Research Of Rectal Cancer Segmentation Algorithms Based On MRI Image

Rectal cancer is one of the most common malignant tumors in the digestive system.However,there is a lack of research on MRI image segmentation based on rectal cancer both domestically and internationally.Additionally,many existing medical image segmentation algorithms overly rely on shallow image features.When there are overlapping noisy regions or unclear boundaries between adjacent tissues,segmentation based solely on these shallow features is more prone to errors.The main challenges for rectal cancer image segmentation are:(1)the cancer area has similar grayscale values to the interior of the rectum and lacks regular shapes and clear boundaries;(2)the size and shape of the rectal region and cancer area vary greatly across different patients’ MRI images,leading to reduced generalization ability of the trained model.To address these challenges,this thesis proposes a rectal cancer segmentation algorithm that first roughly locates the rectal region,then performs rectal cancer segmentation on the processed rectal images,and finally builds a validation platform for verification.Specifically,the following contributions are made:(1)thesis proposes a joint spatial domain-based multi-scale U-Net segmentation model to address the existing limitation that image segmentation models are trained only in one coordinate system without considering the joint impact of different spatial coordinate systems in achieving automatic segmentation of target areas.The model achieves rotation invariance through the use of polar coordinates with a center point and utilizes retraining with the Cartesian coordinate system network to achieve translation invariance in the final segmentation results.To address the lack of attention to multiscale spatial information in existing encoding modules,a multilayer dilated convolution encoding module is designed to achieve multiscale content fusion.Relevant experiments demonstrate the superiority of this network model in rectal region segmentation.Compared with Deep Lab V3+ and other models,the Dice coefficient,m Io U,and precision all reached the highest values,which are 0.9309,8912,and 0.8968,respectively.(2)the common encoder-decoder structure can lead to redundant contextual information,and using a symmetrical encoding-decoding structure may not effectively model long-distance feature dependencies.To overcome these limitations,an attention module that integrates the self-attention mechanism is designed,which can obtain richer contextual information dependencies.Additionally,to better segment fuzzy boundaries in rectal cancer regions,a new multiscale residual feedback network is proposed,which facilitates the network’s improved segmentation of difficult-to-predict pixels.Finally,experiments on rectal cancer dataset demonstrate the superior performance of the multiscale attention residual feedback algorithm.Compared with RF-Net and other models,the Dice coefficient,m Io U,and precision all reached the highest values,which are 0.7139,0.6047,and 0.6896,respectively.(3)Integrate the above algorithms into the hospital rectal cancer data intelligent analysis system.The system implements multiple modules,which can display the entire process of rectal and rectal cancer segmentation with an intuitive interface,standardize the management of patient basic information and image information,and better assist doctors in the diagnosis of related diseases.