Method and system for flexible denoising of images using a pruned feature representation domain

Abstract

Systems and methods for denoising images are provided. A standard image module is configured to generate standard anatomic features and standard noise features from a standard image and to reconstruct the standard image from the standard anatomic features and the standard noise features. A reduced quality image module is configured to generate reduced quality anatomic features and reduced quality noise features from a reduced quality image and to reconstruct the reduced quality image from the reduced quality anatomic features and the reduced quality noise features. A loss calculation module is provided for calculating a loss metric based at least in part on 1) a comparison between the reconstructed standard image and the standard image and 2) a comparison between the reconstructed reduced quality image and the reduced quality image. The standard image module outputs a reconstructed standard transfer image when provided with the reduced quality anatomic features.

Description

Technical Field

[0001] This disclosure generally relates to systems and methods for training and tuning neural network models for providing flexible solutions for denoising low-dose images using domain-agnostic learning with deciphered feature representations. Background Technology

[0002] Typically, in most imaging modalities, artifacts (e.g., noise) exist in the acquisition physics or reconstruction that introduce artifacts into the final image. To train a denoising algorithm (e.g., a neural network model), pairs of noisy and noiseless image samples are presented to the neural network model, and the network attempts to minimize the cost function by denoising the noisy images to recover the corresponding noiseless, ground-state images.

[0003] However, any change in the parameters used to acquire the image can lead to changes in the form or amount of artifacts in the corresponding image. For this reason, denoising models used to denoise standard images are less effective when applied to images acquired using different acquisition parameters, such as reduced radiation doses in the context of computed tomography (CT) scans.

[0004] The increasing use of CT scans in modern medical practice has raised concerns about the associated radiation dose required, and dose reduction has become a clinical goal. However, reducing radiation dose often significantly increases noise and other artifacts in reconstructed images, which can impair diagnostic information. Extensive efforts have been made to reduce noise in low-dose CT scans, thereby converting them into high-quality images.

[0005] In the context of image denoising, machine learning techniques (including those using convolutional neural networks (CNNs)) have been studied. However, existing methods are often specifically tuned for particular noise levels and do not generalize well to noise levels not covered in the training set used to train the corresponding CNN.

[0006] In CT imaging, various factors, including kilovolt-pole peak (kVp), milliampere-second (mAs), slice thickness, and patient size, can affect the noise level in reconstructed images. As a result, changing any one of these imaging parameters can lead to different noise levels or different artifact contours, thus requiring different models to denoise images acquired with such varying imaging parameters under normal circumstances. This limits the applicability of CNN-based methods in practical denoising.

[0007] Therefore, there is a need for methods capable of denoising images acquired using imaging parameters different from those used in the training set for the corresponding method. A single training model is also needed to denoise images with varying noise levels, including CT images acquired with lower radiation doses than those in the training set. Summary of the Invention

[0008] Systems and methods for denoising medical images are provided. In one embodiment, a standard image module is configured to generate standard anatomical features and standard noise features from a standard image, and to reconstruct the standard image based on the standard anatomical features and the standard noise features. A degraded image module is similarly configured to generate degraded anatomical features and degraded noise features from a degraded image, and to reconstruct the degraded image based on the degraded anatomical features and the degraded noise features.

[0009] A loss calculation module is provided to enable training of the system and method. This loss calculation module is typically used to calculate a loss metric based at least in part on 1) a comparison between the reconstructed standard image and the standard image, and 2) a comparison between the reconstructed degraded image and the degraded image.

[0010] The loss metric calculated at the loss calculation module is incorporated into the loss function for use in machine learning to tune the standard image module and the degraded image module. When the degraded anatomical features are provided to the standard image module, the standard image module outputs a reconstructed standard transfer image that includes the degraded anatomical features and a noise level lower than that represented by the degraded noise features.

[0011] In some embodiments, the standard image module includes a standard anatomical encoder, a standard noise encoder, and a standard generator. Upon receiving the standard image, the standard anatomical encoder outputs the standard anatomical features, the standard noise encoder outputs the standard noise features, and the standard generator reconstructs the standard image based on the standard anatomical features and the standard noise features.

[0012] In some such embodiments, the degraded image module may similarly include a degraded anatomical encoder, a degraded noise encoder, and a degraded generator, wherein upon receiving the degraded image, the degraded anatomical encoder outputs the degraded anatomical features, the degraded noise encoder outputs the degraded noise features, and the degraded generator reconstructs the degraded image based on the degraded anatomical features and the degraded noise features.

[0013] In some such embodiments, the loss calculation module calculates a loss metric for the standard generator based at least in part on the comparison between the reconstructed standard image and the standard image. Similarly, the loss calculation module may calculate a loss metric for the degraded generator based at least in part on the comparison between the reconstructed degraded image and the degraded image.

[0014] The loss calculation module can also calculate a loss metric for the standard anatomical encoder based on a comparison with the segmentation markers for the standard image, and the loss calculation module can also calculate a loss metric for the degraded anatomical encoder based on a comparison with the output of the standard anatomical encoder.

