A Hybrid Perception Network for Accurate and Interpretable Brain Tumor Classification Using Multi-Modal Imaging

Objective: Accurate classification of brain tumors is of utmost importance in early diagnosis and treatment, which can be further enhanced through advanced deep learning techniques. This work introduces HyPerNet—a Hybrid Perception Network for brain tumor classification that integrates cutting-edge methods to maximize classification accuracy, interpretability, and efficiency. HyPerNet leverages multi-modal fusion, combining Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and Positron Emission Tomography (PET) scans to extract complementary information from different imaging modalities. Methods: A dynamic attention mechanism adapts to image complexity, improving performance under noisy and artifact-laden conditions. The architecture incorporates explainable AI components using Grad-CAM for clinical interpretability, along with 3D convolutional neural networks for volumetric feature learning. Hybrid feature extraction, integrating Convolutional Neural Networks (CNNs) with graph neural networks (GNNs), enables the model to capture both local and relational tumor characteristics. Domain adaptation via transfer learning allows the model to generalize across diverse datasets, improving robustness. Result: In experiments conducted on a benchmark dataset of 8,256 images, HyPerNet achieved a classification accuracy of 97.14%, precision of 97.09%, recall of 97.09%, F1-score of 97.09%, and specificity of 98.9%, outperforming state-of-the-art models such as DenseNet121, ResNet101, and MobileNetV3. Notably, HyPerNet attains these results with only 1.66 million parameters, making it highly efficient for deployment on mobile and edge devices. Conclusion: These findings establish HyPerNet as a reliable, interpretable, and computationally efficient solution for clinical brain tumor classification.