Dynamic-weighting hierarchical segmentation network for medical images

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Automatic medical image segmentation is vital for disease diagnosis and prognosis.
• Current deep learning models often struggle with diverse feature representations using standard fully connected layers.

Purpose of the Study:

• To introduce a novel Hierarchical Segmentation (HieraSeg) Network with a Hierarchical Fully Connected (HFC) layer for improved medical image segmentation.
• To develop a Dynamic-Weighting HieraSeg (DW-HieraSeg) Network for enhanced robustness and accelerated training.

Main Methods:

• Proposed a Hierarchical Fully Connected (HFC) layer that decouples categories into subcategories using multiple weight vectors.
• Introduced Image-level Weight Net (IWN) and Pixel-level Weight Net (PWN) for data-driven curriculum learning in DW-HieraSeg.
• Implemented a class-balanced loss to prevent overfitting in minority classes.

Main Results:

• HieraSeg and DW-HieraSeg Networks achieved state-of-the-art performance on EndoScene, ISIC, and Decathlon benchmark datasets.
• The HFC layer effectively captures variant characteristics for accurate decision boundaries.
• DW-HieraSeg demonstrated robustness against lesion variability and faster training.

Conclusions:

• The proposed HieraSeg and DW-HieraSeg Networks significantly advance medical image segmentation capabilities.
• The novel HFC layer and dynamic weighting strategies offer effective solutions for complex segmentation tasks.
• These methods show great potential for clinical applications in medical image analysis.