DETAILED ACTION
This Office action is responsive to communications filed on 04/21/2026. Claims 1-2, 9-1, & 13 have been amended. Presently, Claims 1-13 remain pending and are hereinafter examined on the merits.
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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 04/21/2026 has been entered.
Priority
Acknowledgment is made of applicant's claim for foreign priority based on an application filed in Japan on 10/06/2023. It is noted, however, that applicant has not filed a certified copy of the JP2023-174423 application as required by 37 CFR 1.55.
Further note; Should applicant desire to obtain the benefit of foreign priority under 35 U.S.C. 119(a)-(d) prior to declaration of an interference, a certified English translation of the foreign application must be submitted in reply to this action. 37 CFR 41.154(b) and 41.202(e).
Failure to provide a certified translation may result in no benefit being accorded for the non-English application.
The priority documents filed on 02/19/2026 are not a certified English translation of the foreign application.
Response to Arguments
Previous claim objections are withdrawn in view of the amendments filed on 04/21/2026.
Previous interpretations under 35 USC § 112(f) are not withdrawn in view of the amendments filed on 04/21/2026.
Applicant’s arguments with respect to claim(s) 1, 7, & 10 have been considered but are moot because the new ground of rejection does not rely on Zang et al applied under 35 USC § 102 in the prior rejection of record for any teaching or matter specifically challenged in the argument. The new grounds of rejection now relies on Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197), rejected under 35 USC § 102.
Applicant’s arguments with respect to claim(s) 9 & 13 have been considered but are moot because the new ground of rejection does not rely on Zang et al in view of Lazarus et al applied under 35 USC § 103 in the prior rejection of record for any teaching or matter specifically challenged in the argument. The new grounds of rejection now relies on Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197 ), in view of Zhao et al (Gibbs-ringing artifact suppression with knowledge transfer from natural images to MR images. Multimed Tools Appl 79, 33711–33733 (2020). https://doi.org/10.1007/s11042-019-08143-6) in view of Lazarus et al (US 2020/0058106 A1), rejected under 35 USC § 103.
Claim Interpretation
The following is a quotation of 35 U.S.C. 112(f):
(f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The claims in this application are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) is invoked.
As explained in MPEP § 2181, subsection I, claim limitations that meet the following three-prong test will be interpreted under 35 U.S.C. 112(f):
(A) the claim limitation uses the term “means” or “step” or a term used as a substitute for “means” that is a generic placeholder (also called a nonce term or a non-structural term having no specific structural meaning) for performing the claimed function;
(B) the term “means” or “step” or the generic placeholder is modified by functional language, typically, but not always linked by the transition word “for” (e.g., “means for”) or another linking word or phrase, such as “configured to” or “so that”; and
(C) the term “means” or “step” or the generic placeholder is not modified by sufficient structure, material, or acts for performing the claimed function.
Use of the word “means” (or “step”) in a claim with functional language creates a rebuttable presumption that the claim limitation is to be treated in accordance with 35 U.S.C. 112(f). The presumption that the claim limitation is interpreted under 35 U.S.C. 112(f), is rebutted when the claim limitation recites sufficient structure, material, or acts to entirely perform the recited function.
Absence of the word “means” (or “step”) in a claim creates a rebuttable presumption that the claim limitation is not to be treated in accordance with 35 U.S.C. 112(f). The presumption that the claim limitation is not interpreted under 35 U.S.C. 112(f), is rebutted when the claim limitation recites function without reciting sufficient structure, material or acts to entirely perform the recited function.
Conversely, claim limitations in this application that do not use the word “means” (or “step”) are not being interpreted under 35 U.S.C. 112(f) except as otherwise indicated in an Office action. This application includes one or more claim limitations that do not use the word “means,” but are nonetheless being interpreted under 35 U.S.C. 112(f) because the claim limitations use a generic placeholder that is coupled with functional language without reciting sufficient structure to perform the recited function and the generic placeholder is not preceded by a structural modifier. Such claim limitations are:
The nonce term “unit” in the phrase “imaging unit” for collecting measurement data” is used in claim(s) 1 & 9, invokes 35 USC 112(f)
The term, “unit” is a non-structural generic placeholder that does not include any specific structure for performing the accompany functions. See MPEP 2181.I.A: The following is a list of non-structural generic placeholders that may invoke 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, paragraph 6: "mechanism for," "module for," "device for," "unit for," "component for," "element for," "member for," "apparatus for," "machine for," or "system for." Welker Bearing Co., v. PHD, Inc., 550 F.3d 1090, 1096, 89 USPQ2d 1289, 1293-94 (Fed. Cir. 2008); Massachusetts Inst. of Tech. v. Abacus Software, 462 F.3d 1344, 1354, 80 USPQ2d 1225, 1228 (Fed. Cir. 2006); Personalized Media, 161 F.3d at 704, 48 USPQ2d at 1886–87; Mas-Hamilton Group v. LaGard, Inc., 156 F.3d 1206, 1214-1215, 48 USPQ2d 1010, 1017 (Fed. Cir. 1998).
Because these claim limitations are being interpreted under 35 U.S.C. 112(f) they are being interpreted to cover the corresponding structure described in the specification as performing the claimed function, and equivalents thereof.
If applicant does not intend to have these limitations interpreted under 35 U.S.C. 112(f) applicant may: (1) amend the claim limitations to avoid them being interpreted under 35 U.S.C. 112(f) (e.g., by reciting sufficient structure to perform the claimed function); or (2) present a sufficient showing that the claim limitations recite sufficient structure to perform the claimed function so as to avoid it/them being interpreted under 35 U.S.C. 112(f).
Please note that for the purposes of this examination the phrase “imaging unit” is being interpreted to include generic MRI apparatus as described in ¶0039, ‘Since the configuration of the imaging unit 10 is the same as a configuration of a normal MRI apparatus’ in the specification as performing the claimed function, and equivalents thereof.
Claim Objections
The following claims are objected to because of the following informalities and should recite:
Claim 1:
line 5, “one or more processors configured to[[, for]]process the measurement data collected by the imaging unit:’
Claim 9:
line 5, “one or more processors configured to[[, for]]process the measurement data collected by the imaging unit:’
Claim 13:
line 9, ”applying [[a]]the Convolutional Neural Network (CNN),”.
line 10, “[[an]]the input image”.
Appropriate correction is needed.
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 1, 7 & 10 are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197).
Claim 1: Wang discloses, A magnetic resonance imaging apparatus (¶Abstract) comprising:
an imaging unit configured to collect, using a measurement matrix, measurement data comprising nuclear magnetic resonance signals; and one or more processors configured to, for the measurement data collected by the imaging unit:
-[III. Experiments and Results A. Implementation details, pg. 2-3], Wang collects measurement data using an imaging unit (SIEMENS 3T Trio scanner) to gather k-space data in a T2-weighted gradient-echo GRE sequence, using the acquisition parameters of a matrix size 256 x 256 x 192. The processing hardware to execute the steps were implemented on the N VIDIA TITAN X GPU.
