Hierarchical Prototypes (HierProtoPNet)

Pytorch implementation for our IEEE JBHI paper "Progressive Mining and Dynamic Distillation of Hierarchical Prototypes for Disease Classification and Localisation".

Email: chongwangsmu@gmail.com.

Introducation:

Medical image analysis tasks need to handle the complexity of diverse lesion characteristics: considerable size, shape, and appearance variations of the lesion structure.

Method:

This appraoch leverages hierarchical visual prototypes across multiple semantic feature granularities to effectively capture diverse lesion patterns. To increase utility of the prototypes, we devise a prototype mining paradigm to progressively discover semantically distinct prototypes, offering multi-level complementary analysis of complex lesions. Also, we introduce a dynamic knowledge distillation strategy that allows transferring essential classification information across hierarchical levels, thereby improving generalisation performance.

Datasets:

1. Mammographic images (CSAW-S)
2. OCT images (NEH)
3. NIH chest X-rays (NIH ChestX-ray14)

Training:

Training procedures and details can be found in main.py. Our trained chest X-ray model is provided here.

Prototype visualisation:

Prototypes are visualized as their nearest training image patches. HierProtoPNet generates semantically-dissimilar prototypes at different hierarchical levels due to the prototype mining paradigm: high-level prototypes focus on the most salient cancerous areas, the mid-level prototypes localise the difficult (i.e., less conspicuous) cancer-boundary areas, and the low-level prototypes capture sparser and finer cancer regions.

Lesion Localisation:

Breast cancer:

Thoracic disease:

Citation:

@article{wang2025progressive,
  title={Progressive Mining and Dynamic Distillation of Hierarchical Prototypes for Disease Classification and Localisation},
  author={Wang, Chong and Liu, Fengbei and Chen, Yuanhong and Kwok, Chun Fung and Elliott, Michael and Pena-Solorzano, Carlos and McCarthy, Davis James and Frazer, Helen and Carneiro, Gustavo},
  journal={IEEE Journal of Biomedical and Health Informatics},
  year={2025},
  publisher={IEEE}
}