Consistency label-activated region generating network for weakly supervised medical image segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Current medical image auto-segmentation methods struggle with imprecise pathological labeling.
• Clinical practice often relies on broad labels (disease/normal) rather than detailed segmentation masks.
• Weakly supervised learning offers potential for leveraging these common clinical labels.

Purpose of the Study:

• To develop a weakly supervised medical image segmentation model capable of generating accurate lesion masks.
• To utilize class activation maps (CAM) and cycle-consistency for improved segmentation guidance.
• To address challenges in boundary definition for segmentation using weak labels.

Main Methods:

• Proposed a weakly supervised network guided by class activation maps (CAM) and cycle-consistency for label-activated region transfer.
• Introduced a complementary branches fusion module to refine segmentation boundaries and preserve semantic information.
• Employed a joint discrimination strategy to enhance lesion mask precision during image synthesis.

Main Results:

• The model accurately identifies pixel-level regions in medical images while maintaining overall semantic structure.
• The complementary branches fusion module effectively clarifies lesion and non-lesion boundaries.
• Achieved superior performance compared to state-of-the-art methods on BraTs, ISIC, and COVID-19 datasets.

Conclusions:

• Weakly supervised learning, guided by CAM and cycle-consistency, can effectively generate precise medical image segmentation masks.
• The proposed complementary branches fusion module significantly improves boundary definition in segmentation.
• The method demonstrates robust performance across diverse medical imaging datasets, offering a promising approach for clinical applications.