DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 07/02/2024 and 10/18/2025 is/are compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Office Action Summary
Claim(s) 1-7, 9-12, 14-16, and 18-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Hendriksen et al (Noise2Inverse: Self-supervised deep convolutional denoising for tomography) in view of Han et al (Self-Supervised Noise Reduction in Low-Dose Cone Beam Computed Tomography (CBCT) Using the Randomly Dropped Projection Strategy) and Sonke et al (Respiratory correlated cone beam CT), further in view of Madesta et al (Self-contained deep learning-based boosting of 4D cone-beam CT reconstruction).
Claim(s) 8, 13, and 17 is/are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1-7, 9-12, 14-16, and 18-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Hendriksen et al (Noise2Inverse: Self-supervised deep convolutional denoising for tomography) in view of Han et al (Self-Supervised Noise Reduction in Low-Dose Cone Beam Computed Tomography (CBCT) Using the Randomly Dropped Projection Strategy) and Sonke et al (Respiratory correlated cone beam CT), further in view of Madesta et al (Self-contained deep learning-based boosting of 4D cone-beam CT reconstruction).
Regarding claim(s) 1, 19, and 20, Hendriksen teaches a computer implemented method for training an unsupervised Machine Learning (ML) model to reduce noise and aliasing artefacts in a reconstructed four-dimensional (4D) medical image of a patient, wherein the ML model comprises a plurality of trainable parameters, and wherein the method comprises:
i. wherein the obtained 2D projections comprise measurement noise (Page 9, C. Experimental data, 1st Paragraph: “The dataset, Dorthe_F_002, was acquired at the Advanced Photon Source at Argonne National Laboratory, and contained 900 noisy projection images […]”; Page 14, 2nd Paragraph: “Noise2Inverse is well-suited to imaging modalities that permit trading acquisition speed for measurement noise, as it aims to remove measurement noise […]”; and Page 13, V. Discussion, 3rd Paragraph: “As in Noise2Noise, the network is presented with two noisy images during training. In Noise2Inverse, however, these images are sub-sampled reconstructions, and since the artifacts arising from sub-sampling the data are correlated […]”);
ii. repeating, for a plurality of iterations (Page 9, C. Experimental data, 2nd Paragraph: “For Noise2Inverse, an MS-D network was trained with the X:1 strategy and 4 splits for 100 epochs […]”; and Page 7, 5th Paragraph: “The networks were trained for 100 epochs using the ADAM algorithm […]”):
selecting two non-empty, mutually disjoint subsets that form a partition of the projection set (Page 1, Right Col., 5th Paragraph: “In the proposed Noise2Inverse approach, the training regime explicitly takes into account the structure of the noise in the inverse problem. In its simplest form, our method splits the measured data in two parts, from which two reconstructions are computed”; and Page 14, Left Col., Last Paragraph: “[…] The number of parts in which the measured data is split, however, deserves more nuance: when the projection angles are under-sampled, the results indicate that two parts yield the best results”);
for each of the two selected subsets, using a linear reconstruction algorithm to reconstruct a volumetric image of the training patient from the 2D projections contained in the subset (Page 1, Right Col., 5th Paragraph: “In the proposed Noise2Inverse approach, the training regime explicitly takes into account the structure of the noise in the inverse problem. In its simplest form, our method splits the measured data in two parts, from which two reconstructions are computed”); and
wherein the method further comprises:
iii. using training pairs from the training data set to update values of the trainable parameters of the ML model (Page 9, C. Experimental data, 2nd Paragraph: “For Noise2Inverse, an MS-D network was trained with the X:1 strategy and 4 splits for 100 epochs […]”; Page 7, 5th Paragraph: “The networks were trained for 100 epochs using the ADAM algorithm […]”; and Abstract: “Noise2Inverse can train a convolutional neural network solely from noisy indirect measurement, without any additional data”).
Hendriksen fails to teach i. obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a training patient volume, wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the training patient
However, Han teaches wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the training patient (Page 10, 4.3. SPARE Challenge Dataset, 2nd Paragraph: “The CV_P1_T_01 dataset is a 4D CT dataset. Thus, it can be divided into 10 subdatasets, depending on the respiratory phase. Therefore, all image data in the SPARE dataset have a respiratory phase value from 1 to 10 [...]”), and wherein the obtained 2D projections comprise, measurement noise (Page 1, 1. Introduction, 3rd Paragraph: “The radiation dose reduction causes severe photon noise with a Poisson distribution in the projection images, degrading the quality of the 3D reconstructed image”);
ii. repeating, for a plurality of iterations: adding the reconstructed volumetric images of the patient to a training data set as an input volume and corresponding target output volume training pair (Page 7, 3.2. Dropped Projection Strategy, 3rd Paragraph: “we propose a strategy to generate the Bernoulli-sampled instance of the projection image and restore it in the reconstructed image domain. The diagram of the proposed method is presented in Figure 3”; and Page7, Figure 3: “[…] In the training phase, we map each slice of the reconstructed image (reconstructed using dropped projection images) to each slice of the reconstructed image (reconstructed by non-dropped projection images)”); and
wherein the method further comprises: iii. using training pairs from the training data set to update values of the trainable parameters of the ML model (Page 3, 2nd Paragraph: “self-supervised noise reduction method in LDCT that solves the mentioned problems. The proposed method can train a denoising neural network for LDCT without paired images”; and Page 13, 5. Conclusions, 1st Paragraph: “a self-supervised learning-based method that can train a denoising neural network for LDCT without paired images. The proposed method uses Bernoulli sampling to generate degraded versions of the projection images and reconstruct the 3D image, and the denoising neural network is trained to restore the image dropped out by Bernoulli sampling in the projection image domain […]”).