[0015] In some embodiments, the loss metric for the degraded anatomical encoder is an adversarial loss metric.

[0016] In some embodiments, the system implementing the method may further include a segmentation network, and the segmentation mask for the reconstructed standard image may be evaluated based on a comparison with segmentation markers for the standard image.

[0017] In some such embodiments, when standard anatomical features are provided to the de-quality generator, the de-quality generator outputs a reconstructed de-quality transfer image, which includes the standard anatomical features and a noise level higher than that represented by the standard noise features. A segmentation mask for the de-quality transfer image can then be evaluated based on a comparison with the segmentation markers for the standard image.

[0018] In some embodiments, the loss metric for the standard transmitted image is evaluated based on a comparison with the reconstructed standard image, and the loss metric for the standard transmitted image is an adversarial loss metric.

[0019] In some embodiments, the standard image module and the degraded image module are trained simultaneously.

[0020] In other embodiments, the standard image module is trained prior to the training of the degraded image module, and the values ​​of variables developed during the training of the standard image module remain constant during the training of the degraded image module. In some such embodiments, after training the standard image module and the degraded image module, the system further trains the standard generator while maintaining constant values ​​for the standard anatomical encoder, the standard noise encoder, and the degraded anatomical encoder.

[0021] In some embodiments, the standard anatomical features and the reduced-quality anatomical features each correspond to a single anatomical structure. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a system according to one embodiment of the present disclosure.

[0023] Figure 2 An imaging apparatus according to one embodiment of the present disclosure is illustrated.

[0024] Figure 3 This is the training pipeline used in the embodiments of this disclosure.

[0025] Figures 4A-4C The diagram illustrates the use of... Figure 3 An exemplary training method used in a training pipeline.

[0026] Figure 5 The diagram illustrates the use of Figure 3 The training pipeline trains a model to denoise images. Detailed Implementation

[0027] The description of illustrative embodiments based on the principles of this disclosure is intended to be read in conjunction with the accompanying drawings, which are considered an integral part of the entire written description. Any references to directions or orientations in the description of the embodiments disclosed herein are for convenience of description only and are not intended to limit the scope of this disclosure in any way. Related terms such as “below,” “above,” “horizontal,” “vertical,” “above,” “below,” “over,” “below,” “top,” and “bottom,” and their derivatives (e.g., “horizontally,” “downward,” “upward,” etc.) should be interpreted as referring to the orientation described subsequently or shown in the discussed drawings. These related terms are for convenience of description only and do not require the device to be constructed or operated in a particular orientation unless explicitly stated otherwise. Terms such as “attach,” “attach,” “connect,” “couple,” and “interconnect” refer to a relationship in which structures are directly or indirectly fastened or attached to each other through an intermediate structure, and to movable or rigid attachments or relationships, unless explicitly described otherwise. Furthermore, the features and benefits of this disclosure are illustrated by reference to exemplary embodiments. Therefore, this disclosure should not be explicitly limited to illustrating exemplary embodiments of a possible combination of non-limiting features, which may exist alone or in other combinations of features; the scope of this disclosure is defined by the appended claims.

[0028] This disclosure describes one or more best practices currently conceived in carrying out this disclosure. This description is not intended to be construed in a limiting sense, but rather to provide examples of this disclosure presented for illustrative purposes only, with reference to the accompanying drawings, to inform those skilled in the art of the advantages and construction of this disclosure. In the various views of the drawings, the same reference numerals denote the same or similar parts.

[0029] It is important to note that the disclosed embodiments are merely examples of many advantageous uses of the innovative teachings herein. In general, the statements in this application's specification do not necessarily limit any of the various claimed disclosures. Furthermore, some statements may apply to some inventive features but not to others. Generally, unless otherwise stated, the singular element may be plural, and vice versa, without loss of generality.

[0030] Typically, to denoise medical images, an image processor (which can denoise images using algorithms or models) is based on the level and form of noise expected to be present in the corresponding image. This level and form of expected noise is usually based on various parameters used to acquire the image.

[0031] For example, in the context of medical imaging based on computed tomography (CT), different image processors (e.g., machine learning algorithms in the form of convolutional neural networks (CNNs)) can be used to process images. Then, in the case of machine learning algorithms, these image processors are trained on different anatomical regions and structures corresponding to specific noise levels. The noise level in the image may be a function of multiple factors, including kilovolt peak value (kVp), milliampere-second value (mAs), slice thickness, and patient size.

[0032] While denoising image processors (e.g., CNNs) can be based on the expected noise level and on standardized parameters (including standardized radiation dose), the systems and methods disclosed herein can effectively apply such models to images acquired using different acquisition parameters (e.g., reduced radiation dose).

[0033] While the following discussion is specifically about CT-based medical imaging implementations, similar systems and methods can be used in the context of other imaging modalities, such as magnetic resonance imaging (MRI) or positron emission tomography (PET).

[0034] Figure 1 This is a schematic diagram of a system 100 according to one embodiment of the present disclosure. As shown, system 100 typically includes a processing device 110 and an imaging device 120.

