Research On The Assisted Diagnosis Method Of Colorectal Cancer Lymph Node Metastasis Based On Deep Transfer Learning

With high morbidity and mortality colorectal cancer is one of the most common malignant tumors of digestive tract,one of its biological characteristics is the occurrence of lymph node metastasis,naming colorectal cancer lymph node metastasis.Lymph node metastasis of colorectal cancer is the most common metastatic part of colorectal cancer and the main way of metastasis,mainly representing cancer cells through lymphatic metastasis to lymph nodes.Whether lymph node metastasis of colorectal cancer occurs in the diagnosis and treatment of colorectal cancer directly affects the formulation of treatment plan,prognosis evaluation,postoperative local recurrence and 5-year survival rate prediction.Therefore,accurate diagnosis of lymph node status of colorectal cancer is not only helpful for doctors to make corresponding surgical plans,but also can reduce the probability of postoperative recurrence and improve the prognosis and quality of patients’ life.Although doctors can assess the status of tumors and lymph nodes by MRI medical images of colorectal cancer,the diagnostic accuracy still fails to meet the needs of doctors.Excessive false negative or false positive will induce overtreatment or delay treatment,and increase the economic pressure of patients and the risk of related complications or the probability of local recurrence and metastasis.In addition,there are still different opinions on the standard of lymph node metastasis dissection(resection of lymph nodes and some surrounding tissues)in the surgical treatment of colorectal cancer,the main reason is that large-scale lymph node dissection is prone to lead to concurrent reactions.However,total mesorectal resection without lymphadenectomy often leads to a higher recurrence rate.Therefore,in the diagnosis and treatment of colorectal cancer,it is urgent to improve the accuracy of the diagnosis of lymph node metastasis of colorectal cancer.From the perspective of practical clinical application,this paper conducted an in-depth study on the assisted diagnosis of lymph node metastasis of colorectal cancer,which is not only of great significance to the assisted diagnosis of lymph node metastasis of colorectal cancer,but also provides an effective reference for future research at home and abroad.The results of this study have certain theoretical significance of the assisted diagnosis of colorectal cancer lymph node metastasis and have high value of clinical practical,and provide a new assisted diagnosis method of clinical work in the future.In this study,there five aspects are studied in depth:the data augmentation pattern of colorectal cancer and colorectal cancer lymph node metastasis,the nature of the primary foci of colorectal cancer determination method,colorectal cancer lymph node metastasis target detection method,colorectal cancer lymph node metastasis semantic segmentation method and the nature of the colorectal cancer lymph node metastasis classification method.The main research contents are as follows:(1)A cascade dual-model MRI image data augmentation pattern was designed for colorectal cancer and lymph node metastasis of colorectal cancer.MRI image data of colorectal cancer and colorectal cancer lymph node metastasis are difficult to obtain and label in practice,resulting in the number of samples in the data set unable to meet the needs of model training,and such data is the basis of subsequent method research.Therefore,an enhanced MRI image data pattern for colorectal cancer and colorectal cancer lymph node metastasis was designed to generate MRI image data samples of colorectal cancer and colorectal cancer lymph node metastasis that meet the needs of subsequent studies,providing data guarantee for subsequent methodological studies.The proposed data augmentation pattern includes image affine transform module and deep generation adversarial network module.Through image affine transformation,the data samples in the original data set are turned over and noise is injected to generate a new image.The generated new image is robust to the deformation and gray value change of the original image.The new sample data and the original sample data are combined into a new data set as the input data of the deep generative adversal network.After training the deep generative adversal network,a variety of realistic sample data are generated.The designed data augmentation pattern can effectively improve the number of samples and increase the diversity of samples,and provide more abundant data samples for the subsequent method research.(2)A feature-weighted classification method based on Transformer for determining the nature of colorectal cancer primary foci was proposed.Lymph node metastasis of colorectal cancer is one of the biological characteristics of malignant tumors,so the status of lymph nodes of colorectal cancer can be determined by determining the nature of the primary tumor.After the existing convolutional neural network method is used to extract features from MRI image data of colorectal cancer,the pooling operation will lead to the loss of part of feature information in the process of feature transfer layer by layer.In order to solve this problem,this study proposes a method to determine the nature of colorectal cancer primary foci,which can avoid the loss of characteristic information.This method replaces the largest pooling layer in AlexNet network structure with Transformer structure to avoid information loss during feature transfer.When the features extracted by convolutional neural network are transmitted to Transformer structure,features are divided into several visual labels.The self-attention mechanism in Transformer structure assigns different weights to visual labels and transmits them as output to the next layer in network structure to prevent feature information loss.The experimental results show that the accuracy of the proposed method is 0.9095,which is better than the similar convolutional neural network method.