Multi-lesion segmentation guided deep attention network for automated detection of diabetic retinopathy

Accurate multi-lesion segmentation together with automated grading on fundus images played a vital role in diagnosing and treating diabetic retinopathy (DR). Nevertheless, the intrinsic patterns of fundus lesions aggravated challenges in DR detection process. Therefore, we proposed a novel multi-lesion segmentation guided deep attention network (MSGDA-Net) for accurate and automated DR detection, consisting of a DR lesion segmentation pathway as an auxiliary task to produce a lesion regional prior knowledge and a DR grading pathway to extract local fine-grained features and long-range dependency. In DR lesion segmentation pathway, we designed a Multi-Scale Attention Block (MSAB) and a Lesion-Aware Relation Block (LARB) to allow interactions among multi-lesion features for alleviating ambiguity in lesion segmentation, generating lesion regional prior knowledge. As for DR grading pathway, we presented a Spatial-Fusion Block (SFB) to enhance the lesion-related local fine-grained feature representations and eliminate irrelevant noise information under the guidance of the resulting lesion regional priors, while constructed an Enhanced Self-Attention Block (ESAB) to optimally fuse fine-grained features from SFB with long-range global-context information for grading DR. The experimental results showed that our MSGDA-Net not only achieved state-of-the-art performance in the tasks of multi-lesion segmentation and DR grading, reaching up to 49.21 % Dice, 38.05 % IoU and 51.15 % AUPR for DR lesion segmentation on the DDR dataset, as well as accuracy values of 75.00 % and 87.18 % for DR grading on local newly-built VisionDR and publicly available APTOS datasets, but also manifested good generalization and robustness on cross-data evaluation. It could serve as a promising tool for computer aided DR screening and diagnosis.