[0035] Processing device 110 can apply processing routines to a received image. Processing device 110 may include memory 113 and processor circuitry 111. Memory 113 may store multiple instructions. Processor circuitry 111 may be coupled to memory 113 and may be configured to execute instructions. Instructions stored in memory 113 may include processing routines and data associated with various machine learning algorithms, such as various convolutional neural networks for processing images.

[0036] The processing device 110 may further include an input unit 115 and an output unit 117. The input unit 115 may receive information (e.g., images) from the imaging device 120. The output unit 117 may output information to a user or a user interface device. The output unit 117 may include a monitor or display.

[0037] In some embodiments, the processing device 110 may be directly associated with the imaging device 120. In alternative embodiments, the processing device 110 may be different from the imaging device 120, such that it receives images at the input section 115 via a network or other interface for processing.

[0038] In some embodiments, the imaging device 120 may include an image data processing device and an energy spectrum or conventional CT scanning unit for generating CT projection data when scanning an object (e.g., a patient).

[0039] Figure 2 An exemplary imaging apparatus according to one embodiment of this disclosure is illustrated. It should be understood that although a CT imaging apparatus is shown and the following discussion is in the context of CT images, similar methods can be applied in the context of other imaging apparatuses, and images to which these methods can be applied can be acquired in a variety of ways.

[0040] In an imaging apparatus according to embodiments of the present disclosure, the CT scanning unit may be adapted to perform multi-axis and / or helical scanning of an object to generate CT projection data. In an imaging apparatus according to embodiments of the present disclosure, the CT scanning unit may include an energy-resolved photon-counting image detector. The CT scanning unit may include a radiation source that emits radiation traversing the object during the acquisition of projection data.

[0041] Additionally, in the imaging apparatus according to embodiments of the present disclosure, the CT scanning unit can perform a localization scan different from the main scan, thereby generating different images associated with the localization scan and the main scan, wherein these images are different but include the same main content.

[0042] exist Figure 2In the example shown, the CT scanning unit 200 (e.g., a computed tomography (CT) scanner) may include a fixed gantry 202 and a rotating gantry 204, the rotating gantry 204 being rotatably supported by the fixed gantry 202. When acquiring projection data, the rotating gantry 204 may rotate about a longitudinal axis around an examination area 206 for the target. The CT scanning unit 200 may include a support 207 for supporting the patient within the examination area 206 and is configured to pass the patient across the examination area during the imaging process.

[0043] The CT scanning unit 200 may include a radiation source 208 (e.g., an X-ray tube), which may be supported by and configured to rotate with a rotating gantry 204. The radiation source 208 may include an anode and a cathode. A source voltage applied between the anode and cathode can accelerate electrons from the cathode to the anode. The electron flow can provide a current from the cathode to the anode to generate radiation for traversing the examination area 206.

[0044] The CT scanning unit 200 may include a detector 210. The detector 210 may be positioned across the examination area 206 opposite the radiation source 208 at an angular arc. The detector 210 may include an array of one-dimensional or two-dimensional pixels (e.g., direct conversion detector pixels). The detector 210 may be adapted to detect radiation traversing the examination area and generate a signal indicating its energy.

[0045] The CT scanning unit 200 may also include generators 211 and 213. Generator 211 may generate tomographic projection data 209 based on signals from detector 210. Generator 213 may receive the tomographic projection data 209 and generate an original image of the object based on the tomographic projection data 209.

[0046] Figure 3 This is a schematic diagram of a training pipeline used in embodiments of this disclosure. The training pipeline 300 is typically implemented by processing device 110, and the corresponding methods are implemented by processor circuitry 111 based on instructions stored in memory 113. Memory 113 may also include a database structure for storing data required to implement the training pipeline 300 or denoising methods (e.g., implementing a model using data generated by the training pipeline). In some embodiments, the database may be stored externally and accessible by processing device 110 via a network interface.

[0047] The method implemented by the training pipeline 300 includes a training learning algorithm (e.g., CNN) for denoising low-dose CT images using domain-agnostic learning with deciphered feature representations.

[0048] Domain adaptation can be defined as adapting a given dataset (e.g., a source-labeled dataset) from a dataset belonging to the source domain. s Knowledge transfer to the target unlabeled dataset D belonging to a specific known target domain t Domain-agnostic learning is defined in a similar way, the difference being that the target unlabeled dataset and the source dataset can consist of data from multiple domains (e.g., for a target source of {D}). t1 D t2 ,...,D tn}, while for the source domain {D} s1 D s2 ,...,D sn The sample does not have any domain labeling indicating which domain it belongs to. Domain-agnostic learning can be achieved by using deciphered feature representations that separate styles from content.

[0049] Therefore, the training pipeline 300 separates anatomical features a from the noise features n of the CT images processed by the pipeline, where both anatomical features and noise features can be used to reconstruct the underlying image. When evaluating the denoising performance of the model, radiologists can score the images against two qualities: structural fidelity and image noise suppression. Structural fidelity is the ability of an image to accurately depict anatomical structures in the field of view, while image noise manifests as random patterns on the image that degrade its quality. By extracting anatomical features from degraded images and pairing them with typical noise reduction levels of higher-quality images, the model generated by the described training pipeline can thus provide images that score highly on both metrics.

