ImageFlowNet: Forecasting Multiscale Trajectories of Disease Progression with Irregularly-Sampled Longitudinal Medical Images

• DOI: 10.36227/techrxiv.172297920.01199828/v1
• Publication Date: 2024-08-06T21:20:13Z
• Authors: Chen Liu; Ke Xu; Liangbo L Shen; Guillaume Huguet; Zilong Wang; Alexander Tong; Danilo Bzdok; Jay Stewart; Jay C Wang; Lucian V Del Priore; Smita Krishnaswamy
• Keywords: Longitudinal data

The forecasting of disease progression from images is a holy grail for clinical decision making. However, this task complicated by the inherent high dimensionality, temporal sparsity and sampling irregularity in longitudinal image acquisitions. Existing methods often rely on extracting hand-crafted features performing time-series analysis vector space, leading to loss rich spatial information within images. To overcome these challenges, we introduce ImageFlowNet, novel framework that learns latent-space flow fields evolve multiscale representations joint embedding spaces using neural ODEs SDEs model domain. Notably, Image-FlowNet representation combining cohorts patients together so can be transferred between patient samples. dynamics then provide plausible trajectories progression, with SDE providing alternative same starting point. We theoretical insights support our formulation ODEs, motivate regularizations involving high-level visual features, latent space organization, trajectory smoothness. demonstrate ImageFlowNet's effectiveness through empirical evaluations three medical datasets depicting retinal geographic atrophy, multiple sclerosis, glioblastoma. Code available at https://github.com/KrishnaswamyLab/ImageFlowNet.