CCA: Contrastive cluster assignment for supervised and semi-supervised medical image segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Artificial Intelligence

Background:

• Transformer models show promise in semantic segmentation but often overlook crucial cross-attention between pixel and class features.
• This limitation hinders the full potential of transformers in complex vision tasks, particularly in medical imaging.

Purpose of the Study:

• To introduce a novel method, cluster learning and contrastive cluster assignment (CCA), to enhance transformer-based medical image segmentation.
• To improve the integration of object queries for better feature-level clustering and pixel class prediction.

Main Methods:

• Developed cluster learning to align object queries with feature-level cluster centers.
• Introduced contrastive cluster assignment (CCA) to guide pixel class prediction using learned cluster centers.
• Designed plug-in modules for both supervised (decoder enhancement) and semi-supervised (encoder enhancement) segmentation networks.

Main Results:

• Demonstrated consistent outperformance over state-of-the-art models across multiple public medical image datasets (ACDC, LA, Synapse, ISIC2018).
• Validated the effectiveness of CCA in improving pixel-level predictions and feature extraction capabilities.
• Showcased the versatility of CCA as a plug-in module applicable to various transformer-based architectures.

Conclusions:

• The proposed cluster learning and contrastive cluster assignment (CCA) method significantly advances medical image segmentation using transformers.
• CCA effectively addresses the limitations of cross-attention, leading to superior performance in both supervised and semi-supervised settings.
• The plug-in nature of CCA allows for broad applicability and enhancement of existing segmentation models.