Optimizing Alzheimer's diagnosis: Multimodal FDG-PET and MRI feature fusion with explainable deep learning Academic Article in Scopus uri icon

abstract

  • Alzheimer's disease (AD) is a progressive neurodegenerative disorder with a rising global incidence. Traditional diagnostic tools, such as magnetic resonance imaging (MRI) and 18F-Fluorodeoxyglucose Positron Emission Tomography (FDG-PET), offer complementary insights into AD pathology. This study proposes an optimization algorithm based on a multimodal deep learning classifier that integrates MRI and FDG-PET data to enhance diagnostic accuracy compared to single-modality approaches. The focus is on binary classification between AD and cognitively normal (CN) subjects, while acknowledging that early mild cognitive impairment (EMCI) and mild cognitive impairment (MCI) represent intermediate stages in the clinical spectrum. Two pre-trained feature extractors ¿ one per imaging modality ¿ were fine-tuned to identify disease-relevant features. Their outputs were fused into a single enriched feature vector used for classification. The dataset comprised 6558 2D coronal MRI slices and 7379 axial FDG-PET slices, extracted from a total of 267 and 493 subjects, respectively, with balanced distributions of AD and CN cases. Data were divided into training, validation, and test sets with no subject overlap to ensure reliable evaluation. Experimental results showed that the multimodal classifier significantly outperformed single-modality models, achieving a 90% accuracy on the test dataset. VGG19 emerged as the most effective feature extractor for both MRI and FDG-PET, achieving individual accuracies of 71.9% and 80.3%, respectively. Additionally, the multimodal model demonstrated superior performance across all evaluation metrics: for the AD class, it achieved a precision of 0.90, recall of 0.94, and F1 score of 0.92; for the CN class, it reached a precision of 0.89, recall of 0.82, and F1 score of 0.86. This research provides compelling evidence supporting the integration of multimodal imaging and advanced machine learning approaches to facilitate timely and accurate AD diagnosis. © 2026

publication date

  • May 15, 2026