Multiple lesion segmentation in diabetic retinopathy with dual-input attentive RefineNet

To address the issue of complex structure, various sizes and the interclass similarity of different lesions, this paper proposes a dual-input attentive RefineNet (DARNet) for automatic multiple lesion segmentation of diabetic retinopathy. DARNet includes a global image encoder, local image encoder and attention refinement decoder. The whole image and the patch image are used as the dual input and fed into ResNet50 and ResNet101 for down-sampling, respectively. The high-level attention refinement decoder adopts a dual attention mechanism to integrate the same-level features in the two encoders with the output of the low-level attention refinement module for multiscale feature fusion, which focuses the model on the lesion area to generate accurate predictions. We evaluated the segmentation performance of four lesions on three datasets, and the proposed method reached an average accuracy of 0.9582/0.9617/0.9578 and a dice score of 0.9521/0.9637/0.9508 on IDRiD, E-ophtha and DDR. Extensive experimental results demonstrate the proposed DARNet outperforms the state- of-the-art models and has better robustness and accuracy. It not only preserves the contour details and shape features of multiscale lesions, but also overcomes the interference of similar tissues and noises to realize accurate multi-lesion segmentation.

Automated summary

The paper proposes a dual-input attentive RefineNet (DARNet) for automatic multiple lesion segmentation of diabetic retinopathy, which integrates dual input and attention mechanism to achieve accurate multi-scale lesion segmentation. Extensive experimental results demonstrate that DARNet outperforms state-of-the-art models in terms of robustness and accuracy by preserving contour details and shape features of multiscale lesions and overcoming interference of similar tissues and noises.