Explainable dynamic fusion of GCNN and CNN for precise brain tumor analysis from MRI scans

Precise analysis is critical for automated medical image analysis, especially in brain tumor classification, as high precision ensures accurate treatment decisions. Existing literature has primarily focused on using Convolutional Neural Networks (CNNs) for hierarchical feature extraction, often employing static fusion strategies when combining multiple models. However, these approaches overlook the potential of capturing relational structures and interactions among regions of brain images from a graph representation perspective. Additionally, these techniques fail to adapt inference by considering model correlations on individual decisions in complex data environments. We propose a novel approach to brain tumor analysis that leverages the strengths of Convolutional Neural Network (CNN) and Graph Convolutional Neural Network (GCNN). To address the limitations of large, resource-intensive models, we introduce lightweight CNN and GCNN architectures that efficiently capture both local spatial patterns and global relational structures in brain MRI scans. To further enhance performance, we propose a Dynamic Model Fusion that adaptively combines the outputs of the two networks. This approach allows the model to leverage the complementary representation learning strengths of the two architectures, leading to improved generalization and handling of complex tumor variations. We evaluate our model on publicly available MRI datasets, demonstrating significant improvements over state-of-the-art methods. Our results highlight the potential of combining lightweight architectures and dynamic fusion for reliable and accurate brain tumor classification. • Proposed a Dynamic Model Fusion that adaptively combines the CNN and the GCNN to enhance brain tumor analysis. • Proposed lightweight GCNN and CNN architectures to capture spatial relationships and hierarchical features . • Rigorous evaluation on public datasets, using XAI based explanations and benchmarking against SotA methods.