DETAILED ACTION
Response to Amendment
Claims 1-12 are pending. Claims 1-12 are amended directly or by dependency on an amended claim.
Response to Arguments
Applicant’s arguments, see pages 9-13, filed 6 August, 2026, with respect to the 35 USC 103 rejections of claims 1-12 along with accompanying amendments received on the same date, have been considered but are moot because the new ground of rejection does not rely on the combination of references applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
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.
Claim(s) 1, 4-7, 9 and 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Choi et al. (“StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation”) in view of Liu et al. (“FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space”).
Regarding claims 1 and 9, Choi et al. disclose a domain generalization method, applied to a server, the method comprising, and a domain generalization server including one or more processors and a computer readable storage medium storing one or more computer program instructions, when executed by the one of more processors, the computer program instructions implementing a domain generalization method ([0105], [0108]) comprising: acquiring image discrimination information uploaded from multiple clients, wherein the image discrimination information is obtained by discriminators in the clients, the discriminators evaluating, based on an initial image of the respective clients an enhanced image generated by a generator in the server, the initial image including amplitude information (“Several image datasets come with a number of labeled attributes. For instance, the CelebA[18] dataset contains 40 labels related to facial attributes such as hair color, gender,and age, and the RaFD [12] dataset has 8 labels for facial expressions such as ‘happy’, ‘angry’ and ‘sad’. These settings enable us to perform more interesting tasks, namely multi-domain image-to-image translation, where we change images according to attributes from multiple domains”, “We propose StarGAN, a novel generative adversarial network that learns the mappings among multiple domains using only a single generator and a discriminator, training effectively from images of all domains”, p8790, input image x, p8791, An important advantage of StarGAN is that it simultaneously incorporates multiple datasets containing different types of labels, so that StarGAN can control all the labels at the test phase, p8792) [image translation = enhanced image]; and updating the generator based on the image discrimination information to obtain an updated generator (As a solution to such problems we propose StarGAN, a generative adversarial network capable of learning mappings among multiple domains. As demonstrated in Fig. 2(b), our model takes in training data of multiple domains, and learns the mappings between all available domains using only one generator, p8790, “The generator G tries to minimize this objective, while the discriminator D tries to maximize it”, p8792) [learning/training with a loss function = updating]; and sending a multi-domain mixed image generated by the updated generator to the clients (Multi-Domain Image-to-Image Translation, title, StarGAN, a novel and scalable approach that can perform image-to-image translations for multiple domains using only a single model, abstract, translate an input image x into an output image y conditioned on the target domain label c, p8791); and determining domain generalization parameters for classifiers in each client based on model update parameters obtained after the clients update their classifiers using the initial image and the multi-domain mixed image (
PNG
media_image1.png
160
382
media_image1.png
Greyscale
, p8792).
While Choi et al. disclose inputting labeled images, which can be interpreted as “the initial image including amplitude information”, another reference is added to make this more explicit. Choi et al. do not explicitly disclose the initial image being locally stored in the respective clients and including amplitude information, wherein the generator generates the enhanced image without accessing the initial image locally stored in the respective clients. Choi et al. do not explicitly discuss a server.
