Deep learning based retinal OCT segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated segmentation of retinal images is crucial for diagnosing eye diseases.
• Spectral domain optical coherence tomography (OCT) provides high-resolution cross-sectional images of the retina.
• Deep learning (DL) methods have shown promise in medical image analysis.

Purpose of the Study:

• To evaluate a novel deep learning (DL) method for automated fine-grained segmentation of spectral domain optical coherence tomography (OCT) images.
• To compare the performance of the proposed DL method against human clinicians and other machine learning (ML) approaches.

Main Methods:

• A new method combining fully convolutional networks (FCN) with Gaussian Processes for post-processing was developed.
• The approach was applied to segment OCT images of patients with mild non-proliferative diabetic retinopathy.
• Performance was evaluated using mean unsigned error and compared to human and other ML methods.

Main Results:

• The proposed DL method achieved performance comparable to human clinicians.
• The method demonstrated a mean unsigned error of 1.06, outperforming other ML methods (errors 1.17-1.81).
• The DL approach showed comparable or superior results to human error (1.10).

Conclusions:

• The developed DL method offers a robust and accurate solution for automated retinal OCT image segmentation.
• This approach holds potential for improving the diagnosis and management of diabetic retinopathy.
• The combination of FCN and Gaussian Processes presents a promising direction for advanced medical image analysis.