Automated geographic atrophy segmentation for SD-OCT images based on two-stage learning model

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of geographic atrophy in spectral-domain optical coherence tomography (SD-OCT) is crucial for disease monitoring.
• Current segmentation methods face challenges in reliability and automation.

Purpose of the Study:

• To develop an effective and automated deep learning framework for segmenting geographic atrophy in SD-OCT images.
• To improve the accuracy and reliability of geographic atrophy segmentation without relying on retinal layer segmentation.

Main Methods:

• A two-stage deep learning framework utilizing a stacked sparse auto-encoder for offline and self-learning was developed.
• Axial cross-section image data was used as input, and a fusion strategy refined segmentation results.
• The method was evaluated on two independent SD-OCT datasets.

Main Results:

• The proposed method achieved a mean overlap ratio (OR) of 89.85% ± 6.35% and an absolute area difference (AAD) of 4.79% ± 7.16% on the first dataset.
• On the second dataset, the mean OR was 84.48% ± 11.98% and the mean AAD was 11.09% ± 13.61%.
• Experimental results demonstrated superior segmentation accuracy compared to state-of-the-art algorithms.

Conclusions:

• The proposed two-stage deep learning framework provides accurate and reliable segmentation of geographic atrophy from SD-OCT images.
• This method offers a promising automated solution for clinical applications in ophthalmology.
• The approach effectively segments geographic atrophy without requiring prior retinal layer segmentation.