Synthetic training data generation for improved machine learning model generalizability
Systems and techniques that facilitate synthetic training data generation for improved machine learning generalizability are provided. In various embodiments, an element augmentation component can generate a set of preliminary annotated training images based on an annotated source image. In various aspects, a preliminary annotated training image can be formed by inserting at least one element of interest or at least one background element into the annotated source image. In various instances, a modality augmentation component can generate a set of intermediate annotated training images based on the set of preliminary annotated training images. In various cases, an intermediate annotated training image can be formed by varying at least one modality-based characteristic of a preliminary annotated training image. In various aspects, a geometry augmentation component can generate a set of deployable annotated training images based on the set of intermediate annotated training images. In various instances, a deployable annotated training image can be formed by varying at least one geometric characteristic of an intermediate annotated training image. In various embodiments, a training component can train a machine learning model on the set of deployable annotated training images.
1. A system, comprising:
a processor that executes computer-executable instructions stored in a memory, which causes the processor to:
access an annotated source image;
generate a set of preliminary annotated training images based on the annotated source image, wherein each preliminary annotated training image is formed by inserting a respective permutation of visual objects into the annotated source image, wherein such visual objects include medical equipment or biological symptoms;
generate a set of intermediate annotated training images based on the set of preliminary annotated training images, wherein each intermediate annotated training image is formed by applying a respective permutation of modality characteristic variations to a respective preliminary annotated training image, wherein such modality characteristic variations include changes to image properties that depend upon settings or parameters of a medical imaging device that captured or generated the annotated source image; and
generate a set of deployable annotated training images based on the set of intermediate annotated training images, wherein each deployable annotated training image is formed by applying a respective permutation of geometric variations to a respective intermediate annotated training image, wherein such geometric variations include spatial transformations of image pixel grids.
2. The system of claim 1 , wherein execution of the computer-executable instructions further causes the processor to:
train a machine learning model on the set of deployable annotated training images.
3. The system of claim 1 , wherein the system maintains an element catalog that lists a set of images of possible visual objects that are insertable into the annotated source image, wherein the system maintains a list of modality characteristics that are modifiable in the preliminary training images, and wherein the system maintains a list of geometric transformations that are appliable to the intermediate training images.
4. The system of claim 3 , wherein the processor updates the element catalog by including within the element catalog a new image of a visual object, wherein the processor updates the list of modality characteristics by including within the list of modality characteristics new image properties that relate to device modality, and wherein the processor updates the list of geometric transformations by including within the list of geometric transformations new spatial operations that are appliable to images.
5. The system of claim 2 , wherein the visual objects are objects of interest which the machine learning model is configured to detect.
6. The system of claim 1 , wherein the processor randomly localizes the visual objects in a range of biologically-possible locations within the annotated source image.
7. The system of claim 1 , wherein the applying a respective permutation of modality characteristic variations includes varying an image gamma level, varying an image blur level, varying an image brightness level, varying an image contrast level, varying an image noise level, varying an image texture, varying an image resolution, varying an image field of view, or applying a modality artifact.
8. The system of claim 1 , wherein the applying a respective permutation of geometric variations includes rotating about an image axis, reflecting about an image axis, image magnifying, image panning, image tilting, or image distorting.
9. A computer-implemented method, comprising:
accessing, by a device operatively coupled to a processor, an annotated source image;
generating, by the device, a set of preliminary annotated training images based on the annotated source image, wherein each preliminary annotated training image is formed by inserting a respective permutation of visual objects into the annotated source image, wherein such visual objects include medical equipment or biological symptoms;
generating, by the device, a set of intermediate annotated training images based on the set of preliminary annotated training images, wherein each intermediate annotated training image is formed by applying a respective permutation of modality characteristic variations to a respective preliminary annotated training image, wherein such modality characteristic variations include changes to image properties that depend upon settings or parameters of a medical imaging device that captured or generated the annotated source image; and
generating, by the device, a set of deployable annotated training images based on the set of intermediate annotated training images, wherein each deployable annotated training image is formed by applying a respective permutation of geometric variations to a respective intermediate annotated training image, wherein such geometric variations include spatial transformations of image pixel grids.
10. The computer-implemented method of claim 9 , further comprising:
training, by the device, a machine learning model on the set of deployable annotated training images.
11. The computer-implemented method of claim 9 , further comprising:
maintaining, by the device, an element catalog that lists a set of images of possible visual objects that are insertable into the annotated source image;
maintaining, by the device, a list of modality characteristics that are modifiable in the preliminary training images; and
maintaining, by the device, a list of geometric transformations that are appliable to the intermediate training images.
12. The computer-implemented method of claim 11 , further comprising:
updating, by the device, the element catalog by including within the element catalog a new image of a visual object;
updating, by the device, the list of modality characteristics by including within the list of modality characteristics new image properties that relate to device modality; and
updating, by the device, the list of geometric transformations by including within the list of geometric transformations new spatial operations that are appliable to images.
13. The computer-implemented method of claim 10 , wherein the visual objects are objects of interest which the machine learning model is configured to detect.
14. The computer-implemented method of claim 9 , further comprising:
randomly localizing, by the device, the visual objects in a range of biologically-possible locations within the annotated source image.
15. The computer-implemented method of claim 9 , wherein the applying a respective permutation of modality characteristic variations includes varying an image gamma level, varying an image blur level, varying an image brightness level, varying an image contrast level, varying an image noise level, varying an image texture, varying an image resolution, varying an image field of view, or applying a modality artifact.
16. The computer-implemented method of claim 9 , wherein the applying a respective permutation of geometric variations includes rotating about an image axis, reflecting about an image axis, image magnifying, image panning, image tilting, or image distorting.
17. A non-transitory computer program product for facilitating synthetic training data generation for improved machine learning generalizability, the non-transitory computer program product comprising a computer readable memory having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:
access an annotated source image;
generate a set of preliminary annotated training images based on the annotated source image, wherein each preliminary annotated training image is formed by inserting a respective permutation of visual objects into the annotated source image, wherein such visual objects include medical equipment or biological symptoms;
generate a set of intermediate annotated training images based on the set of preliminary annotated training images, wherein each intermediate annotated training image is formed by applying a respective permutation of modality characteristic variations to a respective preliminary annotated training image, wherein such modality characteristic variations include changes to image properties that depend upon settings or parameters of a medical imaging device that captured or generated the annotated source image; and
generate a set of deployable annotated training images based on the set of intermediate annotated training images, wherein each deployable annotated training image is formed by applying a respective permutation of geometric variations to a respective intermediate annotated training image, wherein such geometric variations include spatial transformations of image pixel grids.
18. The non-transitory computer program product of claim 17 , wherein the program instructions are further executable to cause the processor to:
train a machine learning model on the set of deployable annotated training images.
19. The non-transitory computer program product of claim 17 , wherein the program instructions are further executable to cause the processor to:
maintain an element catalog that lists a set of images of possible visual objects that are insertable into the annotated source image;
maintain a list of modality characteristics that are modifiable in the preliminary training images; and
maintain a list of geometric transformations that are appliable to the intermediate training images.
20. The non-transitory computer program product of claim 19 , wherein the program instructions are further executable to cause the processor to:
update the element catalog by including within the element catalog a new image of a visual object;
update the list of modality characteristics by including within the list of modality characteristics new image properties that relate to device modality; and
update the list of geometric transformations by including within the list of geometric transformations new spatial operations that are appliable to images.