Hendriksen teaches self-supervised tomographic denoising by splitting measured tomography data into multiple reconstruction datasets and training a convolutional neural network using noisy reconstruction pairs. Han teaches applying self-supervised noise reduction to sparse-view CBCT imaging, where reduced projection acquisition introduces severe noise and streak artefacts in reconstructed CBCT images. It would have been obvious to one of ordinary skill in the art at the time of the invention to combine the self-supervised reconstruction training framework of Hendriksen with the sparse-view CBCT denoising techniques of Han in order to reduce noise and aliasing artefacts in reconstructed respiratory-phase CBCT images while avoiding the need for fully sampled ground-truth datasets. The combination merely applies the known Hendriksen self-supervised reconstruction methodology to the known sparse-view CBCT reconstruction environment of Han using predictable results.
Therefore, it would have been obvious to one of ordinary skill in the art to combine Hendriksen and Han before the effective filing date of the claimed invention. The motivation for this combination of references would have been to reduce noise and aliasing artefacts and because the sparse-view CBCT image suffers from streak artifacts and noise due to the reduced number of projections. While Hendriksen can train a convolutional neural network solely from noisy indirect measurements, without any additional data. This motivation for the combination of Hendriksen and Han is/are supported by KSR exemplary rationale (G) Some teaching, suggestion, or motivation in the prior art that would have led one of ordinary skill to modify the prior art reference or to combine prior art reference teachings to arrive at the claimed invention. MPEP 2141 (III).
Hendriksen and Han fail to teach i. obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a training patient volume
However, Sonke teaches to i. obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a training patient volume, wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the training patient (Abstract: “This respiratory correlated CBCT procedure consists of retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase. Subsequently, these subsets are reconstructed into a four-dimensional (4D) CBCT dataset”; and Page 1177, B. Respiratory signal, 1st Paragraph: “Respiratory correlated imaging requires a respiratory signal. A CBCT scanner acquires a series of 2D projection data, showing the internal anatomy, including the position of moving structures as a function of time”);
ii. repeating, for a plurality of iterations: for each of the two selected subsets, using a linear reconstruction algorithm to reconstruct a volumetric image of the training patient from the 2D projections contained in the subset (Abstract: “This respiratory correlated CBCT procedure consists of retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase. Subsequently, these subsets are reconstructed into a four-dimensional (4D) CBCT dataset”).
Hendriksen teaches a self-supervised tomographic denoising framework in which measured tomography data are split into multiple statistically independent reconstruction datasets and used to train a convolutional neural network using noisy reconstruction pairs. Han teaches applying self-supervised noise reduction techniques to sparse-view CBCT imaging, where reduced projection sampling causes severe streak artefacts and noise in reconstructed CBCT images. Furthermore, Sonke teaches respiratory-correlated cone beam CT imaging in which projection data are sorted into subsets corresponding to different respiratory phases for four-dimensional (4D) CBCT reconstruction.
Therefore, it would have been obvious to one of ordinary skill in the art at the time of the invention to combine the self-supervised reconstruction training framework of Hendriksen with the sparse-view CBCT denoising techniques of Han and the respiratory-phase CBCT reconstruction techniques of Sonke in order to reduce noise and aliasing artefacts in reconstructed respiratory-phase 4D-CBCT images while avoiding the need for fully sampled ground-truth training datasets. The combination merely applies known self-supervised reconstruction training techniques to a known respiratory-phase CBCT reconstruction environment using predictable results. The motivation for this combination of references would have been to reduce noise and aliasing artefacts in a reconstructed four-dimensional (4D) medical image because the sparse-view CBCT image suffers from streak artifacts and noise due to the reduced number of projections and respiration correlated cone beam CT (RC-CBCT) is obtained by retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase, while, Hendriksen can train a convolutional neural network solely from noisy indirect measurements, without any additional data. This motivation for the combination of Hendriksen, Han, and Sonke is/are supported by KSR exemplary rationale (G) Some teaching, suggestion, or motivation in the prior art that would have led one of ordinary skill to modify the prior art reference or to combine prior art reference teachings to arrive at the claimed invention. MPEP 2141 (III).
