Precision Segmentation of Subretinal Fluids in OCT Using Multiscale Attention-Based U-Net Architecture.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of intraretinal fluid (IRF), subretinal fluid (SRF), and pigment epithelial detachment (PED) is crucial for managing Diabetic Macular Oedema (DME) and Age-related Macular Degeneration (AMD).
• Existing segmentation techniques often lack the precision required for effective clinical decision-making.
• Advanced imaging modalities like Optical Coherence Tomography (OCT) generate complex data necessitating sophisticated analysis methods.

Purpose of the Study:

• To develop and validate a deep-learning architecture for precise segmentation of multiple retinal fluid types in DME and AMD.
• To enhance the accuracy and edge definition of fluid segmentation compared to current methods.
• To provide a robust computational tool for improving the diagnosis and treatment planning of retinal diseases.

Main Methods:

• An encoder-decoder deep-learning network, inspired by U-Net architecture, was designed.
• The model processes enhanced OCT images alongside their corresponding edge maps using a dual-input approach.
• The encoder features Residual and Inception modules, incorporating an autoencoder-based multiscale attention mechanism for detailed feature extraction.

Main Results:

• The proposed architecture demonstrated superior performance across multiple datasets (RETOUCH, OPTIMA, DUKE).
• On the RETOUCH dataset, the model achieved high F1 Scores: 0.82 for IRF, 0.93 for SRF, and 0.94 for PED.
• The model consistently exhibited high precision, recall, and F1 Scores, indicating significant improvements in segmentation accuracy and edge precision.

Conclusions:

• The developed deep-learning architecture offers a significant advancement in segmenting retinal fluids.
• This enhanced segmentation accuracy and edge precision can lead to improved clinical outcomes for patients with DME and AMD.
• The model's sophisticated design, including dual-input processing and multiscale attention, represents a valuable tool for retinal disease management and research.