(i) generate, from the measurement data, a correction target image by zero-filling the measurement data in k-space to expand the measurement data to a reconstruction matrix size and performing an inverse Fourier transform on the zero-filled measurement data;
-[III. Experiments and Results A. Implementation details, pg. 2-3] - “ image of size 512 × 512 was used as ground-truth. 60 images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size. 90% of the images were used for training and the remaining 10% were used for testing. Each image was split to patches of the size of 70 × 70 and there were roughly 140, 000 patches in the training dataset.” – see also ¶Abstract & FIG. 3-4. Wang details how the input images (i.e., the correction target images containing artifacts) are created. Wang notes that “images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size.”. By taking truncated K-space data and padding it with zeros to reach the full matrix size before reconstruction, the system creates the image containing the ringing artifacts. The result of this process refers to as zero-filling reconstruction. Reconstructing an image from k-space requires applying a inverse Fourier transform (IFT).
(ii) apply a Convolutional Neural Network (CNN), trained to output an image in which ringing of an input image is corrected, to the correction target image to generate a CNN output image;
-[pg. 2 Merge], Wang teaches this data merging technique by explaining that after the CNN processes the image, “After training, CNN can map images with Gibbs artifacts to images without Gibbs artifacts. To take the advantage of the sampled K-space data, the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data. Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.” & ¶Abstract, “images with Gibbs artifacts can be input into the trained network to get the Gibbs-free images.”. Note; the CNN of Wang is specifically trained to [Introduction / pg. 1 – right col] - “use[d] CNN to reduce Gibbs artifacts in MRI images. The network was trained to learn the nonlinear mapping between MRI images with and without Gibbs artifacts.”. Hence the Gibbs artifacts manifest as ringing around the sharp edges, this process as described directly describes applying a trained CNN to correct ringing in the input image and outputting a CNN output image, see also [pg. 1 Forward-passing training].
(iii) generate composite k-space data by performing a Fourier transform on the CNN output image to obtain k-space data of the CNN output image and replacing a part of the k-space data of the CNN output image with the measurement data in k-space; and
-Wang teaches that after the CNN process the image, [pg. 2 Merge] - “the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data” to take advantage of the sampled k-space data by creating a composite k-space matrix that merges the CNN generated corrections with the original measurement data.
(iv) generate a ringing-corrected image by performing an inverse Fourier transform on the composite k-space data.
-Wang teaches after the CNN output and original sampled data are merged in the frequency domain, an ¶Abstract, “inverse Fourier transform is applied to the merged K-space to get the final image”, [pg. 2 Merge], “Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.”, [pg. 5 / Conclusion], “This paper proposed a method to reduce Gibbs artifacts in MRI images. CNN was trained to map images with Gibbs artifacts to images without Gibbs artifacts. Afterwards, the CNN-output image was merged with original sampled data to obtain the final image. The experimental results show that the proposed method can effectively reduce Gibbs artifacts and preserve image details at the same time, and thus improve the image quality obviously.” Hence, this operation brings the composite data back into the image domain, resulting in an image where Gibbs artifacts are reduced while the image details are preserved.
Claim 7: Wang discloses all the elements above in claim 1, Wang discloses, wherein the CNN includes a trained model that has been trained to obtain a ringing correction effect for at least a one-dimensional direction of the input image.
-Wang discloses, ¶Abstract [Introduction / pg. 1 – right col], [II. Theory pg. 1-2] & [A. implementation details pg 2], that the CNN includes a trained model, the CNN is trained with a batch of image pairs with and without Gibbs artifacts. This is before the CNN is used on the real-world data, the network specifically, uses under sampled k-space data to learn the mapping relationships between these images with and without Gibbs artifacts. To create the training dataset, “ground truth”. This ground truth image is a fully sampled image that acts as the correct answer and does not contain Gibbs artifacts. Wang states, [A. Implementation Details / pg 2], “We used 30% of phase encoding lines in the K-space data to obtain images with Gibbs artifacts. 90% of images were used for training and the remaining 10% were used for testing.”, see also ¶Abstract. The phase encoding lines represent one-dimensional directional axis with the measurement matrix. By truncating the data along these phase encoding lines, the ringing artifacts along that directional axis are used to train the model.
Claim 10: Wang discloses, An image processing method (¶Abstract) of correcting ringing in measurement data consisting of nuclear magnetic resonance signals collected by a magnetic resonance imaging apparatus (¶Abstract), the image processing method comprising:
-[III. Experiments and Results A. Implementation details, pg. 2-3], Wang collects measurement data using an imaging unit (SIEMENS 3T Trio scanner) to gather k-space data in a T2-weighted gradient-echo GRE sequence, using the acquisition parameters of a matrix size 256 x 256 x 192. The processing hardware to execute the steps were implemented on the N VIDIA TITAN X GPU.
generating, from the measurement data, a correction target image by zero-filling the measurement data in k-space to expand the measurement data to a reconstruction matrix size and performing an inverse Fourier transform on the zero-filled measurement data;
-[III. Experiments and Results A. Implementation details, pg. 2-3] - “ image of size 512 × 512 was used as ground-truth. 60 images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size. 90% of the images were used for training and the remaining 10% were used for testing. Each image was split to patches of the size of 70 × 70 and there were roughly 140, 000 patches in the training dataset.” – see also ¶Abstract & FIG. 3-4. Wang details how the input images (i.e., the correction target images containing artifacts) are created. Wang notes that “images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size.”. By taking truncated K-space data and padding it with zeros to reach the full matrix size before reconstruction, the system creates the image containing the ringing artifacts. The result of this process refers to as zero-filling reconstruction. Reconstructing an image from k-space requires applying a IFT.
applying a Convolutional Neural Network (CNN), trained to output an image in which ringing of an input image is corrected, to the correction target image to generate a CNN output image;
-[pg. 2 Merge], Wang teaches this data merging technique by explaining that after the CNN processes the image, “After training, CNN can map images with Gibbs artifacts to images without Gibbs artifacts. To take the advantage of the sampled K-space data, the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data. Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.” & ¶Abstract, “images with Gibbs artifacts can be input into the trained network to get the Gibbs-free images.”. Note; the CNN of Wang is specifically trained to [Introduction / pg. 1 – right col] - “use[d] CNN to reduce Gibbs artifacts in MRI images. The network was trained to learn the nonlinear mapping between MRI images with and without Gibbs artifacts.”. Hence the Gibbs artifacts manifest as ringing around the sharp edges, this process as described directly describes applying a trained CNN to correct ringing in the input image and outputting a CNN output image, see also [pg. 1 Forward-passing training].
generating composite k-space data by performing a Fourier transform on the CNN output image to obtain k-space data of the CNN output image and replacing a part of the k-space data of the CNN output image with the measurement data in k-space; and
-Wang teaches that after the CNN process the image, [pg. 2 Merge] - “the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data” to take advantage of the sampled k-space data by creating a composite k-space matrix that merges the CNN generated corrections with the original measurement data.
generating a ringing-corrected image by performing an inverse Fourier transform on the composite k-space data.
-Wang teaches after the CNN output and original sampled data are merged in the frequency domain, an ¶Abstract, “inverse Fourier transform is applied to the merged K-space to get the final image”, [pg. 2 Merge], “Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.”, [pg. 5 / Conclusion], “This paper proposed a method to reduce Gibbs artifacts in MRI images. CNN was trained to map images with Gibbs artifacts to images without Gibbs artifacts. Afterwards, the CNN-output image was merged with original sampled data to obtain the final image. The experimental results show that the proposed method can effectively reduce Gibbs artifacts and preserve image details at the same time, and thus improve the image quality obviously.” Hence, this operation brings the composite data back into the image domain, resulting in an image where Gibbs artifacts are reduced while the image details are preserved.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
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 3 & 5 are rejected under 35 U.S.C. 103 as being unpatentable over Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197), as applied to claim 1, in further view of Zhao et al (Gibbs-ringing artifact suppression with knowledge transfer from natural images to MR images. Multimed Tools Appl 79, 33711–33733 (2020). https://doi.org/10.1007/s11042-019-08143-6).
