DETAILED ACTION
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 .
Specification
The abstract of the disclosure is objected to because the first sentence can be implied and its removal would promote clarity in addition to changing the second sentence to begin as “A method” instead of “The method” after removal of the first sentence.
A corrected abstract of the disclosure is required and must be presented on a separate sheet, apart from any other text. See MPEP § 608.01(b).
Claim Objections
Claims 1, 11 and 19 are objected to because of the following informalities: “different sample removal models” should be changed to “different sample removal models of the plurality of sample removal models” to clarify the relationship between the “different preset window range” and the “plurality of sample removal models”.
Claim 10 is objected to because of the following informalities: the term “any” should be removed from the preamble to improve clarity.
Claim 19 is objected to because of the following informalities: “A non-transitory computer-readable storage medium, having at least one program stored herein” should be changed to “A non-transitory computer-readable storage medium, having at least one program stored thereon” to improve clarity.
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
Claims 1-20 are rejected under 35 U.S.C. 112(b) as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor regards as the invention.
Claim 1 recites, “obtaining a reference image and a corresponding artifact image, the reference image being an image generated by scanning a sample test object without an implant, the artifact image being a reference image comprising an artifact, and the artifact being a shadow of the implant during scanning ... determining predicted loss values respectively corresponding to the plurality of sample removal models based on pixel differences between the artifact removal results and the reference image” (emphasis added). Claim 1 includes two reference images, the first “reference image being an image generated by scanning a sample test object without an implant” and the second reference image being “the artifact image”. It is unclear which reference image is used to determine the predicted loss values based on pixel differences with the artifact removal results. Thus, “the reference image” in lines 13-14 of claim 1 lacks antecedent basis. Claims 11 and 19 recite substantially similar limitations as claim 1 and are indefinite for the same reasons. For purposes of applying prior art, “the reference image” in lines 13-14 of claim 1 is assumed to refer to the “reference image being an image generated by scanning a sample test object without an implant”. Claims 2-10 are rejected for inheriting and not curing the deficiencies of claim 1. Claims 12-18 are rejected for inheriting and not curing the deficiencies of claim 11. Claim 20 is rejected for inheriting and not curing the deficiencies of claim 19.
Claim 1 recites, “inputting the artifact image into a plurality of sample removal models to obtain artifact removal results corresponding to the artifact image respectively output by the plurality of sample removal models, different sample removal models corresponding to different preset window ranges, and the sample removal model being configured for removing the artifact in the artifact image based on a corresponding preset window range” (emphasis added). It is unclear which removal model amongst the “plurality of sample removal models” and “different sample removal models” corresponds to “the sample removal model”. Thus, “the sample removal model” lacks antecedent basis. Claims 11 and 19 recite substantially similar limitations as claim 1 and are indefinite for the same reasons. For purposes of applying prior art, “the sample removal model” is assumed to refer to each of the “plurality of sample removal models”. Claims 2-10 are rejected for inheriting and not curing the deficiencies of claim 1. Claims 12-18 are rejected for inheriting and not curing the deficiencies of claim 11. Claim 20 is rejected for inheriting and not curing the deficiencies of claim 19.
Claim 1 recites, “generate weight parameters respectively corresponding to the plurality of predicted loss values, the weight parameter being configured for performing weight adjustment on a parameter update of the sample removal model” (emphasis added). It is unclear which weight parameter amongst the “weight parameters” corresponds to “the weight parameter” because it is unclear how the parameters respectively correspond to the predicted loss values. Thus, “the weight parameter” lacks antecedent basis. Claims 11 and 19 recite substantially similar limitations as claim 1 and are indefinite for the same reasons. For purposes of applying prior art, “the weight parameter” is interpreted as “the weight parameters”. Claims 2-10 are rejected for inheriting and not curing the deficiencies of claim 1. Claims 12-18 are rejected for inheriting and not curing the deficiencies of claim 11. Claim 20 is rejected for inheriting and not curing the deficiencies of claim 19.
