Purchase this article with an account.
Lingling Wang, Jianchun Zhao, Xuan Zou, Xuan Chen, Chunhui Jiang, Xing Liu, Hui Xiao, Xiaoli Song, Yao Zhang, Jun Wu, Jie Wang, Dayong Ding, Junfeng Tan, Yuan Tian, Ningjiang Chen; Deep learning approach to improve the efficiency and effectiveness of optical coherence tomography (OCT) scans. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1514.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
OCT macular exams routinely include a macular cube scan covering the whole macular region followed by high-resolution scans focusing on critical regions. However, accurately locating the critical pathological regions requires experience and is often time-consuming. This study aimed to develop a deep learning approach to automatically identify B-scans containing pathological anomalies in macular cube scans, such that technicians and ophthalmologists can efficiently find the critical information, and the subsequent high-resolution scans can be planned accordingly and automatically.
599 cube scans (32 slices per scan) and 1,965 HD 5-line raster scans (5 slices per scan) centered on the macular were collected from a Primus OCT scanner (Carl Zeiss Meditec, Suzhou, China), yielding an OCT image set consisting of 28,993 B-scan slices. This OCT image set was split into a training set consisting of 23,500 B-scans and a test set consisting of 5,124 B-scans. Each of the B-scan slices in the training set was graded by one of three ophthalmologists with at least 5-year experience, to be 1) classified as no anomaly, 2) with visible anomaly in one of six sub-categories: overall deformation, fovea, above external limiting membrane (ELM), in or below ELM, sub-retinal area, and other, or 3) unable to determine. B-scan slices in the test set were cross-labeled by all three ophthalmologists. The training set was used to train a deep convolutional neural network with a customized loss function, and the trained model was evaluated on the test set. Sensitivity and specificity of the model to identify anomalies were computed.
In the test set, 3,339 B-scans were labeled as no anomaly while 1,785 were labeled as with visible anomaly or unable to determine (combined as abnormal). The trained deep learning network correctly identified 3,032 normal and 1,617 abnormal B-scans, resulting in a sensitivity of 90.8% and specificity of 90.58%.
It is possible to detect pathological anomalies in OCT with high sensitivity and specificity using a deep learning approach. With this technique being integrated in OCT exam workflow, potential anomalies in a macular cube scan can be automatically identified, making the entire OCT exam workflow more effective and efficient.
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