Prediction of T Stage of Rectal Cancer After Neoadjuvant Therapy by Multi-Parameter Magnetic Resonance Radiomics Based on Machine Learning Algorithms.

Introduction: Since the response of patients with rectal cancer (RC) to neoadjuvant therapy is highly variable, there is an urgent need to develop accurate methods to predict the post-treatment T (pT) stage. The purpose of this study was to evaluate the utility of multi-parametric MRI radiomics models and identify the most accurate machine learning (ML) algorithms for predicting pT stage of RC. Method: This retrospective study analyzed pretreatment clinical features of 171 RC patients who underwent 3 T MRI prior to neoadjuvant therapy and subsequent total mesorectal excision. Tumors were manually drawn as regions of interest (ROI) layer by layer on high-resolution T2-weighted image (T2WI) and contrast-enhanced T1-weighted image (CE-T1WI) using ITK-SNAP software. The most relevant features of pT stage from CE-T1WI, T2WI, and fusion features (combination of clinical features, CE-T1WI, and T2WI radiomics features) were extracted by the Least Absolute Shrinkage and Selection Operator method. Clinical, CE-T1WI radiomics, T2WI radiomics, and fusion models were established by ML multiple classifiers. Results: In the clinical model, the LightGBM algorithm demonstrated the highest efficiency, with AUC values of 0.857 and 0.702 for the training and test cohorts, respectively. For the T2WI and CE-T1WI models, the SVM algorithm was the most efficient; AUC = 0.969 and 0.868 in the training cohort, and 0.839 and 0.760 in the test cohort, respectively. The fusion model yielded the highest predictive performance using the LR algorithm; AUC = 0.967 and 0.932 in the training and test cohorts, respectively. Conclusion: Radiomics features extracted from CE-T1WI and T2WI images and clinical features were effective predictors of pT stage in patients with rectal cancer who underwent neoadjuvant therapy. ML-based multi-parameter MRI radiomics model incorporating relevant clinical features can improve the pT stage prediction accuracy of RC. [ABSTRACT FROM AUTHOR]