[0050] As shown in the figure, the training pipeline 300 includes a standard image module 310 and a degraded image module 320. The standard image module 310 includes a standard anatomical encoder E. n CT 330, Standard Noise Encoder E n CT 340 and standard generator G CT 350. Standard image source 360 ​​(which can be...) Figure 2 The CT scanning unit 200 (or it may be an image database) then provides a standard image 370 to the standard image module 310. Upon receiving the standard image 370, the standard image 370 is provided to the standard anatomical encoder 330 and the standard noise encoder 340.

[0051] Then, the standard anatomical encoder 330 outputs the standard anatomical feature a. CT 380, and the standard noise encoder 340 outputs standard noise characteristics n CT390. Then, both the standard noise feature 390 and the standard anatomical feature 380 can be provided to the standard generator 350, which can reconstruct the standard image X based on the provided standard anatomical feature 380 and standard noise feature 390. CT CT 370'. For this reason, the standard image module 310 can decompose the standard image 370 into component features 380, 390, and then reconstruct the standard image 370' based on these component features.

[0052] The de-quality image module 320 includes components that are parallel to those discussed with respect to the standard image module 310. Therefore, the de-quality image module 320 includes a de-quality anatomical encoder E. a LDCT 430. Reduce quality noise encoder E n LDCT 440 and reducing the quality generator G LDCT 450. The degraded image source 460 is then provided to the degraded image module 320. 470. The degraded image source 460 can be... Figure 2 The CT scanning unit 200, wherein the parameters discussed above are reconfigured to reduce image quality. This can be based on a reduced radiation dose compared to the acquisition of the standard image 370, thereby obtaining, for example, a low-dose CT scan (LDCT) image. Alternatively, the reduced-quality image source 460 can be an image database. Upon receiving the reduced-quality image 470, the reduced-quality image 470 is provided to the reduced-quality anatomical encoder 430 and the reduced-quality noise encoder 440.

[0053] The quality reduction anatomical encoder then outputs the quality reduction anatomical feature a. LDCT 480, and reduce quality noise encoder 440 output quality noise feature n LDCT 490. Then, both the reduced quality noise feature 490 and the reduced quality anatomical feature 440 can be provided to the reduced quality generator 450, which can reconstruct the reduced quality image X based on the provided reduced quality anatomical feature 480 and reduced quality noise feature 490. LDCT LDCT 470'. Therefore, the degraded image module 320 can decompose the degraded image 470 into component features 480, 490, and then reconstruct the degraded image 470' based on these component features.

[0054] When training the standard image module 310 and the degraded image module 320, a loss calculation module 500 is provided to compute various loss functions. Therefore, the loss metric for the standard generator 350 can be based at least in part on the comparison between the standard image 370 and the reconstructed standard image 370'. Such a loss metric can be a reconstruction loss 510 for the standard generator 350, and can be used during training to ensure that the reconstructed 370' generated by the standard generator accurately reflects the originally provided corresponding standard image 370.

[0055] Similarly, the loss metric for the de-quality generator 450 can be based at least in part on a comparison between the de-quality image 470 and the reconstructed de-quality image 470' generated by the de-quality generator 450. This loss metric can also be a reconstruction loss 520 for the de-quality generator 450, and can be used during training to ensure that the reconstruction 470' from the de-quality generator accurately reflects the originally provided corresponding de-quality image 470.

[0056] Anatomical encoders 330 and 430 are also evaluated by loss calculation module 500. The loss metric for the standard anatomical encoder 330 can be evaluated based on a comparison between standard anatomical features 380 and segmentation labels 540 generated for the corresponding standard image 370. Such segmentation labels can be manually generated when acquiring the standard image 370 or retrieved from a database. Such a loss metric can be the segmentation loss 530.

[0057] The loss metric for the reduced-quality anatomical encoder 430 is evaluated based on a comparison between the reduced-quality anatomical feature 480 and the standard anatomical feature 380 generated by the standard anatomical encoder 330. This comparison will be discussed below. Figures 4A-4C The training methods will be discussed in more detail, but such loss metrics are usually adversarial loss 550.

[0058] To evaluate the anatomical encoders 330 and 340, a loss calculation module 500 is provided with a segmentation network 560, which creates segmentation masks M for the corresponding anatomical features 380 and 480. a CT 570a and M a LDCT 570b. Then the segmentation mask 570a for standard anatomical feature 480 is compared with the segmentation marker 540, and then the segmentation mask 570b for degraded anatomical feature 480 is compared with the segmentation mask 570a for standard anatomical feature 380.

[0059] The loss metric is incorporated into the loss function for use in machine learning to tune the corresponding generators 350, 450 and anatomical encoders 330, 430. Therefore, the reconstruction losses 510 and 520 for generators 350 and 450 can be used to improve the performance of their respective pipelines by adjusting the variables that determine the outputs of the corresponding modules. This can be accomplished by implementing standard machine learning techniques, or by implementing the methods described in the reference below. Figures 4A-4C The exemplary training method discussed is used to accomplish this.

