Abstract
Purpose :
Building an expert system for automatic detection of conversion from non-exudative age-related macular degeneration (AMD) to exudative neovascular AMD using optical coherence tomography (OCT) imaging data in a real-world setting.
Methods :
907 AMD patients (40 944 OCT stacks) were labelled as either “early/intermediate”, or “advanced non-neovascular”, or “advanced neovascular non-exudative” or “advanced neovascular exudative AMD” based on their first visit in our database. Eyes exhibiting diseases other than AMD were excluded. All images labelled as “advanced neovascular exudative AMD” were categorized as “post conversion”, the remaining labels were categorized as “pre conversion”. Different inputs (single B-scans, sets of B-scans as separate images, sets of B-scans as 3-dimensional volume) to ML models were evaluated focusing on reliable detection of “post conversion” stage. Special attention was devoted to model performance evaluation, where datasets were split into train and test sets ensuring that distribution of critical attributes is preserved between the sets.
Results :
The primary validation dataset comprised of 347’573 B-scans (13 342 OCT stacks) of 304 patients (33.2% male, 66.8% female), with a mean age of 79 years (7.6 SD) at first visit. The mean follow-up time was 3.6 (3.141 SD) years, 12’382 out of 13’342 OCT stacks were labelled as “post conversion”. The model using sets of B-scans as separate images achieved an area-under-curve (AUC) of 0.9138 in differentiating post conversion from pre conversion OCT stacks. 10 465 of 12 382 OCT stacks labelled as post conversion were correctly classified, while 810 of 960 OCT stacks labelled as pre conversion were correctly classified as non-conversed (Figure 1).
Conclusions :
Our ML model using a clinically heterogenous OCT data set can effectively detect conversion from non-exudative to exudative neovascular AMD. Considering the limited amount of data, the relatively simple ML model architecture and the time-efficient labelling process, the achieved ML model performance seems very promising. Further training is required to evaluate whether our proposed ML model can achieve even higher performance and will be able to predict the time to/from conversion.
This is a 2021 ARVO Annual Meeting abstract.