Abstract
Purpose :
With automated machine learning (AutoML) platforms, it is now possible for clinicians without coding experience to develop deep learning image classification models. However, to run these models, most platforms require data to be uploaded online. This raises issues of patient confidentiality, data security, and requires stable internet connectivity. We investigated the creation of a compact local model known as an edge model, which was then deployed via an offline android application. This application was able to classify optical coherence tomography images in real-time using the built-in camera on a mobile device.
Methods :
We used Google Cloud AutoML Vision to train an edge model for classification of fovea centered OCT B-scans. These models were trained, validated and tested from the publicly available Waterloo OCT dataset which contains 572 OCT images in total. The OCT images were labelled with normal n=206, diabetic retinopathy (DR) n=107, central serous retinopathy (CSR) n=102, Macular Hole (MH) n=102 and Age-related macular degeneration (AMD) n=55. We ran the training for 10 node hours with 80:10:10 splits for training, validation and test. The app was developed using an open-source template developed by Google Firebase. The app was modified in-order to deploy the OCT classification edge model.
Results :
The app and the classification model was functional on multiple tested Android devices.The edge deep learning model was able to achieve an overall area under the curve (AUC) of 97.4%, precision of 98.04%, and recall of 89.29. The resulting per-label (positive predictive value, negative predictive value, sensitivity and specificity) were AMD: (100%,95.9%, 60%, 100%),CSR: (100%,93.5%, 70%,100%), DR:(100%,100%,100%,100%) MH: (90%, 97.6%, 90%, 97.6) and Normal: (100%,100%,100%,100%).
Conclusions :
We developed a mobile OCT classification model without coding. This demonstrates the democratization of AI to researchers without access to computer science expertise. Although we used a limited public dataset for model training, we achieved high classification accuracy on a mobile app deployment. This proof of concept entails broad future use cases and has the potential of overcoming constraints such as poor internet access, data privacy, and limited access to AI expertise.
This is a 2020 ARVO Annual Meeting abstract.