MedicoCNN-GNN: A Hybrid Model with Integrated Attention Mechanism for Medical Image Classification

Medical image classification has been rapidly advancing to help accelerate the treatment process. Adoption of hybrid deep learning architectures in this field has enhanced feature extraction abilities, interpretability, and accuracy. This study presents a hybrid deep learning model that leverages amalgamation of Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs) to enhance the multi-class classification of medical images. The model addresses the classification of medical images into three categories: Non-Tumor, Non-Viable-Tumor, and Viable. CNN being the backbone efficiently extracts spatial features from input images. Prioritization of critical spatial and channel dimensions in the CNN have been managed by the integration of Convolutional Spatial and Channel Block Attention (CSBAN) mechanism. Concurrently, to capture the inter-relationships among the features of the image, the GNN is utilized to process the tabular and relational data using graph attention layers. The proposed system, MedicoCNN-GNN has been validated carefully on the medical imaging dataset, providing a test accuracy of 92% with superior performance. To enhance the interpretability for providing insights into model decisions, visualization techniques such as Grad-CAM and SHAP have been considered. The integration of CNNs, GNNs, and attention mechanisms in this research has highlighted the efficacy of the model. The model is beneficial for robust and interpretable medical image classification, paving the way for improved diagnostic tools in clinical practice.