MRI-driven multimodal deep learning approach for predicting pathological complete response after neoadjuvant chemoradiotherapy in locally advanced rectal cancer: A multicenter study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The First People's Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China.
- Department of Radiology, Yuebei People's Hospital, Shaoguan, China.
- Department of MRI, Maoming People's Hospital, Maoming, China.
- Department of Radiology, The First People's Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China. Electronic address: [email protected].
- Department of Radiology, The First People's Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China. Electronic address: [email protected].
Abstract
Achieving a pathological complete response (pCR) following neoadjuvant chemoradiotherapy (nCRT) in patients with locally advanced rectal cancer (LARC) enables organ preservation and may obviate the need for radical surgery. This study aimed to develop an interpretable magnetic resonance imaging (MRI)-based clinical deep learning radiomics (DLR) model to noninvasively predict treatment response before therapy. This retrospective cohort enrolled 180 LARC patients who received surgery following nCRT at a tertiary care center. And then they were assigned into a training cohort (Center 1, n = 92) and an external validation cohort (Centers 2 and 3, n = 88) for model development and independent evaluation. In total, 428 quantitative radiomics features and 320 deep learning (DL) features (from ResNet50, GoogLeNet, ResNet18, and VGG16) were obtained per patient from pre-nCRT and post-nCRT T2-weighted (T2WI) and diffusion-weighted imaging (DWI) sequences. The predictive performance of the model was quantified by the area under the receiver operating characteristic curve (AUC), and differences among models were statistically assessed with the DeLong test. To enhance interpretability, the Shapley Additive explanations (SHAP) analysis and Gradient-weighted Class Activation Mapping (Grad-CAM) approach were employed to validate and visualize the contribution of individual features. The DLR fusion model, which integrated radiomics features with ResNet50-derived deep learning features, demonstrated higher predictive performance for pCR in LARC patients after nCRT, yielding AUCs of 0.838 in the training cohort and 0.786 in the external validation cohort-surpassing the performance of other DLR models. The clinical DLR (CDLR) model achieved higher performance (AUCs: 0.923 and 0.866, respectively). SHAP analysis highlighted DL_26_T2WI_post as a key predictor of pCR. The interpretable CDLR fusion model based on pre-nCRT and post-nCRT multiparametric MRI offers a reliable, noninvasive approach for the prediction of pCR in LARC patients after receiving nCRT and shows potential for guiding individualized clinical decision-making.