Claim 1 recites, “obtain an artifact removal model comprising a plurality of artifact removal sub-models, the artifact removal sub-model” (emphasis added). It is unclear which artifact removal sub-model amongst the “plurality of artifact removal sub-models” corresponds to “the artifact removal sub-model”. Thus, “the artifact removal sub-model” lacks antecedent basis. Claims 11 and 19 recite substantially similar limitations as claim 1 and are indefinite for the same reasons. For purposes of applying prior art, “the artifact removal sub-model” is interpreted as “each [[the]] artifact removal sub-model of the plurality of artifact removal sub-models”. Claims 2-10 are rejected for inheriting and not curing the deficiencies of claim 1. Claims 12-18 are rejected for inheriting and not curing the deficiencies of claim 11. Claim 20 is rejected for inheriting and not curing the deficiencies of claim 19.
Claim 1 recites, “determining predicted loss values respectively corresponding to the plurality of sample removal models based on pixel differences between the artifact removal results and the reference image; inputting the predicted loss values respectively corresponding to the plurality of sample removal models into a sample weight model to generate weight parameters respectively corresponding to the plurality of predicted loss values, the weight parameter being configured for performing weight adjustment on a parameter update of the sample removal model; and training the plurality of sample removal models based on the predicted loss values” (emphasis added). The number of “respectively corresponding” predicted loss values is not necessarily the same as the “plurality of sample removal models”. Thus, it is unclear which of “the predicted loss values” correspond to the “plurality of predicted loss values” and therefore “the plurality of predicted loss values” lacks antecedent basis. Claims 11 and 19 recite substantially similar limitations as claim 1 and are indefinite for the same reasons. For purposes of applying prior art, “the plurality of predicted loss values” is interpreted as “the Claims 2-10 are rejected for inheriting and not curing the deficiencies of claim 1. Claims 12-18 are rejected for inheriting and not curing the deficiencies of claim 11. Claim 20 is rejected for inheriting and not curing the deficiencies of claim 19.
Claim 7 recites, “determining corresponding mapping relationships between the first model parameters and the second model parameters” (emphasis added). Claims 1, 5, 6 and 7 do not recite any “first model parameters”. Therefore, “the first model parameters” lacks antecedent basis. Claim 17 recites substantially similar limitations as claim 7 and is indefinite for the same reasons. For purposes of applying prior art, “the first model parameters” in the above-cited portion of claim 7 and the corresponding portion of claim 17 are interpreted as “[[the]] first model parameters”.
Claim 8 recites, “the adjustment effect condition being representing a limitation on the predicted loss value” (emphasis added). Claim 1 recites more than one predicted loss value. Therefore, “the predicted loss value” lacks antecedent basis. Claim 18 recites substantially similar limitations as claim 8 and is indefinite for the same reasons. For purposes of applying prior art, ““the predicted loss value” is interpreted as ““the predicted loss values”.
Claim 9 recites, “performing gradient descent on the first learning rate based on the first learning attenuation rate to obtain a target learning rate corresponding to the artifact removal model” (emphasis added). This phrase is confusing for two reasons. First, performing gradient descent “on” a learning rate is confusing. Gradient descent is typically performed on model parameters, while the learning rate is a hyperparameter that controls the step size. It is therefore confusing how gradient descent is perform “on” the “first learning rate”. For purposes of applying prior art, “on the first learning rate” is interpreted as “[[on]] using the first learning rate”. Second, gradient descent does not typically operate to produce another learning rate. It is therefore confusing how the “target learning rate” is obtained though “performing gradient descent”. For purposes of applying prior art, the step of obtaining the target learning rate is interpreted as a separate step from the step of performing gradient descent at or using the first learning rate.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Multiple Window Learning for CT Metal Artifact Reduction to Niu et al. (hereinafter “Niu”) in view of Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting to Shu et al. (hereinafter “Shu”).