Hendriksen, Han, and Sonke fail to teach selecting two non-empty, mutually disjoint subsets that form a partition of the projection set, wherein each of the two selected subsets contains 2D projections that are respiratory uncorrelated. However, Madesta teaches to i. obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a training patient volume (Figure 1: “the full projection set p of a 4D CBCT scan is split into Nb disjoint subsets […]”; and Page 5624, Chapter 3.A. 4D CBCT data sets, 1st Paragraph: “Evaluation of the proposed boosting scheme was based on three different 4D CBCT data sets: two in-house data sets, consisting of clinical image data as well as phantom measurements, and a subset of the SPARE challenge data set”), wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the training patient, and wherein the obtained 2D projections comprise measurement noise (Page 5624, 3.A.3. SPARE challenge data set: “A full half-fan beam projection set pfull with about 2400 projections and corresponding respiratory phase ф and phase bins bj, a subsampled version psub of this full projection set with 680 projections and corresponding respiratory phase ф and phase bins bj […]” and Page 5626, 4.B. Image quality, 2nd Paragraph: “we also evaluated the image quality enhancement for the 4D CBCT phantom data set with a 4D CT image acting as ground truth. As can be seen in Fig. 7, the streak artifacts and blurriness are suppressed resulting in a significantly improved image quality”);
selecting two non-empty, mutually disjoint subsets that form a partition of the projection set (Figure 1: “the full projection set p of a 4D CBCT scan is split into Nb disjoint subsets […]”; Page 5622, 1st Paragraph: “Algorithm 1 is carried out for all j є [1, Nb] yielding […] with the breathing phases corresponding to the contained data being almost equally distributed for each individual subset”; and Page 5621 – 5622, 2.B. Projection selection for pseudo-average cone-beam computed tomography generation: “Pseudo-average projection subsets (
p
j
p
a
) are generated by assigning each projection of p to exactly on subset of
p
j
p
a
”), wherein each of the two selected subsets contains 2D projections that are respiratory uncorrelated (Figure 1: “The individual
f
j
~
p
a
are different time-averaged (= blurred) representations of the patient geometry, that is, they do not contain a bias toward a specific breathing phase […]”; and Page 5625, 3.B. Network training, 2nd Paragraph: “the image acquisition starts and ends with arbitrary respiratory phases, the Nb breathing phases are in general not included with exactly equal frequency in the pseudo-average projection data. Thus,
f
j
~
p
a
may be slightly biased toward the real respiratory phase bin j”);
for each of the two selected subsets, using a linear reconstruction algorithm to reconstruct a volumetric image of the training patient from the 2D projections contained in the subset (Figure 1: “Each
p
j
p
a
is then reconstructed using
X
г
,
b
i
n
-
1
, resulting in low-quality, artifact-affected image representations
f
j
~
p
a
[…]”; and Page 5622, 2.C. Image reconstruction) ; and
adding the reconstructed volumetric images of the patient to a training data set as an input volume and corresponding target output volume training pair (Figure 1: “[…] Eventually, the model is trained with (
f
j
~
p
a
,
f
~
) tuples […]; and 2.D. Deep learning-based boosting framework).
Hendriksen teaches a self-supervised tomographic denoising framework in which measured tomography data are split into multiple statistically independent reconstruction datasets and used to train a convolutional neural network using noisy reconstruction pairs. Han teaches applying self-supervised noise reduction techniques to sparse-view CBCT imaging, where reduced projection acquisition introduces severe streak artefacts and noise in reconstructed CBCT images. Sonke teaches respiratory-correlated cone beam CT imaging in which projection data are sorted into subsets corresponding to different respiratory phases for four-dimensional (4D) CBCT reconstruction. Madesta further teaches splitting a 4D-CBCT projection dataset into disjoint pseudo-average projection subsets in which breathing phases are distributed across the subsets such that the subsets are not biased toward a specific respiratory phase.
Therefore, it would have been obvious to one of ordinary skill in the art to combine the self-supervised reconstruction training framework of Hendriksen with the sparse-view CBCT denoising techniques of Han, the respiratory-phase 4D-CBCT reconstruction techniques of Sonke, and the distributed respiratory-phase projection subset strategy of Madesta in order to generate statistically independent respiratory-phase reconstruction pairs for self-supervised training while reducing noise and sparse-view aliasing artefacts in reconstructed 4D-CBCT images using predictable results. The motivation for this combination of references would have been to reduce noise and aliasing artefacts in a reconstructed four-dimensional (4D) medical image because the sparse-view CBCT image suffers from streak artifacts and noise due to the reduced number of projections. Additionally, respiration correlated cone beam CT (RC-CBCT) is obtained by retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase while, breathing phases are almost equally distributed among the pseudo-average projection subsets and because Hendriksen can train a convolutional neural network solely from noisy indirect measurements, without any additional data. This motivation for the combination of Hendriksen, Han, Sonke, and Madesta is supported by KSR exemplary rationale (G) Some teaching, suggestion, or motivation in the prior art that would have led one of ordinary skill to modify the prior art reference or to combine prior art reference teachings to arrive at the claimed invention. MPEP 2141 (III).
Regarding claim(s) 2, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, further comprising: repeating i. and ii. for a plurality of training patients (where Han teaches in Page 9, 4. Experiments and Results, 2nd Paragraph: “the SPARE challenge dataset, which is the clinical sparse-view 4D CBCT dataset,”; and Page 10, 4.3. SPARE Challenge Dataset, 2nd Paragraph: “The CV_P1_T_01 dataset is a 4D CT dataset. Thus, it can be divided into 10 subdatasets, depending on the respiratory phase. Therefore, all image data in the SPARE dataset have a respiratory phase value from 1 to 10 [...]”; and where Madesta teaches in Figure 1: “the full projection set p of a 4D CBCT scan is split into Nb disjoint subsets […]”; and Page 5624, Chapter 3.A. 4D CBCT data sets, 1st Paragraph: “Evaluation of the proposed boosting scheme was based on three different 4D CBCT data sets: two in-house data sets, consisting of clinical image data as well as phantom measurements, and a subset of the SPARE challenge data set”).
