MC-RVAE: Multi-channel recurrent variational autoencoder for multimodal Alzheimer’s disease progression modelling

• Published in: NeuroImage (Orlando, Fla.) Vol. 268; p. 119892
• Main Authors: Martí-Juan, Gerard, Lorenzi, Marco, Piella, Gemma
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.03.2023; Elsevier Limited; Elsevier
• Subjects: Alzheimer Disease - diagnostic imaging; Alzheimer's disease; Biomarkers; Cognitive ability; Disease Progression; Disease progression modelling; Factor analysis; Humans; Longitudinal; Magnetic resonance imaging; Magnetic Resonance Imaging - methods; Medical research; Multimodal; Neural networks; Neural Networks, Computer; Neurodegenerative diseases; Neuroimaging - methods; Positron-Emission Tomography - methods; Recurrent neural network; Variables; Variational autoencoder; Variational autoencoder; Longitudinal; Alzheimer’s disease; Multimodal; Recurrent neural network; Disease progression modelling
• ISSN: 1053-8119; 1095-9572
• DOI: 10.1016/j.neuroimage.2023.119892

•A multi-channel model based on recurrent variational autoencoders was proposed to capture spatial and temporal evolution of AD using multimodal data.•Proposed model was evaluated on synthetic and real datasets.•Model outperforms a set of baselines for missing data reconstruction across modalities. [Display omitted] The progression of neurodegenerative diseases, such as Alzheimer’s Disease, is the result of complex mechanisms interacting across multiple spatial and temporal scales. Understanding and predicting the longitudinal course of the disease requires harnessing the variability across different data modalities and time, which is extremely challenging. In this paper, we propose a model based on recurrent variational autoencoders that is able to capture cross-channel interactions between different modalities and model temporal information. These are achieved thanks to its multi-channel architecture and its shared latent variational space, parametrized with a recurrent neural network. We evaluate our model on both synthetic and real longitudinal datasets, the latter including imaging and non-imaging data, with N=897 subjects. Results show that our multi-channel recurrent variational autoencoder outperforms a set of baselines (KNN, random forest, and group factor analysis) for the task of reconstructing missing modalities, reducing the mean absolute error by 5% (w.r.t. the best baseline) for both subcortical volumes and cortical thickness. Our model is robust to missing features within each modality and is able to generate realistic synthetic imaging biomarkers trajectories from cognitive scores.