Morph-Transformer: A Deep Spatiotemporal Framework for Shape Analysis in Longitudinal Neuroimaging

Morphometric brain change is closely related to neurodegenerative disease progression. Modeling temporal trajectories of complex morphometry under various hypothetical disease interventions - termed the "digital twin" problem - is of great interest. Digital twin models can provide insights for understanding disease pathology and guide treatment development. However, longitudinal neuroimaging data is generally limited in size with heterogeneous intra-subject scan intervals. It is thus imperative to develop a reliable spatio-temporal generative model capable of learning in such settings. To address this challenge, we propose the "Morph-Transformer". Morph-Transformer embeds spatial and temporal mesh information into a high-dimensional vector space with the embedding used in the self-attention learning mechanism of auto-regressive transformer model. We validate our model using the PPMI Parkinson’s disease MRI dataset. Morph-Transformer performance compares favorably to the state-of-the-art auto-encoder based generative model. To the best of our knowledge, Morph-Transformer is the first "auto-encoder-free", "transformer-only" deep generative framework to address the digital neuroimaging twin problem.