Related Experiment Video
Updated: May 19, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Outer retinal band segmentation in healthy subjects: comparative study between human grading and deep convolutional
João Duarte Afonso1, Pedro Camacho2, Bruno Pereira2,3
1ITI/LARSyS, Instituto Superior Técnico, Universidade de Lisboa, Lisbon, Portugal.
Background:
Accurate identification of outer retinal bands (ORBs) on spectral-domain optical coherence tomography (SD-OCT) is clinically important for the assessment and monitoring of several retinal diseases. However, manual segmentation of these ORBs is time-consuming and subject to considerable intra- and inter-observer variability. Deep convolutional neural networks (dCNNs) offer a promising approach for achieving more consistent and reproducible segmentation of retinal microstructures, yet their performance for individual ORBs and their dependence on grader reliability remain insufficiently characterized. The aim of this study was to compare manual grading and different dCNN architectures for the segmentation of individual ORBs in healthy subjects, while assessing the reproducibility of human annotations and the performance of automated methods.
Methods:
A cross-sectional observational study was conducted using foveal-centered SD-OCT scans obtained from healthy participants. Manual segmentation of the second hyperreflective band (Band 2), the third and fourth hyperreflective bands (Band 3+4) and photoreceptor outer segments (POS) was performed by two trained graders using onboard OCT software. Reproducibility was assessed using the intraclass and interclass correlation coefficient (ICC). Human annotations were used to train different dCNNs pretrained with Sparse masked modelling on unlabelled data. The performance of the dCNNs was evaluated using dice score and pixel difference analyses.
Results:
One hundred eyes of healthy subjects were included. The intraclass correlation values were 0.737 (0.689, 0.779), 0.830 (0.795, 0.860) and 0.878 (0.852, 0.900), and the ICCs were 0.104 (-0.049, 0.253), 0.810 (0.740, 0.862) and 0.752 (0.678, 0.812) for Band 2, Band 3+4 and POS, respectively. From the dCNN, DRUNET showed the best performance: 0.916 for Band 2, 0.924 for Band 3+4 and 0.910 for POS. Pixel difference analysis between DRUNET and grader 1 showed a deviation of 0.89±0.44.
Conclusions:
Manual segmentation of ORBs demonstrated moderate to high repeatability, whereas Band 2 segmentation showed low reproducibility. DRUNET outperformed other dCNN architectures, showing great potential to reduce inter-observer variability.