Synaptic Insights: A Holistic Approach to Brain Pathology Classification through Transfer Learning and Ensemble Techniques

Abstract

Abstract Background: Detecting brain diseases like tumors, multiple sclerosis, and strokes at an early stage is challenging due to limited access imaging technologies. Analyzing Magnetic Resonance Imaging (MRI) scans can help spot the disease's progression, which will be especially beneficial for clinicians to plan précised treatment. Methods: This research study proposes a novel framework for classifying brain pathologies from MRI modalities to improve clinicians' decision-making ability. This research study developed three distinct deep learning models: a scratch CNN model, a ResNet 101 model improved with transfer learning(m-ResNet101), and an Inception V3 model enhanced through transfer learning (m-InceptionV3). To further improve the efficacy in pathology classification, Weighted Snapshot Fusion Ensemble (WSFE) algorithm is employed to optimize the performance of the proposed model. The internal dynamics of the proposed model is visualized through GradCAM visualization. Results: m-ResNet101 model built on applying a transfer learning approach outperformed every other model, achieving an accuracy - 98.72%, F1 score - 99.35%, precision - 99.17%, and recall - 99.21%. Snapshot ensembled model on combining snapshots of m-ResNet-101 model achieves an impressive accuracy of 99.23%, F1 score of 99.46%, precision of 99.34%, and recall of 98.63%. Conclusion: The research findings suggest that combining transfer learning and snapshot ensembling will improve the model's performance in classifying brain pathology. Furthermore, the feature maps generated through the GradCAM experiment visually highlight the areas and features within an image that greatly influence the model to make a final classification. Such visuals make the models more transparent and trustworthy, which is critical for deploying AI-based models in healthcare networks.