[0060] In some embodiments, the loss calculation module also generates additional loss metrics. Such loss metrics may include a segmentation loss 580 for the reconstruction of the standard image 370' generated by the standard generator 350. Therefore, the segmentation network 590 can be applied to the reconstructed image 370' to generate a segmentation mask M. CT CT 600, then based on the segmentation markers 540 for the corresponding image 370, the segmentation mask M is... CT CT Evaluation is performed on 600. Any training process used in the training pipeline 500 can consider the segmentation loss 580.

[0061] As shown in the figure, there is a connection between the output anatomical feature 380 of the standard anatomical encoder 330 and the de-quality generator 450, and a connection between the output anatomical feature 480 of the de-quality anatomical encoder 430 and the standard generator 350. When the standard anatomical feature 380 from the standard anatomical encoder 330 is provided, the generator 350 outputs a reconstructed standard image 370' based on the standard anatomical feature and the standard noise feature 390 generated by the standard noise encoder 340. In contrast, when the de-quality anatomical feature 480 generated by the de-quality anatomical encoder 430 is provided, the standard generator 350 outputs a reconstructed standard transfer image X. LDCT CT 470".

[0062] The reconstructed standard transfer image 470 includes reduced-quality anatomical features 480 and a noise level lower than that represented by reduced-quality noise features 490. This can be constructed by generator 350 based on reduced-quality anatomical features 480 and noise features 390 generated by standard noise feature encoder 340. In the case of a transfer image, such noise features 390 can be, for example, the average of standard noise features generated during training based on the corresponding standard image 370.

[0063] Similarly, the de-quality generator 450 can be used to generate a transfer image. Therefore, when the de-quality generator 450 is provided with standard anatomical features 380 from the standard anatomical encoder 330, the generator 350 outputs a reconstructed de-quality transfer image X. CT LDCT 370". The reconstructed degraded transfer image 370" includes standard anatomical features 380 and noise levels based on degraded noise features 490.

[0064] To further improve the quality of any model implemented using the training pipeline 500, the transfer images 370” and 470” can be used to generate additional loss metrics. Thus, the reconstructed standard-quality transfer image 470” can be evaluated using the adversarial loss 610 to compare the transfer image with the reconstructed image 370’ output by the standard generator 350.

[0065] Similarly, a loss metric can be used to evaluate the reconstructed degraded transfer image 370". Because the degraded transfer image 370" includes standard anatomical features 380, the training pipeline 300 typically has access to the corresponding segmentation markers 540. Therefore, the loss metric for the degraded transfer image 370" can be a segmentation loss 620, where the segmentation network 630 generates an appropriate segmentation mask M. CT LDCT 640.

[0066] It should also be noted that in some embodiments, the segmentation networks 560, 590, 630 themselves can be evaluated based on segmentation losses 530, 580, 620, which are the results of a comparison between the obtained segmentation masks 570a, 580, 620 and the segmentation marker 540.

[0067] In some embodiments, the entire training pipeline 300 is trained simultaneously. Therefore, both the standard image module 310 and the degraded image module 320 are provided with a variety of images for training the respective modules. As the network improves, based on loss metrics, anatomical features, and the reconstructed images, the training pipeline will be instructed to generate transfer images 470”, 370”, and then these transfer images will be evaluated in parallel.

[0068] In contrast, in some embodiments, the training pipeline 300 is trained sequentially, as shown in the following reference. Figures 4A-4C The details discussed.

[0069] The described training pipeline 300 can be used to generate a model for generating a transfer image 470, which can then be used to denoise the previously degraded image 470. The training pipeline 300 can be used to create modules for transferring anatomical features from a wide variety of source images obtained using different imaging parameters. For example, different encoders can be trained to transfer images obtained using different doses.

[0070] Furthermore, while the same model created using the training pipeline 300 can be used across different anatomical features, in some embodiments, the use of the pipeline-trained model may be limited to specific anatomical structures. Although the degraded image 470 and the standard image 370 are not for the same patient or the same organ, they will be associated with the same anatomical structure in different images. For example, the model may be trained for images of the head, abdomen, or a specific organ.

[0071] Figures 4A-4C The diagram illustrates the use of... Figure 3 An exemplary training method used in the training pipeline 300.

[0072] In some embodiments, such as Figure 4A As shown, the standard image module 310 is trained before the degraded image module 320. In such an embodiment, as... Figure 4B As shown, when training the degraded image module 320, the variables associated with the standard image module (including variables embedded in the standard image encoders 330 and 340, variables in the standard generator 350, and variables in the segmentation network 560 associated with the standard anatomical features 380) remain constant.

[0073] In such an embodiment, after independently training the standard image module 310 and the degraded image module 320, the training pipeline 300 can then further train the standard generator 350 while maintaining the values ​​for the standard anatomical encoder 330, the standard noise encoder 340, and the degraded anatomical encoder 430.

[0074] Therefore, when training a model using the claimed method, the method of implementing the training pipeline 300 according to this disclosure can initially provide a standard image module 310. The standard image module 310 has a standard anatomical encoder 330 for extracting standard anatomical features 380 from a standard image 370 and a standard noise encoder 340 for extracting standard noise features 390 from the standard image.

