Research On Medical Image Segmentation Technology Based On Deep Learnin

Medical image segmentation technology is an image segmentation technique aimed at organs and lesion areas in medical images.Its segmentation results can assist doctors in making accurate judgments and reduce the workload of medical staff.In recent years,deep learning has been widely applied in the field of medical image segmentation,and its performance has surpassed traditional methods.Although deep learning has advanced the performance of medical image segmentation models,there are still some issues such as inaccurate extraction of model image features,and low semantic utilization of high-dimensional data.To address the above issues,this paper proposes a feature extraction module suitable for medical image segmentation and designs a deep learning model for medical image segmentation as follows:(1)Aiming to address the issues of inaccurate feature extraction and failure to capture multi-scale features in images,a novel feature extraction module named ResX block is proposed.Based on rectangular convolution and dilated convolution,the ResX block has the characteristics of few parameters and wide receptive field.By controlling the size of the convolution kernel to extract multi-scale features,the ResX block can improve the accuracy of feature extraction.This article tests the performance of the ResX block based on the U-Net framework,and verifies the efficiency and accuracy of the ResX block in extracting medical image features through testing results on five public datasets: LiTS2017,3Dircadb,LIDC,LCTSC,and Synapse Multi-organs.The performance of the ResX block is significantly higher than that of Res Net and Res2 Net.In addition,compared to the Transformer,the ResX block has advantages in both parameter quantity and performance.(2)To address the problem of poor overall performance and low semantic utilization of high-dimensional data in the U-Net model,a deep learning model for 3D medical image segmentation called the RA V-Net is proposed,which improves the U-Net basic framework from the encoding part,skip connection,and decoding part.First,in the encoding part,a CofRes Module(Composite Original Feature Residual Module)was proposed based on the conclusion of the ResX block to obtain multi-scale features,which can sensitively capture organ positions in high-dimensional data and focus on low-dimensional data features such as organ contours and edges.Secondly,the skip connection incorporates a CA Module(Channel Attention Module),which attaches high weights to semantically rich channels by matrix multiplication,to improve the semantic utilization of high-dimensional data.Finally,in the decoding part,an AR Module(Attention Recovery Module)is proposed,which provides pixel-level spatial attention to the model through LSTM.The overall performance of RA V-Net is strong,and its segmentation results are verified on the LiTS2017 and 3Dircadb public datasets,with evaluation indicators higher than other deep learning models such as SA U-Net,BCD U-Net,Res U-Net,CS2 Net,and mU-Net.In summary,the ResX block and RA V-Net proposed in this paper have advanced the performance indicators of feature extraction blocks and medical image segmentation models,providing effective references for the subsequent optimization of medical image segmentation models and having important scientific research value and practical significance.