Authors: Bhavesh Avinash Gadekar, Neha Sanjay Gaikwad, Swayam Vijay Bhosale, Prasad Ambadas Bidgar, Ganesh K Gaikwad
Abstract: Dementia diagnosis is a critical challenge in neurology, often relying on time-intensive and subjective manual analysis of MRI scans. This research proposes a novel hybrid AI-based system for early and accurate dementia detection, combining traditional neuroimaging techniques with advanced deep learning models. The system employs a hybrid 2D-3D pipeline that integrates slice-based 2D convolutional models with volumetric 3D CNN architectures, ensuring a balance between computational efficiency and spatial pattern recognition. The 2D models focus on extracting detailed features from individual MRI slices, while the 3D models capture spatial relationships across the entire brain volume. Additionally, clinical metrics such as cognitive scores are integrated with the MRI data to enhance diagnostic accuracy. Attention mechanisms and Grad-CAM visualizations improve model interpretability by highlighting critical brain regions, addressing the need for transparent AI-driven clinical tools. This hybrid approach significantly improves diagnostic accuracy, generalizability, and explainability compared to conventional methods. The system classifies scans into Strongly Demented, Mildly emented, or Non-Demented categories, providing actionable insights for clinicians. By bridging AI with neuroimaging and multimodal data integration, the proposed system aims to revolutionize dementia detection, enabling earlier intervention and improved patient outcomes.
DOI: https://doi.org/10.5281/zenodo.17985501
