An explainable knowledge-driven hybrid CNN-transformer and radiomics framework for multi-modal MRI-based brain tumor analysis

Abstract

The increasing complexity of medical data in modern healthcare environments necessitates intelligent frameworks that not only deliver accurate predictions but also support effective knowledge generation and decision-making. In the context of brain tumor analysis using multi-modal magnetic resonance imaging (MRI), challenges such as tumor heterogeneity, variability in imaging protocols, and limited interpretability of deep learning models hinder their integration into clinical knowledge workflows. This study proposes an explainable, knowledge-driven hybrid CNN–Transformer–Radiomics (HCTR) framework designed to facilitate both predictive performance and clinical knowledge extraction. The framework integrates convolutional neural networks for localized feature learning, transformer-based architectures for global contextual understanding, and radiomic descriptors for structured, domain-relevant feature representation. A cross-attention-based fusion mechanism is employed to combine these heterogeneous knowledge sources into a unified representation. Beyond detection and analysis, the proposed system emphasizes interpretability through Grad-CAM visualizations and feature attribution methods, enabling the transformation of model outputs into clinically meaningful insights. This supports enhanced transparency, trust, and knowledge dissemination within clinical decision-making processes. The proposed framework contributes to the development of intelligent decision support systems by bridging data-driven modeling with knowledge-centric interpretation. It provides a scalable approach for integrating explainable AI into healthcare knowledge management environments, facilitating improved diagnostic reasoning and informed clinical decisions.