Systems and methods to process electronic images for synthetic image generation
Systems and methods are disclosed for generating synthetic medical images, including images presenting rare conditions or morphologies for which sufficient data may be unavailable. In one aspect, style transfer methods may be used. For example, a target medical image, a segmentation mask identifying style(s) to be transferred to area(s) of the target, and source medical image(s) including the style(s) may be received. Using the mask, the target may be divided into tile(s) corresponding to the area(s) and input to a trained machine learning system. For each tile, gradients associated with a content and style of the tile may be output by the system. Pixel(s) of at least one tile of the target may be altered based on the gradients to maintain content of the target while transferring the style(s) of the source(s) to the target. The synthetic medical image may be generated from the target based on the altering.
1. A system for generating a synthetic medical image using style transfer, the system comprising:
a processor; and
a memory coupled to the processor and storing instructions that, when executed by the processor, cause the system to perform operations comprising:
receiving a target medical image;
receiving a segmentation mask;
receiving one or more source medical images;
using the segmentation mask, dividing the target medical image into one or more tiles;
providing the one or more tiles as input to a trained machine learning system;
receiving, for each of the one or more tiles, gradients associated with a content and a style of the respective tile as output of the trained machine learning system;
altering one or more pixels of at least one of the one or more tiles of the target medical image based on the gradients to maintain content of the target medical image while transferring one or more styles from the one or more source medical images to the target medical image; and
generating the synthetic medical image from the target medical image based on the altering.
2. The system of claim 1 , wherein the trained machine learning system includes an image classification network trained to identify global features of medical images associated with content and local features of medical images associated with style.
3. The system of claim 2 , wherein the gradients associated with the content and the style of the respective tile received as output of the trained machine learning system include a first gradient associated with global features of the respective tile identified by the image classification network and a second gradient associated with local features of the respective tile identified by the image classification network.
4. The system of claim 1 , wherein, to maintain the content of the target medical image while transferring the one or more styles of the one or more source medical images to the target medical image, a similarity in content between the target medical image and the synthetic medical image measured by a first similarity metric is maximized and a similarity in style between the synthetic medical image and the one or more source medical images measured by a second similarity metric is maximized.
5. The system of claim 4 , wherein the altering of the one or more pixels of the at least one of the one or more tiles of the target medical image based on the gradients is iterative based on the first similarity metric and the second similarity metric.
6. The system of claim 5 , wherein the altering iterates until a change in one or more of the first similarity metric or the second similarity metric from a previous alteration to a current alteration does not increase.
7. The system of claim 1 , wherein generating the synthetic medical image from the target medical image based on the altering includes replacing a corresponding tile from the one or more tiles of the target medical image with the at least one of the one or more tiles of the target medical image having the one or more pixels altered.
8. The system of claim 1 , wherein the system is caused to perform further operations including removing one or more seams present in the synthetic medical image.
9. The system of claim 1 , wherein the segmentation mask and the one or more source medical images are received as part of a request for a medical image having a data type, and the one or more styles correspond to at least a portion of the requested data type.
10. The system of claim 9 , wherein the requested data type includes at least one of: an image modality, a target anatomical region, a target morphology, a presence or absence of a condition, or a presence or absence of a treatment effect.
11. The system of claim 10 , wherein the image modality includes digital pathology, magnetic resonance imaging (MRI), computed tomography (CT), X-ray, nuclear medicine imaging, or ultrasound.
12. The system of claim 9 , wherein the target medical image is received as supplementary data, and the target medical image does not include the requested data type.
13. The system of claim 9 , wherein the processor is caused to perform further operations including:
storing the synthetic medical image in association with a label indicating the requested data type; and
subsequently providing the synthetic medical image and the label as part of a training dataset for input to a machine learning system to train the machine learning system.
14. A method for generating a synthetic medical image using style transfer, the method comprising:
receiving a target medical image;
receiving a segmentation mask;
receiving one or more source medical images;
using the segmentation mask, dividing the target medical image into one or more tiles;
providing the one or more tiles as input to a trained machine learning system;
receiving, for each of the one or more tiles, gradients associated with a content and a style of the respective tile as output of the trained machine learning system;
altering one or more pixels of at least one of the one or more tiles of the target medical image based on the gradients to maintain content of the target medical image while transferring one or more styles from the one or more source medical images to the target medical image; and
generating the synthetic medical image from the target medical image based on the altering.
15. The method of claim 14 , wherein:
the trained machine learning system includes an image classification network trained to identify global features of medical images associated with content and local features of medical images associated with style, and
the gradients associated with the content and the style of the respective tile received as output of the trained machine learning system include a first gradient associated with global features of the respective tile identified by the image classification network and a second gradient associated with local features of the respective tile identified by the image classification network.
16. The method of claim 14 , wherein, to maintain the content of the target medical image while transferring the one or more styles of the one or more source medical images to the target medical image, the altering is performed iteratively to maximize a similarity in content between the target medical image and the synthetic medical image and maximize a similarity in style between the synthetic medical image and the one or more source medical images.
17. The method of claim 14 , wherein generating the synthetic medical image from the target medical image based on the altering includes replacing a corresponding tile from the one or more tiles of the target medical image with the at least one of the one or more tiles of the target medical image having the one or more pixels altered.
18. The method of claim 14 , wherein:
the segmentation mask and the one or more source medical images are received as part of a request for a medical image having a data type, the one or more styles corresponding to at least a portion of the requested data type, and
the target medical image is received as supplementary data, the target medical image not including the requested data type.
19. The method of claim 18 , further comprising:
storing the synthetic medical image in association with a label indicating the requested data type; and
subsequently providing the synthetic medical image and the label as part of a training dataset for input to a machine learning system to train the machine learning system.
20. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform operations for generating a synthetic image using style transfer, the operations comprising:
receiving a target medical image;
receiving a segmentation mask;
receiving one or more source medical images;
using the segmentation mask, dividing the target medical image into one or more tiles;
providing the one or more tiles as input to a trained machine learning system;
receiving, for each of the one or more tiles, gradients associated with a content and a style of the respective tile as output of the trained machine learning system;
altering one or more pixels of at least one of the one or more tiles of the target medical image based on the gradients to maintain content of the target medical image while transferring one or more styles from the one or more source medical images to the target medical image; and
generating the synthetic medical image from the target medical image based on the altering.