Narges Norouzi

BACKGROUND Forecasting wound healing from longitudinal images is challenging due to irregular follow-up intervals, limited datasets, and the complex biological processes underlying tissue repair. Recent advances in generative artificial intelligence offer an opportunity to model wound progression as an image-sequence prediction problem. OBJECTIVE This study aims to evaluate generative modeling approaches for sequential wound-healing prediction and to characterize tradeoffs between reconstruction fidelity and perceptual realism. METHODS We evaluated three efficient generative frameworks, LLaMA-Adapter, Sparsely Conditional Flow Matching (River CFM), and StyleGAN2-based DyneODE, for sequential wound image forecasting. The models were trained on a porcine wound dataset collected over a 21-day healing period and conditioned on early wound images to generate future trajectories. Performance was assessed across forecast horizons using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Kernel Inception Distance (KID), which is reliable for evaluating performance on small datasets. Reconstruction results are also evaluated through qualitative analysis. RESULTS Across all models, prediction quality degraded as the forecast horizon increased. DyneODE achieved the highest PSNR (18.24) and SSIM (0.368), but qualitative analysis showed that its predictions often collapsed into oversmoothed, visually implausible wound-healing patterns. River CFM produced partially coherent trajectories with an intermediate PSNR (17.11) but showed reduced structural clarity and perceptual quality at longer horizons, as reflected in a higher KID (0.187). In contrast, LLaMA-Adapter attained the lowest KID (0.123) and generated the most visually plausible wound progression patterns, preserving texture and localized wound features despite lower PSNR (16.57). CONCLUSIONS Generative modeling of sequential wound forecasting remains challenging in small-data settings. Although the models often captured coarse temporal trends, longer-horizon predictions frequently degraded, particularly for DyneODE and River CFM, which tended to collapse toward smoother, mean-like appearances. In contrast, LLaMA-Adapter better preserved perceptual realism and produced more visually plausible trajectories. Overall, the findings characterize the strengths and weaknesses of current sequential generative approaches for wound forecasting.