Outer retinal band segmentation in healthy subjects: comparative study between human grading and deep convolutional...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate identification of outer retinal bands (ORBs) on spectral-domain optical coherence tomography (SD-OCT) is crucial for diagnosing and monitoring retinal diseases.
• Manual segmentation of ORBs is time-consuming and prone to significant intra- and inter-observer variability.
• Deep convolutional neural networks (dCNNs) present a potential solution for consistent and reproducible retinal microstructure segmentation.

Purpose of the Study:

• To compare manual grading with various dCNN architectures for segmenting individual ORBs in healthy subjects.
• To assess the reproducibility of human annotations for ORB segmentation.
• To evaluate the performance of automated dCNN methods in comparison to human graders.

Main Methods:

• A cross-sectional observational study using SD-OCT scans from healthy participants.
• Manual segmentation of Band 2, Band 3+4, and photoreceptor outer segments (POS) by two trained graders.
• Reproducibility assessment using intraclass and interclass correlation coefficients (ICC).
• Training and evaluation of dCNNs using human annotations, with performance measured by dice score and pixel difference.

Main Results:

• Manual segmentation showed moderate to high repeatability, but low reproducibility for Band 2.
• The DRUNET dCNN architecture achieved the highest performance: 0.916 for Band 2, 0.924 for Band 3+4, and 0.910 for POS.
• Pixel difference analysis between DRUNET and a human grader showed a deviation of 0.89±0.44.

Conclusions:

• Manual segmentation of outer retinal bands has limitations, particularly for Band 2.
• DRUNET demonstrated superior performance compared to other dCNN architectures for ORB segmentation.
• dCNNs, particularly DRUNET, hold significant potential for reducing inter-observer variability in retinal image analysis.