Texture-aware diffusion modelling for severity-conditioned synthetic progression-like visualization of diabetic foot ulcers from single images

Abstract

Diabetic foot ulcers (DFUs) are serious complications of diabetes associated with delayed healing, infection, and limb amputation. Modelling DFU progression remains challenging because publicly available image datasets generally contain independent single-timepoint images rather than longitudinal follow-up sequences. This study proposes a texture-aware conditional diffusion framework for generating severity-conditioned, progression-like synthetic DFU images from a single RGB input. Instead of claiming patient-specific temporal prediction, the framework models plausible population-level visual transitions across ordinal pseudo-stages derived from clinically relevant severity cues. The full model achieved an SSIM of 0.851 ± 0.020, PSNR of 24.76 ± 1.12 dB, and FID of 44.33 ± 5.02, with higher observed SSIM and PSNR and lower FID values than the evaluated GAN and VAE baselines across structural fidelity and distributional realism. The GAN baseline achieved SSIM 0.782 ± 0.031, PSNR 21.98 ± 1.56 dB, and FID 59.13 ± 6.59, while the VAE baseline achieved SSIM 0.738 ± 0.038, PSNR 19.63 ± 1.83 dB, and FID 69.77 ± 7.44. The framework also obtained a Texture Consistency Score of 0.79 ± 0.04 and a Granulation Tissue Fidelity Score of 0.76 ± 0.05, indicating preservation of fine-grained texture and granulation-related appearance patterns. Ablation results confirmed the contribution of texture conditioning and cross-attention to image fidelity. Advanced severity-state analysis produced SSIM values of 0.674, 0.804, and 0.726 across three pseudo-stages, supporting severity-aligned structural consistency. Since true longitudinal future-state images were unavailable, these metrics indicate synthetic image fidelity and distributional realism, not longitudinal progression accuracy. Longitudinal DFU cohorts and larger clinician-centred validation remain necessary before clinical decision-support use or temporal progression claims can be justified. To assess generalization beyond the primary 493-image dataset, external validation was additionally performed on an independent DFU dataset under a zero-shot setting, demonstrating stable cross-dataset performance with only modest degradation in image fidelity and texture consistency. Therefore, the proposed framework should be interpreted as modelling cross-sectional severity-conditioned appearance interpolation across independent cases, rather than true patient-specific future progression prediction.