Global-Local Transformer Network for Automatic Retinal Pathological Fluid Segmentation in Optical Coherence...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of retinal pathological fluids like intraretinal fluid (IRF), subretinal fluid (SRF), and pigment epithelial detachment (PED) is crucial for diagnosing and managing various retinopathies.
• Challenges in segmentation include variations in fluid characteristics, low contrast, noise, and limitations of traditional deep learning models in capturing global dependencies.

Purpose of the Study:

• To develop an automatic method for precise segmentation and quantitative analysis of multi-type retinal fluids in optical coherence tomography (OCT) images.
• To address the limitations of existing models in capturing global dependencies for improved pathological feature identification.

Main Methods:

• Developed a novel Global-Local Transformer Network (GLTNet) with a U-shape architecture for simultaneous multi-type retinal fluid segmentation.
• Integrated a Global-Local Attention Module (GLAM) into a VGG-19 backbone to enhance feature representation and noise suppression.
• Incorporated a Multi-Scale Transformer Module (MSTM) in the encoder pathway to capture long-term dependencies across multiple scales.

Main Results:

• The GLTNet achieved superior segmentation performance on the Kermany dataset, with Dice coefficients of 0.8395, IoU of 0.7657, Sensitivity of 0.8631, and Precision of 0.8202.
• The model significantly outperformed other state-of-the-art retinal fluid segmentation approaches.
• Experimental results on DUKE and UMN datasets demonstrated satisfactory generalizability of the proposed model.

Conclusions:

• The developed GLTNet significantly improves retinal fluid segmentation accuracy, generalization, and robustness compared to current methods.
• The model shows great potential in assisting ophthalmologists with diagnosing diverse eye disorders and developing tailored therapy regimens.