ACTION++: Improving Semi-supervised Medical Image Segmentation with Adaptive Anatomical Contrast

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical data frequently displays long-tail distributions and severe class imbalance, complicating the classification of minority classes.
• While unsupervised contrastive learning aids semi-supervised medical image segmentation in long-tailed scenarios, its effectiveness on imbalanced labeled data remains uncertain.

Purpose of the Study:

• To introduce ACTION++, an enhanced contrastive learning framework for semi-supervised medical segmentation.
• To address the challenges of class imbalance in medical image segmentation using adaptive anatomical contrast.

Main Methods:

• Developed an adaptive supervised contrastive loss function for online contrastive matching with uniformly distributed class centers.
• Introduced a dynamic temperature parameter, adjusted via a cosine schedule, to improve class separation in long-tailed medical data.

Main Results:

• ACTION++ achieved state-of-the-art performance on the ACDC and LA benchmarks across two semi-supervised settings.
• Empirical evaluations demonstrated the framework's effectiveness in improving segmentation accuracy for imbalanced medical datasets.

Conclusions:

• The proposed adaptive anatomical contrast method offers superior performance and label efficiency in semi-supervised medical segmentation.
• ACTION++ provides a robust solution for segmenting imbalanced medical data, outperforming existing methods.