Research On Segmentation Method Of Rectal Tumor MRI Image Based On Deep Learning

By studying deep learning technology,especially deep convolutional neural network model,this paper realized the high-precision automatic segmentation of rectal tumor Magnetic Resonance Imaging(MRI).In view of the common problems in medical images: first,the amount of available data is small;second,the contrast between the lesion area and other soft tissues in the image is not obvious,and the target and background area categories are imbalanced;third,the lack of easy-to-use automation segmentation system.This paper focuses on the following three aspects of research and innovation:(1)In order to realize the accurate segmentation of rectal tumor area in MRI image and solve the problem of imbalance of image categories,we first designed a new convolutional neural network method ASR-UNet.This method uses U-Net as the basic framework and uses an improved residual structure as the basic convolution module of the segmentation model.This structure with skip connections is conducive to the long-distance propagation of gradient information,thereby improving the trainability of the model.Secondly,in the process of fusing the shallow and deep features of the network,this method introduces a dual attention mechanism of spatial regions and feature channels,so that the model can adaptively select relatively important information from the feature space for fusion,such as a relatively small number of masks area,so as to improve the feature representation ability of the network,better combine context information and improve the segmentation results.Finally,we designed a combined loss function based on the binary cross-entropy loss function and the DICE loss function,and applied it to the segmentation model training process to further improve the class imbalance problem in the training sample.(2)In order to solve the problem of the lack of high-quality rectal tumor annotation image data sets that can be used for training,we first constructed a new rectal tumor MRI image data set,and verified and compared the performance of different automatic segmentation frameworks based on this data set.Secondly,image processing techniques,such as normalization and contrast enhancement,were used in this paper to improve image quality,improve image clarity and contrast of the lesion area.Third,this paper used data enhancement methods to improve the diversity of training samples through geometric transformation.In terms of network training strategy,we used the pre-trained network model obtained on a large-scale training set as the initialization parameter of our segmented network model based on the idea of transfer learning,which can speed up network convergence,improve network generalization ability,and reduce over-fitting risk.(3)In order to visualize the results of the segmentation algorithm and assist doctors in the clinical medical diagnosis of rectal tumors,this paper designed and implemented an automatic segmentation system for rectal tumors MRI images based on Matlab.In summary,the ASR-UNet model proposed in this paper integrates the advantages of U-Net network,residual network,and attention mechanism.The performance comparison experiment on the rectal tumor MRI image data set shows the effectiveness and rationality of the research program proposed in this paper.Based on the proposed new method,this paper finally designed and implemented an automatic segmentation system for rectal tumor MRI images,which provides reliable reference information for ensuring the accuracy of clinical diagnosis by doctors,and has certain practical application value.