Abstract
Purpose :
Optical coherence tomography (OCT) scans of the central retina show the anatomy of the vitreoretinal interface in detail and are widely used by clinicians in the decision-making process for pars plana vitrectomy with epiretinal membrane peeling. In recent years, significant progress has been made in artificial intelligence and computer vision research. We trained a deep convolutional artificial neural network to predict treatment indication for vitrectomy with epiretinal membrane peeling based on macular OCT scans without human intervention.
Methods :
In a retrospective approach we identified a total of 449 vitrectomies with epiretinal membrane removal from the electronic intervention records of a university hospital. OCT images acquired in a 10 week interval before surgical membrane removal were assigned to the treatment group. A random collection of OCT scans without following surgical procedures constituted the control group. After image preprocessing, OCT images were divided into training and test datasets. We trained a deep convolutional neural network of the Inception V4 type using the Google TensorFlow framework and assessed its performance on OCT scans in the test dataset. We calculated prediction accuracy, sensitivity, specifity and receiver operating characteristics.
Results :
A total of 4920 individual OCT B-scans were exported from the institutional image archive and assigned to their respective groups. 4428 images were used to train the neural network classifier. After training the neural network reached a predicition accuracy of 91.9% on the OCT scans of the test dataset. For single retinal B-scans without clinical information a specificity of 95.3% and a sensitivity of 87.8% were achieved. The area under the receiver operating characteristics curve was 0.967 for individual retinal B-scan and 0.978 by averaging over all B-scans of an OCT examination session.
Conclusions :
After training with historical clinical data, artificial intelligence systems are effective in generating treatment propositions. Those methods can offer the clinician support in the decision-making process. Care should be taken not to mistake neural network output as treatment recommendation and to ensure a final thorough evaluation by the treating physician.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.