Regarding claim(s) 3 and 14, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, wherein the obtained 2D projections comprise stochastic measurement noise (where Hendriksen teaches in Page 7, Simulated data, 2nd Paragraph: “The projection images of the foam dataset were corrupted with various levels of Poisson noise. The noise was varied by altering the average absorption of the sample α and the incident photon count per pixel Io”; and “The pixels in the noisy projections where sampled from
p
~
, which for clean pixel value p was distributed as […] a Poisson distribution on the pre-log raw data”; and where Han teaches in Page 1, 1. Introduction, 3rd Paragraph: “The radiation dose reduction causes severe photon noise with a Poisson distribution in the projection images, degrading the quality of the 3D reconstructed image”; Equation (3); and Page 6, 2nd Paragraph: “where
p
^
is a parallel projection with photon noise and n is a random photon noise distribution function observed in the projection images”).
Regarding claim(s) 4 and 15, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, where Hendriksen teaches wherein the obtained 2D projections comprise measurement noise that is elementwise independent and mean-zero (Abstract: “Training a CNN-based denoiser is enabled by exploiting the noise model to compute multiple statistically independent reconstructions. We develop a theoretical framework which shows that such training indeed obtains a denoising CNN, assuming the measured noise is element-wise independent and zero-mean”).
Regarding claim(s) 5 and 16, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, where Hendriksen teaches wherein the ML model is a Convolutional Neural Network (CNN) (Figure 1; and Page 1, Right Col., 5th Paragraph: “In its simplest form, our method splits the measured data in two parts, from which two reconstructions are computed. We train a CNN to transform one reconstruction into the other, and vice versa. The properties of the physical forward model cause the noise in the reconstructed images to be statistically independent. This enables the CNN to perform blind image denoising on the reconstructed images”).
Regarding claim(s) 6, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, wherein for each of the two selected subsets, using a linear reconstruction algorithm to reconstruct a volumetric image of the training patient from the 2D projections contained in the subset comprises using an analytic reconstruction algorithm to reconstruct the volumetric image (where Hendriksen teaches in Page 1, Right Col., 5th Paragraph: “In the proposed Noise2Inverse approach, the training regime explicitly takes into account the structure of the noise in the inverse problem. In its simplest form, our method splits the measured data in two parts, from which two reconstructions are computed”; where Han teaches in Page 7, 3.2. Dropped Projection Strategy, 3rd Paragraph: “we propose a strategy to generate the Bernoulli-sampled instance of the projection image and restore it in the reconstructed image domain. The diagram of the proposed method is presented in Figure 3”; and Page7, Figure 3: “[…] In the training phase, we map each slice of the reconstructed image (reconstructed using dropped projection images) to each slice of the reconstructed image (reconstructed by non-dropped projection images)”; and where Sonke teaches in Abstract: “This respiratory correlated CBCT procedure consists of retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase. Subsequently, these subsets are reconstructed into a four-dimensional (4D) CBCT dataset”; and Page 1177, Left Col., 2nd Paragraph: “An in-house developed implementation of the Feldkamp-Davis-Kress filtered back projection algorithm reconstructs 670 projection images into a 256 volume (1 mm cubic voxel size) […]”).
Regarding claim(s) 7, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, where Madesta teaches wherein for each of the two selected subsets, using a linear reconstruction algorithm to reconstruct a volumetric image of the training patient from the 2D projections contained in the subset comprises:
using an algorithm having a property that an average of reconstructions generated by the algorithm using subsets of an available projection set is approximately equal to the reconstruction generated by the algorithm using all projections of the available projection set (Figure 1: “the full projection set p of a 4D CBCT scan is split into Nb disjoint subsets […] The individual
f
j
~
p
a
are different time-averaged (= blurred) representations of the patient geometry, that is, they do not contain a bias toward a specific breathing phase […]”; Page 5622, 1st Paragraph: “Algorithm 1 is carried out for all j є [1, Nb] yielding […] with the breathing phases corresponding to the contained data being almost equally distributed for each individual subset”; Page 5621 – 5622, 2.B. Projection selection for pseudo-average cone-beam computed tomography generation: “Pseudo-average projection subsets (
p
j
p
a
) are generated by assigning each projection of p to exactly on subset of
p
j
p
a
”; Page 5622, 2.C. Image reconstruction; and Page 5625, 3.B. Network training, 2nd Paragraph: “the image acquisition starts and ends with arbitrary respiratory phases, the Nb breathing phases are in general not included with exactly equal frequency in the pseudo-average projection data. Thus,
f
j
~
p
a
may be slightly biased toward the real respiratory phase bin j”).
