Predicting Lymph Node Metastasis Of Rectal Cancer Based On Magnetic Resonance Imaging

Rectal cancer is a common malignant tumor of gastrointestinal tumor in China,and its morbidity and mortality are increasing year by year,which has become a problem troubling the health of our people.The diagnosis of lymph node metastasis is a key factor in determining the therapeutic effect of patients with rectal cancer which not only determines the treatment of the patient,but also significantly affects the prognosis and survival.Preoperative accurate evaluation of lymph node status is an urgent problem to be solved in this field.In clinical practice,the commonly used diagnostic method is magnetic resonance imaging(MRI).Because of its high soft-tissue resolution,its diagnostic accuracy is better than that of ultrasound,computed tomography,and other imaging techniques,but it still does not meet the needs of clinicians.The high false positive rate will lead to overtreatment of patients,increasing risk of related complications,patients’ economic burden and miss the best time for treatment.while the underdiagnosis caused by high false negative rate will also increase the risk of postoperative recurrence and metastasis because the stage is lower than the actual pathological stage.Therefore,it is urgent to improve the accuracy of diagnosis of local lymph node metastasis of rectal cancer.The image intelligent diagnosis technology can quickly extract quantitative features from tomographic digital images and transform digital features into high-throughput data that can be excavated.The technology can capture images of diseases which can’t be recognized by human eyes helping radiologists make a more accurate diagnosis.In this paper,image recognition technology is used to take a large amount of real clinical rectal cancer magnetic resonance image data as the research object.In view of the poor ability of existing clinical methods to predict lymph node,this paper launches a classified prediction research based on a variety of image intelligent diagnosis models.The purpose of this paper is to build a lymph node metastasis prediction model for different clinical diagnosis and treatment needs and improve the diagnostic ability of rectal cancer lymph node metastasis.The main work is carried out from the following aspects:(1)In this paper,the relationship between the gray characteristics of magnetic resonance images of primary rectal cancer and lymph node metastasis was discussed.Medical image recognition technology gradually shows its advantages in clinical diagnosis,but there is not much research in the field of rectal cancer.To determine whether image features are directly related to lymph node metastasis diagnosis is the premise of this study.In this study,through the analysis of T2WI,DWI and ADC multi-sequences primary tumor images,the grey features(such as skewness,kurtosis,etc.)were extracted in the region of interest.The results of logical regression showed that DWICV,DWIMode,DWIKurtosis,T2WIKurtosis and T2WI-MapP5 were independent predictors of lymph node metastasis.Based on the above gray characteristics,a prediction model was established combining with expert evaluation,which overcomes the problem of lack of discrimination of human eyes and improves the ability of radiologists in predicting and diagnosing lymph node metastasis.It is proved that the computer image analysis method is reasonable and effective in the prediction of lymph node metastasis of rectal cancer,which lays a foundation for the following chapters.(2)Aiming at the prediction ability of lymph node metastasis of primary rectal cancer and reducing the labor intensity of doctors’ diagnosis.A single machine diagnosis method was studied.A neural network for predicting lymph node metastasis based on the combination of texture features and depth features of T2WI images was proposed.First,the U-Net encoder structure was used as the main framework with ResNet auxiliary network to obtain depth features.The texture features were enriched by self-attention-based texture feature fully connected network.The 3D-Fu-MCANet diagnostic model was established by combining these two networks.This study overcame the problem that small data sets could not use deep network to extract high-dimensional features and avoided over-fitting.This model was trained and verified by 572 cases of primary tumor images.The results showed that the single machine algorithm of the model could improve the ability of lymph node metastasis prediction based on colorectal cancer primary tumor lesions(AUC 0.771),and the negative predictive value reached 83.6%,which had important clinical promotion significance.(3)In this paper,a small target image lymph node metastasis predicting model based on radiomics features and multi-objective optimization algorithm was proposed.The small target of the lymph node image leads to the characteristics of few pixels,low resolution,and less information in the whole image.The radiomics method can extract massive quantitative image features with the help of computer methods assisting radiologists to improve their diagnostic ability.However,from the point of view of comprehensively optimizing the diagnostic efficiency,there were still some deficiencies in feature dimensionality reduction,feature screening and optimal solution acquisition.To solve the above shortcomings,the dimension reduction method of Pearson correlation coefficient was proposed to reduce the dimension of mass radiomics features.Then the above features were further optimized and screened by multi-objective optimization immune algorithm to obtain the best model parameters and establish a prediction model.Two prediction models of lymph node metastasis of rectal cancer were constructed based on 21 features obtained by dimensionality reduction and evaluated by clinical and radiological experts.The results showed that the diagnostic ability of the multi-objective optimization prediction model based on the high throughput features of small target lymph node images was significantly better than that of the clinical diagnosis model,and the combination of the two can further improve the sensitivity and specificity of the model and had significant clinical practical value.(4)A prediction model of lymph node metastasis of rectal cancer based on deep migration learning was proposed.In view of the irregular shape of lymph nodes and complex internal signals after treatment,resulting in more difficult to distinguish the details and edges of small target images.It is more difficult to distinguish lymph node metastasis,which brings confusion to the clinical work of radiologists.This part of the study continued to take the lymph node small target image as the research object,carried out the deep migration learning technology to establish the classification model for the small data sample.Based on the GoogleInception-v3 model framework,aiming at the problem of lymph node metastasis of rectal cancer,the technical details of the network pre-training model in Inception module,network size reduction strategy,auxiliary output module and back propagation strategy were adjusted.And the Inception4LNs network were constructed.The results showed that the model had a good classification effect in the small target and small sample data set,and significantly improved the accuracy of lymph node diagnosis after treatment.