Regarding claim 1, Niu teaches a method for training an artifact removal model performed by a computer device (The training and results obtained require a computer. See Niu at section 3 and 4), the method comprising (Niu, abstract, “we propose a multiple-window learning method for CT MAR. The basic idea of multiple-window learning is that the content of large HU values may help improve features of small HU values, and vice versa. Our method can precisely process multiple specified windows through simultaneously and interactively learning to remove metal artifacts within multiple windows.”):
obtaining a reference image (Artifact free training image. See Niu at section 3) and a corresponding artifact image (Training image including metal artifact. See Niu at section 3), the reference image being an image generated by scanning a sample test object without an implant (Niu, pg. 2, “ADN17 was proposed to use clinical unpaired artifact-affected and artifact-free CT images to train a disentanglement network. ... we designed to improve the performance of ADN under the low-dimensional manifold constraint”, section 3, “We used three datasets corresponding to three locations: spine, teeth, and hip, which contain 136 slices (75 with metal artifacts and 61 without metal artifacts), 152 slices (90 with metal artifacts and 62 without metal artifacts), and 100 slices (20 with metal artifacts and 80 without metal artifacts) respectively.” Example metal artifacts include “spine implants”. See Niu at section 1, page 1.), the artifact image being a reference image comprising an artifact, and the artifact being a shadow (e.g., streak) of the implant (a metallic object positioned inside a patient is an implant) during scanning (Niu, section 1, “The main reason for such metal artifacts is that the metal objects in the field of view strongly attenuate x-rays or even completely block them so that reconstructed images from the compromised/incomplete data are corrupted in various ways, which usually show as bright or dark streaks. As a result, CT metal artifacts significantly affect image analysis and subsequent clinical treatment.”. Streaks (shadows) are artifacts. See Niu at Figure 1.);
inputting the artifact image (Niu, Figure 1, “input” image) into a plurality of sample removal models (Niu, Figure 1, “CNN-1”, “CNN-2”, “CNN-3”) to obtain artifact removal results corresponding to the artifact image respectively output by the plurality of sample removal models (Niu, Figure 1, the “HU: [-1000, 2000]” image output by “CNN-1”, the “HU: [-320, 480]” image output by “CNN-2”, and the “HU: [-160, 240]” image output by “CNN-3”), different sample removal models corresponding to different preset window ranges (Niu, Figure 1, “HU: [-1000, 2000]”, “HU: [-320, 480]”, and “HU: [-160, 240]”), and the sample removal model being configured for removing the artifact in the artifact image based on a corresponding preset window range (Niu, abstract, “learning to remove metal artifacts within multiple windows.”);
determining predicted loss values respectively corresponding to the plurality of sample removal models (The overall training loss is calculated as the summed L1 loss of the wth branch. See Niu at equation (2). The parameter w identifies a particular branch/window. Equation (2) is a sum over each branch-specific L1 loss for each predicted image. In the case of three branches, w = 3, and L = L1 + L2 + L3) based on pixel differences between the artifact removal results and the reference image (Niu, section 2, “
I
w
o
u
t
is the predicted image in the wth branch, and
I
w
g
t
is the corresponding ground truth”, equation (2), “|
I
w
o
u
t
-
I
w
g
t
|”. The images are made of arrays of pixel values and the difference between the images is the difference between their respective pixel values.);
generate weight parameters respectively corresponding to the plurality of predicted loss values (Niu, section 4, “the proposed MWLNet method clipped each of the multiple windows and normalized pixel values to the same range so that small HU values in a smaller window received similar weights as large HU values in a larger window in the final loss.”), the weight parameter being configured for performing weight adjustment on a parameter update of the sample removal model (Gradient-based optimization through backpropagation adjusts the internal weights of each CNN. See Niu at section 2, page 3); and
training the plurality of sample removal models based on the predicted loss values and the weight parameters to obtain an artifact removal model comprising a plurality of artifact removal sub-models (Training for MWLNet is performed through gradient-based optimization and backpropagation. See Niu at section 2, page 3), the artifact removal sub-model being configured for performing artifact removal on a target image based on a corresponding preset window range (Results include the trained model (and its CNN sub-models) performing artifact removal for three different preset window ranges. See Niu at Figures 4 and 5.), but does not teach that which is explicitly taught by Shu.
