Dynamic Weighted Fusion for Cross-Patched Interactions for Early Classification of Alzheimer’s Disease using Neuroimaging Data

Convolutional Neural Networks (CNNs) are becoming extensively utilised for image Classification tasks, although it is challenging to distinguish between brain scans associated with Alzheimer’s Disease (AD) and those revealing various stages of Mild Cognitive Impairment (MCI). However, deep learning models including transformers offer advanced capabilities in feature extraction and the analysis of complex patterns in imaging data. This research proposes a novel framework of dynamic weighted fusion for cross-patched modelling for early Classification of AD. The method improves the relationships between image patches by using a unique alternating linear layer structure, enabling autonomous and symmetrical interactions. The core of the proposed approach lies in the dynamic weighted fusion which dynamically adjusts the weighting of features extracted from non-overlapping image patches based on their diagnostic relevance. The proposed approach not only enhances the sensitivity and specificity of AD and MCI detection but also leverages the intrinsic spatial relationships within MRI scans more effectively than state-of-the-art CNN demonstrating significant performance improvements of 11.23% in early-MCI as well as 11.75% for late-MCI scenarios. Overall, the study highlights the importance for the integration of deep learning approaches in enhancing early detection capabilities contributing to better patient outcomes in AD management.