Liu et al. teach a domain generalization server (The paradigm works in a way that each local client (e.g., hospital) learns from their own data, and only aggregates the model parameters at a certain frequency at the central server to generate a global model, in federated paradigm, data are stored distributedly and the learning at each client can only access its local data, enriching the established multi-domain distributions for each local client, part 1, central server then collects the local parameters ɵ k from all clients and aggregates them to update the global model, part 3.1) acquiring image discrimination information uploaded from multiple clients, wherein the image discrimination information is obtained by discriminators in the clients, the discriminators evaluating, based on an initial image of the respective clients, an enhanced image generated by a generator in the server (The central server then collects the local parameters ɵ k from all clients and aggregates them to update the global model, part 3.1
PNG
media_image2.png
260
444
media_image2.png
Greyscale
, transform its image appearance as Eqs. (2-3) To address the restriction of decentralized datasets, the
foundation of our solution is to exchange the distribution information across clients, such that each local client can get access to multi-source data distributions for learning generalizable
parameters, we can enrich the established multidomain distributions for each local client, benefiting from a dedicated dense space with smooth distribution changes, part 3.2) the initial image being locally stored in the respective clients and including amplitude information (
PNG
media_image3.png
272
480
media_image3.png
Greyscale
, Fig. 2,
PNG
media_image4.png
194
434
media_image4.png
Greyscale
, Practically, for each input
we will sample an amplitude spectrum, part 3.2), wherein the generator generates the enhanced image without accessing the initial image locally stored in the respective clients (Federated learning allows distributed medical institutions to collaboratively learn a shared prediction model with privacy protection, abstract,
PNG
media_image5.png
316
432
media_image5.png
Greyscale
, First, the multi-source data in FL are stored distributedly and the learning at each client can only access its individual local distribution, which constrains to make full use of the multi-source distributions to learn generalizable parameters, part 3.1, To address the restriction of decentralized datasets, the
foundation of our solution is to exchange the distribution information across clients, such that each local client can get access to multi-source data distributions for learning generalizable
parameters. Considering that sharing raw images is forbidden, we propose to exploit the information inherent in the frequency space, which enables to separate the distribution (i.e. style) information from the original images to be shared between clients without privacy leakage, part 3.2, We establish the local training as an episodic meta-learning scheme, which learns generalizable model parameters by simulating train/test domain shift explicitly, part 3.3) and determining domain generalization parameters for classifiers in each client based on model update parameters obtained after the clients update their classifiers using the initial image and the multi-domain mixed image (Domain generalization aims to learn a model from multiple source domains such that it can directly generalize to unseen target domains, part 2.2, At each federated round t, every client k will receive the same global model weights ɵ from the central server and update the model with their local data Sk for E epochs. The central server then collects the local parameters ɵ k from all clients and aggregates them to update the global model. This process repeats until the global model converges, part 3.1, calculate loss, part 3.3).
Choi et al. and Liu et al. are in the same art of domain generalization (Choi et al., abstract; Liu et al., abstract). The combination of Liu et al. with Choi et al. will enable using amplitude information and federated learning. It would have been obvious at the time of filing to one of ordinary skill in the art to combine the amplitude and federated learning of Liu et al. with Choi et al. as this was known at the time of filing, the combination would have predictable results, and as Liu et al. indicate “Data collaboration across multiple medical institutions is increasingly desired to build accurate and robust data-driven deep networks for medical image segmentation” and “To solve this FedDG problem, our insight is to enable each client to access multi-source data distributions in a privacy-protecting way” (part 1) thus likely providing an accuracy and security benefit to combining inventions.
Regarding claims 4 and 12, Choi et al. and Liu et al. disclose the method and server according to claims 3 and 11. Liu et al. further teach updating the generator based on the image discrimination information, to obtain the updated generator, includes: aggregating the image discrimination information obtained by the clients after evaluating the initial enhanced image, to obtain an aggregated discriminator loss result; and determining a generator loss result based on the aggregated discriminator loss result; and updating parameters of the generator based on the generator loss result to obtain an initially updated generator; and sending updated enhanced image generated by the initially updated generator to the clients; and if the updated enhanced image does not satisfy discriminator conditions in the clients, acquiring the image discrimination information corresponding to the updated enhanced image from the clients; and updating the parameters of the initially updated generator, based on the image discrimination information corresponding to the updated enhanced image, to obtain a further updated generator; and if the updated enhanced image satisfies the discriminator conditions in the clients, determining the further updated generator as the updated generator (
PNG
media_image6.png
470
440
media_image6.png
Greyscale
PNG
media_image7.png
78
448
media_image7.png
Greyscale
).
Regarding claim 5, Choi et al. disclose a domain generalization method, applied to a client, the method comprising: acquiring an initial image and a multi-domain mixed image, the initial image includes amplitude information and phase information and processing the multi-domain mixed image and the initial image to obtain a multi-domain image (
PNG
media_image8.png
294
762
media_image8.png
Greyscale
PNG
media_image9.png
594
396
media_image9.png
Greyscale
, p8791); and updating parameters of classifiers in the client based on the multi-domain image and the initial image to obtain the model update parameters, and sending the model update parameters to a server (
PNG
media_image1.png
160
382
media_image1.png
Greyscale
, p8792); and acquiring domain generalization parameters determined by the server based on the model update parameters (optimized hyperparameters, p8792, StarGAN generated images of higher visual quality compared to existing methods [15, 22, 32], owing to the generalization capability behind the multi-task learning setting, p8796); and updating the classifier parameters based on the domain generalization parameters to obtain a domain generalization model (As a solution to such problems we propose StarGAN, a generative adversarial network capable of learning mappings among multiple domains. As demonstrated in Fig. 2(b), our model takes in training data of multiple domains, and learns the mappings between all available domains using only one generator, p8790, “The generator G tries to minimize this objective, while the discriminator D tries to maximize it”, p8792).
