Abstract
Purpose :
Retinopathy of prematurity (ROP) is a vasoproliferative disorder that frequently occurs in premature infants of low birth weight. ROP plus disease is characterized by abnormal dilation and tortuosity of the retinal arteries and veins. Plus disease needs immediate attention and treatment. Computer algorithms, including deep learning methods, have been developed to diagnose plus disease automatically. However, few of them took the pathological features as prior knowledge into consideration. The purpose of this study is to develop a fusion neural network, which can fully utilize the prior information, to diagnose the plus disease.
Methods :
We first constructed a basic U-net for the segmentation of the blood vessels and the optic disc. Then we automatically evaluated the four pathological features: tortuosity, vessel width, the fractal dimension of the vessel network, and the vessel density. These features were concatenated to the feature vectors of the network to constructed a fusion deep convolutional neural network.
Results :
We observed significant differences of the pathological features between the plus disease and the healthy condition. By fusing these features to the neural network, the performance of the network was significantly improved compared to the original network.
Conclusions :
We presented a new method for diagnosing ROP plus disease using a fusion deep convolutional neural network. It took into consideration the prior clinical knowledge. Our results demonstrated that the fusion network significantly outperformed the original network.
This is a 2020 ARVO Annual Meeting abstract.