Abstract
Diabetic Retinopathy (DR) lesion segmentation is significantly challenged by extreme variations in lesion size and the loss of fine-grained spatial details during standard neural network downsampling. We present WSRFNet, a novel architecture that employs a recurrent feedback mechanism using lossless Haar Wavelet Transforms (HWT) and scale-specific refinement. By focusing on improving the quality of multi-scale features before fusion, WSRFNet achieves state-of-the-art performance on major benchmarks, significantly outperforming existing methods in segmentation accuracy.
Introduction
Diabetic Retinopathy (DR) is a primary cause of vision loss, marked by lesions such as hemorrhages, exudates, and microaneurysms. A significant challenge in automatic Diabetic Retinopathy Lesion Segmentation (DRLS) is the extreme variance in lesion size; in standard datasets, the largest lesions can be over 70 times the size of the smallest ones.
Conventional approaches often attempt to address this through complex multi-scale feature fusion modules. However, these methods frequently fail because the underlying features lose critical spatial details during standard downsampling operations like pooling or strided convolutions. WSRFNet shifts the focus from fusion complexity to feature quality, arguing that high-quality, refined features are essential for identifying tiny lesions.
Methods
WSRFNet introduces a Wavelet-Based Scale-Specific Recurrent Feedback Network designed to iteratively enhance feature maps. The architecture consists of three core components:
- Recurrent Refinement: The model performs multiple iterations, using high-level global context from previous passes to correct and sharpen features in subsequent loops.
- Wavelet-based Feedback Pyramid Module (WFPM): To prevent information loss during feedback propagation, WSRFNet utilizes the Haar Wavelet Transform (HWT). Unlike standard downsampling, HWT is a reversible and lossless operation, ensuring that fine-grained spatial data is preserved when passing signals to higher-resolution layers.
- Scale-specific Refinement Module (SRM): Recognizing that different layers require different types of correction—low-level layers need semantic context while high-level layers need detail—the SRM generates targeted feedback tailored to each specific scale.
Experimental Results
The model was evaluated on the IDRiD and DDR benchmarks. On the DDR dataset, WSRFNet demonstrated a substantial performance leap, exceeding previous state-of-the-art methods by 14.77% in mAUPR and 12.86% in mDice. On the high-resolution IDRiD dataset, it achieved top-tier results with an mAUPR of 71.05 and mDice of 67.53. Ablation studies confirmed that the wavelet-based feedback mechanism is superior to traditional bilinear interpolation or strided convolutions.
Conclusion
WSRFNet demonstrates that refining feature quality through lossless wavelet-based feedback and scale-specific modules is more effective than increasing fusion complexity. While the recurrent nature of the network introduces a computational trade-off—with four iterations found to be optimal—the gains in segmentation accuracy for multi-scale lesions are significant. Future research will focus on optimizing the efficiency of these recurrent loops and extending the framework to other medical imaging domains.
Summary updated