Author: Agata Mosinska (Switzerland)
Co-authors: Anthony Gigon, Stefanos Apostolopoulos, Carlos Ciller, Irmela Mantel, Sandro De Zanet
Purpose
To develop and validate an automatic Retinal Pigment Epithelial and Outer Retinal Atrophy (RORA) progression prediction model for dry Age-related Macular Degeneration cases in Optical Coherence Tomography scans.
Setting/Venue
A machine learning algorithm development and retrospective analysis of AMD patients from the Department of Ophthalmology at Jules-Gonin Eye Hospital, Lausanne, Switzerland.
Methods
A total of 129 eyes of 119 patients were identified and exported. Inclusion criteria consisted of a confirmed AMD diagnosis, a follow-up duration of at least 1 year with OCTs acquired with Spectralis device and RORA presence in the follow-up period. Eyes with neovascular complications, anti-VEGF treatment, any confounding retinopathy or poor image quality were excluded. 109 eyes were used for algorithm development with the remaining 20 for validation. In the training OCTs, RORA segmentation was obtained using an automated algorithm, while in the test scans it was manually delineated by an experienced clinician. All OCT scans were automatically segmented into 6 retinal layers (RNFL, GCL+IPL, INL+OPL, ONL, RPE+PR, choroid) and drusen, from which en-face thickness and reflectance maps corresponding to each retinal layer were computed. Those served as an input to the neural network, which provided RORA prediction at a future time point. Based on the baseline OCT scan of each test patient, en-face RORA predictions were calculated for subsequent visits. The performance of the algorithm was quantified using Dice score and square root area error compared to the manual ground-truth. The scores were categorized and averaged based on the time elapsed since the baseline and baseline RORA area.
Results
19 out of 20 test patients presented atrophy at the baseline visit. The mean RORA area at the baseline was 5.9±4.7 mm2. The mean follow-up time in the test set was 41±18 months and the mean growth rate 0.33±0.16 mm/year. The average Dice score for segmentations at the baseline was 0.84, 0.82 for predictions up to 1 year since the baseline, 0.79 up to 2 years, 0.73 up to 3 years, 0.76 up to 4 years and 0.78 for predictions more than 4 years since the baseline. The average square root RORA area error for baseline prediction was 0.16mm , 0.19mm for predictions up to 1 year since the baseline, 0.24mm up to 2 years, 0.27 up to 3 years, 0.32 up to 4 years and 0.47 for predictions more than 4 years since the baseline. The performance varied also based on the baseline RORA area with a square root area error of 0.28mm for patients with square root baseline area less than 1mm, 0.24mm for baseline area less than 2.0mm and 0.27mm for baseline area more than 2.0mm.
Conlusions
Automatic retinal layer, drusen and RORA segmentation enabled creating a large training set and developing a machine learning-based method for RORA progression prediction. The algorithm was shown to provide accurate RORA predictions for periods as long as 4 years since the baseline visit, with performance decreasing with time elapsed since the baseline and for smaller baseline RORA area cases. The potential applications of such a method include better understanding of the disease progression by the clinician and patients and, once the treatment is available, informed treatment planning.
Financial Disclosure
RetinAI Medical AG - employee and shareholder
Comments
-