Automated Foveal Avascular Zone Segmentation in Optical Coherence Tomography Angiography Across Multiple Eye Diseases...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated foveal avascular zone (FAZ) segmentation in optical coherence tomography angiography (OCTA) is crucial for diagnosing eye diseases.
• Current algorithms struggle with varying data accessibility and image quality across different pathologies and devices.
• Vision transformers, while powerful, are data-hungry and prone to overfitting on limited OCTA datasets.

Purpose of the Study:

• To develop a robust, multi-condition transformer-based architecture for improved FAZ segmentation in OCTA.
• To enhance model generalization and performance in low-data or imbalanced OCTA datasets.
• To leverage knowledge distillation for cross-dataset and cross-modality feature transfer.

Main Methods:

• Proposed a multi-condition transformer architecture utilizing four teacher encoders for knowledge distillation.
• Employed intra-modality distillation across OCTA datasets (healthy, Alzheimer's, AMD, diabetic retinopathy) and inter-modality distillation with color fundus photographs.
• Evaluated model performance using the Dice Index across various ocular conditions.

Main Results:

• The multi-condition model achieved a mean Dice Index of 83.8% with pretraining, surpassing single-condition models (mean 83.1%).
• Pretraining on color fundus images provided a marginal improvement in Dice Index across most conditions.
• The proposed architecture demonstrated superior generalization across diverse OCTA datasets and imaging settings.

Conclusions:

• The multi-condition transformer architecture effectively improves FAZ segmentation accuracy and generalizability in OCTA.
• Knowledge distillation is a viable strategy to overcome data limitations in ophthalmic imaging analysis.
• This approach holds potential for broader applications in diagnosing ophthalmic and systemic diseases using diverse imaging data.