Multimodal 3D Deep Learning for Early Diagnosis of Alzheimer’s Disease

Alzheimer’s disease (AD) is a neurodegenerative disease that affects the elderly and leads to cognitive decline and memory loss. Treatments for stopping or slowing the progression of AD have not been discovered yet; therefore, delaying the progression of AD is the only option, which makes early diagnosis of AD crucial. Additionally, although
$\text{A}\beta $
plaques and tau proteins are considered the causes of early AD, few studies have used this information to diagnose early AD. In this study, a middle-fusion multimodal model is proposed for the diagnosis of early AD. The proposed multimodal model extracts features without loss using a depthwise separable convolution block without an activation function. Subsequently, middle fusion is applied using mix skip connection and sharing weight convolution blocks, both designed to learn the complex relationships between modalities. In contrast to other studies, the proposed approach has three main novelties. 1) A middle-fusion multimodal model is proposed for the early diagnosis of AD. 2) The proposed model is evaluated using the entire ADNI series, including T1-weighted magnetic resonance imaging (T1w MRI) and 18F-FluoroDeoxyGlucose positron emission tomography (FDG PET) from the ADNI1 dataset, as well as
$\text{A}\beta $
PET and tau protein PET from ADNI2 and ADNI3 datasets. 3) A novel region-of-interest (ROI) extraction method is proposed for the hippocampus, middle temporal, and inferior temporal regions, which are known to be affected in the early stages of AD. In the experimental results, the proposed multimodal model achieved a balanced accuracy of 1.00, for the task of Alzheimer’s disease vs cognitive normal (CN) and 0.76 for the task of mild cognitive impairment vs cognitive normal.