June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Visual Field Prediction using a Deep Bidirectional Gated Recurrent Unit Network Model
Author Affiliations & Notes
  • Hwayeong Kim
    Ophthalmology, Pusan National University Hospital, Busan, Korea (the Republic of)
  • Jiwoong Lee
    Ophthalmology, Pusan National University Hospital, Busan, Korea (the Republic of)
  • Sang Woo Moon
    Ophthalmology, Pusan National University Hospital, Busan, Korea (the Republic of)
  • Eun Ah Kim
    Ophthalmology, Pusan National University Hospital, Busan, Korea (the Republic of)
  • Sung Hyun Jo
    Ophthalmology, Pusan National University Hospital, Busan, Korea (the Republic of)
  • Keunheung Park
    Busan Medical Center, Busan, Busan, Korea (the Republic of)
  • Jeong Rye Park
    Kyungpook National University, Daegu, Daegu, Korea (the Republic of)
  • Sangil Kim
    Pusan National University, Kumjeong-ku, Korea (the Republic of)
  • Taehyeong Kim
    Pusan National University, Kumjeong-ku, Korea (the Republic of)
  • Sang wook Jin
    Dong-A University Medical Center, Busan, Busan, Korea (the Republic of)
  • Jung Lim Kim
    Inje University Busan Paik Hospital, Busan, Busan, Korea (the Republic of)
  • Jonghoon Shin
    Pusan National University Yangsan Hospital, Yangsan, Korea (the Republic of)
  • Seung Uk Lee
    Kosin University Gospel Hospital, Busan, Busan, Korea (the Republic of)
  • Geunsoo Jang
    Pusan National University, Kumjeong-ku, Korea (the Republic of)
  • Yuanmeng Hu
    Pusan National University, Kumjeong-ku, Korea (the Republic of)
  • Footnotes
    Commercial Relationships   Hwayeong Kim None; Jiwoong Lee None; Sang Woo Moon None; Eun Ah Kim None; Sung Hyun Jo None; Keunheung Park None; Jeong Rye Park None; Sangil Kim None; Taehyeong Kim None; Sang wook Jin None; Jung Lim Kim None; Jonghoon Shin None; Seung Uk Lee None; Geunsoo Jang None; Yuanmeng Hu None
  • Footnotes
    Support  Medical Big Data and AI-based early detection of visual dysfunction funded by Busan and managed by Busan Techno Park, and by the Patient-Centered Clinical Research Coordinating Center, funded by the Ministry of Health & Welfare, Republic of Korea (grant no. HI19C0481 and HC19C0276), and by a National Research Foundation (NRF) of Korea grant, funded by the Korean government (NRF-2021R1I1A1A01057767, NRF-2021R1A2B5B03087097, NRF-2017R1A5A1015722M, and NRF-2022R1A5A1033624).
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 356. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Hwayeong Kim, Jiwoong Lee, Sang Woo Moon, Eun Ah Kim, Sung Hyun Jo, Keunheung Park, Jeong Rye Park, Sangil Kim, Taehyeong Kim, Sang wook Jin, Jung Lim Kim, Jonghoon Shin, Seung Uk Lee, Geunsoo Jang, Yuanmeng Hu; Visual Field Prediction using a Deep Bidirectional Gated Recurrent Unit Network Model. Invest. Ophthalmol. Vis. Sci. 2023;64(8):356.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Although deep learning architecture have been applied to deal with sequential data, only a few studies have attempted to detect glaucoma progression using deep learning algorithms.
In this study, we proposed a bidirectional gated recurrent unit (bi-GRU) algorithm to predict future visual field examinations.

Methods : A total of 5,413 eyes from 3,321 patients were used as a train dataset and 1,272 eyes from 1,272 patients were used as a test dataset. Five consecutive visual field examinations were provided as input and 6th visual field examination was compared with the prediction of the bi-GRU. The performance of the bi-GRU was evaluated in comparison with the conventional linear regression (LR) and long short-term memory (LSTM) algorithms.

Results : The prediction performance of bi-GRU was better than those of LR and LSTM. The root mean square error (RMSE) of bi-GRU was 3.71 ± 2.42 dB and that of LR and LSTM was 4.81 ± 3.89 dB and 4.06 ± 2.61 dB, respectively. There were statistically significant differences in the prediction errors among the three models (F = 42.94, P < 0.001). The RMSE of bi-GRU was significantly lower than that of the other two models (both P < 0.001).
In the pointwise prediction, bi-GRU exhibited the lowest prediction error among the three models. Bi-GRU showed significantly better performance at 29 points and 49 points compared to LR and LSTM, respectively (all Ps < 0.001).
Furthermore, bi-GRU was the least affected by worsening of reliability indices and glaucoma severity. The prediction error of bi-GRU was significantly lower in all ranges of false positive rate, false negative rate, and fixation loss percentage compared with the other two models (P ≤ 0.025). As mean deviation decreased, the prediction performance decreased in all three models, but the RMSE in bi-GRU was still the lowest among the three models (P < 0.001).

Conclusions : A deep learning architecture utilizing the bi-GRU algorithm can predict future visual field examinations significantly better than the pointwise LR and LSTM algorithms.
Accurately predicting future visual field examination with the bi-GRU algorithm can help clinicians in treatment decision making.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

(a) The architecture of the LSTM-proposed method.
(b) The architecture of the bi-GRU proposed method.

(a) The architecture of the LSTM-proposed method.
(b) The architecture of the bi-GRU proposed method.

 

Representative examples of visual field prediction according to MD of the first visual field examination.

Representative examples of visual field prediction according to MD of the first visual field examination.

×
×

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.

×