In view of the complex structure of diabetic retinopathy and the large differences in the scales of different lesions,a novel network which integrates deformable convolution and attention mechanism is proposed for automatic diabetic retinopathy multi-lesion segmentation.Specifically,deformable convolution Haar wavelet transform encoder takes place of the original convolutional downsampling encoder to adapt to the irregular shape changes of lesions and extract effective feature information;a dense feature perception and aggregation module is introduced at the bottleneck layer to extract multi-scale features by aggregating multiple receptive fields,thus enhancing deep semantic information;and finally,in order to fully integrate the decoder output and improve the recognition accuracy of edge information,a multi scale adaptive fusion module is used to weight the decoder output of each layer for obtaining the most accurate segmentation feature map.The validation of hard percolation,bleeding point,and soft percolation segmentations on the DDR-RLS dataset reveals that the proposed network shows increases of 0.026 2,0.051 8 and 0.046 5 in IoU coefficient,0.027 1,0.058 1 and 0.050 4 in Dice coefficient,and 0.0423,0.0691 and 0.0734 in AUPR value,as compared with the original Unet.

In view of the complex structure of diabetic retinopathy and the large differences in the scales of different lesions,a novel network which integrates deformable convolution and attention mechanism is proposed for automatic diabetic retinopathy multi-lesion segmentation.Specifically,deformable convolution Haar wavelet transform encoder takes place of the original convolutional downsampling encoder to adapt to the irregular shape changes of lesions and extract effective feature information;a dense feature perception and aggregation module is introduced at the bottleneck layer to extract multi-scale features by aggregating multiple receptive fields,thus enhancing deep semantic information;and finally,in order to fully integrate the decoder output and improve the recognition accuracy of edge information,a multi scale adaptive fusion module is used to weight the decoder output of each layer for obtaining the most accurate segmentation feature map.The validation of hard percolation,bleeding point,and soft percolation segmentations on the DDR-RLS dataset reveals that the proposed network shows increases of 0.026 2,0.051 8 and 0.046 5 in IoU coefficient,0.027 1,0.058 1 and 0.050 4 in Dice coefficient,and 0.0423,0.0691 and 0.0734 in AUPR value,as compared with the original Unet.