SOLO-OCT:Segmentation Model Of Diabetic Macular Edema Based On Improved SOLOv2

Diabetic Macular Edema(DME)is one of the common causes of visual impairment in diabetic patients.The incidence of DME increases with the duration of diabetes,and about 20-30% of diabetic patients have DME.Therefore,early screening and diagnosis of diabetic macular edema is important for the prevention of the disease and the protection of patients’ vision.Optical Coherence Tomography(OCT)is a non-invasive imaging technique that helps to enhance the early detection and prevention of diabetic retinopathy.It can provide high-resolution,high-contrast retinal tomography images.By using OCT technology,it can enhance the monitoring and early detection of diabetic retinopathy by improving the accuracy and sensitivity in the early diagnosis of diabetic retinopathy.Compared with traditional manual examinations,AI technology can detect DME lesions more quickly and accurately,reducing errors in manual judgment while greatly improving diagnostic efficiency.Although many deep learning models have been used for medical image segmentation and good progress has been made,these deep learning models have some limitations and there are some problems that need to be solved,as follows:(1)There is a large amount of scatter noise in the DME region of OCT images.This is due to the scattering of light as it passes through the eye tissue.These scattering noises can interfere with the accurate segmentation and quantitative analysis of the DME region;(2)OCT images contain a large number of small target DME regions.These small target regions also have an impact on the analysis of diabetic macular edema;the existing deep learning models do not handle these small target regions better,and there are also problems such as missed segmentation.To address the above problems,this paper proposes a new segmentation model for diabetic macular edema(SOLO-OCT)based on the SOLOv2 model,including:(1)introducing an improved Dn CNN network;after the original image input,it is directly fed into the Dn CNN network for noise removal,and the noise information in the shallow layers of the image is extracted by using jump connections and adding channel attention mechanisms to clarify the difference between the target image and the noisy image to improve the input image quality;(2)Introducing Feature Pyramid Transformer(FPT);performing self-attention enhancement after feature pyramid output,synthesizing contextual information of different layers,and using multi-layer features for prediction to improve the recognition ability and learning ability of the model for small targets;(3)Improving Non-Maximum Suppression(NMS)algorithm;adding the calculation of the distance between the real frame and the center point of the prediction frame and using log function for non-linear weight penalty in the calculation of the suppression function to alleviate the problem of missed segmentation of small target regions.The SOLO-OCT model is compared with other example segmentation models(including Mask R-CNN,SOLO,and SOLOv2)to evaluate its segmentation performance on DME regions.Compared with Mask R-CNN,SOLO and SOLOv2 models,the SOLOOCT model improves the segmentation accuracy(m AP)of DME regions by 3.3 points and the segmentation accuracy(APs)of small target DME regions by 2.4 percentage points,while the processing time of a single image increases only by 0.0387 s.The experimental results show that the improved model in this paper is comparable to The experimental results show that the improved model in this paper performs well in all evaluation indexes and improves the efficiency and accuracy of segmentation compared with other models mentioned in the paper.The segmentation model of diabetic macular edema(SOLO-OCT)proposed in this paper can be used for large-scale diabetic retinopathy screening.