BT‐XAI: Explainable Deep Learning for Robust Brain Tumour MRI Classification

ABSTRACT

Magnetic resonance imaging (MRI) has become an important tool for brain tumour diagnosis, but reliable multi‐class classification across different clinical settings remains difficult, especially when model interpretability and practical deployment are considered together. In this study, we developed BT‐XAI (explainable artificial intelligence), an interpretable deep learning framework for brain tumour MRI classification. The model uses a modified DenseNet121 backbone with an efficient channel attention module to enhance discriminative feature representation. Grad‐CAM, local interpretable model‐agnostic explanations (LIME), and SHapley additive exPlanations (SHAP) were used to examine model decisions from spatial, local, and feature‐attribution perspectives, and the resulting explanations were reviewed by radiologists. On a public MRI benchmark dataset, BT‐XAI achieved an accuracy of 99.10% using stratified five‐fold cross‐validation. In an independent clinical cohort of 489 cases, the model achieved an accuracy of 94.50%, indicating encouraging but still preliminary generalisation to external data. Image degradation experiments further showed that the model retained acceptable performance under reduced image quality. After conversion to ONNX and dynamic quantisation, the model size was reduced to 8.2 MB, with a CPU inference time of 680 ± 24 ms per image. These findings suggest that BT‐XAI may provide interpretable and efficient support for MRI‐based brain tumour analysis, although prospective multicentre validation is still needed.