Claim 3: Wang discloses all the elements above in claim 1, Wang discloses, wherein the CNN includes a trained model that has been trained by using, as training data, a correct answer image in which ringing has not occurred
-Wang discloses, ¶Abstract [Introduction / pg. 1 – right col], [II. Theory pg. 1-2] & [A. implementation details pg 2], that the CNN includes a trained model, the CNN is trained with a batch of image pairs with and without Gibbs artifacts. This is before the CNN is used on the real-world data, the network specifically, uses undersampled k-space data to learn the mapping relationships between these images with and without Gibbs artifacts. To create the training dataset, “ground truth”. This ground truth image is a fully sampled image that acts as the correct answer and does not contain Gibbs artifacts.
Wang fails to explicitly disclose:
and an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region of a training image is smaller than a matrix size of the correct answer image.
However, Zhao in the context of Gibbs-ringing artifacts suppression using convolutional neural networks discloses: and an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region of a training image is smaller than a matrix size of the correct answer image.
-Zhao generates training data pairs by using artifact-free image (i.e., correct answer image) alongside simulated artifact-degraded image, FIG. 3 & [pg 33713], “To generate training examples, we follow the pipeline shown in Fig. 2 to simulate Gibbs-ringing artifacts (4-fold truncation in k-space), where Fig. 2a and d form a pair of training sample (y, x). It is worth noting that the artifact-degraded image x is generated by zero-padding the k-space to match the size of the artifact-free image y. However, the artifacts may appear in the images produced with or without zero-padding in real MRI scans. Zero-padding just makes the Gibbs-ringing artifacts more visible because the oscillation pattern of the artifacts is amplified by the zero-filling (i.e., sinc) interpolation [...]. In this work, we set the problem to the case of zero-padding [...] so that the artifacts in degraded images are more visible.” [4.1.1 NI dataset / pg 33719], “We follow the process shown in Fig. 2 to simulate the ringing artifacts for both natural and MR images. All the degraded images are produced by 4-fold truncation in k-space so that the image degradation is relatively serious. Therefore, the network recovers the artifact-free image from only 1/16 of the total information.”, [4.2 Training setting pg. 33719], “In our settings, the artifact-degraded image x is 4-fold truncated and zero-padded in k-space so that it has the same size as the output image . All training images are split into sub images of size 48 × 48 before feeding into the model. This is achieved by random extraction of paired artifact-degraded and artifact-free image patches. The data augmentation is realized by random horizontal flips and 90∘ rotations. The number of residual blocks and feature maps is set to 32 and 64 respectively, and kernel size is 3 × 3 for all conv layers. The proposed models are implemented and trained in TensorFlow 1.7.0, and the Adam optimizer [27] is used to minimize the loss function by setting β1 = 0.9, β2 = 0.999 and 𝜖 = 10− 8. Batch size is set to 16. We use piecewise constant decay for learning rate, i.e., it is initialized as 10− 4 for all layers and halved at every 2 × 105 iterations. All models are trained for one million iterations. The Xavier’s method [16] is used to initialize network parameters when the model is trained from scratch.”. The degraded image is created by simulating a 4-fold truncation in k-space, which corresponds to taking a smaller measurement matrix containing only 1/16 of the total original frequency information. To match the matrix size of the artifact-free correct answer image, thus truncated k-space data is then zero-padded before performing an IFT to produce the visible ringing artifacts. This entire process of generating the degraded image happens prior to applying the CNN. The resulting degraded image (i.e., Gibbs input) acts as the starting point that is directly fed into the first convolutional layer of the CNN. See 4.1-4.3.3.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the trained model of Wang such that an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region of a training image is smaller than a matrix size of the correct answer image as taught by Zhao. The motivation to do this yield predictable results such as improving the model performance as suggested by Zhao, pg. 4.3.3 pg 33721.
Claim 5: Wang discloses all the elements above in claim 1, Wang discloses that the CNN includes a trained model that has been trained, Wang discloses, ¶Abstract [Introduction / pg. 1 – right col], [II. Theory pg. 1-2] & [A. implementation details pg 2], that the CNN includes a trained model, the CNN is trained with a batch of image pairs with and without Gibbs artifacts. This is before the CNN is used on the real-world data, the network specifically, uses undersampled k-space data to learn the mapping relationships between these images with and without Gibbs artifacts. To create the training dataset, “ground truth”. This ground truth image is a fully sampled image that acts as the correct answer and does not contain Gibbs artifacts
Wang fails to explicitly discloses, wherein the CNN includes a trained model that has been trained to obtain a ringing correction effect for an input image in which a ratio between the size of the measurement matrix and the reconstruction matrix size is a predetermined value.
However, Zhao in the context of Gibbs-ringing artifacts suppression using convolutional neural networks discloses: wherein the CNN includes a trained model that has been trained to obtain a ringing correction effect for an input image in which a ratio between the size of the measurement matrix and the reconstruction matrix size is a predetermined value. (Specifically, Zhao teaches all artifact-degraded images used to train the model are produced by a 4-fold truncation in k-space, [4.1.1 NI dataset/pg. 33719], “All the degraded images are produced by 4-fold truncation in k-space so that the image degradation is relatively serious. Therefore, the network recovers the artifact-free image from only 1/16 of the total information.”, see also [pg 33713]. Because this data is 4-fold down sampled in both directions, this establishes a predetermined ratio where the network is trained to recover artifact-free image from only 1-16 of the total information.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the trained model of Wang such that it has been trained to obtain a ringing correction effect for an input image in which a ratio between the size of the measurement matrix and the reconstruction matrix size is a predetermined value as taught by Zhao. The motivation to do this yield predictable results such as improving the model performance as suggested by Zhao, pg. 4.3.3 pg 33721.
Claims 2 and 11 are rejected under 35 U.S.C. 103 as being unpatentable over Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197), as applied to claim 1 and 10, in further view of Nielsen et al (US 2024/0074671 A1).
Claim 2: Wang discloses all the elements above in claim 1, Wang fails to disclose: wherein the one or more processors are configured to repeat the function of correcting the ringing via the CNN and generating the composite k-space data at least twice.
However, Nielsen the context of Fourier-induced Gibbs ringing artifacts from MRI images discloses, wherein the one or more processors are configured to repeat the function of correcting the ringing via the CNN and generating the composite data at least twice. (¶0234, ‘The correcting unit may be further configured to iteratively improve the accuracy of the estimated local amplitudes and the quality of the ringing correction by re-estimating the local amplitudes of ringing artifacts and performing a ringing correction twice or multiple times.’)
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the one or more processors of modified Wang to be configured repeat the function of correcting the ringing via the CNN and generating the composite data at least twice as taught by Novikov. The motivation to do this yields predictable results such as improving the quality of the ringing correction, ¶0234 of Nielsen.
Claim 11: Wang discloses all the elements above in claim 10, Wang fails to disclose: wherein applying the CNN and generating the composite k-space data are repeated two or more times.
However, Nielsen the context of Fourier-induced Gibbs ringing artifacts from MRI images discloses, wherein applying the CNN and generating the composite k-space data are repeated two or more times. (¶0234, ‘The correcting unit may be further configured to iteratively improve the accuracy of the estimated local amplitudes and the quality of the ringing correction by re-estimating the local amplitudes of ringing artifacts and performing a ringing correction twice or multiple times.’)
