To develop a predictive model with pre-treatment magnetic resonance imaging (MRI) findings of the structured report template and clinical parameters for poor responses prediction after neoadjuvant chemoradiotherapy (neoCRT) in locally advanced rectal cancers (LARC) patients.
Patients with clinicopathologically confirmed LARC (training and validation datasets, n = 100 and 71, respectively) were enrolled. Patients’ clinical data were retrospectively collected. MRI findings of the structured report template were analysed. The tumour regression grade (TRG) system as proposed by Mandard et al was used. Poor response was defined as TRG 3-5. Univariate logistic regression analysis and a lasso regression model were performed to select the significant predictive features from the training set. A nomogram was constructed based on a multivariable logistic regression analysis. Calibration, discrimination, and clinical usefulness of the nomogram were assessed. The calibrative and discriminative ability of our model were compared with those of models including the tumour-node-metastasis (TNM) stage and clinical factors.
The MRI-reported T4b stage, MRI-reported extramural venous invasion (EMVI) positivity, MRI-detected number of positive mesorectal lymph nodes (LNs) > 0, and preoperative oxaliplatin and capecitabine (CAPOX) chemotherapy regimen were incorporated into our nomogram. The nomogram showed good discrimination, with areas under the receiver operating characteristic (ROC) curves of 0·823 and 0·820 in the training and test sets, respectively, and good calibration in both datasets. The decision curve analysis confirmed that the nomogram was clinically useful. The calibrative and discriminative ability of our model were better than those models including the TNM stage and clinical factors.
A nomogram based on pre-treatment MRI features of the structured report template and clinical risk factors has potential for use as a non-invasive tool to preoperatively predict poor responses in LARC patients after neoCRT.
Copyright © 2020 Elsevier B.V. All rights reserved.