Comparison of deep learning models and machine learning models in pre-operative prediction of mismatch repair system classification for colorectal cancer:based on CT radiomics
WANG Zimeng
WANG Wenjiang
WANG Dawei
CUI Shujun
Abstract：Objective To establish a variety of machine learning models based on CT radiomics features and deep learning models based on convolutional neural networks,to compare the performance of the two methods in predicting the preoperative mismatch repair(MMR)typing model in colorectal cancer(CRC)patients.Methods A retrospective study was conducted on 120 colorectal cancer patients who were randomly divided into 7:3 into the training group and the test group,and all of these cases were from the First Affiliated Hospital of Hebei North University.The region of interest(ROI)was plotted and the radiomics features were extracted to select the optimal set.The machine learning models were builded including random forests,support vector machines and logistic regression algorithms,as well as deep learning convolutional neural network(CNN)structures Vgg16 models.The diagnostic performance of the model was evaluated by the area under the ROC curve(AUC),sensitivity,accuracy,specificity and F1 score.Results The AUC values of the test group of the three machine learning models were 0.82(95%CI:0.75-0.84),0.75(95%CI:0.73-0.81),and 0.71(95%CI:0.59-0.74),and the F1 scores were 0.60,0.82,and 0.57.The test group AUC of the deep learning model was 0.87(95%CI:0.76-0.91)and the F1 score was 0.82.Conclusion The radiomics machine learning and deep learning models based on CT images can effectively identify the MMR typing of colorectal cancer.According to the AUC values obtained by different models,it is found that the deep learning model is more efficient than the machine learning model in identifying the two types of MMR.
Keywords：colorectal cancermismatch repair systemradiomicsdeep learningrandom forest
Publication Date：2025-03-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：8( 315-322 )