Multi-scale adaptive fusion network for retinal layer and fluid segmentation in optical coherence tomography B-scans

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic macular edema (DME), age-related macular degeneration (AMD), and retinal vein occlusion (RVO) are major threats to visual health.
• Accurate diagnosis and interpretation of retinal diseases are crucial for effective treatment.
• Retinal fluid structures complicate disease diagnosis and localization using optical coherence tomography (OCT).

Purpose of the Study:

• To develop an advanced deep learning architecture, the Adaptive Multi-Domain Fusion Network (AMDF-Net), for improved detection of retinal diseases.
• To enhance the accuracy of diagnosing and localizing retinal diseases, particularly those involving retinal fluids.

Main Methods:

• The Adaptive Multi-Domain Fusion Network (AMDF-Net) was developed, incorporating a Hybrid Spectral-Spatial Transformer (HSST) for global and local feature analysis.
• A Dynamic Attention Fusion (DAF) module was integrated to identify features specific to retinal fluids.
• A Disease-Inclusive Segmentation (DIS) module was utilized for accurate primary fluid diagnosis.

Main Results:

• AMDF-Net demonstrated high performance on publicly available and real-time data.
• Achieved a Dice coefficient of 98.87% for segmentation accuracy.
• Attained a classification accuracy of 98.12% for disease detection.

Conclusions:

• AMDF-Net shows significant potential in elevating automated retinal disease analysis.
• The model can provide valuable assistance in treatment-focused decision-making.
• This deep learning approach offers a promising tool for ophthalmologists in diagnosing and managing retinal conditions.