While Choi et al. disclose inputting labeled images, which can be interpreted as “the initial image includes amplitude information and phase information”, Choi et al. do not explicitly disclose “acquiring an initial image and a multi-domain mixed image received from a server, the initial image includes amplitude information and phase information, wherein the multi-domain mixed image is generated by a generator in the server without accessing the initial image” therefore another reference is added.
Liu et al. teach acquiring an initial image and a multi-domain mixed image received from a server (The paradigm works in a way that each local client (e.g., hospital) learns from their own data, and only aggregates the model parameters at a certain frequency at the central
server to generate a global model, in federated paradigm, data are stored distributedly and the learning at each client can only access its local data, enriching the established multi-domain distributions for each local client, part 1, central server then collects the local parameters ɵ k from all clients and aggregates them to update the global model, part 3.1, established multidomain distributions for each local client, part 3.2), the initial image includes amplitude information and phase information (
PNG
media_image3.png
272
480
media_image3.png
Greyscale
, Fig. 2,
PNG
media_image4.png
194
434
media_image4.png
Greyscale
, Practically, for each input
we will sample an amplitude spectrum, part 3.2), wherein the multi-domain mixed image is generated by a generator in the server without accessing the initial image (Federated learning allows distributed medical institutions to collaboratively learn a shared prediction model with privacy protection, abstract,
PNG
media_image5.png
316
432
media_image5.png
Greyscale
, First, the multi-source data in FL are stored distributedly and the learning at each client can only access its individual local distribution, which constrains to make full use of the multi-source distributions to learn generalizable parameters, part 3.1, To address the restriction of decentralized datasets, the foundation of our solution is to exchange the distribution information across clients, such that each local client can get access to multi-source data distributions for learning generalizable parameters. Considering that sharing raw images is forbidden, we propose to exploit the information inherent in the frequency space, which enables to separate the distribution (i.e. style) information from the original images to be shared between clients without privacy leakage, part 3.2, We establish the local training as an episodic meta-learning scheme, which learns generalizable model parameters by simulating train/test domain shift explicitly, part 3.3) and acquiring domain generalization parameters determined by the server based on the model update parameters (Domain generalization aims to learn a model from multiple source domains such that it can directly generalize to unseen target domains, part 2.2, At each federated round t, every client k will receive the same global model weights ɵ from the central server and update the model with their local data Sk for E epochs. The central server then collects the local parameters ɵ k from all clients and aggregates them to update the global model. This process repeats until the global model converges, part 3.1).
Choi et al. and Liu et al. are in the same art of domain generalization (Choi et al., abstract; Liu et al., abstract). The combination of Liu et al. with Choi et al. will enable using amplitude information and federated learning. It would have been obvious at the time of filing to one of ordinary skill in the art to combine the amplitude and federated learning of Liu et al. with Choi et al. as this was known at the time of filing, the combination would have predictable results, and as Liu et al. indicate “Data collaboration across multiple medical institutions is increasingly desired to build accurate and robust data-driven deep networks for medical image segmentation” and “To solve this FedDG problem, our insight is to enable each client to access multi-source data distributions in a privacy-protecting way” (part 1) thus likely providing an accuracy and security benefit to combining inventions.
Regarding claim 6, Choi et al. and Liu et al. disclose the method according to claim 5. Liu et al. further indicate processing the multi-domain mixed image and the initial image to obtain the multi-domain image includes: performing interpolation calculation on the amplitude information obtained by Fourier decomposition of the initial image and the multi-domain mixed image to obtain an interpolated image; and performing inverse Fourier transform on the interpolated image and the phase information obtained by Fourier decomposition of the initial image to obtain the multi-domain image (
PNG
media_image10.png
258
430
media_image10.png
Greyscale
PNG
media_image11.png
192
454
media_image11.png
Greyscale
, part 3.2).
