OCT Image Classification of Glaucoma Using AutoML – A Code-Free Deep Learning Platform

Abstract

Purpose : Deep learning technology has been successfully implemented in medical image analysis, including ocular imaging. However, effective use of this technology requires expertise in machine learning and coding skills. AutoML provides a code free option for people without such skills to train deep learning models as an investigational tool. This study aimed to evaluate the effectiveness of AutoML in detecting glaucomatous eyes from optical coherence tomography (OCT) image dataset.

Methods : We used Google’s AutoML on its Vertex AI platform. Clinical glaucoma dataset was obtained from New York University, which included 1805 3D OCT (Cirrus HD-OCT, 200x200 optic nerve head (ONH) scans, Zeiss, Dublin, CA) images (1001 scans from 166 healthy and 804 scans from 353 glaucoma cases). A balanced dataset with approximately equal number of images in each class was created using multiple scans from the same individual, especially for healthy subjects. Resampled circum-papillary images at 1.7 mm radius from the center of the ONH were extracted and used as input to the AutoML. The entire dataset was divided into 80% training, 10% validation and 10% testing. AutoML model was trained 10 times (training and validation data randomly sampled each time). Results reported are for the same test data. Model performance was calculated from the confusion matrix provided by Vertex AI. As a reference, performance of a previously published custom convolutional neural network model trained on OCT volumes from the same patient cohort is provided.

Results : AutoML performance is shown in Table, with the performance of a published custom deep learning model as a reference. AutoML performed reasonably well for accuracy, precision and specificity, but not for area under the receiver operating characteristic curve (AUC) and sensitivity.

Conclusions : Glaucoma classification performance of AutoML was moderate despite its reasonable accuracy, precision, and specificity. A drawback of AutoML is its black-box nature, which limits interpretability and the ability to determine sources of performance differences between models. However, AutoML may still be useful for testing feasibility of deep learning applications in clinical studies without the need for coding. Future advances in AutoML might provide similar performance to custom models with increased interpretability, thus increasing its clinical utility.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.