Regarding claim(s) 9, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, where Madesta teaches wherein using training pairs from the training data set to update values of the trainable parameters of the ML model comprises repeating, until a convergence (Page 5625, 3.B. Network training, 2nd Paragraph: “The whole network was trained for 50 epochs until convergence. Corresponding plots for all tested values of a are available in the supplementary material (Fig. S1)”) condition is satisfied:
inputting an input of a training pair from the training data set to the ML model, wherein the ML model processes the input in accordance with current values of the trainable parameters of the ML model and generates an ML model output (Figure 1: “[…] Eventually, the model is trained with (
f
j
~
p
a
,
f
~
) tuples […]; 2.D. Deep learning-based boosting framework; and Page 5625, 3.C. Boosting/network application: “To boost a reconstructed phase image
f
~
bj, the trained network B was applied axial slice-wise until the whole volume was processed, that is, about 220 forward passes were needed to cover the whole scan range”);
comparing the ML model output to the target output of the training pair (Figure 1: “[…] Eventually, the model is trained with (
f
j
~
p
a
,
f
~
) tuples […]; and 2.D. Deep learning-based boosting framework; where Han teaches in Figure 3: “map each slice of the reconstructed image (reconstructed using dropped projection images) to each slice of the reconstructed image (reconstructed by non-dropped projection images)”); and
updating trainable parameters of the ML model to optimize a function of the comparison (where Madesta teaches in Page 5625, 3.B. Network training, 2nd Paragraph: “all Nb artifact-affected slices corresponding to one high quality image slice were considered an individual mini-batch during stochastic gradient decent (Adam optimizer), ensuring an unbiased gradient calculation and update of the network parameters during training. Furthermore, the training scheme include a learning rate decay […] to improve convergence”).
Regarding claim(s) 10, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 1, where Madesta teaches wherein using training pairs from the training data set to update values of the trainable parameters of the ML model comprises:
using corresponding slices from corresponding dimensions of each of the input and target output volumes (Page 5625, 3.B. Network training, 1st Paragraph: “Network training was based on the axial slices (with respect to the CBCT rotation axis) of the (pseudo-)time-average CBCT data, since these contain continuous streak artifacts that are in the focus of the intended image quality boosting”; Page 5625, 3.B. Network training, 2nd Paragraph: “all Nb artifact-affected slices corresponding to one high quality image slice were considered an individual mini-batch during stochastic gradient decent (Adam optimizer), ensuring an unbiased gradient calculation and update of the network parameters during training. Furthermore, the training scheme include a learning rate decay […] to improve convergence”; and Page 5625, 3.C. Boosting/network application: “To boost a reconstructed phase image
f
~
bj, the trained network B was applied axial slice-wise until the whole volume was processed, that is, about 220 forward passes were needed to cover the whole scan range”).
Regarding claim(s) 11 and 20, Hendriksen teaches a computer implemented method for reducing noise and aliasing artefacts in a reconstructed four-dimensional, 4D, medical image of a patient, the method comprising:
wherein the obtained 2D projections comprise measurement noise (Page 9, C. Experimental data, 1st Paragraph: “The dataset, Dorthe_F_002, was acquired at the Advanced Photon Source at Argonne National Laboratory, and contained 900 noisy projection images […]”; Page 14, 2nd Paragraph: “Noise2Inverse is well-suited to imaging modalities that permit trading acquisition speed for measurement noise, as it aims to remove measurement noise […]”; and Page 13, V. Discussion, 3rd Paragraph: “As in Noise2Noise, the network is presented with two noisy images during training. In Noise2Inverse, however, these images are sub-sampled reconstructions, and since the artifacts arising from sub-sampling the data are correlated […]”);
inputting the generated reconstruction to an unsupervised Machine Learning (ML) model (Abstract: “Noise2Inverse can train a convolutional neural network solely from noisy indirect measurement, without any additional data”; Page 1, Right Col., 5th Paragraph: “In the proposed Noise2Inverse approach, the training regime explicitly takes into account the structure of the noise in the inverse problem. In its simplest form, our method splits the measured data in two parts, from which two reconstructions are computed”; Page 7, 5th Paragraph: “The networks were trained for 100 epochs using the ADAM algorithm […]”; and Page 9, C. Experimental data, 2nd Paragraph: “For Noise2Inverse, an MS-D network was trained with the X:1 strategy and 4 splits for 100 epochs […]”).
Hendriksen fails to teach obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a patient volume, wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the patient, and dividing the plurality of 2D projections in the projection set into a plurality of phase sets, wherein each phase set comprises 2D projections that are respiratory correlated, corresponding to a single phase of the respiratory cycle of the patient; for each phase of the respiratory cycle of the patient, generating a reconstructed volumetric image of the patient from the 2D projections contained in the corresponding phase set; and wherein the ML model is operable to process the input generated reconstructed volumetric images according to trained parameters of the ML model and to output volumetric images of the patient having reduced noise and aliasing artefacts, and wherein the ML model is trained using reconstructed volumetric images that are respiratory uncorrelated.
However, Han teaches wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the patient (Page 10, 4.3. SPARE Challenge Dataset, 2nd Paragraph: “The CV_P1_T_01 dataset is a 4D CT dataset. Thus, it can be divided into 10 subdatasets, depending on the respiratory phase. Therefore, all image data in the SPARE dataset have a respiratory phase value from 1 to 10 [...]”), and wherein the obtained 2D projections comprise measurement noise (Page 1, 1. Introduction, 3rd Paragraph: “The radiation dose reduction causes severe photon noise with a Poisson distribution in the projection images, degrading the quality of the 3D reconstructed image”);
inputting the generated reconstruction to an unsupervised Machine Learning (ML) model, wherein the ML model is operable to process the input generated reconstructed volumetric images according to trained parameters of the ML model and to output volumetric images of the patient having reduced noise and aliasing artefacts (Figure 3; Page 1, 1. Introduction, 3rd Paragraph: “The radiation dose reduction causes severe photon noise with a Poisson distribution in the projection images, degrading the quality of the 3D reconstructed image”; Page 3, 2nd Paragraph: “self-supervised noise reduction method in LDCT that solves the mentioned problems. The proposed method can train a denoising neural network for LDCT without paired images”; and Page 13, 5. Conclusions, 1st Paragraph: “a self-supervised learning-based method that can train a denoising neural network for LDCT without paired images. The proposed method uses Bernoulli sampling to generate degraded versions of the projection images and reconstruct the 3D image, and the denoising neural network is trained to restore the image dropped out by Bernoulli sampling in the projection image domain […]”).
