of the method of Mormont, Finnie, and Moradi to predict the cardiac disease based on the specific feature of ejection fraction falls within the scope of the method of Mormont, Finnie, and Moradi.
Regarding Claim 15, it is a method claim that recites similar corresponding limitations to those of Claim 2. Therefore, it is rejected for the same reasons as Claim 2 above.
Response to Arguments
The Examiner acknowledges the Applicant’s amendments to Claims 1-4, 10, 14, and 20.
Applicant’s arguments, filed 02/20/2025, regarding the rejections of Claims 2-3, 5, and 8-11 under 35U.S.C. § 112(b) have been fully considered and are persuasive. The rejections have been withdrawn.
Applicant’s arguments A and C, filed 09/23/2025, regarding the rejections of Claims 1-4, 10, 12-15, 17, and 20 under 35 U.S.C. § 103 have been fully considered and are found moot in light of the new grounds of rejection (see rejection above). Argument B has been fully considered but is not persuasive.
Applicant alleges, on Page 9 of the Remarks, that Mormont’s disclosure of applying “simple random vertical and horizontal flips as well as extraction of a random square crop” is basic data augmentation that does not encompass the required limitations of previously rejection Claims 9 and 10. Examiner respectfully disagrees. Though the applicant is entitled to act as their own lexicographer, the strategies for generating synthetic data samples mentioned in the specification, for example GANs or mask transformations, are simply recited in the specification as desirable alternatives or specific embodiments and the specification does not suggest that the generation of synthetic data samples is specifically defined by one of these methods, thereby failing to constitute a deviation from the plain meaning of "synthetic examples" in the field of machine learning. As such, it would be improper to import these limitations from the specification into the claims and the recited claim limitations must be given their plain meaning under BRI consistent with the specification, which is "additional training data... derived from the data samples of actual patients and/or simulation" that augment the training data.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MEGAN E HWANG whose telephone number is (703)756-1377. The examiner can normally be reached Monday-Thursday 10:00AM-7:30PM ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Jennifer Welch can be reached at (571) 272-7212. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/M.E.H./Examiner, Art Unit 2143
/JENNIFER N WELCH/Supervisory Patent Examiner, Art Unit 2143
Prosecution Insights