Regarding claim 7, Choi et al. and Liu et al. disclose the method according to claim 5. Choi et al. further indicate updating the classifier parameters in the client based on the multi-domain image and the initial image to obtain model update parameters, includes: classifying the multi-domain image and the initial image separately using the classifiers in the client to obtain a first classification result and a second classification result; and determining a classifier loss result based on the first classification result, the second classification result, and a classifier loss function; and updating the classifier parameters based on the loss result to obtain the model update parameters (
PNG
media_image12.png
226
410
media_image12.png
Greyscale
PNG
media_image13.png
998
394
media_image13.png
Greyscale
p8791-8792).
Claim(s) 2, 8, 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Choi et al. (“StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation”) and Liu et al. (“FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space”) as applied to claim 1, 7, and 9 above, further in view of Mohamed et al. (US 11544796 B1).
Regarding claims 2 and 10, Choi et al. and Liu et al. disclose the method and server according to claims 1 and 9. Choi et al. and Liu et al. do not disclose determining a number of samples of initial images corresponding to each client; and determining model weights of each client based on the number of samples from each client and a total number of samples from all clients; and determining the domain generalization parameters for the classifiers in each client, based on the model update parameters obtained after the clients update their classifiers using the initial image and the multi-domain mixed image, comprising: acquiring the model update parameters of each client after the client updates its classifiers using the initial image and the multi-domain mixed image; and performing weighted calculation on the model update parameters of each client based on the model weights of each client, to obtain the domain generalization parameters.
Mohamed et al. teach determining a number of samples of initial images corresponding to each client; and determining model weights of each client based on the number of samples from each client and a total number of samples from all clients; and determining the domain generalization parameters for the classifiers in each client, based on the model update parameters obtained after the clients update their classifiers using the initial image and the multi-domain mixed image, comprising: acquiring the model update parameters of each client after the client updates its classifiers using the initial image and the multi-domain mixed image; and performing weighted calculation on the model update parameters of each client based on the model weights of each client, to obtain the domain generalization parameters (The parameters (e.g., weights and/or biases) of the machine learning model may be updated to minimize (or maximize) the cost. For example, the machine learning model may use a gradient descent (or ascent) algorithm to incrementally adjust the weights to cause the most rapid decrease (or increase) to the output of the loss function, col. 2, lines 20-30
PNG
media_image14.png
404
334
media_image14.png
Greyscale
col. 8 lines 40 – 67
optimize the hyper-parameters for each domain-specific model independently (resulting in multiple machine learning models per cluster), Accordingly, the number of domains for a cluster may be a hyperparameter that may be optimized col. 9 lines 1-25).
While Mohamed et al. do not disclose “domain generalization parameters for the classifiers in each client”, as Choi et al. and Liu et al. indicate those concepts in claim 1 (see for instance Choi et al. classification of images, Fig. 3, domain classification, hyperparameters, p8792; Chen et al., regularize a model to improve its generalizability, [0041]-[0045], image classifier, [0072], hyperparameters, [0078]), together the references disclose the whole claim.
Choi et al. and Liu et al. and Mohamed et al. are in the same art of optimizing hyperparameters (Choi et al., p8792; Liu et al., part 3.3; Mohamed et al., col. 9 lines 1-25). The combination of Mohamed et al. with Choi et al. and Liu et al. will enable using weights of each client based on the number of samples. It would have been obvious at the time of filing to one of ordinary skill in the art to combine the weights of Mohamed et al. with Choi et al. and Liu et al. as this was known at the time of filing, the combination would have predictable results, and as Mohamed et al. indicate, “However, as can be seen in FIG. 1, various other domains (e.g., Domain N) have a relatively small number of samples available for machine learning model training relative to Domain 1. Accordingly, a machine learning model (and, in particular, deep learning models) trained for prediction using a relatively small amount of data available for that domain (e.g., Domain N) may suffer from over-fitting and may have relatively poor performance relative to a domain with more training samples available (e.g., Domain 1). Accordingly, the cross-domain learning techniques described herein can be used to increase performance of domains with relatively small amounts of training data available, as described below” (col. 4, lines 50-65) indicating a benefit of requiring less training data in combination.
Regarding claim 8, Choi et al. and Liu et al. disclose the method according to claim 7. Choi et al. further indicate classifying the multi-domain image and the initial image separately using the classifiers in the client to obtain the first classification result and the second classification result, comprising: classifying the multi-domain image and the initial image separately using the classifiers and the model weights, and outputting the first classification result and the second classification result, the first and second classification results carrying the model weights ((
PNG
media_image12.png
226
410
media_image12.png
Greyscale
PNG
media_image13.png
998
394
media_image13.png
Greyscale
p8791-8792).