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the one or more processors of modified Zhang to be configured to ringing correction processing via the CNN and combining processing for generating composite data at least twice as taught by Novikov. The motivation to do this yields predictable results such as improving the quality of the ringing correction, ¶0234 of Nielsen.
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197 ) in view of Zhao et al (Gibbs-ringing artifact suppression with knowledge transfer from natural images to MR images. Multimed Tools Appl 79, 33711–33733 (2020). https://doi.org/10.1007/s11042-019-08143-6), as applied to claim 3, in further view of Lazarus et al (US 2020/0058106 A1).
Claim 4: Wang as modified discloses all the elements above in claim 3, Wang fails to explicitly disclose: wherein the trained model has been trained by increasing a weight of an error in the measurement matrix region of the training image to be larger than a weight of an error in a region other than the measurement matrix region of the training image in a loss evaluation function during training.
However, Lazarus in the context of removing artifacts in MR data using CNN discloses, wherein the trained model has been trained by increasing a weight of an error in the measurement matrix region of the training image to be larger than a weight of an error in a region other than the measurement matrix region of the training image in a loss evaluation function during training.
-L2 weighted loss between output and target data as one of the loss function that may be employed for training the CNN, ¶0041, ‘generating an MR image from input MR data at least in part by using a neural network model (e.g., a model comprising one or more convolutional layers) to suppress at least one artefact in the input MR data’; ¶[0125-0132, ‘In some embodiments, one or a linear combination of multiple loss functions may be employed to train the neural network models described herein: [0126] L2 loss between output and target data [0127] L1 loss between output and target data [0128] L2 weighted loss between output and target data. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (the farther from the center of k-space), the higher the weight. Using such weights causes the resulting model to keep the high spatial frequencies which are noisier than the low frequencies [0129] L1 weighted regularization on the output. A sparse prior may be enforce on the output of the neural network by using the l.sub.1 norm, optionally after weighting. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (far from the center of k-space), the smaller the weight. This encourages sparsity. [0130] Generative Adversarial Nets loss [0131] Structured similarity index loss [0132] Any of the other loss functions described herein including in connection with FIGS. 1A-1E and 2A-2B.’
-The k-space coordinates used for applying these weights amounts to the measurement matrix region. K-space is the spatial frequency domain where the MR signals are acquired, which represent the raw measurements. Therefore, increasing the weight of errors in regions of higher spatial frequency (i.e., father from the center of k-space) during training, as it differentially weights errors based on their location within the measurement domain (k-space). This means errors in the outer, high frequency parts of k-space are given more important than errors in the inner, lower-frequency parts.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the trained model of modified Wang to include the teachings of Lazarus for the advantage of providing and improved apparatus for causing the model to keep the high frequencies which are noisier than the low frequencies, for capturing and preserving find anatomical details and sharp edges in the images, as suggested by Lazarus, ¶0100, ¶0123-0132.
Claims 6 are rejected under 35 U.S.C. 103 as being unpatentable over Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197) in view of Zhao et al (Gibbs-ringing artifact suppression with knowledge transfer from natural images to MR images. Multimed Tools Appl 79, 33711–33733 (2020). https://doi.org/10.1007/s11042-019-08143-6), as applied to claim 5, in further of Zhang Q et al ( MRI Gibbs-ringing artifact reduction by means of machine learning using convolutional neural networks. Magn Reson Med. 2019 ( a copy of which was made of record on 10/04/2024).
Claim 6: Wang as modified discloses all the elements above in claim 5, Wang fails to disclose: the following taught by Zhang in the context of MRI Gibbs ringing artifact reduction using convolutional neural networks, Zhang discloses, wherein the CNN includes a plurality of CNNs that have been trained to obtain a ringing correction effect for input images having different matrix ratios, and
-Zhang teaches a GRA-CNN model was trained and validated using data with a single under sampling of 50%, pg. 2136- 2.2.3| Performance evaluation, left col ¶1.-“One was trained and validated using data with a single undersampling ratio of 50%.”. This means the high-frequency k-space data was truncated to a predetermined level, affecting the “measurement matrix size” relative to the full reconstruction size.
-In another model, the GRA-CNN-m, was trained an validated using mixed data with varying undersampling ratios from 30% to 50%, pg. 2136- 2.2.3| Performance evaluation, left col ¶1.-“The other was trained and validated using mixed data with varying undersampling ratios from 30% to 50%.”. While varying, these were specific predetermined ratios (30%, 40%, 50%) used in the training data, (i.e., not a continuous changing ratio) during a single training instance, FIG. 8, pg 2139 3.5 | Brain images with varying undersampling levels.
the one or more processors are configured to select at least one of the plurality of CNNs to perform the function of correcting the ringing.
-pg. 2136 2.2.2 | Implementation details, “All experiments were implemented using MATLAB (MATLAB 2014b, Mathworks) on a 64-bit Windows 8 workstation (Intel Xeon CPU and 128 GB RAM).” The Intel Xeon CPU is a type of processor.
-For the test cases involving varying undersampling levels (30%, 40%, 50%), the GRA-CNN-m model was applied, which demonstrates removing ringing consistently across these levels; however, single identical undersampling level was slightly higher in PSNR reduction, “the real undersampling level, i.e., the appearance of Gibbs artifact in MR images, may not be accurately obtained. To address this problem, we trained and validated a GRA-CNN-m model by using mixed data with multiple undersampling levels. This model could consistently remove ringing artifact in MR images under varying sampling levels from 30% to 50%. However, it resulted in a slight PSNR reduction compared with the model trained using the data of single identical undersampling level.”-pg.2142 | Discussion left col ¶2. This indicates a decision process, where depending on the undersampling level of the image, an appropriate, pre-trained CNN model is chosen for correction.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the CNN of modified Wang such that it includes a plurality of CNNs that have been trained to obtain a ringing correction effect for input images having different matrix ratios, and the one or more processors are configured to select at least one of the plurality of CNNs to perform the function of correcting the ringing as taught by Zhang. The motivation to do this yield predictable results such as improving the model performance as suggested by Zhang [Discussion pg. 2140-2143.
Claims 8 are rejected under 35 U.S.C. 103 as being unpatentable over Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197), as applied to claim 1, in further of Zhang Q et al ( MRI Gibbs-ringing artifact reduction by means of machine learning using convolutional neural networks. Magn Reson Med. 2019 ( a copy of which was made of record on 10/04/2024).
Claim 8: Wang as modified discloses all the elements above in claim 1, Wang fails to disclose: the following taught by Zhang in the context of MRI Gibbs ringing artifact reduction using convolutional neural networks, Zhang discloses, wherein the CNN includes a plurality of CNNs that have been trained to obtain a ringing correction effect for input images having different sampling patterns, for each sampling pattern of the measurement data, and
-Zhang teaches a GRA-CNN model was trained and validated using data with a single under sampling of 50%, pg. 2136- 2.2.3| Performance evaluation, left col ¶1.-“One was trained and validated using data with a single undersampling ratio of 50%.”. This means the high-frequency k-space data was truncated to a predetermined level, affecting the “measurement matrix size” relative to the full reconstruction size.
