BACKGROUND: To explore the feasibility of artificial intelligence technology based on deep learning to automatically recognize the properties of vitreous opacities in ophthalmic ultrasound images.
METHODS: The normal and three typical vitreous opacities confirmed as physiological vitreous opacity (VO), asteroid hyalosis (AH) and vitreous hemorrhage (VH),were selected and marked from 2000 gray scale Color Doppler ultrasound images for each lesion. Five residual networks (ResNet) and two GoogLeNet models were trained to recognize the vitreous lesions. 75% images were randomly selected as the training set, the remaining 25% as a test set. The accuracy and parameters were recorded and compared among these seven different deep learning (DL) models. The precision, recall, FI score and the area under the receiver operating characteristic curves (AUC) values of recognizing the vitreous lesions were calculated with the most accurate DL model.
RESULTS: There were significant statistical differences in the accuracy and parameters among these seven DL models. GoogleNet inception V1 achieved the highest accuracy (95.5%) and the least parameters (10315580) in recognizing the vitreous lesions. GoogleNet inception V1 achieved 0.94, 0.94, 0.96, and 0.96 precision;0.94, 0.93, 0.97and 0.98 recall ;0.94, 0.93, 0.96 and 0.97 F1Score in recognizing normal, VO, AH, and VH. The AUC values of these four vitreous lesions were 0.99, 1.0, 0.99 and 0.99, respectively.
CONCLUSIONS: GoogLeNet inception V1 has shown promising results in recognizing the ophthalmic ultrasound image. With more and more ultrasound image data, a wide variety of hidden information in the eye diseases can be clearly detected automatically by the artificial intelligence technology based on deep learning.