Automated Deep Learning Classification of Retinal Pathologies from OCT Videos

Author Block: Samir Touma  , Badr Ait Hammou, Fares Antaki, Renaud Duval.  University of Montreal.

Author Disclosure Block: S. Touma:    All other investments or relationships that could be seen by a reasonable, well-informed participant as having the potential to influence the content of the educational activity; Name of for-profit or not-for-profit organization(s); Bayer Inc.. All other investments or relationships that could be seen by a reasonable, well-informed participant as having the potential to influence the content of the educational activity; Description of relationship(s); Research grant.  B. Ait Hammou:   None.  F. Antaki:    All other investments or relationships that could be seen by a reasonable, well-informed participant as having the potential to influence the content of the educational activity; Name of for-profit or not-for-profit organization(s); Bayer Inc.. All other investments or relationships that could be seen by a reasonable, well-informed participant as having the potential to influence the content of the educational activity; Description of relationship(s); Research grant.  R. Duval:    Funded grants or Clinical Trials; Name of for-profit or not-for-profit organization(s); Alcon, Bayer, Novartis, Roche.. Funded grants or Clinical Trials; Description of relationship(s); Consulting and research grants.

Abstract Title: Automated Deep Learning Classification of Retinal Pathologies from OCT Videos

Abstract Body: Purpose:   Optical coherence tomography (OCT) is a very effective imaging modality for the evaluation of various retinal pathologies. Automated machine learning (AutoML) enables the production of artificial intelligence algorithms without requiring code or programming knowledge. The goal of the project is to develop a deep learning algorithm via Google AutoML Video Intelligence to automatically classify various retinal pathologies from OCT videos. Also, we want to evaluate if the performance of the algorithm is better when it uses complete OCT videos to do the classification versus a single image (focused on the fovea). We will also compare the performance of this algorithm produced without code with models produced manually by an expert in artificial intelligence.   Study Design:   Unicentric prospective observational study   Methods:   We used 1173 OCT videos to produce and test the different models. They had to classify the OCT videos among five different categories: normal retina, diabetic macular edema, wet age-related macular degeneration, macular hole and epiretinal membrane. A fovea-centered image was extracted from each video. The same database was used to train models via AutoML (image and videos) as well as those produced manually. About 80% of the videos/images were used to train (train) the models and 20% to validate them.   Results:   The video AutoML model demonstrated excellent discriminative performance, even outperforming bespoke deep learning model. The area under the precision-recall curve was 0.984, the precision 94.1% and the sensitivity 94.1%. Accuracy was also 94.1%. The best manually produced algorithm had an area under the precision-recall curve of 0.910, a precision of 94.2%, a sensitivity of 94.2% and an accuracy of 96.7%.  The AutoML image model performed better than the one using videos. The overall performance measures were: area under the accuracy-recall curve 0.99, accuracy 96.6%, recall 96.6% and 96.6% accuracy.   Conclusions:   A deep learning model was produced by ophthalmology residents with no coding experience to accurately classify various retinal pathologies via OCT videos with a performance comparable to or better than models developed by an expert. AutoML can play a significant role in the democratization of artificial intelligence among researchers and clinicians who do not have the resources or knowledge to produce deep learning models.