5 September 2024
OCT scans can reveal early signs of retinal atrophy, but differences between scanners make automated assessment difficult. In our MACUSTAR paper, we developed a model to detect incomplete and complete retinal pigment epithelium and outer retinal atrophy (iRORA and cRORA) across devices.
The pipeline combines domain adaptation with segmentation. It first reduces differences in image appearance, then detects the granular biomarkers used to identify iRORA and cRORA. These predictions are combined into atrophy segmentations.
We compared the algorithm with manual annotations in 168 patients with intermediate age-related macular degeneration. For the three-way distinction between no atrophy, iRORA and cRORA, quadratic weighted kappa was 0.73 on Spectralis and 0.37 on ZEISS scans. Removing domain adaptation reduced the Spectralis score to 0.63.
It forms part of my PhD thesis on robust medical imaging AI.
Please use this when referring to the post or publication:
C. de Vente, P. Valmaggia, C.B. Hoyng, F.G. Holz, M.M. Islam, C.C. Klaver, C.J. Boon, S. Schmitz-Valckenberg, A. Tufail, M. Saßmannshausen and others. "Generalizable Deep Learning for the Detection of Incomplete and Complete Retinal Pigment Epithelium and Outer Retinal Atrophy: A MACUSTAR Report", Translational Vision Science & Technology, 2024;13(9):11-11.