[0075] The standard image module 310 also has a standard generator 350 for generating a reconstructed standard image 370' based on standard anatomical features 380 and standard noise features 390.

[0076] The method can then provide a degraded image module 320. The degraded image module 320 has a degraded anatomical encoder 430 for extracting degraded anatomical features 480 from the degraded image 470 and a degraded noise encoder 440 for extracting degraded noise features 490 from the standard image.

[0077] The degraded image module 410 also has a degraded generator 450 for generating a reconstructed degraded image 470' based on the degraded anatomical features 480 and the degraded noise features 490.

[0078] The standard image module 310 is then trained by receiving multiple standard images 370 at the standard image module. These images are received from an image source 360, which may be a CT scan unit 200 or alternatively an image database. The system then compares the reconstructed standard image 370' output by the standard image generator 350 for each received standard image 370 with the corresponding standard image to extract a standard reconstruction loss metric. The variables of the standard image module 310 are then updated based on this loss metric, thereby tuning the standard encoders 330 and 340 and the standard generator 350.

[0079] While the standard image module 310 is still being trained, the method generates a segmentation mask 570a at the segmentation network 560 corresponding to the standard anatomical features 380 extracted from each standard image 370 by the standard anatomical encoder 330. The segmentation mask 570a for each standard image 370 is then compared with the segmentation label 540 associated with the corresponding standard image 370 to generate a standard anatomical loss metric 530. At least one variable of the standard anatomical encoder 330 is then updated based on the standard anatomical loss metric.

[0080] like Figure 4A As shown, these initial training steps can be performed independently of the de-quality image module 320 on a portion of the training pipeline 300 associated with the standard image 370. During the first part of training, additional training elements can be implemented to improve the models used for the standard encoders 330, 340, and the standard generator 350. For example, the reconstructed standard image 370' can be submitted to an additional segmentation network 590 to generate a segmentation mask 600, which can then be compared with a known segmentation label 540 for the corresponding image 370 to generate additional segmentation loss metrics.

[0081] After the standard image module 510 has been trained and provides acceptable results, the degraded image module 320 can be trained, such as... Figure 4BAs shown. Therefore, multiple degraded images 470 are received from image source 460 at the degraded image module 320. As discussed above, the image source can be CT scan unit 200, an image database, or some combination thereof.

[0082] The system then compares the reconstructed degraded image 470' output by the degraded image generator 450 for each received degraded image 470 with the corresponding degraded image to extract the degraded reconstruction loss 520. The variables of the standard image module 310 are then kept constant, while the variables of the degraded image module 320 are updated based on the loss metric, thereby tuning the degraded encoders 430 and 440 and the degraded generator 450.

[0083] The segmentation network 560, initially trained when training the standard image module 310, remains similarly constant when training the degraded image module 320. The segmentation network 560 is then used to generate segmentation masks 570b corresponding to the degraded anatomical features 480 extracted from each degraded image 470 by the degraded anatomy encoder 430. The segmentation masks 570b are then compared with at least one segmentation mask 570a previously generated by the standard image module 310 (typically the average of segmentation masks 570a) to generate an adversarial loss metric.

[0084] The adversarial loss metric is then used to update at least one variable of the degraded anatomical encoder 430.

[0085] Once the degraded image module 320 has been trained in this manner, the degraded anatomical features 480 generated by the degraded anatomical encoder 430 can be provided to the standard generator 350 in the standard image module 310. The standard generator 350 can then output a reconstructed standard transfer image 470 based on the degraded anatomical features 480 and at least one standard quality noise feature 390 (in some cases, the average of the standard quality noise features 390).

[0086] The reconstructed standard transfer image 470” can then be evaluated by comparing the reconstructed standard transfer image 470” with the average result of a standard image reconstruction 370’ or multiple such reconstructions to generate an adversarial loss 610. While the standard generator 350 remains constant and therefore is not tuned based on this loss metric, the quality reduction anatomical encoder 430 can be further tuned on this basis.

[0087] Additionally, as discussed above, the de-quality generator 450 can be used to generate a reconstructed de-quality transfer image 370. This transfer image 370 can be based on standard anatomical features 380 generated by the standard image module 320 and can be parsed by the segmentation network 630 to generate a corresponding segmentation map 640. The segmentation map 640 can then be evaluated against segmentation markers 540 to generate a segmentation loss 620. This loss metric can then be used to further tune the de-quality module 320.

[0088] In some embodiments, the model created by the described training pipeline 300 can be used to create a standard transfer image 470", which can then be denoised. In other embodiments (e.g., Figure 4C In the illustrated embodiment, the training pipeline can then be further trained. As shown, the relevant encoders 330, 340, and 430 can remain constant while the standard generator 350 is further tuned. This is done by creating an additional reconstructed standard transfer image 470", which is then compared with one or more standard reconstructions 370' to generate additional adversarial loss data 610, which can then be used to further tune the standard generator 350.

[0089] By training the described training pipeline 300 in this manner, there are multiple checks in the design that encourage anatomical embedding to be domain-agnostic and anatomical-preserving.

