Abstract
Purpose :
The image quality of optical coherence tomography (OCT) scans is crucial for the accurate and reliable interpretation of retinal structures. Traditionally, signal strength is used as an index to include or exclude OCT scans for further analysis. However, it is insufficient to assess other image quality issues which require specialized knowledge in OCT for such assessment. Our study aims to develop and evaluate a deep learning system (DLS) as an automated tool for filtering out ungradable OCT volumetric scans.
Methods :
We proposed a two-stage method to improve the effectiveness of training and the generalization of DLS. In a total of 3205 optic disc OCT volumetric scans were extracted from Cirrus HD-OCT (Carl Zeiss Meditec, Inc., Dublin, CA, USA), in which contains 778 eyes collected from a tertiary eye hospital in Hong Kong for training (80%) and testing (20%). Each scan was labeled as gradable or ungradable by a trained grader and then by a senior grader to mitigate the dissonance. OCT volumes fulfilling either the criteria: 1) missing data, 2) motion artifacts, 3) blurry, 4) signal loss, or 5) poor centration were defined as ungradable. We developed a3D model based on RestNet18 structure to discriminate ungradable optic disc OCT scans from gradable ones. We further evaluated our system with two independent datasets of 1009 scans from 803 eyes collected from another two eye clinics. Receiver Operation Characteristics (ROC) curve, the area under the ROC curve (AUC), sensitivity and specificity were used to evaluate the performance.
Results :
A total of 4214 volumes were used. In the internal validation, the DLS achieved an AUC of 0.918 (95%CI, 0.895 to 0.942). The optimal sensitivity, specificity, and accuracy were 0.876, 0.846, and 0.856, respectively. In the two external validation datasets, the AUC were 0.777 (95%CI, 0.738 to 0.816) and 0.806 (95%CI, 0.769 to 0.864), with sensitivity of 0.787 and 0.718, specificity of 0.699 and 0.815, accuracy of 0.727 and 0.782, respectively.
Conclusions :
The presented DLS achieved a good performance to discriminate the ungradable from gradable OCT volumes in both internal and external validation.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.