-In another model, the GRA-CNN-m, was trained an validated using mixed data with varying undersampling ratios from 30% to 50%, pg. 2136- 2.2.3| Performance evaluation, left col ¶1.-“The other was trained and validated using mixed data with varying undersampling ratios from 30% to 50%.”. While varying, these were specific predetermined ratios (30%, 40%, 50%) used in the training data, (i.e., not a continuous changing ratio) during a single training instance, FIG. 8, pg 2139 3.5 | Brain images with varying undersampling levels.
the one or more processors are configured to select at least one CNN of the plurality of CNNs to perform the function of correcting the ringing.
-pg. 2136 2.2.2 | Implementation details, “All experiments were implemented using MATLAB (MATLAB 2014b, Mathworks) on a 64-bit Windows 8 workstation (Intel Xeon CPU and 128 GB RAM).” The Intel Xeon CPU is a type of processor.
-For the test cases involving varying undersampling levels (30%, 40%, 50%), the GRA-CNN-m model was applied, which demonstrates removing ringing consistently across these levels; however, single identical undersampling level was slightly higher in PSNR reduction, “the real undersampling level, i.e., the appearance of Gibbs artifact in MR images, may not be accurately obtained. To address this problem, we trained and validated a GRA-CNN-m model by using mixed data with multiple undersampling levels. This model could consistently remove ringing artifact in MR images under varying sampling levels from 30% to 50%. However, it resulted in a slight PSNR reduction compared with the model trained using the data of single identical undersampling level.”-pg.2142 | Discussion left col ¶2. This indicates a decision process, where depending on the undersampling level of the image, an appropriate, pre-trained CNN model is chosen for correction.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the CNN of modified Wang such that it includes a plurality of CNNs that have been trained to obtain a ringing correction effect for input images having different sampling patterns, for each sampling pattern of the measurement data, and the one or more processors are configured to select at least one CNN of the plurality of CNNs to perform the function of correcting the ringing, as taught by Zhang. The motivation to do this yield predictable results such as improving the model performance as suggested by Zhang [Discussion pg. 2140-2143.
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197 ), as applied to claim 10, in further view of Zhao et al (Gibbs-ringing artifact suppression with knowledge transfer from natural images to MR images. Multimed Tools Appl 79, 33711–33733 (2020). https://doi.org/10.1007/s11042-019-08143-6) in view of Lazarus et al (US 2020/0058106 A1).
Claim 12: Wang discloses all the elements above in claim 10, Wang discloses: further comprising: a step of training the CNN using, as training data, a correct answer image in which ringing has not occurred
-Wang discloses, ¶Abstract [Introduction / pg. 1 – right col], [II. Theory pg. 1-2] & [A. implementation details pg 2], that the CNN includes a trained model, the CNN is trained with a batch of image pairs with and without Gibbs artifacts. This is before the CNN is used on the real-world data, the network specifically, uses undersampled k-space data to learn the mapping relationships between these images with and without Gibbs artifacts. To create the training dataset, “ground truth”. This ground truth image is a fully sampled image that acts as the correct answer and does not contain Gibbs artifacts.
Wang fails to explicitly disclose:
and an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region of a training image is smaller than a matrix size of the correct answer image,
However, Zhao in the context of Gibbs-ringing artifacts suppression using convolutional neural networks discloses: and an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region of a training image is smaller than a matrix size of the correct answer image,
-Zhao generates training data pairs by using artifact-free image (i.e., correct answer image) alongside simulated artifact-degraded image, FIG. 3 & [pg 33713], “To generate training examples, we follow the pipeline shown in Fig. 2 to simulate Gibbs-ringing artifacts (4-fold truncation in k-space), where Fig. 2a and d form a pair of training sample (y, x). It is worth noting that the artifact-degraded image x is generated by zero-padding the k-space to match the size of the artifact-free image y. However, the artifacts may appear in the images produced with or without zero-padding in real MRI scans. Zero-padding just makes the Gibbs-ringing artifacts more visible because the oscillation pattern of the artifacts is amplified by the zero-filling (i.e., sinc) interpolation [...]. In this work, we set the problem to the case of zero-padding [...] so that the artifacts in degraded images are more visible.” [4.1.1 NI dataset / pg 33719], “We follow the process shown in Fig. 2 to simulate the ringing artifacts for both natural and MR images. All the degraded images are produced by 4-fold truncation in k-space so that the image degradation is relatively serious. Therefore, the network recovers the artifact-free image from only 1/16 of the total information.”, [4.2 Training setting pg. 33719], “In our settings, the artifact-degraded image x is 4-fold truncated and zero-padded in k-space so that it has the same size as the output image . All training images are split into sub images of size 48 × 48 before feeding into the model. This is achieved by random extraction of paired artifact-degraded and artifact-free image patches. The data augmentation is realized by random horizontal flips and 90∘ rotations. The number of residual blocks and feature maps is set to 32 and 64 respectively, and kernel size is 3 × 3 for all conv layers. The proposed models are implemented and trained in TensorFlow 1.7.0, and the Adam optimizer [27] is used to minimize the loss function by setting β1 = 0.9, β2 = 0.999 and 𝜖 = 10− 8. Batch size is set to 16. We use piecewise constant decay for learning rate, i.e., it is initialized as 10− 4 for all layers and halved at every 2 × 105 iterations. All models are trained for one million iterations. The Xavier’s method [16] is used to initialize network parameters when the model is trained from scratch.”. The degraded image is created by simulating a 4-fold truncation in k-space, which corresponds to taking a smaller measurement matrix containing only 1/16 of the total original frequency information. To match the matrix size of the artifact-free correct answer image, thus truncated k-space data is then zero-padded before performing an IFT to produce the visible ringing artifacts. This entire process of generating the degraded image happens prior to applying the CNN. The resulting degraded image (i.e., Gibbs input) acts as the starting point that is directly fed into the first convolutional layer of the CNN. See 4.1-4.3.3.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the trained model of Wang such that an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region of a training image is smaller than a matrix size of the correct answer image, as taught by Zhao. The motivation to do this yield predictable results such as improving the model performance as suggested by Zhao, pg. 4.3.3 pg 33721.
Wang fails to disclose:
wherein, in the step of training the CNN, a weight of an error in the measurement matrix region is increased in a loss evaluation function during training.
However, Lazarus in the context of removing artifacts in MR data using CNN discloses, wherein, in the step of training the CNN, a weight of an error in the measurement matrix region is increased in a loss evaluation function during training.
-L2 weighted loss between output and target data as one of the loss function that may be employed for training the CNN, ¶0041, ‘generating an MR image from input MR data at least in part by using a neural network model (e.g., a model comprising one or more convolutional layers) to suppress at least one artefact in the input MR data’; ¶[0125-0132, ‘In some embodiments, one or a linear combination of multiple loss functions may be employed to train the neural network models described herein: [0126] L2 loss between output and target data [0127] L1 loss between output and target data [0128] L2 weighted loss between output and target data. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (the farther from the center of k-space), the higher the weight. Using such weights causes the resulting model to keep the high spatial frequencies which are noisier than the low frequencies [0129] L1 weighted regularization on the output. A sparse prior may be enforce on the output of the neural network by using the l.sub.1 norm, optionally after weighting. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (far from the center of k-space), the smaller the weight. This encourages sparsity. [0130] Generative Adversarial Nets loss [0131] Structured similarity index loss [0132] Any of the other loss functions described herein including in connection with FIGS. 1A-1E and 2A-2B.’
