Research On Image Segmentation Technology Of Rectal Cancer Based On Deep Learning

With the development of information technology and the emergence of digital medical equipment,the connotation and capacity of medical information have been greatly enriched,and medical imaging has gradually developed into an increasingly important clinical diagnosis method.To meet the public’s demand for medical services,it is of great significance to use the powerful processing power of computers to automate the processing of medical images.Convolutional neural networks have powerful feature extraction capabilities,and fully convolutional neural networks are among the best.Symmetrically structured fully convolutional neural networks can integrate features of different scales.Channel concatenating can broaden the network and richer the features.Residual structures can be combined deep and shallow features and eliminate the degradation of the network.Attention mechanism is very beneficial to achieve a special medical image segmentation task.This thesis aims to use deep learning infrastructure to study techniques that can enhance the feature extraction capabilities of neural network block and apply them to the segmentation of rectal cancer images.The main contents and contributions are as follows:（1）Aiming at the problem of the rough segmentation results of the existing rectal cancer automatic segmentation technology.Residual structure and channel splicing structure are introduced into the fully convolutional neural network with symmetric structure to strengthen the feature extraction ability of the network model.The simple convolution block is optimized into a encoder block that can reasonably extract features and has better feature extraction effect.A atrous convolution structure with a small amount of parameters is introduced in the skip connection,which can improve the receptive field without changing the image size.Finally,the accuracy of rectal cancer image segmentation is 83.48%.（2）Aiming at the problem of improving the feature extraction ability of the model and bringing many network model parameters.The attention mechanism can make the network model learn to ignore invalid information and focus on efficient information.This thesis proposes a spatial selective attention block and an adaptive multi-scale channel attention block,which are interspersed in the downsampling stage and skip connection stage of the fully convolutional neural network,respectively.Through experiments,it is proved that the spatial attention block,the channel attention block and the attention module combined in the form of targeted interspersing can improve the segmentation effect of the network model.The segmentation effect on rectal cancer reaches 83.27% with only 8.39 MB of parameters introduced.（3）In the process of acquiring rectal cancer images,to ensure that the lesions can be completely included,a large amount of background information will appear redundant.Although certain image processing methods can be used to reduce this irrelevant information,the location of rectal cancer images does not have regularity,there is still redundant background information.Aiming to solve this problem,this thesis combines cross-entropy loss and dice loss to design a loss function that can consider both foreground and background information,and can adjust the training weights according to different datasets.This thesis also proves the effectiveness of the loss function through experiments,and selects the optimal training weight for the rectal cancer data set used in this thesis.Using this loss function for training,the accuracy of rectal cancer segmentation reaches 82.87%.