Conditional Score-Based Diffusion Model for Cortical Thickness Trajectory Prediction

Presentation Time: 08:00 AM - 08:15 AM

Abstract Keywords: Clinical Decision Support, Machine Learning, Deep Learning
Working Group: Biomedical Imaging Informatics Working Group
Primary Track: Applications
Programmatic Theme: Clinical Informatics

Accurate prediction of cortical thickness (CTh) trajectories in Alzheimer's Disease (AD) is crucial for early diagnosis and intervention. This study proposes a conditional score-based diffusion model to model the CTh progression based on only baseline information, including age, sex, and initial imaging results. The model utilizes a 1-D Attention U-net architecture to encode the demographic and CTh data, then learns forward and reverse diffusion processes to generate prediction results from noise, conditioned on the given prior. The model is trained and validated on the TADPOLE Challenge cohort, considering data from 898 participants across five time points. The proposed method outperforms other models in predicting longitudinal CTh across the entire testing cohort and specific subgroups (CN, MCI, and AD). The diffusion model's ability to quantify prediction uncertainty provides more reliable support for clinical decision-making processes. The results demonstrate the potential of the proposed method in predicting longitudinal CTh, which may benefit diagnosis and intervention strategies in the preclinical stages of AD.

Speaker(s):
Xiang Li, PhD
Massachusetts General Hospital and Harvard Medical School