Abstract
Purpose :
Build a detection model for atrophy-related biomarkers in OCT Bscans of AMD patients and explore multiple training strategies using multi-grader annotations.
Methods :
Retrospective cohort of 1738 OCT volumes from 1738 distinct eyes and 1471 patients. Those volumes were automatically selected based on inferred biomarkers and the pathology prediction suggested by our previous work. The target biomarkers are related to atrophy: Drusen, Reticular Drusen, Double Layer Sign, Drusenoid PED, cRORA, iRORA, Subretinal Fibrosis and three escape options (Healthy, Not in the list, and Cannot grade). The simultaneous presence of multiple biomarkers is possible. For training, 7690 Bscans from 1538 volumes were annotated, each Bscan by three randomly selected graders. For testing, 9800 Bscans from 200 volumes were annotated by all five graders. Bscans with no grader majority were discarded. Figure 1 shows the biomarker distributions on both sets. We considered two architectures: a model trained to simulate the grading majority vote and an ensemble model composed of 5 models, each trained to simulate one grader. For the ensemble architecture, we combined three outputting and two model selection strategies, as listed in Figure 2. All models use a Resnet50 pre-trained model. We evaluated all models using Cohen’s kappa and compared them to graders on a newly computed majority with the model output as an additional grader.
Results :
All 7 trained models present similar solid performances, as shown in Figure 2. The majority model achieves the highest macro-averaged Cohen’s kappa K=0.71±0.18 and appears to be the best training strategy. On average, the proposed model outperforms all graders for the given task, achieving K=0.78 against K=0.75 of the best grader on the newly computed majority.
Conclusions :
Deep learning can estimate the presence of atrophy-related biomarkers in AMD patients with solid performances. Such a detector can be of a great help to assist clinical practices and clinical development of new medications and therapies for atrophy development.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.