Incorporating medical domain knowledge into data-driven method: A vessel attention guided multi-granularity network for automatic cataract classification

Abstract

Early cataract screening is essential to help patients prevent irreversible vision loss. Existing data-driven methods can automatically extract deep fundus features for automatic cataract classification. However, due to the lack of medical domain knowledge, these deep features may contain large pathological irrelevant information, thus reducing the classification accuracy of data-driven methods. To fully take the advantages of domain knowledge to enhance the discriminative representation and avoid the irrelevant impact, we comprehensively investigate how to incorporate it into data-driven models. Considering the vessel information and multi-granularity representation are two key points of domain knowledge for cataract diagnosis, we propose a vessel attention guided multi-granularity network (VAM-Net), which contains three main components. Specifically, the vessel-level segmentation subnet is firstly designed by leveraging U-shape segmentation network. The refined vessel information obtained in this step can be regarded as visual attention to enable deep network extract discriminative features. With the guidance of vessel attention, the global-local classification subnet is further proposed to eliminate the interference of irrelevant information and enhance the multi-granularity feature learning (including global structural features and local subtle features). Finally, the multiple levels predictions are integrated into the final cataract diagnosis decision via constructing multi-granularity ensemble network. In this manner, both domain knowledge and deep networks are organically combined for pursuing better diagnosis performance. By setting extensive comparative and ablation experiments, we evaluate the effectiveness of the proposed VAM-Net on the real-world cataract dataset (92.78% detection accuracy and 87.68% grading accuracy).