RSAPower: Random Style Augmentation Driven Structure Perception Network for Generalized Retinal OCT Fluid...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Optical Coherence Tomography (OCT) is crucial for non-invasive diagnosis of retinal diseases.
• Accurate segmentation of retinal fluid in OCT images is vital for disease quantification and clinical decisions.
• Current segmentation methods struggle with pathological variations, unclear boundaries, low contrast, and domain shift due to varying OCT image styles.

Purpose of the Study:

• To propose a novel method, RSAPower, to enhance the generalization ability of fluid segmentation networks.
• To address the challenge of domain variability in OCT image styles for improved fluid segmentation.

Main Methods:

• Introduced RSAPower, a method combining a random style transform augmentation (RSTAug) module and a fluid perception network (FLPNet).
• RSTAug generates diverse, realistic style-augmented data from the source domain.
• FLPNet features a hybrid structure attention (HSA) module for spatial and long-range feature perception and a saliency-guided multi-scale attention (SGMA) block for adaptation to augmented data.

Main Results:

• RSAPower demonstrated superior generalization ability compared to state-of-the-art methods.
• The method achieved high effectiveness in retinal fluid segmentation on the Retouch and Kermany datasets.
• Experimental validation confirmed the robustness of RSAPower against domain variations.

Conclusions:

• RSAPower effectively enhances the generalization of fluid segmentation networks through style augmentation.
• The proposed method offers a promising solution for accurate and robust fluid segmentation in diverse OCT imaging conditions.
• This approach aids in more reliable quantification of retinal fluid-associated diseases.