(Sunday, 7/12/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room: Track 1
Purpose: develop an unsupervised domain adaptation (CT to T1w, T2w MRI)-based multiple MRI sequence segmentation without expert segmented MRI datasets for MR-guided radiotherapy.
Methods: developed and validated an unsupervised adversarial deep domain adaptation approach for multi-organ segmentation from multiple MRI sequences (T1w, T2w) lacking expert segmented training sets by leveraging expert segmented CT datasets from unrelated set of patients. Our approach combines a one-to-many modality domain adaptation with separate modality Unet segmentation models all trained jointly and end-to-end. Our feature disentanglement approach used a single universal image content feature encoder (R?????}; H is height, W is width, C is number of channels), which was combined with target modalities’ specific style encoders (R¹??) and target code (R?; d is number of modalities) to produce image-to-image (I2I) transformation. The I2I translations were constrained to prevent modality hallucination or loss of structures of interest by using a multi-domain joint density structure discriminator that penalized image translations by combining the translated images and the voxel-wise segmentation probability maps generated by the segmentation networks as a joint density. We trained our approach on 30 CT, 20 T2w and T1w MRIs for segmenting abdominal organs including the liver, spleen, right and left kidneys. Segmentation accuracy was evaluated by comparing against clinical delineations using the Dice similarity coefficient (DSC) and performance comparisons done against two other methods.
Results: method produced the highest accuracy with an average DSC of 0.85 on T1w (organ specific DSC: liver 0.90; spleen 0.84, left kidney 0.81; right kidney 0.83) and 0.90 on T2w (organ specific DSC: liver 0.90, spleen 0.86, left kidney 0.92, right kidney 0.90) on the test set of 20 T1w, T2w MRI.
Conclusion: developed a new one-to-many unsupervised domain adaptation approach to segment from multiple MRI sequences without requiring expert-segmented MRI sets.
Keywords
Taxonomy
IM- MRI : Machine learning, computer vision
Contact Email
