Abstract

Study Design: Experimental Research

Purpose: To make a codeless machine learning model using transfer learning for fundus and optical coherence tomography (OCT) image classification

Methods: Publicly available datasets of over 400 fundus images and 572 OCT scans were utilised to train machine learning models on AML platforms such as Apple’s CreateML, Google’s AutoML, Clarifai and Medicmind.

Results: Google’s AutoML resulted in a model with an average accuracy of 0.979 with a precision and recall of 98.21%. Apple's CreateML showed an accuracy of 64%. Model trained on Clarifai showed an average accuracy of 0.951 with precision of 0.72 and recall of 0.56. Model trained on Medicmind for diabetic retinopathy grading with inceptionV3 architecture showed a receiver operating characteristic (ROC) area under the curve (AUC) of 0.75.

Conclusion: AI models trained using transfer learning and AML produced competent models with reasonable accuracy in classifying fundus and OCT images.