Evaluation of an aided diagnosis system for vitreous and retinal diseases by analyzing B-scan ultrasound images based on deep convolutional neural network / 中华实验眼科杂志

ABSTRACT

Objective:To explore the clinical value of a diagnostic system of ophthalmic B-scan ultrasound images based on deep convolutional neural network.Methods:A total of 3 600 ophthalmic B-scan ultrasound images of 1 278 patients with an average age of (49.32±7.69) years at the Eye Center of Renmin Hospital of Wuhan University from January 2018 to October 2020 were collected to build an image database.These B-scan images were labeled by three ophthalmologists.The database was divided into the training dataset of 2 812 images and the testing dataset of 788 images.The deep learning algorithm was used to build a diagnostic model, which can identify retinal detachment (RD), vitreous hemorrhage (VH) and posterior vitreous detachment (PVD), and the accuracy of the model was evaluated.Another 120 B-scan ultrasound images were collected for the human-computer comparison between the model and 3 senior ophthalmologists.Eight junior clinicians were selected to evaluate another 150 B-scan images with and without the assistance of the model, and the accuracy was analyzed to evaluate the effect of the model.This study adhered to the Declaration of Helsinki and the study protocol was approved by Renmin Hospital of Wuhan University (No.WDRY2020K-192).Results:The accuracy of the model for identifying normal fundus, RD, VH, PVD and other diseases were 0.954, 0.909, 0.881, 0.990 and 0.920, respectively.The accuracy of the model was similar to that of senior doctors, and the time the model used was almost half that of doctors.With the assistance of the model, the diagnostic accuracy of the 8 junior clinicians who participated in the training was significantly improved ( P<0.01). Conclusions:The accuracy of RD, VH and PVD identification of the intelligent evaluation system is good, and the system can improve the accuracy and efficiency of clinical examinations, and can better assist clinicians in clinical evaluation.