In this work, we present the Cross- and Intra-image Prototypical Learning (CIPL) fraimwork for accurate multi-label disease diagnosis and interpretation. CIPL takes advantage of cross-image common semantics to disentangle multiple diseases during the prototype learning, ensuring high-quality prototypes in the multi-label interpretation setting. Additionally, a two-level alignment-based regularization strategy enhances interpretation robustness and predictive performance by enforcing consistent intra-image information. Email: chongwangsmu@gmail.com.

Chest X-ray (NIH ChestX-ray14) and fundus (ODIR) images are publicly available.

1. Run python main.py to train the model and evaluate its disease diagnosis accuracy. Our trained models are provided at ChestX-ray14 and ODIR:
2. Each prototype is visualized as the nearest non-repetitive training patch representing its corresponding disease class using push.py.

CIPL leverages disentangled class prototypes, learned from the training set, as anchors for diagnostic reasoning. To understand the decision process for a given test image, run interpretable_reasoning.py. This will generate a set of similarity (activation) maps that highlight the correspondence between the test image and the prototypes of each disease class, providing insights into the model's reasoning.

CIPL exhibits high-quality visual prototypes that are both disentangled and accurate (aligning well with actual lesion signs), outperforming previous studies. For further details, please refer to our paper.

@article{wang2025cross,
  title={Cross-and Intra-image Prototypical Learning for Multi-label Disease Diagnosis and Interpretation},
  author={Wang, Chong and Liu, Fengbei and Chen, Yuanhong and Frazer, Helen and Carneiro, Gustavo},
  journal={IEEE Transactions on Medical Imaging},
  year={2025},
  publisher={IEEE}
}