A multicenter-validated interpretable transformer model for pituitary microadenoma detection on non-contrast multiparametric MRI

BACKGROUND: Detecting pituitary microadenomas using non-contrast multi-parametric magnetic resonance imaging (MRI) is challenging yet essential for clinical decisions. This study aimed to develop a transformer deep learning (DL) model for detecting pituitary microadenomas based on non-contrast multiparametric MRI and explore the explainability techniques to enhance transparency in convolutional neural network (CNN)-based classification. The primary research question addressed is how to improve the accuracy, generalization, and interpretability of CNNs for microadenomas detection. METHODS: Non-contrast multiparametric MRI sella area scans of 590 patients were retrospectively collected from three hospitals. The development and comparison of 2D_DL, 2.5D_DL, 2D_multichannel, and transformer models for classification. By incorporating Explainable AI (XAI), including Gradient-weighted Class Activation Mapping(Grad-CAM) and SHapley Additive exPlanations (SHAP), we improve model interpretability. RESULTS: The performance of the 2D_multichannel model, with an area under the curve (AUC) of 0.893, was better to that of the 2D_T1SAG_DL, 2D_T1COR_DL, 2D_T2COR_DL (AUC, 0.884, 0.779, and 0.846, respectively). The performance of the transformer model, with an area under the curve (AUC) of 0.985, was superior to that of the 2.5D_T1SAG_DL, 2.5D_T1COR_DL, 2.5D_T2COR_DL (AUC, 0.763, 0.863, and 0.835, respectively). The non-contrast MRI-based 2.5D_DL transformer model all shows outperforming performance in the internal and two external test sets (AUC, 0.874, 0.829, and 0.819, respectively). CONCLUSIONS: Given its robust diagnostic performance and enhanced interpretability, this model demonstrates significant potential for clinical translation as a decision-support tool in the detection of pituitary microadenomas.