Diversity-Enhanced Consistency Learning for Label-Efficient Cardiac Image Segmentation

Cardiovascular diseases are among the leading causes of death worldwide, highlighting the need for automated cardiac structure analysis on MRI to improve diagnostic accuracy. While supervised learning has significantly advanced medical image segmentation, relying on annotated datasets continues to be a problem, particularly in cardiac images, where the availability of labelled is limited. To overcome this limitation, we present an innovative label-efficient approach: the multi-teacher Consistency Model, specifically tailored for label-efficient cardiac image segmentation. The methods focus on maximizing the utility of unlabelled data to achieve high performance while minimizing dependence on labels. This approach utilizes a teacher-student paradigm and incorporates a diverse teacher design to capture varying viewpoints. This addresses the inherent variability in cardiac structures and anatomical features while avoiding local optima. The method integrates a hybrid loss function that combines supervised components with consistency-based constraints to align the outputs of the student and teachers. Additionally, we employed multiple strategies to reinforce the stability and reliability of the student-teacher architecture. Assessment using the MM-WHS Cardiac image database reveals that the introduced method exceeds the performance of leading label-efficient approaches, especially in cases with minimal labels.