Automated detection of retinopathy of prematurity by deep attention network

• Published in: Multimedia tools and applications 2021-12, Vol.80 (30), p.36341-36360
• Main Authors: Lei, Baiying, Zeng, Xianlu, Huang, Shan, Zhang, Rugang, Chen, Guozhen, Zhao, Jinfeng, Wang, Tianfu, Wang, Jiantao, Zhang, Guoming
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Automation; Blindness; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Localization; Medical imaging; Multimedia Information Systems; Retinal images; Special Purpose and Application-Based Systems; Test sets
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-021-11208-0

Retinopathy of prematurity (ROP) is a retinal vascular proliferative disease principally observed in infants born prematurely with low birth weight. ROP is the leading cause of childhood blindness. Early screening and timely treatment are crucial in preventing ROP blindness. Previous ROP diagnosis lacks clear understanding of the underlying factors and properties that supports the final decision. For this reason, a deep convolutional neural network (DCNN) is developed for automated ROP detection using wide-angle retinal images. Specifically, we first choose ResNet50 as our base architecture and improve the ResNet by adding a channel and a spatial attention module. Then, we utilize a class-discriminative localization technique (i.e., gradient-weighted class activation mapping (Grad-CAM)) to visualize the trained models and realize pathological structure localization. The efficacy of the proposed network is evaluated on two test datasets. Our method obtains a sensitivity of 94.84   % and a specificity of 99.49   % on test set 1 while a sensitivity of 98.03   % and a specificity of 94.55   % on test set 2. Also, the model successfully detects the pathological structures of ROP (e.g., demarcation lines or ridges) in the retina images.