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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Public cystoscopic benchmarks suffer from limited annotations, class imbalance, and weak structural supervision.
• Generic data augmentation techniques often fail to preserve crucial elements like anatomy, color fidelity, and pathology variations simultaneously.

Purpose of the Study:

• To introduce Uro-PPLD, a latent diffusion framework designed to overcome the limitations of existing cystoscopic benchmarks and augmentation methods.
• To improve the quality and utility of synthetic cystoscopic data for downstream machine learning tasks.

Main Methods:

• Uro-PPLD integrates chroma-luminance decoupled attention, anatomy-guided ControlNet, continuous filling-state conditioning, and a residual pathology adapter within a denoising backbone.
• Topology-aware and color-consistency constraints are employed to ensure structural plausibility and appearance fidelity.
• Two task-specific variants are developed: one for classification augmentation and another for segmentation augmentation.

Main Results:

• Experiments conducted on EBTC, CystoDS, and BlaVeS datasets demonstrate Uro-PPLD's superior generation quality compared to existing methods.
• The framework yields consistent downstream performance improvements in both cystoscopic image classification and segmentation tasks.
• Uro-PPLD effectively addresses challenges related to scarce annotations and weak structural supervision in public benchmarks.

Conclusions:

• Uro-PPLD presents a robust solution for generating high-fidelity synthetic cystoscopic images.
• The proposed framework significantly enhances the utility of generated data for training and improving machine learning models in urology.
• This work advances the development of reliable benchmarks for cystoscopic image analysis.