Shu teaches inputting predicted loss value of a DNN-based model (Shu, section 1, pg. 2, “Due to the universal approximation capability of this weight net, it can finely fit a wide range of weighting functions including those used in conventional research.”, section 2.1, pg. 3, “The optimal parameter w is calculated by minimizing the following weighted loss: [equation (1)] ... The optimal parameter Θ* can be obtained by minimizing the following meta-loss: [equation (2)]”.) into a sample weight model (Shu - The architecture of MW-Net (Meta-Weight-Net) is shown in Figure 1(c) with loss as the input and weight as the output. The learning is performed iteratively using gradient ).
Niu discloses a differentiable cascaded network (MWLNet) producing multiple window-specific outputs and corresponding loss terms for a plurality of models, which are summed with equal weight during joint end-to-end training. Niu further discloses identifying insufficient relative weighting of HU information as a problem in training such a network through gradient-based optimization. training problem. See Niu at section 1, page 2. Thus, Niu shows that it was known in the art before the effective filing date of the claimed invention to train each model having a different target range of window values with a respective loss function (L = L1 + L2 + L3 for w = 1, 2, 3, where the value), which is analogous to the claimed invention in that it is pertinent to the problem being solved by the claimed invention, preventing loss contributions from causing useful information to dominate or be overwhelmed during neural network training. Shu discloses a learned loss-to-weight function (Shu, section 2.1, “V(Li(w);Θ)”) parameterized by the Meta-Weight-Net MLP that receives an individual scalar training loss, outputs a corresponding scalar weight, and uses the weighted loss in gradient-based update of a deep neural network. A gradient-based update of the MW-Net parameters is performed with equation (4) to obtain the current sample-weight model and the primary classifier is updated along the descent direction of the loss using equation (5) and SGD (stochastic gradient descent) to thereby use the current sample-weight model to adjust the parameters of the primary model. See Shu at page 4. Shu further discloses that the gradient computations can be generalized to any deep learning classifier network. See Shu at section 2.2, page 4. Thus, Shu shows that it was known in the art before the effective filing date of the claimed invention to adaptively weight loss contributions during neural network training using a sample weight model to prevent useful information from receiving the wrong amount of weight/influence, which is analogous to the claimed invention in that it is pertinent to the problem being solved by the claimed invention, preventing loss contributions from causing useful information to dominate or be overwhelmed during neural network training.
A person of ordinary skill in the art would have been motivated to insert Shu’s MW-Net into Niu’s training loss path by adding Shu’s loss-to-weight MLP after Niu calculates the three losses for the different preset HU ranges and updating Niu’s CNNs and the added MW-Net based on Shu’s Algorithm 1, to thereby replace Niu’s unweighted loss terms with weighted loss terms using Shu’s loss-to-weight MLP to calculate the weight of each respective loss term as a weighted sum, which would then be backpropagated through Niu’s cascaded CNN structure to update the parameters of each CNN through stochastic gradient descent. Based on the foregoing, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have made such modification according to known methods to yield the predictable results to have the benefit of preventing one branch from overly or underly-influencing another and adaptively balancing the influence of each objective.
