First Author: S.Kuwayama JAPAN
Co Author(s): T. Yasukawa A. Kato H. Usui Y. Ogura
Back to previous
Optical coherence tomography (OCT) is a useful tool for diagnosis of macular diseases and assessment of treatment efficacy. Nevertheless, interpretation of various findings on OCT requires expertise. Recently, machine learning systems such as a convolutional neural network (CNN), Artificial Intelligence (AI) have been developing. The main purpose of this study is to assess the feasibility of machine learning of OCT images to support the diagnosis of macular diseases.
Retrospective data analysis at a single facility. Patients who underwent OCT from 2010 to 2015 were enrolled. Cirrus HD-OCT, Model 4000 (Carl Zeiss Meditec AG, Jena, Germany) was used. Patients, whose OCT images were obscure because of hazy media (e.g.; severe cataract), nystagmus, or fixation failure, were excluded.
Both horizontal and sagittal OCT images passing through the fovea in 600 eyes of 300 patients were collected. In total, 1200 images were read by a masked experienced doctor. Then 1100 images were used for machine learning training. Thereafter, remaining 100 images were diagnosed by a trained CNN model, provided with top 5 candidates of diagnosis. The diagnoses by AI and a doctor were compared.
The diagnosis of 1100 images by a doctor involved ‘normal’ (n=570, 47.5%), ‘age-related macular degeneration’ (n=136, 11.3%), diabetic retinopathy (n=104, 8.7%), epiretinal membrane (n=90, 7.5%), and other 19 diseases (n=200, 18.2%). Automated detection by a trained model on 100 images showed that actual diagnosis corresponded to top 1 candidate in 83 images (83.0%), top 2 candidate in 7 (7.0%), and top 3 candidate in 4 (4.0%). Ninety-two images (92.0%) were diagnosed within top 3 candidates and at the probability more than 10% for the correct answer.
A trained CNN model could diagnose correctly in 83.0% of OCT images, suggesting the feasibility of AI to apply to automated detection of macular diseases. Further researched will be needed to improve the precision and recall.