MRI radiomics combined with delta-radiomics model for predicting pathological complete response in locally advanced rectal cancer patients after neoadjuvant chemoradiotherapy: A multi-institutional study

Le résumé fourni par la source

PURPOSE: To construct and validate a magnetic resonance imaging (MRI) radiomics combined with delta-radiomics and clinical information (C) model for predicting pathological complete response (pCR) in patients with locally advanced rectal cancer (LARC) after neoadjuvant chemoradiotherapy (nCRT). METHODS: A total of 198 patients with LARC who underwent MRI before and after nCRT were retrospectively enrolled in this multi-institutional retrospective study. MRI radiomics features were extracted from pre- and post-nCRT diffusion weighted imaging (DWI) and T2-weighted imaging (T2WI) images. The least absolute shrinkage and selection operator (LASSO) and principal component analysis (PCA) algorithm were used to select the optimal predictive features. We constructed the following models, four single-modal radiomics models: DWI-post, DWI-pre, T2-post, T2-pre, two delta-radiomics models: DWI-delta, T2-delta and four multi-modal fusion models: DWI-post + DWI-pre, DWI-post + DWI-delta, DWI-post + DWI-delta + T2-delta, DWI-post + DWI-delta + T2-delta + C. The models were developed using four machine learning classifiers, including Decision Tree (DT), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost). RESULTS: The multi-modal fusion model DWI-post + DWI-delta + T2-delta achieved the best performance with an area under the curve (AUC) of 0.879 for predicting pCR, which was significantly higher than that of the single-modal model DWI-post (optimal AUC = 0.824), DWI-pre (optimal AUC = 0.836) and the delta-radiomics model DWI-delta (optimal AUC = 0.841), T2-delta (optimal AUC = 0.837) in the internal validation sets. XGBoost classifier showed better prediction performance than the other classifiers in the most models. The DWI-post + DWI-pre model with DT classifier and PCA feature selection achieved the highest AUC of 0.754 and the DWI-post + DWI-delta + T2-delta + C model with SVM classifier and LASSO feature selection achieved the suboptimal AUC of 0.734 in the external validation sets. CONCLUSION: The multi-modal fusion model significantly outperforms conventional single-modal prediction models. The model could be used as a reliable and noninvasive tool for the personalized therapy in LARC patients.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.