Progressive Mining and Dynamic Distillation of Hierarchical Prototypes for Disease Classification and Localisation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Effective lesion representation is crucial for disease classification and localization in medical imaging.
• Existing prototype-based models struggle with lesion complexity due to fixed-size, single-level prototypes and redundancy.

Purpose of the Study:

• To develop a novel prototype-based framework, HierProtoPNet, for enhanced lesion representation in medical images.
• To address the limitations of current models in handling diverse lesion characteristics and prototype redundancy.

Main Methods:

• HierProtoPNet utilizes hierarchical visual prototypes across different semantic feature granularities.
• A novel prototype mining paradigm progressively discovers semantically distinct prototypes to prevent redundancy.
• A dynamic knowledge distillation strategy facilitates information transfer across hierarchical levels.

Main Results:

• HierProtoPNet achieved state-of-the-art classification performance on binary breast cancer screening, multi-class retinal disease diagnosis, and multi-label chest X-ray classification benchmarks.
• The framework demonstrated significant advantages in weakly-supervised disease localization and segmentation tasks.

Conclusions:

• HierProtoPNet effectively captures diverse lesion patterns using hierarchical prototypes and a novel mining paradigm.
• The proposed dynamic knowledge distillation enhances generalization, leading to superior performance in classification, localization, and segmentation.