Research And Implementation Of Pathological Myopia Recognition And Lesion Segmentation Technology Based On Deep Learning

Myopia is a global public health and social problem,and its prevalence has significantly increased in recent years.Pathological myopia is a severe form of myopia,with various types of lesions at different stages and irreversible damage to the fundus of the eye.Therefore,accurate identification of lesion morphology is crucial for the clinical diagnosis and grading treatment of pathological myopia.Nowadays,computer technology can effectively assist in the diagnosis of pathological myopia,but due to the complex retinal structure and diverse lesion types,the identification and segmentation of pathological myopia are still challenging.In recent years,deep learning technology has shown outstanding performance in medical image classification and segmentation.Therefore,based on existing deep learning networks,this thesis proposes a deep learning-based algorithm for the identification and lesion segmentation of pathological myopia,with a focus on improving recognition efficiency and segmentation accuracy.The specific research contributions are as follows:(1)The thesis proposes a deep learning model called Rep VGG-DAPPM for identifying pathological myopia from retinal images.Compared to commonly used feature extraction networks such as VGG,Google Net,and Res Net,the Rep VGG network has higher accuracy,fewer parameters,and faster inference speed.Therefore,the Rep VGG network is chosen as the feature extraction backbone and a Block＿A block is added to enhance its feature extraction capability.Then,the Deep Aggregation Pyramid Pooling Module(DAPPM)is applied to further process the output features of the backbone network and fuse multi-scale contextual information in a cascaded way,aiming to extract more semantic information while minimizing the increase in inference time.Experimental results show that compared to other deep learning models,Rep VGG-DAPPM performs better and can effectively identify pathological myopia from retinal images.(2)Aiming at the problem that pathological myopia has multiple lesion types and significant morphological differences,which leads to difficulties in segmentation,this thesis proposes an improved segmentation model based on U-Net.This model replaces the original convolutions in U-Net with Series Deformable Convolution(SDC),which can learn lesion features more flexibly and improve segmentation accuracy.A Triple Attention Fusion(TAF)module is introduced into the deep feature extraction part of the encoder to enhance the semantic information on the deep feature map,thereby improving the model’s segmentation ability for small lesions.At the same time,the Feature Guided Attention Module(FGAM)is used in shallow feature layers to promote the model’s intrinsic feature representation capability and enhance its ability to segment lesion edges.In addition,this thesis presents an improved loss function that optimizes the network from area overlap and shape similarity of the segmentation results,which further improves the segmentation performance of the network.The experimental results show that the proposed model improves the segmentation accuracy of small lesions and can clearly segment lesion edges,demonstrating an overall superior segmentation performance.(3)Based on the models designed in this study,combined with Py Qt5-GUI design and My SQL database,an online system for automatic recognition and segmentation of pathological myopia is established,which can automatically identify pathological images and accurately segment various types of lesions in fundus color images.In the future,this system can be further explored in clinical applications to assist doctors in better diagnosing and treating pathological myopia.