Collaborative Multiscale and Wavelet-Based Fusion Network for Leakage Area Semantic Segmentation of Ultrawide Field...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, affecting 30-40% of diabetes patients.
• Vascular leakage in DR, visualized by ultra-widefield fluorescein angiography (UWFA), is difficult to segment due to irregular morphology and high-resolution image complexity.

Purpose of the Study:

• To develop and validate a deep learning framework for accurate and efficient segmentation of diabetic retinopathy leakage in UWFA images.
• To address the challenges of irregular lesion morphology and high computational demands in UWFA image analysis.

Main Methods:

• A deep learning framework combining multiscale sampling, 2D wavelet transforms, and an exponential moving average (EMA) mechanism for feature fusion.
• Implementation of a cross-guided neighborhood refinement strategy to improve segmentation boundary accuracy.

Main Results:

• The proposed model achieved optimal performance with an EMA parameter of 0.3.
• The UNet-Wavelet network significantly outperformed traditional segmentation networks.
• The multiscale fusion framework demonstrated superior robustness compared to non-framework approaches.

Conclusions:

• The developed deep learning method enables efficient and accurate segmentation of leakage regions in UWFA images.
• This approach enhances segmentation accuracy and computational efficiency, facilitating objective clinical quantification of DR-related leakage.
• The findings support earlier intervention for patients with diabetic retinopathy by enabling timely quantification of disease progression.