July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Utilizing Deep Learning to Improve Prediction Outcomes of Vision Symptoms
Author Affiliations & Notes
  • Guangming Dai
    Research and Development, Johnson & Johnson Vision, Milpitas, California, United States
  • Cynthia Gong
    Research and Development, Johnson & Johnson Vision, Milpitas, California, United States
  • Footnotes
    Commercial Relationships   Guangming Dai, Johnson and Johnson (E); Cynthia Gong, Johnson and Johnson (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 5062. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Guangming Dai, Cynthia Gong; Utilizing Deep Learning to Improve Prediction Outcomes of Vision Symptoms. Invest. Ophthalmol. Vis. Sci. 2019;60(9):5062.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : To create a deep learning model to predict the severity of vision symptoms that include glares, halos, ghosting and difficulty driving at night.

Methods : Vision symptoms are important factors to consider in refractive and cataract surgeries. It is important to be able to associate with or to predict from wavefront aberrations, such as Zernike coefficients. In this study, we collected vision symptoms and wavefront aberrations for 334 myopic eyes from a US IDE study. Data from pre-operative and 1M, 3M, 6M, 9M, and 12M post-operative visits were used, resulting in 1686 pairs of dataset for 5 mm pupil size and 1240 pairs of dataset for 6 mm pupil size. For the vision symptom scores, we scaled them from 1 (least) to 5 (most severe) to 0-1 (least to most severe) to optimize for the neural networks. We constructed a convolutional neural network (CNN) model using TensorFlow in Google colab platform. The tool we used was Keras Sequential, which uses a linear stack of layers that can be tuned. 80% of randomly selected dataset pairs were used for training and the rest for testing. For comparison purpose, we also ran a multivariate regression on the same dataset to obtain the statistics.

Results : For the five vision symptoms studied, day glare and ghosting appear to have the lower root mean square (RMS) error than night glare, halo, and difficulty driving at night. Night glare has the highest mean square error among the five symptoms, and it is consistent with both CNN and multivariate regression analysis. This conclusion is also true for both 5 mm pupil and 6 mm pupil sub-groups. However, 6 mm pupil sub-group shows a significantly lower RMS error than the 5 mm pupil sub-group, indicating that it is more useful to have larger area of the wavefront information to predict vision symptoms. Further studies may be useful to associate a particular vision symptom with one or more Zernike coefficients.

Conclusions : Preliminary results from a deep learning approach shows that a convolutional neural network can predict vision symptoms better than a multivariate regression.

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

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×