Uro-PPLD: Physiology and Pathology-Aware Latent Diffusion for Cystoscopic Image Generation and Augmentation.

Abstract

Public cystoscopic benchmarks are limited by scarce annotations, class imbalance, and weak structural supervision. Generic augmentation often fails to preserve anatomy, color fidelity, and pathology-related variation simultaneously. To address these challenges, Uro-PPLD is proposed-a latent diffusion framework that integrates chroma-luminance decoupled attention, anatomy-guided ControlNet, continuous filling-state conditioning, and a residual pathology adapter into one denoising backbone. Topology-aware and color-consistency constraints further enforce structural plausibility and appearance fidelity. Two task-specific variants are introduced: a weakly conditioned variant for classification augmentation and a vessel-aware variant for segmentation augmentation. Experiments on EBTC, CystoDS, and BlaVeS under a unified public-benchmark protocol show that Uro-PPLD improves generation quality and yields consistent downstream gains in both classification and segmentation.