Artificial intelligence in ophthalmology - the value of artificial neural networks

In the past few years, ophthalmology has become a major research research focus within the field of artificial intelligence (AI). (AI), also known as artificial intelligence (AI). The underlying research field is extensive and deals with every section of the deals with every section of the eye.

The authors of the study "Deep Learning for Predicting Refractive Error From Retinal Fundus Images" by Avinash V. Varadarajan, Ryan Poplin, Katy Blumer et al (2018)¹ focus their research on refractive error. They evaluate the extent to which AI can be applied to infer refractive deficit based on fundus photographs and OCT images. As a basis, they use knowledge from previous studies that refractive error, particularly axial ametropia, is associated with characteristic changes in the fundus. To accomplish the study, they use 24,007 records from the Age-Related Eye Disease Study (AREDS) from the United States. The data  included subjective refraction values and retinal images (30° section of the macula) using fundus photography, from participants aged 55 to 80 years. A further 15,750 records were from the observational study, UK Biobank from the United Kingdom. These included objective refraction and OCT images (45° section of macula) with participants aged 40 to 69 years.

The data was used to build an artificial neural network. Deep learning is the process of learning the correct parameter values ("training") the AI so that it performs a specific task, for example, generating a prediction from the pixel values in a retinal image. Deep learning involves building an artificial neural network (algorithm). In this case, 10% of the nearly 40,000 data sets could be validated to train the AI to interpret the results. Through the attention technique used, new image features were visualized and identified by the AI. The remaining datasets were used for optimization and predictability testing.

The resulting algorithm provided the refractive deficit based on the spherical equivalent and OCT images with a mean absolute error of 0.56 dpt. The mean absolute error for the calculation based on fundus photography was higher and was 0.91 dpt. The attentional features used by the artificial neural networks for prediction were distributed over the entire 30°-45° section of the macula. However, the evaluation showed that the fovea was one of the most important areas used by the algorithm for this prediction.

The study group concludes that in the future, AI will have the ability to estimate refractive errors with high accuracy from imaging data. The deep learning model has shown high accuracy in predicting the spherical equivalent, but not in predicting the cylindrical value. This was already expected since astigmatism is the result of the toricity of the cornea and/or the crystalline lens. This information was not available to the model used.

In conclusion, AI is advancing in rapid steps in the medical sector and its potential is manifold. AI assists in the diagnosis of rare diseases by specifically differentiating dysfunctions or structural changes with very similar appearances. This allows early therapy to be initiated and, for the first time, classification of specific pathologies, as demonstrated by the examples of papilledema² and myopic macular degeneration³. Another field of research is the use of AI and Deep Learning in the automated recognition process. Applied applications include the detection of diabetic retinopathy and macular edema⁴ and keratoconus.⁵ As a predictive model, AI is used, for example, in myopia development in adolescents.⁶

The AI characterizes features that medical experts normally cannot extract from images alone, such as age, gender, blood pressure, and other cardiovascular health factors.⁷ Users cannot always determine on the basis of which manifestations the AI has recognized these correlations. Whether the refractive deficit calculated by the artificial neural network will one day be relevant in practice cannot be estimated today.

References

[1] Varadarajan, A. V., Poplin, R., Blumer, K., Angermueller, C., Ledsam, J., Chopra, R., Keane, P. A., Corrado, G. S., Peng, L., Webster, D. R. (2018).
Deep Learning for Predicting Refractive Error From Retinal Fundus Images. Invest. Ophthalmol. Vis. Sci., 59, 2861-2868

[2] Vasseneix, C., Najjar, R. P., Xu, X., Tang, Z., Loo, J.L., Singhal, S., Tow, S., Milea, L., Wei Sw Ting, D. Shu,Liu, Y., Wong, T. Y., Newman, N. J., Biousse, V., Milea, D. (2021). Accuracy of a Deep Learning System for Classification of Papilledema Severity on Ocular Fundus Photographs. Neurology.

[3] Flitcroft, D. I., He, M., Jonas, J. B., Jong, M., Naidoo, K., Ohno-Matsui, K., Rahi, J., Resnikoff, S., Vitale, S., Yannuzzi, L. (2019). IMI - Defining and Classifying Myopia: A Proposed Set of Standards for Clinical and Epidemiologic Studies. Invest. Ophthalmol. Vis. Sci., 60, M20-M30.

[4] Cheung, C. Y., Tang, F., Ting, D. S. W., Tan, G. S. W., Wong, T. Y. (2019). Artificial Intelligence in Diabetic Eye Disease Screening. Asia Pac. J. Ophthalmol. (Phila).

[5] Feng, R., Xu, Z., Zheng, X., Hu, H.,Jin, X., Chen, D. Ziyi,Yao, K., Wu, J. (2021). KerNet: A Novel Deep Learning Approach for Keratoconus and Sub-clinical Keratoconus Detection Based on Raw Data of the Pentacam System. IEEE J. Biomed. Health Inform.

[6] Yang, X., Chen, G., Qian, Y., Wang, Y., Zhai, Y., Fan, D.,Xu, Y. (2020). Prediction of Myopia in Adolescents through Machine Learning Methods. Int. J. Environ Res. Public Health, 17

[7] Poplin, R., Varadarajan, A. V., Blumer, K., Liu, Y., McConnell, M. V., Corrado, G. S., Peng, L.,Webster, D. R. (2018). Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat. Biomed. Eng.,158-164.