Adversarially robust brain tumor detection in multi-modal MRI using hybrid deep learning and radiomic features

Abstract

Deep learning has significantly improved automated brain tumor detection from multi-modal magnetic resonance imaging (MRI). However, many existing models remain vulnerable to adversarial perturbations, noise, and domain shifts, which can reduce diagnostic reliability in real-world clinical environments. To address these challenges, this study proposes HCTR-R, a robustness-oriented hybrid computational intelligence framework that integrates convolutional neural networks, transformer-based contextual modeling, and radiomic feature representations for reliable tumor detection. The framework incorporates defense-aware learning strategies, including noise injection, adversarial sample generation using the Fast Gradient Sign Method (FGSM), and adaptive data augmentation to enhance model stability under perturbations. A cross-attention fusion mechanism is employed to combine deep and radiomic features, enabling more effective representation learning from multi-modal MRI data. Unlike conventional hybrid architectures, the proposed approach explicitly focuses on robustness and cross-dataset generalization in medical imaging. Experimental evaluation was conducted using the combined BraTS 2019–2021 datasets for training and internal validation, while the TCGA-GBM dataset was used for external validation to assess generalization across institutions. The results indicate that the proposed framework achieves strong and consistent performance under noisy and adversarial conditions, with high accuracy, F1-score, and AUC values. Explainability analysis using Grad-CAM and SHAP further demonstrates meaningful correspondence between model predictions and tumor regions. The proposed approach supports the development of robust and interpretable computational intelligence systems for AI-driven clinical decision support in healthcare.