CAT: Constrained Adversarial Training for Anatomically-Plausible Semi-Supervised Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels in semi-supervised medical image segmentation, but often produces anatomically implausible results.
• Incorporating complex anatomical constraints (e.g., connectivity, convexity) into standard deep learning is difficult due to their non-differentiable nature.

Purpose of the Study:

• To develop a novel method, Constrained Adversarial Training (CAT), for generating anatomically plausible medical image segmentations.
• To address the challenge of non-differentiable anatomical constraints in deep learning segmentation models.

Main Methods:

• Proposed Constrained Adversarial Training (CAT) method using a Reinforce algorithm to handle non-differentiable constraints.
• Employed an adversarial training strategy to generate constraint-violating examples, maximizing constraint loss for gradient acquisition.
• Updated the segmentation network to be robust against these adversarial examples, ensuring anatomical plausibility.

Main Results:

• Demonstrated effectiveness of CAT on synthetic and four clinically-relevant datasets.
• Achieved improved segmentation accuracy and enhanced anatomical plausibility compared to standard methods.
• Validated the generic applicability of CAT across different segmentation networks.

Conclusions:

• Constrained Adversarial Training (CAT) provides an effective and efficient approach for incorporating complex anatomical constraints into medical image segmentation.
• The method successfully generates anatomically plausible segmentations, addressing a key limitation of current deep learning models.