Automated Stage Analysis of Retinopathy of Prematurity Using Joint Segmentation and Multi-instance Learning

Abstract

Retinopathy of prematurity (ROP) is the primary cause of childhood blindness. Prior works have demonstrated the remarkable performances of deep learning (DL) in detecting plus disease and classification between ROP or Normal with retinal images. However, few studies are focused on identifying the “stage” of ROP disease, which is an important factor to evaluate the severity of the disease. In general, only a small region (typical less than 5% of the image) of a fundus image contributes its being classified as different stages of ROP. Therefore, traditional convolutional neural network (CNN) classifier may be ineffective when it is applied to a global feature extraction while the ROP features are localized with a limited number of labeled images. To address this issue, we combine the segmentation and staging, using both fully convolutional network (FCN) and multi-instance learning (MIL) to achieve integrated task of ROP staging and lesions localization. The proposed network is evaluated on 7330 retinal images (2000 Normal, 630 Stage1, 980 Stage2, 870 Stage3 and 250 Stage4) obtained by RetCam3. Experimental results show that the proposed network achieves 0.93 area under the curve (AUC) on the test dataset (accuracy 92.25%, sensitivity 90.53% and specificity 92.35%), and ROP lesions such as demarcation lines, ridges can be accurately located in the fundus images.

This work was supported partly by National Natural Science Foundation of China (Nos. 61871274, 61801305 and 81571758), National Natural Science Foundation of Guangdong Province (No. 2017A030313377), Guangdong Pearl River Talents Plan (2016ZT06S220), Shenzhen Peacock Plan (Nos. KQTD2016053112051497 and KQTD2015033016 104926), and Shenzhen Key Basic Research Project (Nos. JCYJ20170413152804728, JCYJ20170817112542555, JCYJ20180507184647636, JCYJ20170818142347251 and JCYJ20170818094109846).