A multimodal ConvNeXt-Tiny deep learning model for simultaneous prediction of IDH mutation and Ki-67 expression in gliomas

Abstract

Objective To construct and validate a multi-task deep learning model based on ConvNeXt-Tiny for synchronous prediction of isocitrate dehydrogenase (IDH) mutation status and Ki-67 expression level in gliomas. Materials and Methods This retrospective multicenter study included adult patients with diffuse glioma from two hospitals, which served as the training set and the independent validation set, respectively. All patients underwent multimodal MRI examinations within 4 weeks before surgery, including T2-weighted imaging (T2WI), T2-fluid attenuated inversion recovery (T2-FLAIR), contrast-enhanced T1-weighted imaging (T1CE), and three-dimensional arterial spin labeling imaging (ASL). MRI data were uniformly preprocessed to extract traditional radiomics features, and a dual-task deep learning model based on ConvNeXt-Tiny was constructed. Clinical characteristics, radiomics features, and deep learning features were multimodally fused to establish the multimodal model and were compared with other prediction models. Model performance and clinical utility were evaluated using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis, and the net reclassification index (NRI). SHapley additive explanations (SHAP) analysis and gradient-weighted class activation mapping (Grad-CAM) were applied for visual interpretation of model decisions. Results The multimodal model demonstrated the best diagnostic performance for both IDH mutation status and Ki-67 expression level prediction. For IDH prediction, the multimodal model achieved AUCs of 0.901 and 0.883 in the training set and test set, respectively, while for Ki-67 prediction, the area under the curves (AUCs) reached 0.939 and 0.924. NRI analysis further confirmed that the multimodal model significantly improved case reclassification in both tasks. Calibration curves and the Hosmer–Lemeshow test indicated good model fit, and decision curve analysis showed a higher net clinical benefit. SHAP analysis revealed that model predictions mainly relied on deep learning features, particularly those derived from T1CE, ADC, and cerebral blood flow (CBF) images, whereas radiomics features and clinical variables contributed less. Grad-CAM visualizations demonstrated that the model primarily focused on tumor-related regions. Conclusion This study developed a ConvNeXt-Tiny-based multi-task model that enables preoperative synchronous prediction of IDH mutation and Ki-67 status in adult diffuse glioma and exhibits robust performance superior to single-task and traditional models.