患者のOCT画像を入力とするCNNを用いた注射後の視力推定精度の検証

抄録

The effect of anti-VEGF therapy on macular edema due to Branch Retinal Vein Occlusion（BRVO）varies depending on patients and is therefore difficult to predict the prognosis in advance. In this paper, we present neural networks that predict LogMAR visual acuity scores improved by injecting Aflibercept from pre-treated OCT images of BRVO patients obtained before the injection. We tested two types of neural nets. The details are as follows. One neural net is a fine-tuned model whose input is only a vertical cross-sectional image of a fundus and another net is the unique CNN model with its own architecture whose input is two images : horizontal and vertical cross sections of a fundus. The training images and test images are taken using different kinds of OCT apparatuses. As a result, the fine-tuned model can predict LogMAR visual acuity scores within an error of 0.3 for 65% of the test images ? the unique CNN for 66%. These results demonstrate that both the presented nets can predict visual acuity scores even for unlearned OCT images with sufficiently high accuracy.