sCT image generation using CycleGAN with deformable layers

1. A computer-implemented method for generating a synthetic computed tomography (sCT) image from a cone-beam computed tomography (CBCT) image, the method comprising:

receiving a CBCT image of a subject;

generating, using a first generative model, a sCT image corresponding to the CBCT image, the first generative model trained based on one or more deformable offset layers in a generative adversarial network (GAN) to process the CBCT image as an input and provide the sCT image as an output, the first generative model trained in a cycle generative adversarial network (CycleGAN) comprising a second generative model, wherein the second generative model:

receives a real computed tomography (CT) image and outputs first and second synthetic CBCT (sCBCT) images, the real CT image being paired with a training CBCT image, the first sCBCT image generated with application of a set of deformable offset layers and the second sCBCT image generated without application of the set of deformable offset layers, the first generative model to receive the first and second sCBCT images to generate first and second cycle CT images for calculating cycle-consistency losses, the first cycle CT image generated with application of the one or more deformable offset lavers and the second cycle CT image generated without application of the one or more deformable offset layers; and

generating a display of the sCT image for medical analysis of the subject.

2. The method of claim 1 , wherein:

the generative adversarial network is configured to train the first generative model using a discriminative model;

values applied by the first generative model and the discriminative model are established using adversarial training between the discriminative model and the first generative model; and

the first generative model and the discriminative model comprise respective convolutional neural networks.

3. The method of claim 2 , wherein:

the adversarial training comprises:

training the first generative model to generate a first sCT image from a given CBCT image by applying a first set of the one or more deformable offset layers to the given CBCT image;

training the first generative model to generate a second sCT image from the given CBCT image without applying the first set of the one or more deformable offset layers to the given CBCT image; and

training the discriminative model to classify the first sCT image as a synthetic or a real CI image, and

the output of the first generative model is used for training the discriminative model and an output of the discriminative model is used for training the first generative model.

4. The method of claim 3 , wherein the CycleGAN comprises the first generative model and the discriminative model, wherein the discriminative model is a first discriminative model, wherein the CycleGAN further comprises:

the second generative model trained to:

process a given CT image as an input;

provide a first synthetic sCBCT image as an output by applying a second set of the one or more deformable offset layers to the given CT image; and

provide a second sCBCT image as an output without applying the second set of the one or more deformable offset layers to the given CT image; and

a second discriminative model trained to classify the first synthetic sCBCT image as a synthetic or a real CBCT image.

5. The method of claim 4 , wherein the CycleGAN comprises a first portion to train the first generative model, wherein the first generative model includes first and second input interfaces and a first shared generator portion, wherein the second generative model includes third and fourth input interfaces and a second shared generator portion, the first portion being trained to:

transmit the training CBCT image to an input of the first generative model via first and second paths to output the first and second sCT images, respectively, the first path comprising the first input interface including the first set of the one or more deformable offset layers and a first set of one or more convolution layers, the second path comprising second input interface including the first set of the one or more convolution layers without the first set of the one or more deformable offset layers;

receive the first sCT image at the input of the first discriminative model to classify the first sCT image as the synthetic or the real CT image; and

receive the first and second sCT images at an input of the second generative model via third and fourth paths to generate first and second cycle CBCT images, respectively, for calculating cycle-consistency losses, the third path comprising the third input interface including a second set of the one or more deformable offset layers and a second set of the one or more convolution layers, the fourth path comprising the fourth input interface including the second set of the one or more convolution layers without the second set of the one or more deformable offset layers.

6. The method of claim 5 , wherein the CycleGAN comprises a second portion that is trained to:

receive the first synthetic CBCT image at the input of the second discriminative model to classify the first synthetic CBCT image as a synthetic or real CBCT image.

7. The method of claim 6 , wherein:

the cycle-consistency losses being generated based on a comparison of the first and second cycle CBCT images with the training CBCT image and a comparison of the first and second cycle CT images with the real CT image;

the first generative model is trained to minimize or reduce a first pixel-based loss term using the second sCT image, the first pixel-based loss term represents an expectation of difference between a plurality of synthetic CT images and respectively paired real CT images; and

the second generative model is trained to minimize or reduce a second pixel-based loss term using the second sCBCT image, the second pixel-based loss term representing an expectation of difference between a plurality of sCBCT images and respectively paired real CBCT images.

