Purchase this article with an account.
Anran RAN, Xi Wang, Luyang Luo, Poemen Chan, Robert Chang, Suria Sudhakaran Mannil, Hao Chen, Pheng-Ann Heng, Clement C Y Tham, Carol Yim-lui Cheung; A 3D Deep Learning System for Detecting Glaucomatous Optic Neuropathy from Volumetric and En Face Optical Coherence Tomography Scans. Invest. Ophthalmol. Vis. Sci. 2019;60(9):5571.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop and validate a deep learning system (DLS) using full 3D optical coherence tomography (OCT) volume combined with 2D line scanning ophthalmoscope (LSO) en face images for an automated binary glaucomatous optic neuropathy (GON) classification.
OCT and LSO images of optic disc cube performed with Cirrus HD-OCT (Carl Zeiss Meditec, Inc., Dublin, CA, USA) were taken from a total of 1,660 eyes in a tertiary eye hospital (Hong Kong) as training (80%) and testing (20%). Each cube scan was labeled as Yes/No GON according to the criteria of retinal nerve fiber layer (RNFL) thinning on reliable SDOCT images, with a structural defect that correlated in position with the visual field defect. Two independent deep networks, a 3D ResNet37 (Figure 1A) and a 2D ResNet34 (Figure 1B) were employed to feed OCT volume and LSO image as input, respectively, with an output of yes/no GON. External validation was performed using two independent datasets with similar GON definition: (1) Prince of Wales Hospital, Hong Kong (419 eyes); and (2) Stanford University (565 eyes). Receiver Operation Characteristics (ROC) curve, the area under the ROC curve (AUC), sensitivity and specificity were used to evaluate the performance of the DLS. Heatmaps were generated by class activation map for further qualitative evaluation.
A total of 7,083 paired volumetric and LSO images were used. In the internal validation, the DLS achieved an AUC of 0.935 (95%CI, 0.935 to 0.936). The overall accuracy, sensitivity, and specificity were 86.9%, 88.0%, and 85.0%, respectively. In the two external validation data sets, the AUC were 0.830 (95%CI, 0.828 to 0.838) and 0.832 (95%CI, 0.830 to 0.835), with accuracy of 71.7% and 76.3%, sensitivity of 70.5% and 77.9%, specificity of 77.2% and 73.2%, respectively. Almost the whole retina, not just RNFL, was taken as a region of interest by the DLS for learning and interpreting GON classification, as demonstrated in the heatmap (Figure 2). The main reasons for misclassification were optic disc tilting and presence of peripapillary atrophy.
The presented DLS, to the best of our knowledge, is the first attempt to use 3D OCT volume combined with 2D LSO image as the input, and it achieved a good diagnostic performance for binary GON discrimination, in both internal and external validation.
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