Leveraging CapsNet for enhanced classification of 3D MRI images for Alzheimer's diagnosis

Dementia is a condition of an acute yet disruptive deprivation of cognitive abilities, usually endemic in older individuals, resulting from an extreme dysfunctionality of brain cells, i.e., a neurologic disorder called Alzheimer's Disease (AD). Early diagnosis and prediction of AD is essential for effective treatment and intervention. Most of the previous studies employ deep learning techniques (such as CNNs and Simple CapsNet), which often lead to information loss during feature extraction and limited extraction of scalar feature information. This ultimately leads to erroneous identification and classification of AD, particularly in the early stage. In this paper, we introduce a framework that enhances the classification of AD in its early stages, utilizing a Hybrid CapsNet-3D model and precise hippocampal volume measurement through the Automatic Segmentation of the Hippocampal Subfield (ASHS) approach. The proposed hybrid model combines CapsNet with 3D convolutional layers. It captures spatial hierarchies in 3D data while retaining the raw feature distribution, potentially enhancing the preservation of subtle details in 3D MRI images. The proposed framework is robust in terms of rotation and affine transformation. It classifies AD in the early stage, also with the help of the volume estimation approach, where the brain is segmented through ASHS with Joint Label Fusion (JLF). Pre-processing techniques and down-sampling are applied to isolate and segmentalize the hippocampus region using a hybrid model on a dataset sourced from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. By integrating Hybrid CapsNet-3D in the classification framework, we are able to streamline the estimation of hippocampal volume. Our results demonstrate that the training, validation, and testing accuracies of AD classification are 98.11%, 75.29%, and 73.45% respectively, which is significantly improved performance compared to the state-of-the-art.