[0090] Figure 5 The diagram illustrates the use of Figure 3 The model, trained in the training pipeline, denoises the images. In doing so, the model first encodes low-dose anatomical data from the degraded image 470 into degraded anatomical features 480 using a degraded anatomical encoder 430. Then, the degraded anatomical features 480, along with standard noise features 390 or the average of such noise features, are fed into a standard generator 350 to generate the reconstructed standard transfer image 470.

[0091] This training pipeline 300 can be used to train a conversion from low-dose CT or other degraded images to multiple normal-dose CT or standard images with different noise levels. In the case of CT images, these noise levels can represent the overall average result of normal-dose CT data, or the average result of a set of CT data with certain characteristics (e.g., those CT data reconstructed using a specific algorithm, such as filtered backprojection or iterative reconstruction). To obtain these noise features, the trained CT noise encoder E... n CT330 is applied to all or a subset of normal-dose CT images 370 in the training set to encode them into CT noise features 390. Then, the average of these features is taken within each group to obtain their own representative CT noise features n. CT1 n CT2 Each of these CT noise features can be used as the average noise feature 390 discussed above. To denoise low-dose CT data, users can select from the predefined CT noise features 390 based on specific needs. The training pipeline 300 can also perform interpolation of CT noise features extracted from different groups, allowing users to continuously adjust the CT noise.

[0092] It should be understood that although the method described herein is presented in the context of CT scan images, various imaging techniques (including various medical imaging techniques) can be anticipated, and the method described herein can effectively denoise images generated using a wide variety of imaging techniques.

[0093] The methods according to this disclosure can be implemented on a computer as a computer-implemented method, or implemented in dedicated hardware, or implemented in a combination of both. Executable code for the methods according to this disclosure can be stored on a computer program product. Examples of computer program products include storage devices, optical storage devices, integrated circuits, servers, online software, etc. Preferably, the computer program product may include non-transient program code stored on a computer-readable medium for performing the methods according to this disclosure when the program product is run on a computer. In one embodiment, the computer program may include computer program code adapted to perform all steps of the methods according to this disclosure when the computer program is run on a computer. The computer program may be embodied on a computer-readable medium.

[0094] Although this disclosure has been described with respect to several embodiments in considerable length and specificity, it is not intended to limit this disclosure to any such details or embodiments or any particular embodiment, but rather to be interpreted with reference to the appended claims in order to provide the broadest possible interpretation of such claims in consideration of the prior art, and thus effectively cover the intended scope of this disclosure.

[0095] All examples and conditional language described herein are intended for pedagogical purposes to aid the reader in understanding the principles of this disclosure and the ideas contributed by the inventors to further the development of the art, and should be construed as not being limited to such specific examples and conditions. Furthermore, all statements and specific examples of the principles, aspects, and embodiments of this disclosure described herein are intended to cover their structural and functional equivalents. Additionally, it is intended that such equivalents include both currently known equivalents and those developed in the future (i.e., any element developed that performs the same function, regardless of its structure).

Claims

1. A system for denoising medical images, comprising: A standard image module is configured to: generate standard anatomical features and standard noise features based on a standard image, and reconstruct the standard image based on the standard anatomical features and the standard noise features to generate a reconstructed standard image; A degraded image module is configured to: generate degraded anatomical features and degraded noise features based on a degraded image, and reconstruct the degraded image based on the degraded anatomical features and the degraded noise features to generate a reconstructed degraded image; A loss calculation module is used to calculate a loss metric based at least in part on 1) a comparison between the reconstructed standard image and the standard image and 2) a comparison between the reconstructed degraded image and the degraded image; The loss metric is incorporated into a loss function for using machine learning to tune the standard image module and the degraded image module, and wherein, when the degraded anatomical features are provided to the standard image module, the standard image module outputs a reconstructed standard transfer image, the reconstructed standard transfer image including the degraded anatomical features and a noise level lower than the noise level represented by the degraded noise features.

2. The system according to claim 1, wherein: The standard image module includes a standard anatomical encoder, a standard noise encoder, and a standard generator. Upon receiving the standard image, the standard anatomical encoder outputs the standard anatomical features, the standard noise encoder outputs the standard noise features, and the standard generator reconstructs the standard image based on the standard anatomical features and the standard noise features. The degraded image module includes a degraded anatomical encoder, a degraded noise encoder, and a degraded generator. Upon receiving the degraded image, the degraded anatomical encoder outputs the degraded anatomical features, the degraded noise encoder outputs the degraded noise features, and the degraded generator reconstructs the degraded image based on the degraded anatomical features and the degraded noise features. The loss calculation module calculates a loss metric for the standard generator based at least in part on the comparison between the reconstructed standard image and the standard image. The loss calculation module calculates a loss metric for the degraded image generator based at least in part on the comparison between the reconstructed degraded image and the degraded image. The loss calculation module calculates a loss metric for the standard anatomical encoder based on a comparison with segmentation markers for the standard image. The loss calculation module calculates a loss metric for the degraded anatomical encoder based on a comparison with the output of the standard anatomical encoder.

