Interpretable Semantic Medical Image Segmentation with Style and Confidence.
Authors
Abstract
The scarcity of semantically labelled data presents major challenges for medical image segmentation using deep learning models, and the "black-box" nature of these models inherently limits their interpretability during clinical deployment. To address these issues, we propose Generative Adaptable Segmentation Evolution (GASE), an end-to-end, style-based generative adversarial framework for robust and interpretable medical image segmentation under extreme data scarcity. Trained on limited single-sequence Magnetic Resonance (MR) data, GASE self-adapts to unseen acquisition-level image variations-arising from differences in scanning protocols, imaging sequences, and patient demographics-while simultaneously improving interpretability through automatic input validity assessment and output reliability estimation. During adversarial training, the style-learning generator captures a manifold-like representation of input style features for valid input interpretation and diversifies labelled training pairs through style interpolation to simulate acquisition-level image variations. The segmentation-based discriminator, trained on these diversified samples, further improves adaptation to unseen styles and automatically estimates the reliability of predicted masks via confidence learning. Extensive experiments on knee and pelvis MR image datasets demonstrate that GASE achieves high segmentation accuracy for assessments of multiple tissues under examination, strong adaptability to unseen acquisition-level image variations, and intrinsic interpretability, underscoring its potential for reliable clinical deployment.