Multiparametric MRI-based machine learning models for preoperatively predicting rectal adenoma with canceration

Area of Science:

• Radiology
• Oncology
• Machine Learning

Background:

• Rectal adenoma with canceration requires accurate preoperative diagnosis.
• Multiparametric MRI (mpMRI) offers detailed tissue characterization.
• Machine learning (ML) can analyze complex imaging data for predictive modeling.

Purpose of the Study:

• To develop and evaluate mpMRI-based ML models for predicting rectal adenoma with canceration.
• To compare the performance of models using different MRI sequences and combined data.

Main Methods:

• Retrospective study of 53 patients with rectal adenoma or adenoma with canceration.
• Extraction of 1396 radiomics features from high-resolution T2-weighted imaging (HR-T2WI) and diffusion-weighted imaging (DWI).
• Feature selection using LASSO, followed by model construction with logistic regression (LR) and support vector machine (SVM) on training and test cohorts.

Main Results:

• Optimal features were selected from HR-T2WI, DWI, and combined sequences.
• The combined model (Model_combined) outperformed models using single sequences (Model_T2, Model_DWI).
• Model_combined achieved AUCs of 0.867 (LR) and 0.900 (SVM) in the test cohort, with no significant difference between algorithms.

Conclusions:

• mpMRI-based ML models show potential for preoperative prediction of rectal adenoma with canceration.
• The combined multiparametric approach yields the best predictive performance.
• Both LR and SVM algorithms demonstrate excellent and comparable performance for model construction.