Hendriksen teaches a self-supervised tomographic denoising framework in which measured tomography data are split into statistically independent reconstruction datasets and used to train a convolutional neural network using noisy reconstruction pairs. Han teaches applying self-supervised CNN-based denoising to sparse-view CBCT imaging in which reduced projection acquisition causes severe streak artefacts and photon noise in reconstructed CBCT images.
Therefore, it would have been obvious to one of ordinary skill in the art to combine the self-supervised reconstruction training framework of Hendriksen with the sparse-view CBCT denoising techniques of Han in order to apply known self-supervised tomographic denoising techniques to sparse-view respiratory CBCT reconstruction for reduction of noise and view-aliasing artefacts using predictable results. The motivation for this combination of references would have been to reduce noise and aliasing artefacts in a reconstructed four-dimensional (4D) medical image because the sparse-view CBCT image suffers from streak artifacts and noise due to the reduced number of projections, while Hendriksen can train a convolutional neural network solely from noisy indirect measurements, without any additional data. This motivation for the combination of Hendriksen and Han is supported by KSR exemplary rationale (G) Some teaching, suggestion, or motivation in the prior art that would have led one of ordinary skill to modify the prior art reference or to combine prior art reference teachings to arrive at the claimed invention. MPEP 2141 (III).
Hendriksen and Han fails to teach to obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a patient volume; dividing the plurality of 2D projections in the projection set into a plurality of phase sets, wherein each phase set comprises 2D projections that are respiratory correlated, corresponding to a single phase of the respiratory cycle of the patient; for each phase of the respiratory cycle of the patient, generating a reconstructed volumetric image of the patient from the 2D projections contained in the corresponding phase set; and wherein the ML model is trained using reconstructed volumetric images that are respiratory uncorrelated.
However, Sonke teaches to obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a patient volume, wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the patient (Abstract: “This respiratory correlated CBCT procedure consists of retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase. Subsequently, these subsets are reconstructed into a four-dimensional (4D) CBCT dataset”; and Page 1177, B. Respiratory signal, 1st Paragraph: “Respiratory correlated imaging requires a respiratory signal. A CBCT scanner acquires a series of 2D projection data, showing the internal anatomy, including the position of moving structures as a function of time”);
dividing the plurality of 2D projections in the projection set into a plurality of phase sets, wherein each phase set comprises 2D projections that are respiratory correlated, corresponding to a single phase of the respiratory cycle of the patient (Page 1178, Left Col., 1st Paragraph: “reconstructed volumes at multiple respiratory phases can be obtained by retrospective sorting in projection space. That is, cone beam projections are snapshots (recorded with, e.g., 25 ms x-ray pulses) representing a certain respiratory phase while different projections represent different respiratory phases. By sorting the breathing signal and the corresponding projections into several phase bins and subsequently feeding each subset of projections to the reconstruction algorithm, a 4D CBCT dataset is generated”; Abstract: “This respiratory correlated CBCT procedure consists of retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase. Subsequently, these subsets are reconstructed into a four-dimensional (4D) CBCT dataset”; and Page 1177, B. Respiratory signal, 1st Paragraph: “Respiratory correlated imaging requires a respiratory signal. A CBCT scanner acquires a series of 2D projection data, showing the internal anatomy, including the position of moving structures as a function of time”); and
for each phase of the respiratory cycle of the patient, generating a reconstructed volumetric image of the patient from the 2D projections contained in the corresponding phase set (Abstract: “This respiratory correlated CBCT procedure consists of retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase. Subsequently, these subsets are reconstructed into a four-dimensional (4D) CBCT dataset”).
Hendriksen teaches a self-supervised tomographic denoising framework using noisy reconstruction pairs generated from split tomography measurements. Han teaches self-supervised CNN-based sparse-view CBCT denoising for reducing streak artefacts and photon noise caused by reduced projection sampling. Sonke teaches respiratory correlated cone beam CT reconstruction in which projection data are sorted into subsets corresponding to different respiratory phases for generation of a 4D-CBCT dataset.
Therefore, it would have been obvious to one of ordinary skill in the art to combine the self-supervised denoising framework of Hendriksen with the sparse-view CBCT denoising techniques of Han and the respiratory-phase reconstruction framework of Sonke in order to reduce noise and sparse-view aliasing artefacts in respiratory-phase 4D-CBCT reconstructions while maintaining known respiratory correlated phase reconstruction workflows. The motivation for this combination of references would have been to reduce noise and aliasing artefacts in a reconstructed four-dimensional (4D) medical image because respiration correlated cone beam CT (RC-CBCT) is obtained by retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase and because the sparse-view CBCT image suffers from streak artifacts and noise due to the reduced number of projections, while Hendriksen can train a convolutional neural network solely from noisy indirect measurements, without any additional data. This motivation for the combination of Hendriksen, Han and Sonke is/are supported by KSR exemplary rationale (G) Some teaching, suggestion, or motivation in the prior art that would have led one of ordinary skill to modify the prior art reference or to combine prior art reference teachings to arrive at the claimed invention. MPEP 2141 (III).