-The k-space coordinates used for applying these weights amounts to the measurement matrix region. K-space is the spatial frequency domain where the MR signals are acquired, which represent the raw measurements. Therefore, increasing the weight of errors in regions of higher spatial frequency (i.e., father from the center of k-space) during training, as it differentially weights errors based on their location within the measurement domain (k-space). This means errors in the outer, high frequency parts of k-space are given more important than errors in the inner, lower-frequency parts.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the CNN of modified Zhang to include the teachings of Lazarus for the advantage of providing and improved apparatus for causing the model to keep the high frequencies which are noisier than the low frequencies, for capturing and preserving find anatomical details and sharp edges in the images, as suggested by Lazarus, ¶0100, ¶0123-0132.
Claims 9 & 13 are rejected under 35 U.S.C. 103 as being unpatentable over Y. Wang et al, ("Reduction of Gibbs artifacts in magnetic resonance imaging based on Convolutional Neural Network," 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Shanghai, China, 2017, pp. 1-5, doi: 10.1109/CISP-BMEI.2017.8302197 ), in view of Zhao et al (Gibbs-ringing artifact suppression with knowledge transfer from natural images to MR images. Multimed Tools Appl 79, 33711–33733 (2020). https://doi.org/10.1007/s11042-019-08143-6) in view of Lazarus et al (US 2020/0058106 A1).
Claim 9: Wang discloses: A magnetic resonance imaging apparatus (¶Abstract) comprising:
an imaging unit configured to collect, using a measurement matrix, measurement data comprising nuclear magnetic resonance signals; and one or more processors configured to, for the measurement data collected by the imaging unit:
-[III. Experiments and Results A. Implementation details, pg. 2-3], Wang collects measurement data using an imaging unit (SIEMENS 3T Trio scanner) to gather k-space data in a T2-weighted gradient-echo GRE sequence, using the acquisition parameters of a matrix size 256 x 256 x 192. The processing hardware to execute the steps were implemented on the N VIDIA TITAN X GPU.
(i) generate, from the measurement data, a correction target image by zero-filling the measurement data in k-space to expand the measurement data to a reconstruction matrix size and performing an inverse Fourier transform on the zero-filled measurement data;
-[III. Experiments and Results A. Implementation details, pg. 2-3] - “ image of size 512 × 512 was used as ground-truth. 60 images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size. 90% of the images were used for training and the remaining 10% were used for testing. Each image was split to patches of the size of 70 × 70 and there were roughly 140, 000 patches in the training dataset.” – see also ¶Abstract & FIG. 3-4. Wang details how the input images (i.e., the correction target images containing artifacts) are created. Wang notes that “images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size.”. By taking truncated K-space data and padding it with zeros to reach the full matrix size before reconstruction, the system creates the image containing the ringing artifacts. The result of this process refers to as zero-filling reconstruction. Reconstructing an image from k-space requires applying a IFT.
(ii) apply a Convolutional Neural Network (CNN), trained to output an image in which ringing of an input image is corrected, to the correction target image to generate a CNN output image;
-[pg. 2 Merge], Wang teaches this data merging technique by explaining that after the CNN processes the image, “After training, CNN can map images with Gibbs artifacts to images without Gibbs artifacts. To take the advantage of the sampled K-space data, the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data. Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.” & ¶Abstract, “images with Gibbs artifacts can be input into the trained network to get the Gibbs-free images.”. Note; the CNN of Wang is specifically trained to [Introduction / pg. 1 – right col] - “use[d] CNN to reduce Gibbs artifacts in MRI images. The network was trained to learn the nonlinear mapping between MRI images with and without Gibbs artifacts.”. Hence the Gibbs artifacts manifest as ringing around the sharp edges, this process as described directly describes applying a trained CNN to correct ringing in the input image and outputting a CNN output image, see also [pg. 1 Forward-passing training].
(iii) generate composite k-space data by performing a Fourier transform on the CNN output image to obtain k-space data of the CNN output image and replacing a part of the k-space data of the CNN output image with the measurement data in k-space; and
-Wang teaches that after the CNN process the image, [pg. 2 Merge] - “the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data” to take advantage of the sampled k-space data by creating a composite k-space matrix that merges the CNN generated corrections with the original measurement data.
(iv) generate a ringing-corrected image by performing an inverse Fourier transform on the composite k-space data,
-Wang teaches after the CNN output and original sampled data are merged in the frequency domain, an ¶Abstract, “inverse Fourier transform is applied to the merged K-space to get the final image”, [pg. 2 Merge], “Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.”, [pg. 5 / Conclusion], “This paper proposed a method to reduce Gibbs artifacts in MRI images. CNN was trained to map images with Gibbs artifacts to images without Gibbs artifacts. Afterwards, the CNN-output image was merged with original sampled data to obtain the final image. The experimental results show that the proposed method can effectively reduce Gibbs artifacts and preserve image details at the same time, and thus improve the image quality obviously.” Hence, this operation brings the composite data back into the image domain, resulting in an image where Gibbs artifacts are reduced while the image details are preserved.
wherein the CNN has been trained by using, as training data, a correct answer image in which ringing has not occurred
-Wang discloses, ¶Abstract [Introduction / pg. 1 – right col], [II. Theory pg. 1-2] & [A. implementation details pg 2], that the CNN includes a trained model, the CNN is trained with a batch of image pairs with and without Gibbs artifacts. This is before the CNN is used on the real-world data, the network specifically, uses undersampled k-space data to learn the mapping relationships between these images with and without Gibbs artifacts. To create the training dataset, “ground truth”. This ground truth image is a fully sampled image that acts as the correct answer and does not contain Gibbs artifacts.
Wang fails to disclose:
and an image obtained performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region is smaller than a matrix size of the correct answer image,
However, Zhao in the context of Gibbs-ringing artifacts suppression using convolutional neural networks discloses: an image obtained performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region is smaller than a matrix size of the correct answer image,
-Zhao generates training data pairs by using artifact-free image (i.e., correct answer image) alongside simulated artifact-degraded image, FIG. 3 & [pg 33713], “To generate training examples, we follow the pipeline shown in Fig. 2 to simulate Gibbs-ringing artifacts (4-fold truncation in k-space), where Fig. 2a and d form a pair of training sample (y, x). It is worth noting that the artifact-degraded image x is generated by zero-padding the k-space to match the size of the artifact-free image y. However, the artifacts may appear in the images produced with or without zero-padding in real MRI scans. Zero-padding just makes the Gibbs-ringing artifacts more visible because the oscillation pattern of the artifacts is amplified by the zero-filling (i.e., sinc) interpolation [...]. In this work, we set the problem to the case of zero-padding [...] so that the artifacts in degraded images are more visible.” [4.1.1 NI dataset / pg 33719], “We follow the process shown in Fig. 2 to simulate the ringing artifacts for both natural and MR images. All the degraded images are produced by 4-fold truncation in k-space so that the image degradation is relatively serious. Therefore, the network recovers the artifact-free image from only 1/16 of the total information.”, [4.2 Training setting pg. 33719], “In our settings, the artifact-degraded image x is 4-fold truncated and zero-padded in k-space so that it has the same size as the output image . All training images are split into sub images of size 48 × 48 before feeding into the model. This is achieved by random extraction of paired artifact-degraded and artifact-free image patches. The data augmentation is realized by random horizontal flips and 90∘ rotations. The number of residual blocks and feature maps is set to 32 and 64 respectively, and kernel size is 3 × 3 for all conv layers. The proposed models are implemented and trained in TensorFlow 1.7.0, and the Adam optimizer [27] is used to minimize the loss function by setting β1 = 0.9, β2 = 0.999 and 𝜖 = 10− 8. Batch size is set to 16. We use piecewise constant decay for learning rate, i.e., it is initialized as 10− 4 for all layers and halved at every 2 × 105 iterations. All models are trained for one million iterations. The Xavier’s method [16] is used to initialize network parameters when the model is trained from scratch.”. The degraded image is created by simulating a 4-fold truncation in k-space, which corresponds to taking a smaller measurement matrix containing only 1/16 of the total original frequency information. To match the matrix size of the artifact-free correct answer image, thus truncated k-space data is then zero-padded before performing an IFT to produce the visible ringing artifacts. This entire process of generating the degraded image happens prior to applying the CNN. The resulting degraded image (i.e., Gibbs input) acts as the starting point that is directly fed into the first convolutional layer of the CNN. See 4.1-4.3.3.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the trained model of Wang such that an image obtained performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region is smaller than a matrix size of the correct answer image, as taught by Zhao. The motivation to do this yield predictable results such as improving the model performance as suggested by Zhao, pg. 4.3.3 pg 33721.
