CNN-Based Device-Agnostic Feature Extraction From ONH OCT Scans

Purpose: Optical coherence tomography (OCT)-derived measurements of the optic nerve head (ONH) from different devices are not interchangeable. This poses challenges to patient follow-up and collaborative studies. Here, we present a device-agnostic method for the extraction of OCT biomarkers using artificial intelligence. Methods: ONH-centered OCT volumes from the Heidelberg SPECTRALIS, ZEISS CIRRUS HD-OCT 5000, and Topcon 3D OCT-1000 Mark I/II and 3D OCT-2000 devices were annotated by trained graders. A convolutional neural network (CNN) was trained on these segmented B-scans and utilized to obtain several ONH biomarkers, such as the retinal nerve fiber layer (RNFL) and the minimal rim width (MRW). The CNN results were compared between different devices and to the manufacturer-reported values using an independent test set. Results: The intraclass correlation coefficient (ICC) for the circumpapillary retinal nerve fiber layer (cpRNFL) at 3.4 mm reported by the CIRRUS and 3D OCT-2000 was 0.590 (95% confidence interval [CI], -0.079 to 0.901), and our CNN resulted in a cpRNFL ICC of 0.667 (95% CI, -0.035 to 0.939). The cpRNFL at 3.5 mm on the CIRRUS, 3D OCT-2000, and SPECTRALIS generated by the CNN resulted in an ICC of 0.656 (95% CI, 0.055-0.922). Comparing the global mean MRWs from the SPECTRALIS between CNN and manufacturer yielded an ICC of 0.983 (95% CI, 0.917-0.997). The CNN ICC for the MRW among the CIRRUS, 3D OCT-2000, and SPECTRALIS was 0.917 (95% CI, 0.947- 0.981). Conclusions: Our device-agnostic feature extraction from ONH OCT scans showed a higher reliability than the measures generated by the manufacturers for cpRNFL. MRW measurements compared very well among the manufacturers.