Context-Aware Generative-Convolutional Network (CAGCN) for Enhanced Brain Tumour Segmentation and Classification in Multimodal MRI Imaging

Abstract

Brain tumour segmentation and classification using MRI images present significant challenges due to variations in tumour morphology, imaging artifacts, and limited labelled data. To address these challenges, we propose a novel deep learning framework, CAGCN (Contrastive Attention-Driven Graph Convolutional Network), which enhances segmentation accuracy and classification performance. CAGCN integrates four key modules: Contrastive Multi-Scale Generative Enhancers (CMGE) for reconstructing missing or degraded image regions, Context-Aware Blocks (CAB) for spatial feature enhancement, Feature Self-Supervised Modules (FSSM) for leveraging global spatial relationships, and a Contrastive Attention Transformer (CAT) for improved representation learning. The model is trained using a contrastive self-supervised approach, followed by fine-tuning with cross-entropy and reconstruction loss. Additionally, Character-CAM is employed for interpretability, highlighting critical tumor regions for improved visualization. Experimental results on benchmark MRI datasets demonstrate that CAGCN outperforms existing models, achieving superior segmentation accuracy and classification performance, particularly in the presence of missing or noisy data. The full CAGCN model achieves 92.3%, 92.3%, 92.8%, and 92.6% across different evaluation metrics. Ablation studies reveal the contributions of each module: removing CAB reduces performance to 89.4%, 88.1%, 90.5%, and 89.3%, removing CMGE results in 90.8%, 90.2%, 91.1%, and 90.6%, removing CAT lowers accuracy to 88.7%, 86.9%, 89.2%, and 88.0%, and removing FSSM leads to 91.1%, 90.4%, 91.9%, and 91.1%. These findings highlight the effectiveness of CAGCN in enhancing segmentation accuracy and classification robustness. By reconstructing incomplete scans and leveraging unlabelled MRI images, CAGCN proves to be highly effective for real-world clinical applications, assisting radiologists in precise tumour diagnosis and analysis.