Retinal Image Classification via Vasculature-Guided Sequential Attention

Mengliu Zhao, Ghassan Hamarneh; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0

Age-related macular degeneration and diabetic retinopathy are diseases of increasing prevalence globally in recent years. Traditionally, diagnosing these diseases relied on manual visual inspection by experts, which was costly, time-consuming and laborious as it required closely examining high-resolution color fundus images. More recently, deep learning networks have shown great potential in predicting diseases from retinal images. However, being purely data-driven, these networks are susceptible to overfitting and their training requires large annotated data. In this paper, we propose to enrich deep learning-based fundus image classifiers with prior knowledge on special structures in the retina implicated with the disease. In particular, we leverage vessel priors to guide the attention mechanism of deep learning architectures. In addition, we leverage a bi-directional dual-layer LSTM module to learn the inter-dependencies between a sequence of prior-guided attention maps deployed across the depth of the disease classification network. Results on the clinical datasets show the proposed method could bring performance improvement by as much as 8%?