Image Segmentation Of Endoscopic Cataract Based On Semantic Segmentation

Medical imaging technology has become one of the indispensable computer-aided intervention methods in the clinical diagnosis and treatment of diseases.It helps to identify and locate the lesion area,detect and segment the lesions of different modalities of tissues and organs,and is widely used in clinical examination and medical aided diagnosis.Medical image segmentation is the basis of medical image analysis and is a very important research topic.Endoscopic cataract image segmentation is a branch of medical image segmentation.It is mainly used for the assistance and guidance of cataract surgery.Its segmentation performance has a direct impact on improving the efficiency and effect of surgery.Therefore,it has received certain attention in recent years.In MICCAI,the top international conference in the field of medical image analysis,related competitions are often held.The segmentation challenges of endoscopic cataract images mainly include: difficulties in labeling and data acquisition of surgical video datasets,distribution of instances of cataract images,and unbalanced pixel distribution.This thesis focuses on in-depth research on the segmentation of endoscopic cataract images,especially for the problem of difficult to obtain labeled data.Based on the framework of semantic segmentation,a method combining self-supervision and attention mechanism is proposed to improve the segmentation performance.The main work of this thesis is as follows:(1)The research is based on weakly supervised learning to achieve semantic segmentation of surgical instruments in endoscopic cataract images.The class activation map is used to segment the surgical instrument class of cataract images,and the hole convolution is introduced into the class activation map.At the same time,a feature integration module is designed to enhance the hole class activation map to improve the dense seed area and provide it for subsequent The segmentation network is trained to improve the segmentation performance.Postprocessing of the network model output uses fully connected conditional random fields to finely segment the edges of surgical instruments.(2)Research on self-supervised endoscopic cataract coordinate attention network,propose CA-PSP model,first use BYOL self-supervised model to train parameters for cataract dataset,conduct autonomous learning without supervision,and transfer the trained model parameters to A new model is used to help training;secondly,Resnet50 is used as the backbone network to extract features,a coordinate attention mechanism module is added between each residual module in the backbone network,and Resnet is combined with the improved enhanced hole class activation map to help The network focuses on subtle category segmentation such as surgical instruments,and uses the PSPNet pyramid model to obtain contextual information between instances to improve the model’s segmentation performance on the cataract dataset.(3)Comparing the self-supervised coordinate attention network CA-PSP with the four methods of U-net,PSPNet,DANet and Deeplabv3+ in three types of experimental settings,the experimental results show that adding self-supervision and coordinate The network after the attention module achieves an m Io U accuracy of0.861 on the Ca DIS-8 class,and the 17-class and 25-class experiments both outperform other networks,verifying the effectiveness of the proposed method.