Abstract
Purpose :
Artificial intelligence (AI)-assisted detection of myopia has the potential to maximize detection rates. This particularly benefits the Chinese community, considering the growing numbers of myopes and the demand for widespread myopic screening to allow earlier interventions. Automated machine learning (AutoML) is a new form of AI that provides readily available machine learning algorithms in a code-free manner. Yet, it has not been broadly used for myopic screening. This study aimed to assess the discriminative performance of AutoML in detecting myopes from normal eyes on fundus images of the Chinese population.
Methods :
Using Google Vertex AI AutoML, we developed a deep-learning classification model to classify fundus images into normal and myopic eyes, including high myopia and pathological myopia. We used an open-access dataset provided by Shanggong Medical Technology Co. Ltd, with images collected from multiple hospitals in China (a total of 5000 fundus images). According to the consensus criteria established for gradable images, the Shanggong dataset was curated into a smaller dataset of myopic and normal images, with 80% of the images being randomly split into training, 10% into validation, and 10% into testing. External validation was performed on 399 fundus images retrieved from another dataset, in which the images were retrospectively collected from the Chinese population at the Zhongshan Ophthalmic Center.
Results :
The AutoML model achieved an area under the precision-recall curve (AUPRC) of 0.997. At the 0.5 confidence threshold cut-off, the overall performance metrics were as follows: F1-score (0.995), precision (99.5%), recall (99.5%), and sensitivity (99.0%). The AutoML model was able to demonstrate excellent discriminative performance and comparable performance to bespoke deep-learning classification models hand-coded by AI experts. During external validation, the AUPRC was 0.788, the F1 score was 0.784, the precision was 78.45%, the recall was 78.54%, and the sensitivity was 64%.
Conclusions :
The AutoML algorithm displayed satisfactory performance in confirming myopia using routinely collected fundus images in practice. It could potentially be a useful tool for advancing myopic screening across the Chinese community. Further research aimed at improving the algorithm’s generalizability through the training dataset (e.g. larger-sized sets of community-collected images) is to be conducted.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.