| GL F-029 |
| A Novel Visual Field Prediction Using Recurrent Neural Network Architecture |
| Department of Ophthalmology, Pusan National University College of Medicine, Busan, Korea |
| Keunheung Park, Jiwoong Lee |
|
목적 : To develop a reliable visual field prediction algorithm using a state-of-art deep learning algorithm, recurrent neural network (RNN), and evaluate its performance compared with conventional pointwise ordinary linear regression (OLR) method.
방법 : An RNN utilizing long short-term memory (LSTM) cells was constructed. 5 visual field tests were provided to the RNN as input and the 6th visual field test was compared with output of RNN. A total of 1408 eyes were used to training dataset and another 281 eyes were used for the test dataset. Root mean square error (RMSE) of every prediction was calculated to compare overall prediction accuracy. Pointwise mean absolute error (MAE) was also calculated to evaluate the spatial distribution of prediction accuracy.
결과 : The prediction RMSE was 4.31 ± 2.54 dB / 4.96 ± 2.76 dB (RNN / OLR respectively) and this was significantly different (P < 0.001). Pointwise MAE of RNN was smaller than OLR in most areas and superotemporal, superonasal, inferotemporal, and inferonasal areas where is known to be vulnerable to glaucomatous damage were significantly different. Among the reliability indices of visual field, false negative rate (FNR) was significantly affected to the both RNN and OLR but RNN showed smaller and more slowly increasing RMSE as FNR increased.
결론 : The RNN predicted future visual field more accurately than conventional linear regression method. RNN was more robust to the worsening of reliability of visual field examination.
* This research was supported by the Bio & Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (No. NRF-2018M3A9E8066253)
|
| |