8. The method of claim 7 , wherein:

the CycleGAN is trained to apply a metric to the first and second pixel-based loss terms, the metric generated based on a map having a same size as a pair of a CBCT images and real CT images, such that each pixel value in the map represents a similarity level between a given CBCT image and a given real CT image that is paired with the given CBCT image; and

the CycleGAN is trained to apply a threshold to the metric such that when the similarity level exceeds the threshold, the metric is applied to the first and second pixel-based loss terms and otherwise a zero value is applied to the first and second pixel-based loss terms.

9. The method of claim 8 , wherein the CycleGAN is trained to apply one of a plurality of metrics to the first and second pixel-based loss terms, the metrics being generated using low-pass filtering and down-sampling of the paired CBCT and CT images at different image resolutions or view levels.

10. The method of claim 9 , wherein the one or more deformable offset layers are trained based on adversarial training to change a sampling amount, introduce coordinate offsets, and resample images using interpolation in order store or absorb deformed structure information between the paired CBCT and CT images.

11. A computer-implemented method for training a model to generate a synthetic computed tomography (sCT) image from a cone-beam computed tomography (CBCT) image, the method comprising:

receiving a CBCT image of a subject as an input of a first generative model; and

training the first generative model, via first and second paths, in a generative adversarial network (GAN) to process the CBCT image to provide first and second synthetic computed tomography (sCT) images corresponding to the CBCT image as outputs of the first generative model, the first path comprising a first set of one or more deformable offset layers and a first set of one or more convolution layers, the second path comprising the first set of the one or more convolution layers without the first set of the one or more deformable offset layers, the first generative model trained in a cycle generative adversarial network (CycleGAN) comprising a second generative model, wherein the second generative model:

receives a real computed tomography (CT) image and outputs first and second synthetic CBCT (sCBCT) images, the real CT image being paired with a training CBCT image, the first sCBCT image generated with application of a set of deformable offset layers and the second sCBCT image generated without application of the set of deformable offset layers, the first generative model to receive the first and second sCBCT images to generate first and second cycle CT images for calculating cycle-consistency losses, the first cycle CT image generated with application of the one or more deformable offset layers and the second cycle CT image generated without application of the one or more deformable offset layers.

12. The method of claim 11 , wherein the CycleGAN comprises the first generative model and a first discriminative model, further comprising:

training the second generative model to process produced first and second sCT images as inputs and provide first and second cycle-CBCT images as outputs via third and fourth paths, respectively, the third path comprising a second set of the one or more deformable offset layers and a second set of the one or more convolution layers, the fourth path comprising the second set of the one or more convolution layers without the second set of the one or more deformable offset layers; and

training a second discriminative model to classify the first cycle-CBCT image as a synthetic or a real CBCT image.

13. The method of claim 12 , wherein the CycleGAN comprises first and second portions to train the first generative model, further comprising:

transmitting the training CBCT image to the input of the first generative model via the first and second paths to output the first and second sCT images;

receiving the first sCT image at the input of the first discriminative model;

classifying, with the first discriminative model, the first sCT image as a synthetic or real CT image;

receiving the first and second sCT images at the input of the second generative model via the third and fourth paths to generate the first and second cycle CBCT images for calculating cycle-consistency losses;

receiving the first synthetic training CBCT image at the input of the second discriminative model;

classifying, with the second discriminative model, the first synthetic training CBCT image as a synthetic or real CBCT image;

training the first generative model to minimize or reduce a first pixel-based loss term using the second sCT image, the first pixel-based loss term representing an expectation of a difference between a plurality of sCT images and respectively paired real CT images; and

training the second generative model to minimize or reduce a second pixel-based loss using the second synthetic training image, the second pixel-based loss term representing an expectation of a difference between a plurality of synthetic CBCT images and respectively paired real CBCT images.

