Predicting the results of visual field tests using deep learning

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K16903
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 56060:Ophthalmology-related
• Research Institution: Jichi Medical University
• Principal Investigator: Inoda Satoru 自治医科大学, 医学部, 講師 (60741098)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 眼科 / AI / 緑内障

Outline of Final Research Achievements

We trained the model using 871 HFA cases and their fundus color photographs taken within a year before and after the HFA, in 1:1 pairs. For the HFA results, we adopted the central 24-2 threshold map, and excluded data with an eye movement abnormality rate, false positive rate, or false negative rate of 1/3 or more. We used the ResNet18 network architecture for training, and predicted the HFA threshold map, MD value, and PSD value from the fundus images for evaluation.
The average root mean square error of the threshold map for each image was 6.82dB, and the average measured value of MD was -7.52±8.87dB, while the AI predicted value was -5.16±0.47dB. The average error between the measured and predicted values was 6.48±6.37dB.

Free Research Field

眼科

Academic Significance and Societal Importance of the Research Achievements

視野検査は煩雑で時間と患者さん個人の集中力が要求されるが、眼底写真は非侵襲的に短時間で患者さんの集中力や理解力によらず撮影が可能である。今回、眼底写真から視野検査結果が予測できた。まだ誤差は大きいが、一人の患者さんを経時的に予測する場合、その誤差が臨床上許容可能範囲まで学習精度を上昇させることができる可能性が高い。本研究は、失明原因第一位の緑内障の早期診断とその後の視野フォローに有用であることが示された