Deep learning-based detection of retinal detachment with vitreous hemorrhage in ocular ultrasound images.
Authors
Affiliations (3)
Affiliations (3)
- Department of Ophthalmology, Faculty of Medicine, University of Miyazaki, 5200 Kihara, Kiyotake, Miyazaki, 889-1692, Japan.
- Electrical and Electronic Engineering Program, Faculty of Engineering, University of Miyazaki, 1-1 Gakuen Kibanadai-Nishi, Miyazaki, 889-2192, Japan. [email protected].
- Department of Ophthalmology, Faculty of Medicine, University of Miyazaki, 5200 Kihara, Kiyotake, Miyazaki, 889-1692, Japan. [email protected].
Abstract
Retinal detachment (RD) is a serious ocular disease that can lead to permanent vision loss. In cases with fundus-obscuring vitreous hemorrhage (VH), it is difficult to detect RD even using ocular ultrasonography. We developed a convolutional neural network (CNN) based on the You Only Look Once version 5 (YOLOv5) architecture to detect RD and VH on B-scan ultrasound images. The model was trained using 2,188 images and validated using 1,042 images. We applied image enhancement techniques, including unsharp masking (UM), to improve the detection accuracy. The final model (Incorporating fivefold cross-validation along with previous techniques) achieved overall accuracies of 96.6%, 99.2%, and 98.0% for RD, VH, and RD with VH, respectively. Our deep-learning algorithm showed high accuracy in detecting RD and VH on ocular ultrasound images. In cases with fundus-obscuring VH, our deep-learning algorithm might be useful for detecting RD as a supportive tool on ocular ultrasound images.