Hendriksen, Han and Sonke fail to teach to inputting the generated reconstruction to an unsupervised Machine Learning (ML) model, wherein the ML model is operable to process the input generated reconstructed volumetric images according to trained parameters of the ML model and to output volumetric images of the patient having reduced noise and aliasing artefacts, and wherein the ML model is trained using reconstructed volumetric images that are respiratory uncorrelated.
However, Madesta teaches to obtaining a projection set comprising a plurality of two-dimensional (2D) projections of a patient volume, wherein the obtained 2D projections represent a plurality of phases of a respiratory cycle of the patient (Figure 1: “the full projection set p of a 4D CBCT scan is split into Nb disjoint subsets […]”; and Page 5624, Chapter 3.A. 4D CBCT data sets, 1st Paragraph: “Evaluation of the proposed boosting scheme was based on three different 4D CBCT data sets: two in-house data sets, consisting of clinical image data as well as phantom measurements, and a subset of the SPARE challenge data set”), and wherein the obtained 2D projections comprise measurement noise (Page 5624, 3.A.3. SPARE challenge data set: “A full half-fan beam projection set pfull with about 2400 projections and corresponding respiratory phase ф and phase bins bj, a subsampled version psub of this full projection set with 680 projections and corresponding respiratory phase ф and phase bins bj […]” and Page 5626, 4.B. Image quality, 2nd Paragraph: “we also evaluated the image quality enhancement for the 4D CBCT phantom data set with a 4D CT image acting as ground truth. As can be seen in Fig. 7, the streak artifacts and blurriness are suppressed resulting in a significantly improved image quality”);
dividing the plurality of 2D projections in the projection set into a plurality of phase sets (Figure 1: “the full projection set p of a 4D CBCT scan is split into Nb disjoint subsets […]”; Page 5622, 1st Paragraph: “Algorithm 1 is carried out for all j є [1, Nb] yielding […] with the breathing phases corresponding to the contained data being almost equally distributed for each individual subset”; and Page 5621 – 5622, 2.B. Projection selection for pseudo-average cone-beam computed tomography generation: “Pseudo-average projection subsets (
p
j
p
a
) are generated by assigning each projection of p to exactly on subset of
p
j
p
a
”), wherein each phase set comprises 2D projections that are respiratory correlated, corresponding to a single phase of the respiratory cycle of the patient (Figure 1: “The individual
f
j
~
p
a
are different time-averaged (= blurred) representations of the patient geometry, that is, they do not contain a bias toward a specific breathing phase […]”; and Page 5625, 3.B. Network training, 2nd Paragraph: “the image acquisition starts and ends with arbitrary respiratory phases, the Nb breathing phases are in general not included with exactly equal frequency in the pseudo-average projection data. Thus,
f
j
~
p
a
may be slightly biased toward the real respiratory phase bin j”);
for each phase of the respiratory cycle of the patient, generating a reconstructed volumetric image of the patient from the 2D projections contained in the corresponding phase set (Figure 1: “Each
p
j
p
a
is then reconstructed using
X
г
,
b
i
n
-
1
, resulting in low-quality, artifact-affected image representations
f
j
~
p
a
[…]”; and Page 5622, 2.C. Image reconstruction); and
inputting the generated reconstruction to an unsupervised Machine Learning (ML) model, wherein the ML model is operable to process the input generated reconstructed volumetric images according to trained parameters of the ML model and to output volumetric images of the patient having reduced noise and aliasing artefacts, and wherein the ML model is trained using reconstructed volumetric images that are respiratory uncorrelated (Page 5622, 1st Paragraph: “Algorithm 1 is carried out for all j є [1, Nb] yielding […] with the breathing phases corresponding to the contained data being almost equally distributed for each individual subset”; Page 5625, 3.B. Network training, 1st Paragraph: “Network training was based on the axial slices (with respect to the CBCT rotation axis) of the (pseudo-)time-average CBCT data, since these contain continuous streak artifacts that are in the focus of the intended image quality boosting”; Page 5625, 3.B. Network training, 2nd Paragraph: “all Nb artifact-affected slices corresponding to one high quality image slice were considered an individual mini-batch during stochastic gradient decent (Adam optimizer), ensuring an unbiased gradient calculation and update of the network parameters during training. Furthermore, the training scheme include a learning rate decay […] to improve convergence”; and Page 5625, 3.C. Boosting/network application: “To boost a reconstructed phase image
f
~
bj, the trained network B was applied axial slice-wise until the whole volume was processed, that is, about 220 forward passes were needed to cover the whole scan range”).
Hendriksen teaches a self-supervised tomographic denoising framework in which tomography measurements are partitioned into statistically independent reconstruction datasets for CNN training. Han teaches self-supervised sparse-view CBCT denoising using CNN-based reconstruction training to reduce streak artefacts and projection noise. Sonke teaches respiratory correlated cone beam CT reconstruction using respiratory phase-binned projection subsets corresponding to different breathing phases. Madesta further teaches generating pseudo-average projection subsets in which breathing phases are distributed among the subsets such that the resulting reconstructed images are not biased toward a single respiratory phase and may be used for deep-learning-based image boosting and network training.
