Value of MRI-based machine learning models in predicting the efficacy of neoadjuvant chemaradiotherapy for mid-low rectal cancer
ZHAO Fengshu
ZHANG Rui
QIN Jiaming
LIU Tianqi
DONG Wenjin
CHEN Mengxin
MA Xiangyun
WANG Wenhong
Abstract：Objective To explore a noninvasive method for evaluating the efficacy of neoadjuvant chemoradiotherapy(nCRT)in patients with mid-low rectal cancer and to provide new insights for selecting optimal treatment strategies.Methods A total of 212 patients who underwent preoperative nCRT and radical resection with pathologically confirmed were retrospectively enrolled.Based on postoperative pathology,patients were divided into a pathological complete response(pCR)group(30 cases)and a non-pCR(npCR)group(182 cases).All patients were randomly assigned to a training set(170 cases;pCR 25 cases,npCR 145 cases)and a validation set(42 cases;pCR 5 cases,npCR 37 cases)at an 8∶2 ratio.Clinical features in the training set were screened using the Boruta algorithm,and selected features were combined with imaging indicators[tumor stage,tumor length,circumferential resection margin(CRM),extramural vascular invasion(EMVI),and MRI tumor regression grade(mrTRG)]to build three supervised machine learning classification models—logistic regression(LR),naïve Bayes(NB),and neural network(NN)—for predicting nCRT efficacy.Model performance was evaluated on the validation set.Interobserver agreement for imaging indicators was assessed using intraclass correlation coefficients(ICC)and Kappa statistics.Predictive performance was assessed with receiver operating characteristic(ROC)curves,area under the curve(AUC),accuracy,sensitivity,and specificity.Results Interobserver agreement between two radiologists for imaging indicators was excellent(ICC＞0.9,all κ＞0.8).Among the three models,the NN model showed the best predictive performance(training set:AUC 0.931,accuracy 0.900,sensitivity 0.890,specificity 0.875;validation set:AUC 0.687,accuracy 0.855,sensitivity 0.866,specificity 0.723).Conclusion Machine learning models constructed through feature selection can effectively predict the efficacy of nCRT in mid-low rectal cancer.Combining multiple indicators improves the preoperative predictive accuracy of mrTRG.
Keywords：Rectal cancerNeoadjuvant chemoradiotherapyMagnetic resonance imagingTumor regression gradeMachine learning
Publication Date：2025-09-15
Online Publishing Date：2025-10-13(First online date of this platform, not the publication date of the document)
Pages：7( 541-547 )