A Three-Dimensional Convolutional Neural Network Model for Estimating the Glaucomatous Visual Field from Optical Coherence Tomography
DOI:
https://doi.org/10.51094/jxiv.170Keywords:
optical coherence tomography, glaucoma, visual field, Deep learningAbstract
Purpose: To estimate glaucomatous visual field (VF) from optical coherence tomography (OCT) images more accurately, we trained a segmentation-free, three-dimensional convolutional neural network (3DCNN) model instead of using the conventional two-dimensional model, which makes it difficult to increase the number of cases.
Methods: This was a retrospective, single-clinic all-case survey. We investigated 3,416 consecutive patients (6,356 eyes) who underwent OCT and Humphrey Field Analyzer (HFA) 24-2 or 10-2 tests, including patients with all types of glaucoma or suspected glaucoma. The supervisory signals to be learned—those of the VF thresholds and their slopes—were calculated from the pointwise linear regression of the measured VF thresholds. The training and test subjects were separated and divided into 4 groups according to the presence of a low signal strength index (SSI) and intraocular disease at training by fivefold cross-validation using the 3DCNN model. Then, tenfold cross-validation was performed under the training conditions that yielded the best test results.
Results: The best estimation performance was obtained by the fivefold cross-validation model trained, including all SSIs and intraocular diseases. The root mean square error per paired data for glaucoma patients was [mean ± standard deviation] 3.27 ± 2.20 dB for the HFA24-2 values and 3.02 ± 2.28 dB for the HFA10-2 values by the tenfold cross-validation model.
Conclusions: The model achieved high VF estimation accuracy. Future multicenter studies are expected to easily increase caseloads.
Translational Relevance: This model can reduce the burden by estimating the VF with high accuracy from OCT images.
Downloads *Displays the aggregated results up to the previous day.
References
Weinreb, R. N. et al. Primary open-angle glaucoma. Nature Reviews Disease Primers vol. 2 Preprint at https://doi.org/10.1038/nrdp.2016.67 (2016).
Allison, K., Patel, D. & Alabi, O. Epidemiology of Glaucoma: The Past, Present, and Predictions for the Future. Cureus (2020) doi:10.7759/cureus.11686.
Altangerel, U., Spaeth, G. L. & Rhee, D. J. Visual function, disability, and psychological impact of glaucoma. Curr Opin Ophthalmol vol. 14 (2003).
Wang, M. et al. Characterization of Central Visual Field Loss in End-stage Glaucoma by Unsupervised Artificial Intelligence. JAMA Ophthalmol 138, (2020).
Blumberg, D. M. et al. Association between undetected 10-2 visual field damage and vision-related quality of life in patients with glaucoma. JAMA Ophthalmol 135, (2017).
Bansal, M., Arora, T. & Dada, T. Effect of a novel computer software simulating humphrey visual field (HVF) on patient performance of HVF. Invest Ophthalmol Vis Sci 55, (2014).
Hashimoto, Y. et al. Predicting 10-2 Visual Field From Optical Coherence Tomography in Glaucoma Using Deep Learning Corrected With 24-2/30-2 Visual Field. Transl Vis Sci Technol 10, (2021).
Asaoka, R. et al. A Joint Multitask Learning Model for Cross-sectional and Longitudinal Predictions of Visual Field Using OCT. Ophthalmology Science 1, 100055 (2021).
Cirafici, P. et al. Point-wise correlations between 10-2 Humphrey visual field and OCT data in open angle glaucoma. Eye (Basingstoke) 35, 868–876 (2021).
Shin, J., Kim, S., Kim, J. & Park, K. Visual field inference from optical coherence tomography using deep learning algorithms: A comparison between devices. Transl Vis Sci Technol 10, (2021).
Sugiura, H. et al. Estimating glaucomatous visual sensitivity from retinal thickness with pattern-based regularization and visualization. in Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 783–792 (Association for Computing Machinery, 2018). doi:10.1145/3219819.3219866.
Xu, L. et al. Improving Visual Field Trend Analysis with OCT and Deeply Regularized Latent-Space Linear Regression. Ophthalmol Glaucoma 4, (2021).
Xu, L. et al. PAMI: A Computational Module for Joint Estimation and Progression Prediction of Glaucoma. in Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 3826–3834 (Association for Computing Machinery, 2021). doi:10.1145/3447548.3467195.
Xu, L. et al. Predicting the Glaucomatous Central 10-Degree Visual Field From Optical Coherence Tomography Using Deep Learning and Tensor Regression. Am J Ophthalmol 218, (2020).
Hashimoto, Y. et al. Deep learning model to predict visual field in central 10° from optical coherence tomography measurement in glaucoma. British Journal of Ophthalmology 105, (2021).
Asano, S. et al. Predicting the central 10 degrees visual field in glaucoma by applying a deep learning algorithm to optical coherence tomography images. Sci Rep 11, (2021).
Çallı, E., Sogancioglu, E., van Ginneken, B., van Leeuwen, K. G. & Murphy, K. Deep learning for chest X-ray analysis: A survey. Medical Image Analysis vol. 72 Preprint at https://doi.org/10.1016/j.media.2021.102125 (2021).
Harwerth, R. S., Wheat, J. L., Fredette, M. J. & Anderson, D. R. Linking structure and function in glaucoma. Progress in Retinal and Eye Research vol. 29 Preprint at https://doi.org/10.1016/j.preteyeres.2010.02.001 (2010).