(3)An Attention-based dual-mode target detection method for lymph node metastasis in colorectal cancer was proposed.When lymph node metastasis of colorectal cancer is confirmed,it is necessary to conduct target detection of lymph node to determine its specific location,so as to provide data for subsequent studies.Due to the need for rapid and accurate target detection methods in clinical application,the performance of target detection methods should be further improved.However,in the process of improving performance,the method significantly increases the computational overhead.To solve this problem,this paper proposes a target detection method for colorectal cancer lymph node metastasis that can improve detection performance and maintain the original computational overhead.Based on the YOLOv4 network structure,this method combines the channel attention mechanism and the spatial attention mechanism in the attention mechanism.In the process of feature processing,the content information and location information of the target expressed by feature are respectively concerned,and more feature information of the target is obtained in this way.In addition,in order to improve the diversity of features and avoid the loss of part of feature information,the method of feature fusion was adopted to series all features to generate new features with more semantic information,so as to improve the expression ability of feature semantic information for target detection of colorectal cancer lymph node metastasis.Experimental results show that the mAP of the proposed method is 0.9592,which is better than the mainstream method.(4)A semantic segmentation method for colorectal cancer lymph node metastasis with feature space invariant based on STN was proposed.In order to obtain accurate lesions,it is necessary to achieve target segmentation of metastatic colorectal cancer lymph nodes,determine the final lesions,and assist doctors in further diagnosis.Current segmentation methods lack the ability to maintain spatial invariance of data,and pooling units cannot guarantee spatial invariance of large size feature maps in the process of feature transfer.Therefore,this paper proposes a spatially invariant segmentation method for colorectal cancer lymph node metastasis that can maintain the data feature map.In this method,the spatial transformer network is embedded in the input end of U-Net network to keep the spatial invariance of input data.Because U-Net network pooling layer of the pooling of cell size is limited,which can’t deal with the characteristics of the large size figure and lead to the loss of part of the space information.So the use of space transform network replace U-Net pooling in the network layer,improving the network characteristics of large size figure processing,ensuring the feature information keep unchanged space,at the same time in order to extract the characteristic information of the small target,the feature pyramid network is inserted between the encoding and decoding of the network structure to improve the transmission of feature information.Experimental results show that the proposed method,whose MIoU is 0.7535,can effectively and accurately complete target segmentation,and its performance is superior to that of the mainstream semantic segmentation methods.(5)A classification method for colorectal cancer lymph node metastasis based on feature multiple connections was proposed.The nature of lymph node metastasis in clinical practice is directly related to the 5-year survival rate of patients,so determining the nature of lymph node metastasis in colorectal cancer is an important step.In the process of transferring the features extracted by convolutional neural network from low level to high level,the low-level features cannot be fully applied to the final result decision.Therefore,this paper proposes a classification method for lymph node metastasis of colorectal cancer that can combine low-level and high-level features.This method proposes a feature multiple connection method in AlexNet network structure,which can transfer the features of the current layer layer by layer and can transfer the low-level features to the higher network structure,making full use of the location and spatial details of the low-level features.Finally,the feature fusion layer was used to fully integrate low-level and high-level feature information,so as to enhance the diversity of classification decision-making features and realize the classification of colorectal cancer lymph node metastasis.The proposed method was compared with 10 representative classification algorithms and 4 radiologists with more than 5 years of clinical experience.The experimental results showed that without increasing the depth and width of the model,the accuracy of the proposed method was 0.8569,which was better than the mainstream classification methods and radiologists’ classification results.