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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Causal Inference

Background:

• Clinical decision-making requires causal reasoning and longitudinal analysis of diverse patient data, including tabular variables and medical images.
• Synthesizing longitudinal medical images, like brain MRIs for Alzheimer's disease (AD) progression, is challenging due to data complexities.
• Existing correlation-based models struggle with predicting medical image evolution under hypothetical clinical variable changes.

Purpose of the Study:

• To develop a novel Causality-inspired Longitudinal Image Synthesis (CLIS) model.
• To address challenges in synthesizing longitudinal medical images, including dimensionality mismatch, inconsistent intervals, and complex causal mechanisms.
• To enable hypothetical scenario analysis for AD progression and characterization.

Main Methods:

• Proposed a CLIS model integrating generative imaging, continuous-time modeling, and structural causal models with neural networks.
• Utilized tabular causal graphs (TCG) and tabular-visual causal graphs (TVCG) to model dependencies between tabular data and visual data.
• Introduced an independent variable to explicitly model time intervals for longitudinal analysis.

Main Results:

• The CLIS model successfully synthesized high-quality and interpretable brain MRIs.
• Evaluated on the ADNI dataset and two additional AD datasets, demonstrating robust performance.
• Generated MRIs provided valuable insights for Alzheimer's disease characterization.

Conclusions:

• The CLIS model effectively synthesizes longitudinal medical images in a causality-inspired manner.
• The model demonstrates significant potential for clinical applications in diagnosis and treatment follow-up for AD.
• This approach enhances causal reasoning in medical image analysis.