Wang fails to disclose:
and increasing a weight of an error in the measurement matrix region that to be larger than a weight of an error in a region other than the measurement matrix region in a loss evaluation function during training.
However, Lazarus in the context of removing artifacts in MR data using CNN discloses, increasing a weight of an error in the measurement matrix region that to be larger than a weight of an error in a region other than the measurement matrix region in a loss evaluation function during training.
-L2 weighted loss between output and target data as one of the loss function that may be employed for training the CNN, ¶0041, ‘generating an MR image from input MR data at least in part by using a neural network model (e.g., a model comprising one or more convolutional layers) to suppress at least one artefact in the input MR data’; ¶[0125-0132, ‘In some embodiments, one or a linear combination of multiple loss functions may be employed to train the neural network models described herein: [0126] L2 loss between output and target data [0127] L1 loss between output and target data [0128] L2 weighted loss between output and target data. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (the farther from the center of k-space), the higher the weight. Using such weights causes the resulting model to keep the high spatial frequencies which are noisier than the low frequencies [0129] L1 weighted regularization on the output. A sparse prior may be enforce on the output of the neural network by using the l.sub.1 norm, optionally after weighting. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (far from the center of k-space), the smaller the weight. This encourages sparsity. [0130] Generative Adversarial Nets loss [0131] Structured similarity index loss [0132] Any of the other loss functions described herein including in connection with FIGS. 1A-1E and 2A-2B.’
-The k-space coordinates used for applying these weights amounts to the measurement matrix region. K-space is the spatial frequency domain where the MR signals are acquired, which represent the raw measurements. Therefore, increasing the weight of errors in regions of higher spatial frequency (i.e., father from the center of k-space) during training, as it differentially weights errors based on their location within the measurement domain (k-space). This means errors in the outer, high frequency parts of k-space are given more important than errors in the inner, lower-frequency parts.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the CNN of modified Wang to include the teachings of Lazarus for the advantage of providing and improved apparatus for causing the model to keep the high frequencies which are noisier than the low frequencies, for capturing and preserving find anatomical details and sharp edges in the images, as suggested by Lazarus, ¶0100, ¶0123-0132.
Claim 13: Wang discloses: An image processing method (¶Abstract) of correcting ringing in measurement data comprising nuclear magnetic resonance signals collected by a magnetic resonance imaging apparatus (¶Abstract), the image processing method comprising:
-[III. Experiments and Results A. Implementation details, pg. 2-3], Wang collects measurement data using an imaging unit (SIEMENS 3T Trio scanner) to gather k-space data in a T2-weighted gradient-echo GRE sequence, using the acquisition parameters of a matrix size 256 x 256 x 192. The processing hardware to execute the steps were implemented on the N VIDIA TITAN X GPU.
training a Convolutional Neural Network (CNN) to generate an image in which ringing of an input image is corrected;
-[pg. 2 Merge], Wang teaches this data merging technique by explaining that after the CNN processes the image, “After training, CNN can map images with Gibbs artifacts to images without Gibbs artifacts. To take the advantage of the sampled K-space data, the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data. Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.” & ¶Abstract, “images with Gibbs artifacts can be input into the trained network to get the Gibbs-free images.”. Note; the CNN of Wang is specifically trained to [Introduction / pg. 1 – right col] - “use[d] CNN to reduce Gibbs artifacts in MRI images. The network was trained to learn the nonlinear mapping between MRI images with and without Gibbs artifacts.”. Hence the Gibbs artifacts manifest as ringing around the sharp edges, this process as described directly describes applying a trained CNN to correct ringing in the input image and outputting a CNN output image, see also [pg. 1 Forward-passing training].
generating, from the measurement data, a correction target image by zero-filling the measurement data in k-space to expand the measurement data to a reconstruction matrix size and performing an inverse Fourier transform on the zero-filled measurement data;
-[III. Experiments and Results A. Implementation details, pg. 2-3] - “ image of size 512 × 512 was used as ground-truth. 60 images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size. 90% of the images were used for training and the remaining 10% were used for testing. Each image was split to patches of the size of 70 × 70 and there were roughly 140, 000 patches in the training dataset.” – see also ¶Abstract & FIG. 3-4. Wang details how the input images (i.e., the correction target images containing artifacts) are created. Wang notes that “images with Gibbs artifacts were acquired by acquiring center K-space data (from 161 × 161 to 220 × 220 with step 1) with zero-filling to 512 × 512 matrix size.”. By taking truncated K-space data and padding it with zeros to reach the full matrix size before reconstruction, the system creates the image containing the ringing artifacts. The result of this process refers to as zero-filling reconstruction. Reconstructing an image from k-space requires applying a IFT.
applying a Convolutional Neural Network (CNN), trained to output an image in which ringing of an input image is corrected, to the correction target image to generate a CNN output image;
-[pg. 2 Merge], Wang teaches this data merging technique by explaining that after the CNN processes the image, “After training, CNN can map images with Gibbs artifacts to images without Gibbs artifacts. To take the advantage of the sampled K-space data, the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data. Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.” & ¶Abstract, “images with Gibbs artifacts can be input into the trained network to get the Gibbs-free images.”. Note; the CNN of Wang is specifically trained to [Introduction / pg. 1 – right col] - “use[d] CNN to reduce Gibbs artifacts in MRI images. The network was trained to learn the nonlinear mapping between MRI images with and without Gibbs artifacts.”. Hence the Gibbs artifacts manifest as ringing around the sharp edges, this process as described directly describes applying a trained CNN to correct ringing in the input image and outputting a CNN output image, see also [pg. 1 Forward-passing training].
generating composite k-space data by performing a Fourier transform on the CNN output image to obtain k-space data of the CNN output image and replacing a part of the k-space data of the CNN output image with the measurement data in k-space; and
-Wang teaches that after the CNN process the image, [pg. 2 Merge] - “the output image is Fourier transformed into K-space, and then all data that have been sampled are replaced with original sampled data” to take advantage of the sampled k-space data by creating a composite k-space matrix that merges the CNN generated corrections with the original measurement data.
