Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2018 Nov 12;8(1):16685.
doi: 10.1038/s41598-018-35044-9.

Performance of Deep Learning Architectures and Transfer Learning for Detecting Glaucomatous Optic Neuropathy in Fundus Photographs

Affiliations
Free PMC article

Performance of Deep Learning Architectures and Transfer Learning for Detecting Glaucomatous Optic Neuropathy in Fundus Photographs

Mark Christopher et al. Sci Rep. .
Free PMC article

Abstract

The ability of deep learning architectures to identify glaucomatous optic neuropathy (GON) in fundus photographs was evaluated. A large database of fundus photographs (n = 14,822) from a racially and ethnically diverse group of individuals (over 33% of African descent) was evaluated by expert reviewers and classified as GON or healthy. Several deep learning architectures and the impact of transfer learning were evaluated. The best performing model achieved an overall area under receiver operating characteristic (AUC) of 0.91 in distinguishing GON eyes from healthy eyes. It also achieved an AUC of 0.97 for identifying GON eyes with moderate-to-severe functional loss and 0.89 for GON eyes with mild functional loss. A sensitivity of 88% at a set 95% specificity was achieved in detecting moderate-to-severe GON. In all cases, transfer improved performance and reduced training time. Model visualizations indicate that these deep learning models relied on, in part, anatomical features in the inferior and superior regions of the optic disc, areas commonly used by clinicians to diagnose GON. The results suggest that deep learning-based assessment of fundus images could be useful in clinical decision support systems and in the automation of large-scale glaucoma detection and screening programs.

Conflict of interest statement

Dr. Robert N. Weinreb received compensation as a consultant from Aerie Pharmaceuticals and Allergan. He received research support from Carl Zeiss Meditec, CenterVue, Genentech, Heidelberg Engineering, Konan Medical, the National Eye Institute, Optos, Optovue, and Research to Prevent Blindness. Dr. Christopher A. Girkin received research support from EyeSight Foundation of Alabama, Heidelberg Engineering, the National Eye Institute, and Research to Prevent Blindness. Dr. Jeffrey M. Liebman received compensation as a consultant from Alcon, Allergan, Bausch & Lomb, Carl Zeiss Meditec, Heidelberg Engineering, Reichert Technologies, and Valeant Pharmaceuticals. He received research support from Bausch & Lomb, Carl Zeiss Meditec, Heidelberg Engineering, the National Eye Institute, Optovue, Reichert Technologies, and Research to Prevent Blindness. Dr. Linda M. Zangwill received research support from Carl Zeiss Meditec, Heidelberg Engineering, the National Eye Institute, Optovue, and Topcon Medical Systems. Dr. Mark Christopher, Akram Belghith, Christopher Bowd, and Michael H. Goldbaum have no competing interests to declare.

Figures

Figure 1
Figure 1
Example input ONH image (top left) and images resulting from data augmentation. Each input image was augmented by horizontal mirroring to mimic alternative OD/OS orientation and random cropping to produce translations of the image.
Figure 2
Figure 2
(A) Schematics of the three CNN architectures evaluated here. These architectures included VGG16, Inception, and ResNet50. The Inception (B) and Residual (C) are used as building blocks for the Inception and ResNet50 architectures, respectively.
Figure 3
Figure 3
Model performance on the validation dataset by iteration for the first 50,000 training iterations for the VGG16 (top), Inception (middle), and ResNet50 (bottom) architectures.
Figure 4
Figure 4
ROC curves in predicting glaucomatous optic neuropathy (GON) for each deep learning architecture initialized using transfer learning.
Figure 5
Figure 5
Predictions of the best performing model compared to expert truth. Examples of correct, incorrect, and borderline predictions are shown along with the expert truth and model glaucomatous optic neuropathy (GON) prediction probabilities (0 = healthy, 1 = GON).
Figure 6
Figure 6
Mean occlusion testing maps showing most significant regions for distinguishing healthy and glaucomatous optic neuropathy (GON) images. In these images, bright pink regions indicate a large impact on model predictions while dark blue regions indicate a very limited impact on predictions. The maps were generated by averaging occlusion testing maps of the healthy and GON testing images in right eye orientation. The heat maps (bottom) are shown overlaid on a representative healthy and GON eye (top).

Similar articles

See all similar articles

Cited by 12 articles

See all "Cited by" articles

References

    1. Quigley H, Broman AT. The number of people with glaucoma worldwide in 2010 and 2020. British Journal of Ophthalmology. 2006;90:262–267. doi: 10.1136/bjo.2005.081224. - DOI - PMC - PubMed
    1. Tham YC, et al. Global prevalence of glaucoma and projections of glaucoma burden through 2040: A systematic review and meta-analysis. Ophthalmology. 2014;121:2081–2090. doi: 10.1016/j.ophtha.2014.05.013. - DOI - PubMed
    1. Schwartz JT, Reuling FH, Feinleib M. Size of the Physiologic Cup of the Optic Nerve Head: Hereditary and Environmental Factors. Arch. Ophthalmol. 1975;93:776–780. doi: 10.1001/archopht.1975.01010020670002. - DOI - PubMed
    1. Leske MC, Connell AMS, Schachat AP, Hyman L. The Barbados Eye Study. Arch. Ophthalmol. 1994;112:821. doi: 10.1001/archopht.1994.01090180121046. - DOI - PubMed
    1. Chua J, et al. Prevalence, risk factors, and visual features of undiagnosed glaucoma: The Singapore epidemiology of eye diseases study. JAMA Ophthalmol. 2015;133:938–946. doi: 10.1001/jamaophthalmol.2015.1478. - DOI - PubMed

Publication types

Feedback