2021年度 人工知能学会全国大会(第35回)

講演情報

IEEE CYBCONF

IEEE CYBCONF » IEEE CYBCONF

[2M4-CC] Late Breaking Research Session - B

2021年6月9日(水) 15:20 〜 17:20 M会場 (CybConf会場)

Emi Yuda

17:00 〜 17:20

[2M4-CC-06] A research on Visual Outcome Prediction in Macular Hole Using Optical Coherence Tomography Images

Kazuya Imamura1, Hiroharu Kawanaka1, Yoko Mase2, Yoshitsugu Matsui2 (1. Mie University, 2. Mie University Hospital Ophthalmology)

In clinical Ophthalmology, it takes a lot of time to improve a patient’s vision when surgical treatment is required. Most patients usually worry about his/her prognosis, and enough clinical explanations are required. Currently, ophthalmologists estimate a patient’s prognosis with medi-cal reports; however, the decision-making is based on experience-based. From the viewpoint of informed consent, accurate prediction of a patient’s prognosis is essential. This paper aims to make a mathematical model that can accurately predict a patient’s visual outcome. We focus on Macular Hole (MH) and Optical Coherence Tomography (OCT) images obtained from MH pa-tients. MH is aretinal disease, and OCT is usually used to evaluate the progression of MH, be-cause there are significant differences in retinal morphology among healthy, pre and post-operation. Regression analysis is applied to discuss whether it works well or not. We col-lected OCT images from 54 patient and their clinical information as experimental materials. In this paper, 53 features were extracted from the pre-operative OCT images and 5 features from pa-tient records. After this, the methods for feature selection and prevention of multicollinearity were applied to the given data, and finally 5 features were employed as an explanatory variable. In re-gression analysis, Ordinary Least Squares (OLS) are employed. The constructed model showed Adjusted R2 of 0.545, MAE of 0.105, and RMSE of 0.130. Also, the percentage of errors within 0.1 was 79.6%. The obtained performance is not extremely low, but this model should be im-proved for practical use. Also, the given data was imbalanced, as a result the model did not work well in poor vision range. On the other hand, we obtained a new insight that the morphological and structural features of the retina were suggested would be effective for prediction.

講演PDFパスワード認証
論文PDFの閲覧にはログインが必要です。参加登録者の方は「参加者用ログイン」画面からログインしてください。あるいは論文PDF閲覧用のパスワードを以下にご入力ください。

パスワード