ProCNS: Progressive Prototype Calibration and Noise Suppression for Weakly-Supervised Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Weakly-supervised segmentation (WSS) uses sparse annotations to reduce annotation costs in medical imaging.
• Existing WSS methods struggle with ambiguous boundaries and noisy regions, leading to performance degradation.
• There's a need for robust WSS methods that handle annotation noise effectively.

Purpose of the Study:

• To introduce ProCNS, a novel WSS framework designed to improve segmentation accuracy.
• To address limitations in current WSS approaches, particularly regarding noisy pseudo-labels and boundary ambiguity.
• To develop a method that leverages prototype calibration and noise suppression for enhanced medical image segmentation.

Main Methods:

• Proposing a Prototype-based Regional Spatial Affinity (PRSA) loss to refine spatial and semantic element affinities.
• Introducing an Adaptive Noise Perception and Masking (ANPM) module for robust prototype representation and noise handling.
• Generating specialized soft pseudo-labels for noisy regions to provide supplementary supervision.

Main Results:

• The ProCNS framework significantly outperforms state-of-the-art methods in medical image segmentation.
• Experiments across six diverse medical imaging tasks and modalities validate the proposed approach.
• The synergistic modules effectively improve segmentation performance by mitigating errors from noisy regions.

Conclusions:

• ProCNS offers a robust solution for weakly-supervised medical image segmentation.
• The proposed PRSA loss and ANPM module effectively handle noisy annotations and ambiguous boundaries.
• This framework demonstrates significant advancements in WSS, particularly for complex medical imaging applications.