Deep learning-based segmentation of OCT images for choroidal thickness

To develop and validate a custom deep learning-based automated segmentation for choroidal thickness of optical coherence tomography (OCT) scans.

An in-house automated algorithm was trained on a Deeplabv3+ network, based on ResNet50, using a training set of 10,798 manually segmented OCT scans (accuracy 99.25% and loss 0.0229). A test set of 130 unique scans was segmented using manual and in-house automated methods. For manual segmentation, the choroid-sclera border was delineated by the user. For in-house automated segmentation, all borders were automatically detected by the program and manually inspected. Bland-Altman analysis, intraclass correlation coefficient (ICC), and Deming regression compared the central 1-mm diameter and 3-mm and 6-mm annuli for the two methods. The in-house method was also compared with an open-source algorithm for the test set of 130 scans.

Mean choroidal thicknesses obtained with manual and in-house automated methods were not significantly different for the three regions (P > 0.05 for all). The fixed bias between methods ranged from -2.41 to 3.49 µm. Proportional bias ranged from -0.04 to -0.12 (P < 0.05 for all). The two methods demonstrated excellent agreement across regions (ICC: 0.96 to 0.98, P < 0.001 for all). The open-source automated method consistently resulted in thinner choroidal thickness compared to manual and in-house automated methods.

Custom in-house deep learning automated choroid segmentation demonstrated excellent agreement and strong positive linear relationship with manual segmentation. The automated approach holds distinct advantages for estimating choroidal thickness, being more objective and efficient than the manual approach.