Hierarchical Multi-Class Group Correlation Learning Network for Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Machine learning
• Computer vision

Background:

• Hierarchical methods are effective for multi-label segmentation but often neglect deep label space relationships.
• Existing approaches may impose constraints only on shallow layers, limiting segmentation accuracy.

Purpose of the Study:

• To introduce a novel hierarchical multi-class group correlation learning (HMGC) method.
• To overcome limitations of existing hierarchical approaches by considering deep relationships in the label space.
• To improve segmentation accuracy in medical imaging tasks.

Main Methods:

• Transformed regional constraints into voxel vector correlations in a high-dimensional space.
• Computed a voxel vector correlation matrix to group voxel vectors and reduce disparities.
• Introduced two loss functions: intra-class group loss and inter-class group loss.

Main Results:

• Demonstrated effectiveness on BraTS2018, BraTS2019, and BraTS2020 datasets for brain tumor segmentation.
• Achieved first place in overall score on the BraTS2020 dataset.
• Showcased competitive results in cardiac segmentation on the ACDC MICCAI'17 Challenge Dataset.

Conclusions:

• HMGC effectively mitigates bias propagation and enhances segmentation accuracy.
• The method shows strong performance and generalization capabilities across different medical imaging datasets.
• HMGC represents a significant advancement in hierarchical multi-label segmentation for medical applications.