AD-Deep Fusion: A Deep Feature Fusion and Ensemble Learning Framework for Early Alzheimer’s Disease Diagnosis

Abstract

Alzheimer's disease (AD) is a chronic neurodegenerative disease in which early diagnosis is crucial for early intervention and the appropriate management of patients. But the patho-octical changes in early stages, such as mild cognitive impairment (MCI), have been hard to detect with traditional diagnostic methods. This paper introduces a deep feature fusion and ensemble learning based automatic early AD diagnosis system using multimodal magnetic resonance imaging (MRI) and positron emission tomography (PET) images. The proposed framework utilizes a set of separate deep-learning encoders to obtain the modalitiespecific structural and metabolic representations, which are then processed by attentionbased feature enhancement to highlight the diagnostically relevant features. An adaptive fusion mechanism combines complementary MRI and PET attributes and feature optimization minimizes redundancy and enhances discriminative representation. Support-Vector-Machine, Random-Forest, XGBoost and a neural-network-classifier are then applied to the optimized features to classify them, where the probabilities of the classifiers are fused via weighted soft voting. The framework assesses CN, MCI and AD diagnosis, focusing on CN-MCI and MCI-AD differentiation. This work is illustrated by experimental results showing the potential gains of the proposed framework with respect to unimodal and individual-classifier approaches. The research offers a promising basis for the development of strong, multimodal and interpretable AI-powered early diagnosis of AD.