3. The system according to claim 2, wherein, The loss metric for the reduced-quality anatomical encoder is an adversarial loss metric.

4. The system of claim 2, further comprising a segmentation network, wherein, The segmentation mask for the reconstructed standard image is evaluated based on a comparison with the segmentation markers for the standard image.

5. The system according to claim 4, wherein, When standard anatomical features are provided to the de-quality generator, the de-quality generator outputs a reconstructed de-quality transfer image, the de-quality transfer image including the standard anatomical features and a noise level higher than that represented by the standard noise features, and wherein the segmentation mask for the de-quality transfer image is evaluated based on a comparison with the segmentation markers for the standard image.

6. The system according to claim 1, wherein, The loss metric for the standard transmitted image is evaluated based on a comparison with the reconstructed standard image, and wherein the loss metric for the standard transmitted image is an adversarial loss metric.

7. The system according to claim 1, wherein, The standard image module and the degraded image module are trained simultaneously.

8. The system according to claim 1, wherein, The standard image module is trained prior to the training of the degraded image module, and the values ​​of variables developed during the training of the standard image module remain constant during the training of the degraded image module.

9. The system according to claim 8, wherein, After training the standard image module and the degraded image module, the system further trains the standard generator while maintaining constant values ​​for the standard anatomical encoder, the standard noise encoder, and the degraded anatomical encoder.

10. The system according to claim 1, wherein, The standard anatomical features and the reduced-quality anatomical features each correspond to a single anatomical structure.

11. A method for denoising medical images, comprising: The standard image module generates standard anatomical features and standard noise features based on standard images; The standard image is reconstructed by the standard image module based on the standard anatomical features and the standard noise features to generate a reconstructed standard image; The degraded image module generates degraded anatomical features and degraded noise features based on the degraded image; The degraded image module reconstructs the degraded image based on the degraded anatomical features and the degraded noise features to generate a reconstructed degraded image; and The loss metric is calculated by the loss calculation module based at least in part on 1) a comparison between the reconstructed standard image and the standard image and 2) a comparison between the reconstructed degraded image and the degraded image; The loss metric is incorporated into a loss function for using machine learning to tune the standard image module and the degraded image module, and wherein, when the degraded anatomical features are provided to the standard image module, the standard image module outputs a reconstructed standard transfer image, the reconstructed standard transfer image including the degraded anatomical features and a noise level lower than the noise level represented by the degraded noise features.

12. The method according to claim 11, wherein: The standard image module includes a standard anatomical encoder, a standard noise encoder, and a standard generator. Upon receiving the standard image, the standard anatomical encoder outputs the standard anatomical features, the standard noise encoder outputs the standard noise features, and the standard generator reconstructs the standard image based on the standard anatomical features and the standard noise features. The degraded image module includes a degraded anatomical encoder, a degraded noise encoder, and a degraded generator. Upon receiving the degraded image, the degraded anatomical encoder outputs the degraded anatomical features, the degraded noise encoder outputs the degraded noise features, and the degraded generator reconstructs the degraded image based on the degraded anatomical features and the degraded noise features. The loss calculation module calculates a loss metric for the standard generator based at least in part on the comparison between the reconstructed standard image and the standard image. The loss calculation module calculates a loss metric for the degraded image generator based at least in part on the comparison between the reconstructed degraded image and the degraded image. The loss calculation module calculates a loss metric for the standard anatomical encoder based on a comparison with segmentation markers for the standard image. The loss calculation module calculates a loss metric for the degraded anatomical encoder based on a comparison with the output of the standard anatomical encoder.

13. The method according to claim 12, wherein, The loss metric for the reduced-quality anatomical encoder is an adversarial loss metric.

14. The method of claim 12, further comprising: A segmentation mask for the reconstructed standard image is generated at the segmentation network, and the segmentation mask is evaluated based on a comparison with the segmentation labels for the standard image.

15. The method according to claim 14, wherein, When standard anatomical features are provided to the de-quality generator, the de-quality generator outputs a reconstructed de-quality transfer image, the de-quality transfer image including the standard anatomical features and a noise level higher than that represented by the standard noise features, and wherein the segmentation mask for the de-quality transfer image is evaluated based on a comparison with the segmentation markers for the standard image.

16. The method according to claim 11, wherein, The loss metric for the standard transmitted image is evaluated based on a comparison with the reconstructed standard image, and wherein the loss metric for the standard transmitted image is an adversarial loss metric.

17. The method according to claim 11, wherein, The standard image module and the degraded image module are trained simultaneously.

18. The method according to claim 11, wherein, The standard image module is trained prior to the training of the degraded image module, and the values ​​of variables developed during the training of the standard image module remain constant during the training of the degraded image module.

19. The method according to claim 18, wherein, After training the standard image module and the degraded image module, the method further trains the standard generator while maintaining constant values ​​for the standard anatomical encoder, the standard noise encoder, and the degraded anatomical encoder.

20. The method according to claim 11, wherein, The standard anatomical features and the reduced-quality anatomical features each correspond to a single anatomical structure.

Images