Garcia-Medina, J. J., Rotolo, M., Rubio-Velazquez, E., Pinazo-Duran, M. D. & Del-Rio-vellosillo, M. Macular structure–function relationships of all retinal layers in primary open-angle glaucoma assessed by microperimetry and 8 × 8 posterior pole analysis of oct. J Clin Med 10, (2021).
Akashi, A. et al. The ability of SD-OCT to differentiate early glaucoma with high myopia from highly myopic controls and nonhighly myopic controls. Invest Ophthalmol Vis Sci 56, (2015).
Sezgin Akcay, B. I., Gunay, B. O., Kardes, E., Unlu, C. & Ergin, A. Evaluation of the Ganglion Cell Complex and Retinal Nerve Fiber Layer in Low, Moderate, and High Myopia: A Study by RTVue Spectral Domain Optical Coherence Tomography. Semin Ophthalmol 32, (2017).
Anderson, A. J. Significant glaucomatous visual field progression in the first two years: What does it mean? Transl Vis Sci Technol 5, (2016).
Murata, H., Araie, M. & Asaoka, R. A new approach to measure visual field progression in glaucoma patients using variational bayes linear regression. Invest Ophthalmol Vis Sci 55, (2014).
Yu, S. et al. Comparison of SITA Faster 24-2C test times to legacy SITA tests. Invest Ophthalmol Vis Sci 60, (2019).
L.J., S., R.A., R. & D.P., C. Standard or Fast?-Differences in precision between SITA Standard and SITA Fast testing algorithms and their utility for detecting visual field deterioration. Invest Ophthalmol Vis Sci 55, (2014).
Saunders, L. J., Russell, R. A. & Crabb, D. P. Measurement precision in a series of visual fields acquired by the standard and fast versions of the swedish interactive thresholding algorithm analysis of large-scale data from clinics. JAMA Ophthalmol 133, (2015).
Anderson DR, P. V. Automated Static Perimetry, 2nd edtion. (Mosby, St. Louis, 1999).
Nakatani, Y., Higashide, T., Ohkubo, S., Takeda, H. & Sugiyama, K. Effect of cataract and its removal on ganglion cell complex thickness and peripapillary retinal nerve fiber layer thickness measurements by fourier-domain optical coherence tomography. J Glaucoma 22, 447–455 (2013).
Kim, M., Eom, Y., Song, J. S. & Kim, H. M. Effect of Cataract Grade according to Wide-Field Fundus Images on Measurement of Macular Thickness in Cataract Patients. Korean Journal of Ophthalmology 32, (2018).
shijianjian. EfficientNet-PyTorch-3D. github.com. Updated July 20, 2021. Accessed May 10, 2022. https://github.com/shijianjian/EfficientNet-PyTorch-3D (2021).
Tan, M. & Le, Q. v. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv preprint:1905.11946 (2019).
Kingma, D. P. & Ba, J. Adam: A Method for Stochastic Optimization. (2014).
Konar, J., Khandelwal, P. & Tripathi, R. Comparison of Various Learning Rate Scheduling Techniques on Convolutional Neural Network. in 2020 IEEE International Students’ Conference on Electrical, Electronics and Computer Science, SCEECS 2020 (Institute of Electrical and Electronics Engineers Inc., 2020). doi:10.1109/SCEECS48394.2020.94.
Liu, L. et al. On the Variance of the Adaptive Learning Rate and Beyond. (2019).
Qian, X. & Klabjan, D. The Impact of the Mini-batch Size on the Variance of Gradients in Stochastic Gradient Descent. (2020).
Wu, H. & Gu, X. Towards dropout training for convolutional neural networks. Neural Networks 71, (2015).
Loshchilov, I. & Hutter, F. Decoupled Weight Decay Regularization. (2017).
Zhuang, J. et al. AdaBelief optimizer: adapting stepsizes by the belief in observed gradients. NIPS’20: Proceedings of the 34th International Conference on Neural Information Processing Systems 18795–18806 (2020).
Qian, N. On the momentum term in gradient descent learning algorithms. https://www.sciencedirect.com/science/article/abs/pii/S0893608098001166 (1999) doi:https://doi.org/10.1016/S0893-6080(98)00116-6.
Foret, P., Kleiner, A., Mobahi, H. & Neyshabur, B. Sharpness-Aware Minimization for Efficiently Improving Generalization. (2020).
Kwon, J., Kim, J., Park, H. & Choi, I. K. ASAM: Adaptive Sharpness-Aware Minimization for Scale-Invariant Learning of Deep Neural Networks. (2021).
Xie, Q., Luong, M.-T., Hovy, E. & Le, Q. v. Self-training with Noisy Student improves ImageNet classification. (2019).
He, K., Zhang, X., Ren, S. & Sun, J. Identity Mappings in Deep Residual Networks. (2016).
He, K., Zhang, X., Ren, S. & Sun, J. Deep Residual Learning for Image Recognition. (2015).
Downloads
Posted
Submitted: 2022-09-20 04:54:45 UTC
Published: 2022-09-27 00:12:15 UTC — Updated on 2022-11-10 23:40:10 UTC
Versions
Reason(s) for revision
A flow of participants was added. Table 2 was corrected by recounting the diseases in Table 2 because it included some cases before the 33% cutoff for HFA and some cases in Table 1. In addition, the overall text has been carefully reorganized and redescribed because the explanation was insufficient.License
Copyright (c) 2022
Makoto Koyama
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
