A Study on Constructing Machine Learning Models to Predict Perineural Invasion in Rectal Cancer Using Combined MR Radiomics and Clinical Features

Motivation: Non-invasive accurate prediction of peripheral nerve invasion in rectal cancer is challenging yet essential for treatment planning. Goal(s): To establish machine learning prediction models based on MR radiomics features combined with clinical features, and to compare the performance of models constructed by eight different machine learning algorithms in predicting peripheral nerve invasion of rectal cancer. Approach: Collecting preoperative MR images and clinical data, extracting T2WI/DWI radiomics features, selecting key features and clinical independent risk factors, constructed models using eight algorithms, evaluated performance with ROC. Results: SVM-based combined model achieved the highest AUC (training: 0.894; test: 0.854), outperforming radiomics-only and clinical models. Impact: This clinical-radiomics model combining T2WI and DWI imaging offers a powerful tool for preoperative prediction of perineural invasion in rectal cancer, aiding personalized treatment planning and prognosis assessment for better clinical decision-making.