Alzheimer’s disease (AD) is the predominant form of dementia for which no curative treatment currently exists. The accelerated aging progression necessitates precise detection of initial AD for effective patient management and illness delay. Earlier research generally used two-dimensional (2D) imaging, which used a single slice that caused loss of spatial information. Most of the previous techniques concentrated on binary classification; however, they encountered difficulties. Which ultimately leads to more parameters and higher computational costs. Compared to binary classification, little work has been done with multi-class classification with 3D images, but that research had low accuracies. To address these limitations, this research proposes 3D-MobiBrainNet, a novel deep learning framework designed to enhance the multi-class classification of AD by leveraging 3D MRI and multi-plane feature fusion. The model processes volumetric data across the axial, coronal, and sagittal planes, ensuring a more comprehensive understanding of brain abnormalities. This method comprised three main steps: (i) Plane-specific extraction of features employs a bottleneck block which comprises depth-wise separable convolutions for every MRI plane to optimize feature extraction and reduce computation complexities; (ii) feature enhancement and selection utilized a feature recalibration strategy to emphasizes important characteristics and a ReLU6 (Rectified Linear Unit) activation function to improve computing efficiency; and (iii) 3D feature integration and classification combine features from each of the three planes into a unified 3D space of features. Experimental results demonstrate that 3D-MobiBrainNet achieves state-of-the-art classification performance using Alzheimer’s disease Neuroimaging Initiative (ADNI) dataset with an accuracy of 97.33 %, recall of 97.33 %, F1-score of 97.33 %, and an area under the curve (AUC) of 99.92 %. Another metric under evaluation was the model’s parameters. Compared to other implemented techniques, the proposed model had fewer parameters (34,145,099), enhancing its prediction performance and requiring fewer processing resources and memory. Additionally, the five-fold cross-validation method was used to check the model’s ability to work well on unseen data and make sure it does not over fit. The results were promising, with a 90.162 % success rate, which showed the good generalizability performance of the model.
