Research On Diabetic Retinopathy Lesion Segmentation Method In Fundus Images

Diabetic retinopathy,as one of the main complications of diabetes,endangers the health of many people.Diabetic retinopathy can cause vision loss in patients and blindness in severe cases.Early diagnosis of diabetic retinopathy helps patients seek medical treatment early,thereby reducing the risk of diabetic retinopathy.Fundus image analysis based on traditional methods has been gradually phased out due to low feature expression ability and complex pre-processing and post-processing.In recent years,with the development of deep learning,the recognition accuracy of image classification,object detection,semantic segmentation and other fields has repeatedly reached new highs,deep learning have been used for the early diagnosis of diabetic retina.However,when deep learning methods are applied to the diagnosis of diabetic retina,these methods have not been optimized for diabetic retinal lesions,so their performance is poor.Therefore,this paper analyzes the characteristics of diabetic retinopathy lesions,studies semantic segmentation based on deep learning technology,and proposes two new methods for diabetic retinopathy lesion segmentation.The main work is as follows:When the semantic segmentation methods proposed for the natural image is directly transferred to the fundus image,their performance are very general,because of small but dense lesions in fundus image and the gap between the fundus images and the natural images.Therefore,this paper proposes Label Dilation,which expands the area of the lesion label,increases the supervision signal of the lesion,and makes the model more sensitive to the lesion.Then,this paper proposes a coarse-to-fine training strategy with Label Dilation.In detail,the first training stage utilises coarse label maps obtained by label dilation on original fine label maps to supervise the learning of coarse lesion segmentation model such that small lesions are identified but with inaccurate lesion boundaries.To refine lesion boundaries,the second training stage utilises original fine label maps to further fine-tuned the coarse segmentation model.Experimental results on two publicly available datasets demonstrate that our proposed two-stage training strategy is able to improve segmentation methods consistently and substantially.For the downsampling in the semantic segmentation model may cause the loss of detailed features of lesions but general upsampling is difficult to recover these features,this paper proposes a modified HRNet V2 segmentation model for lesion segmentation in fundus images.By replacing CARAFES proposed in this paper with upsampling method in HRNet V2,this model can better restore the detailed features of lesions and achieve higher accuracy.CARAFES mainly consists of three modules: the kernel prediction module,the scaling factor prediction module,and the contentaware reassembly module.The kernel prediction module predicts an upsampling kernel for each target pixel based on content.The scaling factor prediction module predicts the scaling factor for each kernel.Then the scaling factors are used to zoom the reassembly kernel.Finally,the scaled upsampling kernel are used to reassemble the feature map.Experimental results show that the modified HRNet V2 outperforms all the state-of-the-art methods.Also,CARAFES is compared with other up-sampling methods to verify the effectiveness of CARAFES.Visual analysis further demonstrates that the proposed method can better restore the features of lesions.