Regarding claim 2, Niu in view of Shu teaches the method according to claim 1, wherein the training the plurality of sample removal models based on the predicted loss values and the weight parameters to obtain an artifact removal model comprising a plurality of artifact removal sub-models comprises:
determining weighted (Shu, section 2.1, “V(Li(w);Θ)”) loss values respectively corresponding to the plurality of sample removal models (The loss of each branch forms a respective loss value. See Niu at equation (2).) based on the plurality of predicted loss values and weight parameters respectively corresponding to a plurality of loss values (A gradient-based update of the MW-Net parameters is performed with equation (4) to obtain the current sample-weight model. See Shu at page 4.);
respectively adjusting first model parameters (A first iteration of training) of the plurality of sample removal models based on the weighted loss values respectively corresponding to the plurality of sample removal models to obtain the plurality of artifact removal sub-models (Shu, page 4, “As general network training tricks, we employ SGD to optimize the training loss”. The primary classifier/model is updated along the descent direction of the loss using equation (5) and SGD (stochastic gradient descent) to thereby use the current sample-weight model to adjust the parameters of the primary model. See Shu at page 4.); and
using the plurality of artifact removal sub-models as the artifact removal model (Generating an output. See Niu at Figure 1).
The rationale for obviousness is the same as provided for claim 1.
Regarding claim 3, Niu in view of Shu teaches the method according to claim 2, wherein the determining weighted loss values based on the predicted loss values and the weight parameters comprises:
during sth training iteration (Shu, equation (4), s = “t+1” or “t” = s - 1), determining weighted loss values (weight*L) corresponding to sth training iteration based on a predicted loss value (Shu, section 2.1, pg. 3, “imposing weight ... on the i-th sample loss”) obtained in (s–1)th (Shu, equation (4), “t”.) training iteration and weight parameters obtained in the sth training iteration (Shu - The update for a current cycle of the training loop is based on the previous cycle’s parameter state wt(Θ). See lines 5-7 of Algorithm 1. The parameter state depends on the weight parameters determined by V. See equation (3).); and
the adjusting first model parameters of the sample removal models based on the weighted loss values to obtain the artifact removal model comprises:
performing, based on weighted loss values corresponding to the sth training iteration (Shu’s equation (5) uses the MW-Net parameter Θ. See Shu at page 4), gradient adjustment on first model parameters obtained in the (s–1)th training iteration of the sample removal models to obtain first model parameters corresponding to the sth training iteration (Shu, page 4, “As general network training tricks, we employ SGD to optimize the training loss”. The primary classifier/model is updated along the descent direction of the loss using equation (5) and SGD (stochastic gradient descent) to thereby use the current sample-weight model to adjust the parameters of the primary model. See Shu at page 4.), and performing (s+1)th cyclical adjustment (i.e., repeatedly cycling through an iterative training loop) until training of the artifact removal model ends, s being an integer greater than or equal to 1 (The next adjustment cycle, i.e., t+2, until the loop ends. See Shu at page 4, Algorithm 1.).
The rationale for obviousness is the same as provided for claim 1.
Regarding claim 4, Niu in view of Shu teaches the method according to claim 3, wherein the inputting the predicted loss values into a sample weight model to generate weight parameters comprises:
inputting the predicted loss values obtained in the (s–1)th training iteration (Shu, section 2.2, pg. 4, equation (5). The first value of V is the loss calculated using the preceding primary model parameter state w(t).) into a sample weight model obtained in the sth training iteration to generate weight parameters corresponding to the sth training iteration (Shu, section 2.2, pg. 4, equation (5). The first value of V is the loss calculated using the preceding primary model parameter state w(t). Θ(t + 1), the second value of V, is the newly-updated MW-Net parameter state. The output of V is the current sample weight used to update w(t) into w(t + 1).).
The rationale for obviousness is the same as provided for claim 1.
Regarding claim 5, Niu in view of Shu teaches the method according to claim 1, wherein before the inputting the predicted loss values into a sample weight model to generate weight parameters, the method further comprises:
obtaining a verification reference image and a verification artifact image with matching image content (Niu, section 3, “To obtain the paired training images, CatSim19 was used to simulate metal artifacts by inserting metals into the artifact-free images.”);
inputting the verification artifact image into a plurality of sample removal models (See Niu, pages 4-5, Results section) to respectively generate verification removal results corresponding to the verification artifact image (See Niu, Figures 4 and 5 which depict results applied to clinical hip images and clinical tooth images respectively.), but does not teach that which is explicitly further taught by Shu.
