Abstract
Purpose :
Retinal detachments (RD) are highly treatable with timely and appropriate intervention. Thus, early diagnosis is essential when managing this condition. Ultra-wide-field fundus ophthalmoscopy is a recent technology that enables the prompt visualization of a broad fundus area. In this study we applied machine learning to develop and validate and algorithm that automatically detects RD in ultra-wide-field fundus ophthalmoscopy.
Methods :
We used an open database of images from RD patients and normal controls that were obtained between October 2011 and September 2018 at Himeji, Japan (Tsukazaki Optos Public program - available at https://tsukazaki-ai.github.io/optos_dataset/) with the Optos200Tx (Optos®, Dunfermline, UK). 80% of images were used for training, 10% for hyper-parameter tuning and/or to decide when to stop training and 10% of images were used for evaluating the model. We developed an automated model using transfer learning, where a model developed for a specific task is repurposed and leveraged as a starting point for training on a novel task (Google Cloud AutoML Vision). Precision, recall and area under the precision-recall curve were estimated and used to evaluate the performance of the model.
Results :
In total, 673 ultra-wide-field images from RD patients and 624 ultra-wide-field images from normal controls were used to train and evaluate the algorithm. A broad spectrum of RD was included, namely: central, peripheral, macula on and macula-off, tractional, and rhegmatogenous). The total time of training was 16 computing hours. In the validation phase, the algorithm demonstrated high sensivity and specificity for detecting a RD in ultra-wide-field fundus ophthalmoscopy (precision= 89.92%, recall=89.92%, area under the curve=0.953).
Conclusions :
Our model can detect various types of RD with high sensivity and specificity in ultra-wide-field fundus ophthalmoscopy. The automated detection of RD can improve medical care in areas where clinics are not promptly available. Implementing this strategy may promote the early recognition and intervention in this condition, thus preventing further visual morbidity.
This is a 2020 Imaging in the Eye Conference abstract.