TSNet: Task-specific network for joint diabetic retinopathy grading and lesion segmentation of ultra-wide optical coherence tomography angiography images

Diabetic retinopathy (DR) is a common complication of diabetes which may lead to blindness. Early diagnosis can effectively prevent the deterioration of the disease and enable timely treatment. Ophthalmologists diagnose DR by observing ultra-wide optical coherence tomography angiography (UW-OCTA) images, which visualize unprecedented detail of DR lesions. In this paper, we propose an end-to-end task-specific network (TSNet) for joint DR grading and lesion segmentation of UW-OCTA images. Specifically, we design task-specific attention block to generate task-specific feature maps for respective segmentation and classification tasks. Furthermore, we devise task-specific fusion block to fuse the original task-specific feature map and augmented task-specific feature map for the following segmentation and classification decoders to generate DR lesion predictive mask and DR grading predictive result. Experiments on a public-available UW-OCTA dataset demonstrate that our model outperforms state-of-the-art (SOTA) multi-task models and achieves promising results on both DR lesion segmentation and DR grading classification tasks