Shu further teaches determining verification loss values respectively corresponding to the plurality of sample removal models based on pixel differences between the verification removal results and the verification reference image (Shu, pg. 2, “use a small unbiased validation set (meta-data) to guide the training of all its parameters”, section 2.2, “in each iteration of training, a mini-batch of training samples ... is sampled”); and
training the sample weight model based on the verification loss values (Shu - See equation (3) in Algorithm 1.).
Niu and Shu are analogous to the claimed invention for the reasons provided above. Shu further discloses reserving an unbiased meta-data set to be processed with a provisionally-updated primary model (equation (3)), determining meta losses from the primary model’s predictions and ground truth, and using those losses to update the MW-Net parameters using equation (4). Thus, Shu further shows that it was known in the art before the effective filing date of the claimed invention to determine verification loss values corresponding to the primary model to train the sample weight model, which is analogous to the claimed invention in that it is pertinent to the problem being solved by the claimed invention, preventing loss contributions from causing useful information to dominate or be overwhelmed during neural network training.
A person of ordinary skill in the art would have been motivated to train the parameters of the sample-weight model of Niu in view of Shu by using the supplied image-pairs of artifact and artifact free images disclosed by Niu as a relatively unbiased meta-data set to update the sample-weight model as further disclosed by Shu, to thereby provide a supervisory signal to train the sample-weight model. Based on the foregoing, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have made such modification according to known methods to yield the predictable results to have the benefit of more precise sample weights.
Regarding claim 6, Niu in view of Shu teaches the method according to claim 5, wherein the training the sample weight model based on the verification loss value comprises:
during sth training iteration, performing gradient adjustment on second model parameters of the sample weight model based on verification loss values obtained in the (s–1)th training iteration to obtain a sample weight model corresponding to the sth training iteration (Shu’s equation (4) updates the MW-Net parameters by performing a gradient adjustment on the model parameters of Θ based on the verification/meta losses determined using the provisionally updated primary model from equation (3), thereby obtaining the updated sample-weight model parameterized by Θ(t +1). ).
The rationale for obviousness is the same as provided for claim 5.
Regarding claim 7, Niu in view of Shu teaches the method according to claim 6, wherein before the performing gradient adjustment on second model parameters of the sample weight model based on verification loss values obtained in the (s–1)th training iteration to obtain a sample weight model corresponding to the sth training iteration, the method further comprises:
determining corresponding mapping relationships between the first model parameters and the second model parameters during the (s–1)th training iteration based on the first model parameters obtained in the (s–1)th training iteration (Shu’s Algorithm 1, specifically line 5 and equation (3), define a mapping between the current primary model parameters w(t) and the sample-weight model parameters Θ.); and
determining the verification loss values obtained in the (s–1)th training iteration based on the mapping relationships (Shu’s Algorithm 1, equation (4), calculates the meta/verification losses using the above mapping.).
The rationale for obviousness is the same as provided for claim 5.
Regarding claim 8, Niu in view of Shu teaches the method according to claim 1, further comprising:
determining a first model parameter (Shu, Algorithm 1, lines 5-7 determine model parameters.) obtained in a most recent adjustment as a first parameter in response to a number of cyclical iterative adjustment times of the first model parameter reaching a number-of-times threshold (Shu, Algorithm 1, “for t = 0 to T - 1 do ... end for”).
The rationale for obviousness is the same as provided for claim 5.
Regarding claim 9, Niu in view of Shu teaches the method according to claim 1, but does not teach that which is explicitly further taught by Shu.
