Deep Learning Model to Predict Postoperative Visual Acuity from Preoperative Multimedia Ophthalmic Data

Ryo Otsuki, Osamu Sugiyama, Yuki Mori, Masahiro Miyake, Shusuke Hiragi, Goshiro Yamamoto, Luciano Santos, Yuta Nakanishi, Yoshikatsu Hosoda, Hiroshi Tamura, Shigemi Matsumoto, Akitaka Tsujikawa, Tomohiro Kuroda
Vol. 9 (2020) p.241-248

Age-related macular degeneration (AMD) causes visual acuity (VA) loss in people aged ≥ 50 years. Common treatments include intravitreal injection of anti-vascular endothelial growth factor agents such as aflibercept. However, lack of response in some patients makes prediction of posttreatment VA difficult. In this paper, we propose a deep neural network model to predict posttreatment VA using pretreatment medical imaging and patient profile data. The proposed model works with image data (optical coherence tomography and color fundus photograph) and patient profile data including gender, age, affected side and pretreatment decimal visual acuity. The model was tested by comparing mean square errors (MSE) between actual and predicted visual acuity obtained from input of image data alone, input of patient profile data alone, and input of both types of data. When examining the concatenation effectiveness of input of both types of data, the outcomes of concatenation conditions 100:100 and 500:500 were compared. For concatenation condition 100:100, MSE was 0.081 for input of image data alone, 0.052 for input of patient profile data alone, and 0.058 for input of both types of data. For concatenation condition 500:500, the MSE values were 0.081, 0.052, and 0.047, respectively. The model proposed provides highly accurate prediction of posttreatment VA and indication of recovery to physicians and patients. The method can handle incomplete images and patient profile data usually collected from patients before treatment.