WDT-MD: Wavelet Diffusion Transformers for Microaneurysm Detection in Fundus Images

TL;DR

This work addresses the challenge of early microaneurysm detection in fundus images by proposing WDT-MD, a Wavelet Diffusion Transformer that combines noise-encoded image conditioning, pseudo-normal pattern synthesis via inpainting, and a wavelet-domain diffusion process. The method operates in the wavelet space to preserve multi-scale vascular details while avoiding identity mapping, and provides pixel-level supervision to reduce false positives. Extensive experiments on IDRiD and e-ophtha MA demonstrate state-of-the-art performance in both pixel- and image-level MA detection, highlighting the approach's potential for improving early diabetic retinopathy screening. Overall, WDT-MD offers a robust, efficient framework for accurate MA localization and classification with clear clinical relevance.

Abstract

Microaneurysms (MAs), the earliest pathognomonic signs of Diabetic Retinopathy (DR), present as sub-60 $μm$ lesions in fundus images with highly variable photometric and morphological characteristics, rendering manual screening not only labor-intensive but inherently error-prone. While diffusion-based anomaly detection has emerged as a promising approach for automated MA screening, its clinical application is hindered by three fundamental limitations. First, these models often fall prey to "identity mapping", where they inadvertently replicate the input image. Second, they struggle to distinguish MAs from other anomalies, leading to high false positives. Third, their suboptimal reconstruction of normal features hampers overall performance. To address these challenges, we propose a Wavelet Diffusion Transformer framework for MA Detection (WDT-MD), which features three key innovations: a noise-encoded image conditioning mechanism to avoid "identity mapping" by perturbing image conditions during training; pseudo-normal pattern synthesis via inpainting to introduce pixel-level supervision, enabling discrimination between MAs and other anomalies; and a wavelet diffusion Transformer architecture that combines the global modeling capability of diffusion Transformers with multi-scale wavelet analysis to enhance reconstruction of normal retinal features. Comprehensive experiments on the IDRiD and e-ophtha MA datasets demonstrate that WDT-MD outperforms state-of-the-art methods in both pixel-level and image-level MA detection. This advancement holds significant promise for improving early DR screening.

Figures (6)

• Figure 1: An illustration of MAs in fundus images. (a) is sampled from the IDRiD dataset porwal2018indian, and (b) is from the e-ophtha MA dataset decenciere2013teleophta. Three columns in each sub-figure depict the fundus image, patches zooming in MAs, and MA areas marked in red, respectively. Most MAs are within 60 $\mu m$ in diameter, close to 6 pixels in a fundus image with 10 $\mu m$ pixel spacing.
• Figure 2: Overview of our proposed WDT-MD method. It is a supervised DiT-based AD framework operating in the wavelet domain, which focuses on MA detection in fundus images. By synthesizing the normal pattern and subtracting the input from it, the model obtains an anomaly map, which further outputs both pixel-level and image-level predictions. (std: standard deviation)
• Figure 3: Visualization of HSV channel decomposition on (a) IDRiD and (b) e-ophtha MA. The H and S channels carry little effective information but notable noise, while V contains almost all crucial structural and textural features.
• Figure 4: The training process of our WDT-MD.
• Figure 5: The qualitative results for WDT-MD compared with other state-of-the-art methods on IDRiD and e-ophtha MA.