BioFlow: A Patient-Specific 3D Imaging World Model via Conditional Flow Matching from Irregular Longitudinal Histories

Forecasting a patient's future anatomical state is central to high-stakes health applications, from tracking disease progression to providing a reference trajectory against which candidate interventions can be contrasted. Yet longitudinal 3D forecasting is challenging because scans share substantial anatomical structure, while the changes of interest can be subtle, localized, and patient-specific. Existing methods largely target disease-specific settings with structured progression and large longitudinal cohorts, whereas real-world clinical data often provide limited patient-level observations at sparse and irregular time points. We introduce BioFlow, a multi-scan 3D conditional flow-matching framework that serves as a patient-specific imaging world model, predicting a complete future volume from a variable-length, irregularly sampled observation history without assuming a fixed, disease-specific progression pattern. To the best of our knowledge, BioFlow is the first multi-scan 3D flow-matching framework to decouple clinical time from generative flow dynamics by conditioning an evolving prediction state on a fixed longitudinal history. It incorporates relative clinical time intervals and uses cross-attention to extract both global and spatially aligned local information from the history. Across three synthetic benchmarks and three clinical MRI/CT cohorts, BioFlow outperforms existing baselines across SSIM, PSNR, NRMSE, and MSE, with the largest gains in regions exhibiting anatomical change.

Authors: Ming Liu, Xiaoyuan Cao, Naichen Shi, Hairong Wang, Ke Ren