Cross domain medical image segmentation
Systems and method are described for medical image segmentation. A medical image of a patient in a first domain is received. The medical image comprises one or more anatomical structures. A synthesized image in a second domain is generated from the medical image of the patient in the first domain using a generator of a task driven generative adversarial network. The one or more anatomical structures are segmented from the synthesized image in the second domain using a dense image-to-image network of the task driven generative adversarial network. Results of the segmenting of the one or more anatomical structures from the synthesized image in the second domain represent a segmentation of the one or more anatomical structures in the medical image of the patient in the first domain.
1. A method for medical image segmentation, comprising:
receiving a medical image of a patient in a first domain, the medical image comprising one or more anatomical structures;
generating a synthesized image in a second domain from the medical image of the patient in the first domain using a generator of a task driven generative adversarial network; and
segmenting the one or more anatomical structures from the synthesized image in the second domain using a trained dense image-to-image network of the task driven generative adversarial network, wherein results of the segmenting of the one or more anatomical structures from the synthesized image in the second domain represent a segmentation of the one or more anatomical structures in the medical image of the patient in the first domain,
wherein the generator of the task driven generative adversarial network is trained based on predicted segmentation labels of a reconstructed training image in the second domain and annotated segmentation labels of a real training image in the second domain, wherein the reconstructed training image in the second domain is generated from a synthesized training image in the first domain, which is generated from the real training image in the second domain.
2. The method of claim 1 , further comprising:
training the generator of the task driven generative adversarial network based on a comparison between 1) a synthesized training image in the second domain and its predicted segmentation labels and 2) a second real training image in the second domain and its annotated segmentation labels, wherein the synthesized training image in the second domain is generated by the generator from a real training image in the first domain.
3. The method of claim 2 , further comprising:
determining the predicted segmentation labels of the synthesized training image in the second domain using the trained dense image-to-image network.
4. The method of claim 1 , further comprising: determining the predicted segmentation labels of the reconstructed training image in the second domain using the trained dense image-to-image network.
5. The method of claim 1 , further comprising:
training the task driven generative adversarial network using unlabeled training images in the first domain and labeled training images in the second domain.
6. The method of claim 1 , wherein the trained dense image-to-image network is trained to segment the one or more anatomical structures from the synthesized image in the second domain using labeled training images in the second domain, without training images in the first domain.
7. The method of claim 2 , wherein training the generator of the task driven generative adversarial network is further based on:
a comparison between 1) a second synthesized training image in the first domain and its predicted segmentation labels and 2) a real training image in the first domain and its annotated segmentation labels, wherein the second synthesized training image in the first domain is generated by another generator of the task driven generative adversarial network from the second real training image in the second domain.
8. The method of claim 1 , wherein the medical image of the patient in the first domain is an x-ray medical image and the synthesized image in the second domain is a digitally reconstructed radiograph image.
9. An apparatus for medical image segmentation, comprising:
means for receiving a medical image of a patient in a first domain, the medical image comprising one or more anatomical structures;
means for generating a synthesized image in a second domain from the medical image of the patient in the first domain using a generator of a task driven generative adversarial network; and
means for segmenting the one or more anatomical structures from the synthesized image in the second domain using a dense image-to-image network of the task driven generative adversarial network, wherein results of the segmenting of the one or more anatomical structures from the synthesized image in the second domain represent a segmentation of the one or more anatomical structures in the medical image of the patient in the first domain.
10. The apparatus of claim 9 , further comprising:
means for training the generator of the task driven generative adversarial network based on a comparison between 1) a synthesized training image in the second domain and its predicted segmentation labels and 2) a real training image in the second domain and its annotated segmentation labels, wherein the synthesized training image in the second domain is generated by the generator from a real training image in the first domain.
11. The apparatus of claim 10 , further comprising:
means for determining the predicted segmentation labels of the synthesized training image in the second domain using the dense image-to-image network.
12. The apparatus of claim 9 , further comprising:
means for training the generator of the task driven generative adversarial network based on predicted segmentation labels of a reconstructed training image in the second domain and annotated segmentation labels of a real training image in the second domain, wherein the reconstructed training image in the second domain is generated from a synthesized training image in the first domain, which is generated from the real training image in the second domain.
13. The apparatus of claim 12 , further comprising:
means for determining the predicted segmentation labels of the reconstructed training image in the second domain using the dense image-to-image network.
14. A non-transitory computer readable medium storing computer program instructions for medical image segmentation, the computer program instructions when executed by a processor cause the processor to perform operations comprising:
receiving a medical image of a patient in a first domain, the medical image comprising one or more anatomical structures;
generating a synthesized image in a second domain from the medical image of the patient in the first domain using a generator of a task driven generative adversarial network; and
segmenting the one or more anatomical structures from the synthesized image in the second domain using a trained dense image-to-image network of the task driven generative adversarial network, wherein results of the segmenting of the one or more anatomical structures from the synthesized image in the second domain represent a segmentation of the one or more anatomical structures in the medical image of the patient in the first domain,
wherein the generator of the task driven generative adversarial network is trained based on predicted segmentation labels of a reconstructed training image in the second domain and annotated segmentation labels of a real training image in the second domain, wherein the reconstructed training image in the second domain is generated from a synthesized training image in the first domain, which is generated from the real training image in the second domain.
15. The non-transitory computer readable medium of claim 14 , the operations further comprising:
training the generator of the task driven generative adversarial network based on a comparison between 1) a synthesized training image in the second domain and its predicted segmentation labels and 2) a second real training image in the second domain and its annotated segmentation labels, wherein the synthesized training image in the second domain is generated by the generator from a real training image in the first domain.
16. The non-transitory computer readable medium of claim 14 , the operations further comprising:
training the task driven generative adversarial network using unlabeled training images in the first domain and labeled training images in the second domain.
17. The non-transitory computer readable medium of claim 14 , wherein the trained dense image-to-image network is trained to segment the one or more anatomical structures from the synthesized image in the second domain using labeled training images in the second domain, without training images in the first domain.
18. The non-transitory computer readable medium of claim 15 , wherein training the generator of the task driven generative adversarial network is further based on: a comparison between 1) a second synthesized training image in the first domain and its predicted segmentation labels and 2) a real training image in the first domain and its annotated segmentation labels, wherein the second synthesized training image in the first domain is generated by the second generator of the task driven generative adversarial network from another real training image in the second domain.
19. The non-transitory computer readable medium of claim 14 , wherein the medical image of the patient in the first domain is an x-ray medical image and the synthesized image in the second domain is a digitally reconstructed radiograph image.