Purchase this article with an account.
Pearse Andrew Keane, Livia Faes, Siegfried Wagner, Dun Jack Fu, Joseph R Ledsam, Reena Chopra, Christoph Kern, Gabriella Moraes, NIKOLAS PONTIKOS, Martin K Schmid, Lucas M Bachmann, Dawn A Sim, Konstantinos Balaskas; Automated Development of Deep Learning Models to Diagnose Retinal Disease from Fundus and Optical Coherence Tomography Images. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1453.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Deep learning (DL) has the potential to transform healthcare. But the highly specialized technical expertise that is necessary for its adoption is a considerable obstacle to its implementation. In this study, healthcare professionals (HCP) without any DL expertise sought to explore the feasibility of automated DL model development and to investigate their performance in diagnosing retinal disease from fundus and optical coherence tomography (OCT) images.
We used two open-source repositories, including retinal fundus (MESSIDOR) and OCT images (provided by Kermany et al.), to develop two separate DL models that predict retinal disease. Both datasets were fed into a neural architectural search framework that automatically curated the most suitable DL architecture and developed a DL model using a random split-sample validation. The DL models were trained to distinguish images showing diabetic retinopathy (DR) from healthy fundus (MESSIDOR), and neovascular age-related macular degeneration (nvAMD), diabetic macular oedema (DMO), drusen, and healthy OCT-images (Kermany et al.). Recall and precision were used to evaluate the diagnostic properties and the area under the curve (AUC) of plotted receiver operating characteristics curves were used to assess the discriminative performance.
The MESSIDOR dataset involved 1187 images, with 153 images showing mild, 247 moderate, and 254 severe DR. The AUC was 0.87, best accuracy was reached at a cut-off value of 0.5 with a recall of 73.3% and a precision of 73.3%. The dataset provided by Kermany et al. involved 101,418 images of 5,761 patients. 31882 images depicted OCT changes related to nvAMD, 11165 to DMO, 8061 depicted drusen, and 50310 were normal. The AUC was 0.99, best accuracy was reached at a cut-off value of 0.5 with a recall of 97.3% and a precision of 97.7%.
In this preliminary study, automatically developed DL models showed high levels of diagnostic performance and discrimination for the diagnosis of common retinal diseases. This suggests that development of DL prediction models is feasible for (HCP) without any DL expertise. However, the translation of this technological success to meaningful clinical impact constitutes the next great challenge.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only