CLIS: Causality-inspired Longitudinal Image Synthesis and its application to Alzheimer's disease characterization

Clinical decision-making relies heavily on causal reasoning and longitudinal analysis of clinical variables, which include demographic variables, biomarkers, measurements, etc., often stored in a tabular format, and visual medical images. For example, for a patient with Alzheimer's disease (AD), how might brain gray matter atrophy evolve over a year under a hypothetical change in the Aβ₄₂ biomarker level in cerebrospinal fluid? Answering such hypothetical questions is important for diagnosis and follow-up treatment, yet these medical images are neither readily acquired nor effectively predicted by correlation-based image synthesis models. Hence, a Causality-inspired Longitudinal Image Synthesis (CLIS) model is valuable. Building such a CLIS model faces three primary challenges: mismatched dimensionality between high-dimensional images and low-dimensional tabular variables, inconsistent intervals in follow-up data, and the complexity of medical causal mechanisms. In this paper, we propose a CLIS model that addresses these challenges via a novel integration of generative imaging, continuous-time modeling, and structural causal models combined with a neural network. Specifically, we first depict dependencies among tabular variables - including demographics, clinical biomarkers, and brain volumes - using a tabular causal graph (TCG), and then extend this to a tabular-visual causal graph (TVCG) to synthesize brain MRIs in a causality-inspired manner. An independent variable is also introduced to explicitly model time intervals. We train our CLIS on the ADNI dataset and evaluate it on two additional AD datasets, demonstrating that the synthesized images are both high-quality and interpretable. Furthermore, the generated MRIs provide insights for AD characterization, illustrating the model's potential utility in clinical applications.