Reproduce “WDT-MD: Wavelet Diffusion Transformers for Microaneurysm Detection in Fundus Images” on OpenResearch

About this paper

Yifei Sun, Yuzhi He, Junhao Jia, Jinhong Wang, Ruiquan Ge, Changmiao Wang, Hongxia Xu

A Wavelet Diffusion Transformer framework (WDT-MD) improves microaneurysm detection in fundus images by integrating multi-scale wavelet analysis, a noise-encoded image conditioning mechanism, and pseudo-normal pattern synthesis for robust early Diabetic Retinopathy screening. The method achieves pixel-level AUCs of 82.80% on IDRiD and 81.08% on e-ophtha MA datasets, outperforming existing state-of-the-art approaches across various metrics.

Key insights

• Integrating Discrete Wavelet Transformation (DWT) with Diffusion Transformers offers an efficient and effective way to model multi-scale image features, crucial for detecting tiny lesions, while significantly reducing computational overhead.
• Dynamically perturbing the image condition during diffusion model training prevents the model from simply copying anomalies present in the input (identity mapping), thereby improving the detection of true anomalies.
• Leveraging pixel-level supervision derived from pseudo-normal patterns, generated by inpainting anomalous regions, is effective in reducing false positives and enabling generative models to precisely discriminate between true anomalies and confounding factors.

Summary by alphaXiv.