Shu further teaches obtaining a first learning attenuation rate (The schedule for the learning rate is a stepwise decay that begins with an initial learning rate of 0.1, then is later divided by 10, and again divided by 10 after that. See Shu at section 4.1), the first learning attenuation rate being configured for adjusting a first learning rate in a form of attenuation based on a number of iterations (Shu, section 4.1, “The learning rate of ResNet-32 is divided by 10 after 80 and 90 epoch (for a total 100 epochs)”), the first learning rate being a preset update step for training the plurality of sample removal models (The learning rate begins at a preset value. See Shu at section 4.1); and
during training for the plurality of sample removal models, performing gradient descent on the first learning rate based on the first learning attenuation rate (Shu, section 4.1, “We trained ResNet-32 [61] with softmax cross-entropy loss by SGD with a momentum 0.9, a weight decay 5×10−4, an initial learning rate 0.1. The learning rate of ResNet-32 is divided by 10 after 80 and 90 epoch (for a total 100 epochs), and the learning rate of WN-Net is fixed as 10−5.”) to obtain a target learning rate (Shu - a final learning rate after the decay) corresponding to the artifact removal model (Shu - The training completes after the final epoch and the resulting parameters are the trained model(s)).
Niu and Shu are analogous to the claimed invention for the reasons provided above. Shu further discloses a stepwise decay (step decay) learning rate schedule for the primary model and a fixed learning rate for the WN-Net model. Thus, Shu further shows that it was known in the art before the effective filing date of the claimed invention to use a step decay learning rate schedule when training a neural-network for a primary image-processing task that works together with a sample-weight model at a fixed learning rate, which is analogous to the claimed invention in that it is pertinent to the problem being solved by the claimed invention, preventing loss contributions from causing useful information to dominate or be overwhelmed during neural network training.
A person of ordinary skill in the art would have been motivated to train the parameters of the sample-weight model (via the parameters of the sub-models) of Niu in view of Shu by using the step decay learning rate schedule further disclosed by Shu, to thereby give the models large parameter updates early in training and smaller, more precise updates as the number of epochs increments towards a maximum number to mark the end of training. Based on the foregoing, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have made such modification according to known methods to yield the predictable results to have the benefit of preserving the simultaneous and interactive multi-window training while allowing large early parameter adjustment and progressively finer late-stage refinement without large updates disrupting relationships already learned among the branches.
Regarding claim 10, Niu in view of Shu teaches the method according to any claim 1, further comprising:
determining a window range corresponding to an ith sample removal model, i being a positive integer (Niu - determining the HU window associated with the second stage, [-320, 480], for example. See Niu at Figure 1.); and
performing window conversion on the artifact image (Niu - the original artifact-affected CT image is converted to the smaller window range of the second branch via WT [-320, 480] and CNN-2) and an (i–1)th artifact removal result (Niu - the output of CNN-1. See Figure 1) to obtain a window conversion result (WT [-320, 480]. See Niu at Figure 1) corresponding to both the artifact image and the (i–1)th artifact removal result as a model input of the ith sample removal model (WT [-320, 480] receives the original artifact-affected image and output of CNN-1 as inputs. See Niu at Figure 1).
Claims 11-18 substantially correspond to claims 1-8 by reciting a computer device (The training and results obtained require a computer. See Niu at section 3 and 4), comprising a processor and a memory (A computer includes a processor and storage for data and instructions. See Niu at section 3 and 4), the memory storing at least one program, the at least one program being loaded and executed by the processor to implement the method of claims 1-8.
The rationale for obviousness of each of claims 11-18 is the same as provided for each corresponding claim 1-8.
Claims 19 and 20 substantially correspond to claims 1 and 2 by reciting a non-transitory computer-readable storage medium (A computer includes a processor and storage for data and instructions. See Niu at section 3 and 4), having at least one program stored herein, the at least one program being loaded and executed by a processor to implement the method of claims 1 and 2.
The rationale for obviousness of each of claims 19 and 20 is the same as provided for each corresponding claim 1 and 2.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to RYAN P POTTS whose telephone number is (571)272-6351. The examiner can normally be reached M-F, 9am-5pm EST.
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, Sumati Lefkowitz can be reached at 571-272-3638. 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.
/RYAN P POTTS/Examiner, Art Unit 2672
Prosecution Insights