Abstract
Purpose :
To establish an optical coherence tomography angiography (OCTA) based artificial intelligence (AI) machine learning system for detecting different disease status in age-related macular degeneration (AMD) patients.
Methods :
A total of 395 OCTA images were randomly divided into two datasets, in which 356 images formed a training dataset, and the remaining images formed a validation dataset. All OCTA images were defined into 4 different categories: normal, dry AMD (drusen), and wet AMD with either active or inactive choroidal neovascularization (CNV), respectively. ResNet34 was applied to establish the AI model in detecting different diseases status of AMD. Moreover, augmentation skill has been applied to overcome the limitation of the data size and improvement the performance of our AI model. The augmentation skill contained horizontal & vertical flip, random rotation, lighting & contrast change.
Results :
The choroid capillary layer was selected to establish the AI model. (Fig.1)
The heat map and confusion matrix were illustrated in Fig. 2. The accuracy of this AI model was 87.2%. The specificity and sensitivity in most categories ranged from 80% to 100%. However, the sensitivity for drusen was only 50%.
Conclusions :
We successfully developed a quick AI screening system in detecting different status in AMD patients by using OCTA images. Despite the unsatisfied sensitivity for detecting druse, this AI model showed good performance in all the other categories and thus provided a useful screening tool for AMD patients.
This is a 2020 ARVO Annual Meeting abstract.