Therefore, it would have been obvious to one of ordinary skill in the art to combine the self-supervised reconstruction training techniques of Hendriksen with the sparse-view CBCT denoising techniques of Han, the respiratory-phase reconstruction workflow of Sonke, and the pseudo-average respiratory-uncorrelated projection subset techniques of Madesta in order to generate statistically independent respiratory-phase reconstruction pairs for self-supervised denoising and artefact reduction in 4D-CBCT reconstruction systems using predictable results. The motivation for this combination of references would have been to reduce noise and aliasing artefacts in a reconstructed four-dimensional (4D) medical image because respiration correlated cone beam CT (RC-CBCT) is obtained by retrospective sorting in projection space, yielding subsets of projections that each corresponds to a certain breathing phase, while the sparse-view CBCT image suffers from streak artifacts and noise due to the reduced number of projections and because breathing phases are almost equally distributed among the pseudo-average projection subsets, while Hendriksen can train a convolutional neural network solely from noisy indirect measurements, without any additional data. This motivation for the combination of Hendriksen, Han, Sonke, and Madesta is/are supported by KSR exemplary rationale (G) Some teaching, suggestion, or motivation in the prior art that would have led one of ordinary skill to modify the prior art reference or to combine prior art reference teachings to arrive at the claimed invention. MPEP 2141 (III).
Regarding claim(s) 12, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 11, where Madesta teaches wherein the ML model is trained using training pairs of input and target output reconstructed volumetric images (Figure 1: “[…] Eventually, the model is trained with (
f
j
~
p
a
,
f
~
) tuples […]; 2.D. Deep learning-based boosting framework; and Page 5625, 3.C. Boosting/network application: “To boost a reconstructed phase image
f
~
bj, the trained network B was applied axial slice-wise until the whole volume was processed, that is, about 220 forward passes were needed to cover the whole scan range”), and wherein both the input and target output reconstructed volumetric images of a training pair are respiratory uncorrelated (Figure 1: “The individual
f
j
~
p
a
are different time-averaged (= blurred) representations of the patient geometry, that is, they do not contain a bias toward a specific breathing phase […]”; and Page 5625, 3.B. Network training, 2nd Paragraph: “the image acquisition starts and ends with arbitrary respiratory phases, the Nb breathing phases are in general not included with exactly equal frequency in the pseudo-average projection data. Thus,
f
j
~
p
a
may be slightly biased toward the real respiratory phase bin j”).
Regarding claim(s) 18, Hendriksen as modified by Han, Sonke, and Madesta teaches the computer implemented method as claimed in claim 11, wherein inputting the generated reconstruction to an ML model comprises: inputting slices from the generated reconstruction to the ML model (where Madesta teaches in Page 5625, 3.B. Network training, 1st Paragraph: “Network training was based on the axial slices (with respect to the CBCT rotation axis) of the (pseudo-)time-average CBCT data, since these contain continuous streak artifacts that are in the focus of the intended image quality boosting”; Page 5625, 3.B. Network training, 2nd Paragraph: “all Nb artifact-affected slices corresponding to one high quality image slice were considered an individual mini-batch during stochastic gradient decent (Adam optimizer), ensuring an unbiased gradient calculation and update of the network parameters during training. Furthermore, the training scheme include a learning rate decay […] to improve convergence”; and Page 5625, 3.C. Boosting/network application: “To boost a reconstructed phase image
f
~
bj, the trained network B was applied axial slice-wise until the whole volume was processed, that is, about 220 forward passes were needed to cover the whole scan range”; and where Han teaches in Page 7, 3.2. Dropped Projection Strategy, 3rd Paragraph: “we propose a strategy to generate the Bernoulli-sampled instance of the projection image and restore it in the reconstructed image domain. The diagram of the proposed method is presented in Figure 3”; and Page7, Figure 3: “[…] In the training phase, we map each slice of the reconstructed image (reconstructed using dropped projection images) to each slice of the reconstructed image (reconstructed by non-dropped projection images)”).
Allowable Subject Matter
Claim(s) 8, 13, and 17 is/are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Relevant Prior Art Directed to State of Art
Shen et al (US 2023/0024401 A1) are relevant prior art not applied in the rejection(s) above. Shen discloses a method for diagnostic imaging reconstruction comprising: storing a prior image x.sup.pr from a scan of a subject, comprising image intensity at each coordinate in image space; initializing parameters of a neural network using the prior image x.sup.pr; wherein the neural network maps coordinates in image space to corresponding intensity values in the prior image; wherein initializing the parameters comprises minimizing an objective function representing a difference between intensity values of the prior image and predicted intensity values output from the neural network, thereby creating an implicit neural representation of the prior image; performing a scan to acquire subsampled (sparse) measurements y of the subject; training the neural network using the measurements y to learn a neural representation of a reconstructed image x, wherein the training comprises minimizing an objective function representing a difference between the measurements y and a forward model applied to predicted image intensity values output from the neural network; computing image intensity values output from the trained neural network from coordinates in image space input to the trained neural network to produce predicted image intensity values.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JONGBONG NAH whose telephone number is (571) 272-1361. The examiner can normally be reached M - F: 9:00 AM - 5:30 PM.
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, ONEAL MISTRY can be reached on 313-446-4912. 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.
/JONGBONG NAH/Examiner, Art Unit 2674
/ONEAL R MISTRY/Supervisory Patent Examiner, Art Unit 2674
Prosecution Insights