Hybrid Distance-Aimed Triplet Loss for Automated Alzheimer’s Disease Classification Using Feature Embeddings

Abstract This paper presents a novel approach to Alzheimer’s disease (AD) classification utilising MRI images. Central to our methodology is a Feature Embedding Network (FEN) designed to extract and refine discriminative features essential for accurate disease classification. The FEN incorporates a hybrid distance function that integrates Euclidean distance, cosine similarity, and Earth Mover’s Distance, enhancing the network’s ability to capture nuanced differences in brain structure indicative of AD progression. Through a carefully crafted triplet loss framework, the FEN is trained to optimise the embeddings of anchor, positive, and negative examples, facilitating the clustering of similar images while separating dissimilar ones. Experimental results demonstrate the efficacy of our approach, achieving a remarkable 94.89% accuracy on the test set, with corresponding precision, recall, and F1-score values of 91.5%, 90% and 91.5%, respectively. Comparative analyses against existing models show the superiority of our proposed method in AD classification tasks. Moreover, our study contributes to better interpretability of feature embeddings, revealing distinct patterns associated with disease stages that align with clinical diagnoses. The robustness and performance of our model highlight its potential application in the early diagnosis and monitoring of AD, offering clinicians a valuable tool for more accurate patient classification and intervention planning.