Choi et al. and Liu et al. do not disclose acquiring model weights determined by the server, based on a number of samples of the initial images from each client.
Mohamed et al. teach acquiring model weights determined by the server, based on a number of samples of the initial images from each client (The parameters (e.g., weights and/or biases) of the machine learning model may be updated to minimize (or maximize) the cost. For example, the machine learning model may use a gradient descent (or ascent) algorithm to incrementally adjust the weights to cause the most rapid decrease (or increase) to the output of the loss function, col. 2, lines 20-30
PNG
media_image14.png
404
334
media_image14.png
Greyscale
col. 8 lines 40 – 67
optimize the hyper-parameters for each domain-specific model independently (resulting in multiple machine learning models per cluster), Accordingly, the number of domains for a cluster may be a hyperparameter that may be optimized col. 9 lines 1-25).
Choi et al. and Liu et al. and Mohamed et al. are in the same art of optimizing hyperparameters (Choi et al., p8792; Liu et al., part 3.3; Mohamed et al., col. 9 lines 1-25). The combination of Mohamed et al. with Choi et al. and Liu et al. will enable using weights of each client based on the number of samples. It would have been obvious at the time of filing to one of ordinary skill in the art to combine the weights of Mohamed et al. with Choi et al. and Liu et al. as this was known at the time of filing, the combination would have predictable results, and as Mohamed et al. indicate, “However, as can be seen in FIG. 1, various other domains (e.g., Domain N) have a relatively small number of samples available for machine learning model training relative to Domain 1. Accordingly, a machine learning model (and, in particular, deep learning models) trained for prediction using a relatively small amount of data available for that domain (e.g., Domain N) may suffer from over-fitting and may have relatively poor performance relative to a domain with more training samples available (e.g., Domain 1). Accordingly, the cross-domain learning techniques described herein can be used to increase performance of domains with relatively small amounts of training data available, as described below” (col. 4, lines 50-65) indicating a benefit of requiring less training data in combination.
Claim(s) 3 and 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Choi et al. (“StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation”) and Liu et al. (“FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space”) as applied to claims 1 and 9 above, further in view of Chen et al. (US 20250046068 A1).
Regarding claims 3 and 11, Choi et al. and Liu et al. disclose the method and server according to claims 1 and 9. Choi et al. and Liu et al. do not disclose acquiring random Gaussian noise; and generating initial enhanced image based on the generator and the random Gaussian noise, and sending the initial enhanced image to the clients.
Chen et al. teach acquiring random Gaussian noise; and generating initial enhanced image based on the generator and the random Gaussian noise, and sending the initial enhanced image to the clients (Gaussian augmentation, [0004], Perturbed images are generated by adding the Fourier basis noise, [0050]).
Choi et al. and Chen et al. are in the same art of using an adversarial network (Choi et al.,
abstract; Chen et al., abstract). The combination of Chen et al. with Choi et al. and Liu et al. will enable using noise. It would have been obvious at the time of filing to one of ordinary skill
in the art to combine the noise of Chen et al. with Choi et al. and Liu et al. as this was known at the time of filing, the combination would have predictable results, and as Chen et al. indicate “Such adversarial operation of the image amplitude spectrum allows augmenting samples with more appearance variation in most cases. The generated samples 224 are then mixed with the
original samples 212 to finetune the original model and generate SADA trained model 228.
Compared with other methods, SADA trained models, such as the SADA trained model 228, present less sensitivity to domain shift across the frequency space, thus guaranteeing better
generalization performance. Based on such observation, a quantitative measure is disclosed,
which helps predict model generalizability.” ([0045]) and “Aiming to suppress the model
spectral sensitivity, exemplary embodiments instead adversarially augment the image amplitude
spectrum. Exemplary embodiments were experimentally compared to typical baselines under a
single-DG setting and were found to demonstrate superior performance.” ([0048]) suggesting a
performance benefit that will result from combining inventions.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MICHELLE ENTEZARI whose telephone number is (571)270-5084. The examiner can normally be reached 10-7 M-F.
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, Vincent M Rudolph can be reached at (571) 272-8243. 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.
/MICHELLE M ENTEZARI HAUSMANN/Primary Examiner, Art Unit 2671
Prosecution Insights