a step of generating a ringing-corrected image by performing an inverse Fourier transform on the composite k-space data,
-Wang teaches after the CNN output and original sampled data are merged in the frequency domain, an ¶Abstract, “inverse Fourier transform is applied to the merged K-space to get the final image”, [pg. 2 Merge], “Finally, inverse Fourier Transformed is applied to the merged K-space to get the final image.”, [pg. 5 / Conclusion], “This paper proposed a method to reduce Gibbs artifacts in MRI images. CNN was trained to map images with Gibbs artifacts to images without Gibbs artifacts. Afterwards, the CNN-output image was merged with original sampled data to obtain the final image. The experimental results show that the proposed method can effectively reduce Gibbs artifacts and preserve image details at the same time, and thus improve the image quality obviously.” Hence, this operation brings the composite data back into the image domain, resulting in an image where Gibbs artifacts are reduced while the image details are preserved.
wherein, in the step of training the CNN, the CNN is trained by using, as training data, a correct answer image in which ringing has not occurred
-Wang discloses, ¶Abstract [Introduction / pg. 1 – right col], [II. Theory pg. 1-2] & [A. implementation details pg 2], that the CNN includes a trained model, the CNN is trained with a batch of image pairs with and without Gibbs artifacts. This is before the CNN is used on the real-world data, the network specifically, uses undersampled k-space data to learn the mapping relationships between these images with and without Gibbs artifacts. To create the training dataset, “ground truth”. This ground truth image is a fully sampled image that acts as the correct answer and does not contain Gibbs artifacts.
Wang fails to disclose: and an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region is smaller than a matrix size of the correct answer image,
However, Zhao in the context of Gibbs-ringing artifacts suppression using convolutional neural networks discloses: an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region is smaller than a matrix size of the correct answer image,
-Zhao generates training data pairs by using artifact-free image (i.e., correct answer image) alongside simulated artifact-degraded image, FIG. 3 & [pg 33713], “To generate training examples, we follow the pipeline shown in Fig. 2 to simulate Gibbs-ringing artifacts (4-fold truncation in k-space), where Fig. 2a and d form a pair of training sample (y, x). It is worth noting that the artifact-degraded image x is generated by zero-padding the k-space to match the size of the artifact-free image y. However, the artifacts may appear in the images produced with or without zero-padding in real MRI scans. Zero-padding just makes the Gibbs-ringing artifacts more visible because the oscillation pattern of the artifacts is amplified by the zero-filling (i.e., sinc) interpolation [...]. In this work, we set the problem to the case of zero-padding [...] so that the artifacts in degraded images are more visible.” [4.1.1 NI dataset / pg 33719], “We follow the process shown in Fig. 2 to simulate the ringing artifacts for both natural and MR images. All the degraded images are produced by 4-fold truncation in k-space so that the image degradation is relatively serious. Therefore, the network recovers the artifact-free image from only 1/16 of the total information.”, [4.2 Training setting pg. 33719], “In our settings, the artifact-degraded image x is 4-fold truncated and zero-padded in k-space so that it has the same size as the output image . All training images are split into sub images of size 48 × 48 before feeding into the model. This is achieved by random extraction of paired artifact-degraded and artifact-free image patches. The data augmentation is realized by random horizontal flips and 90∘ rotations. The number of residual blocks and feature maps is set to 32 and 64 respectively, and kernel size is 3 × 3 for all conv layers. The proposed models are implemented and trained in TensorFlow 1.7.0, and the Adam optimizer [27] is used to minimize the loss function by setting β1 = 0.9, β2 = 0.999 and 𝜖 = 10− 8. Batch size is set to 16. We use piecewise constant decay for learning rate, i.e., it is initialized as 10− 4 for all layers and halved at every 2 × 105 iterations. All models are trained for one million iterations. The Xavier’s method [16] is used to initialize network parameters when the model is trained from scratch.”. The degraded image is created by simulating a 4-fold truncation in k-space, which corresponds to taking a smaller measurement matrix containing only 1/16 of the total original frequency information. To match the matrix size of the artifact-free correct answer image, thus truncated k-space data is then zero-padded before performing an IFT to produce the visible ringing artifacts. This entire process of generating the degraded image happens prior to applying the CNN. The resulting degraded image (i.e., Gibbs input) acts as the starting point that is directly fed into the first convolutional layer of the CNN. See 4.1-4.3.3.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the trained model of Wang such that an image obtained by performing an inverse Fourier transform on k-space data in which a size of a measurement matrix region is smaller than a matrix size of the correct answer image, as taught by Zhao. The motivation to do this yield predictable results such as improving the model performance as suggested by Zhao, pg. 4.3.3 pg 33721.
Wang fails to disclose:
and increasing a weight of an error in the measurement matrix region in a loss evaluation function during training.
However, Lazarus in the context of removing artifacts in MR data using CNN discloses, increasing a weight of an error in the measurement matrix region in a loss evaluation function during training.
-L2 weighted loss between output and target data as one of the loss function that may be employed for training the CNN, ¶0041, ‘generating an MR image from input MR data at least in part by using a neural network model (e.g., a model comprising one or more convolutional layers) to suppress at least one artefact in the input MR data’; ¶[0125-0132, ‘In some embodiments, one or a linear combination of multiple loss functions may be employed to train the neural network models described herein: [0126] L2 loss between output and target data [0127] L1 loss between output and target data [0128] L2 weighted loss between output and target data. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (the farther from the center of k-space), the higher the weight. Using such weights causes the resulting model to keep the high spatial frequencies which are noisier than the low frequencies [0129] L1 weighted regularization on the output. A sparse prior may be enforce on the output of the neural network by using the l.sub.1 norm, optionally after weighting. The weights may be calculated based on the k-space coordinates. The higher the spatial frequency (far from the center of k-space), the smaller the weight. This encourages sparsity. [0130] Generative Adversarial Nets loss [0131] Structured similarity index loss [0132] Any of the other loss functions described herein including in connection with FIGS. 1A-1E and 2A-2B.’
-The k-space coordinates used for applying these weights amounts to the measurement matrix region. K-space is the spatial frequency domain where the MR signals are acquired, which represent the raw measurements. Therefore, increasing the weight of errors in regions of higher spatial frequency (i.e., father from the center of k-space) during training, as it differentially weights errors based on their location within the measurement domain (k-space). This means errors in the outer, high frequency parts of k-space are given more important than errors in the inner, lower-frequency parts.
It would have been obvious to one of ordinary skilled in the art before the effective filing date of the claimed invention to modify the CNN of modified Wang to include the teachings of Lazarus for the advantage of providing and improved apparatus for causing the model to keep the high frequencies which are noisier than the low frequencies, for capturing and preserving find anatomical details and sharp edges in the images, as suggested by Lazarus, ¶0100, ¶0123-0132.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Kamada et al (US 2014/0197835 A1) in the context of magnetic resonance artifacts causes by non-inform data density in k-space reduction teaches this principle, discloses, generate a ringing-corrected image by performing an inverse Fourier transform on the composite data. (¶0014, ‘a signal rearrangement step of rearranging the unit k-space data after correction in Cartesian coordinate system k-space; and a final imaging step of reconstructing an image by performing an inverse Fourier transform of data after rearrangement by the rearrangement unit’)
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Nicholas Robinson whose telephone number is (571)272-9019. The examiner can normally be reached M-F 9:00AM-5:00PM 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, Pascal Bui-Pho can be reached at (571) 272-2714. 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.
/N.A.R./Examiner, Art Unit 3798
/PASCAL M BUI PHO/Supervisory Patent Examiner, Art Unit 3798
Prosecution Insights