14. A system for generating a synthetic computed tomography (sCT) image from a cone-beam computed tomography (CBCT) image, the system comprising:

processing circuitry comprising at least one processor; and

a storage medium comprising instructions, which when executed by the at least one processor, cause the processor to perform operations comprising:

receiving a CBCT image of a subject;

generating, using a first generative model, a sCT image corresponding to the CBCT image, the first generative model trained based on one or more deformable offset layers in a generative adversarial network (GAN) to process the CBCT image as an input and provide the sCT image as an output, the first generative model trained in a cycle generative adversarial network (CycleGAN) comprising a second generative model, the second generative model:

receives a real computed tomography (CT) image and outputs first and second synthetic CBCT (sCBCT) images, the real CT image being paired with a training CBCT image, the first sCBCT image generated with application of a set of deformable offset layers and the second sCBCT image generated without application of the set of deformable offset layers, the first generative model to receive the first and second sCBCT images to generate first and second cycle CT images for calculating cycle-consistency losses, the first cycle CT image generated with application of the one or more deformable offset layers and the second cycle CT image generated without application of the one or more deformable offset layers; and

generating a display of the sCT image for medical analysis of the subject.

15. The system of claim 14 , wherein:

the one or more deformable offset layers is a first set of one or more of the deformable offset layers;

the generative adversarial network is configured to train the first generative model using a discriminative model;

values applied by the first generative model and the discriminative model are established using adversarial training between the discriminative model and the first generative model; and

the first generative model and the discriminative model comprise respective convolutional neural networks;

wherein the adversarial training comprises:

training the first generative model to generate a first sCT image from a given CBCT image by applying the first set of the one or more deformable offset layers to the given CBCT image,

training the first generative model to generate a second sCT image from the given CBCT image without applying the first set of the one or more deformable offset layers to the given CBCT image, and

training the discriminative model to classify the first sCT image as a synthetic or a real computed tomography (CT) image, and

the output of the first generative model is used for training the discriminative model and an output of the discriminative model is used for training the first generative model.

16. The system of claim 15 , wherein the discriminative model is a first discriminative model, wherein the CycleGAN further comprises:

the second generative model trained to:

process a given CT image as an input;

provide a first synthetic CBCT (sCBCT) image as an output by applying a second set of the one or more deformable offset layers to the given CT image; and

provide a second sCBCT image as an output without applying the second set of the one or more deformable offset layers to the given CT image; and

a second discriminative model trained to classify the first synthetic sCBCT image as a synthetic or a real CBCT image.

17. The system of claim 16 , wherein the CycleGAN comprises a first portion to train the first generative model, the first portion being trained to:

transmit the training CBCT image to the input of the first generative model via first and second paths to output the first and second sCT images, respectively, the first path comprising the first set of the one or more deformable offset layers and a first set of the one or more convolution layers, the second path comprising the first set of the one or more convolution layers without the first set of the one or more deformable offset layers;

receive the first sCT image at the input of the first discriminative model to classify the first sCT image as the synthetic or the real CT image; and

receive the first and second sCT images at the input of the second generative model via third and fourth paths to generate first and second cycle CBCT images, respectively, for calculating cycle-consistency losses, the third path comprising the second set of the one or more deformable offset layers and the second set of the one or more convolution layers, the fourth path comprising the second set of the one or more convolution layers without the second set of the one or more deformable offset layer.

18. The system of claim 17 , wherein the CycleGAN comprises a second portion that is trained to:

receive the first synthetic CBCT image at the input of the second discriminative model to classify the first synthetic CBCT image as a synthetic or real CBCT image; and.

19. The system of claim 18 , wherein:

the cycle-consistency losses being generated based on a comparison of the first and second cycle CBCT images with the training CBCT image and a comparison of the first and second cycle CT images with the real CT image;

the first generative model is trained to minimize or reduce a first pixel-based loss term using the second sCT image, the first pixel-based loss term represents an expectation of difference between a plurality of synthetic CT images and respectively paired real CT images; and

the second generative model is trained to minimize or reduce a second pixel-based loss term using the second sCBCT image, the second pixel-based loss term represents an expectation of difference between a plurality of synthetic CBCT images and respectively paired real CBCT images.

20. The system of claim 14 , wherein the one or more deformable offset layers include at least one of one or more modules based on adversarial training to change a sampling amount, introduce coordinate offsets, and resample images using interpolation, one or more spatial transformers, one or more convolutional layers, or one or more modules configured to store deformed structure information for an image.