Intratumoral and peritumoral radiomics of MRI predict pathological differentiation in patients with rectal cancer.

The histopathological differentiation of rectal cancer (RC) is a key determinant of treatment strategy and prognosis, as tumors with varying differentiation demonstrate notable differences in therapeutic approach selection and survival outcomes. Consequently, accurate preoperative prediction of tumor differentiation is clinically essential for formulating initial surgical plans and adjuvant therapy strategies. The aim of the present study was to develop a magnetic resonance imaging (MRI)-based radiomics strategy that integrates features from both intratumoral and peritumoral (margin, 5 mm) regions to construct a non-invasive predictive model capable of distinguishing well differentiated RC (glandular structures ≥95%) from non-well differentiated RC. A retrospective analysis was performed using data from 224 patients with RC who underwent preoperative MRI. Radiomic features were extracted from T1-weighted images (T1WI), T2-weighted images (T2WI) and diffusion-weighted imaging (DWI) of the tumor and peritumoral 5 mm. The adaptive synthetic technique was used to address class imbalance, and ReliefF was applied to select 20 features. A total of three machine learning algorithms, logistic regression, Light Gradient-Boosting Machine (LightGBM) classifier and Gaussian Naive Bayes, were used to build MRI radiomic models combining the tumor and the peritumoral areas to predict pathological differentiation of RC. The models were evaluated by comparing the area under the curve (AUC) and binary classification metrics, with the best performing T1-LightGBM model selected. Subsequently, intratumoral and peritumoral and intratumoral (T1-LightGBM) models were established. Discriminative ability, calibration and clinical applicability were assessed using receiver operating characteristic curves, calibration curves and decision curves, with Shapley Additive explanations (SHAP) analysis employed for interpretation. Among the nine models constructed using T1WI, T2WI and DWI from the tumor and 5-mm surrounding region, the AUC range was 0.510-0.756 across all validation sets. The T1-LightGBM model incorporating both tumor and peritumoral areas demonstrated the best performance for predicting pathological differentiation in RC, with an AUC of 0.756, accuracy of 0.700 and sensitivity of 0.707. SHAP analysis identified wavelet_HHH_glszm_Size_Zone_Non_Uniformity_Normalized as the most significant feature for predicting pathological differentiation and evaluating non-well differentiated patients. In conclusion, the integrated model employing the LightGBM algorithm on T1WI, which combined both intratumoral and 5-mm peritumoral features, demonstrated promising predictive potential for individualized RC differentiation.