Development and Evaluation of a Deep Learning System for Screening Real-world Multiple Abnormal Findings Based on Ultra-Widefield Fundus Images

Haodong Xiao¹Lie Ju^2,3Zupeng Lu⁴Shiguang Zhang⁵Yan Jiang⁶Yuan Yang¹Xuerui Zhang¹Wenting Zhang¹Huanyu Liu¹Tingyi Liang¹Jianing Ren¹Jiawei Yin¹Xiaoyu Liu⁷Ma Tong²Lin Wang²Wei Feng^2,3Kaimin Song²Yuzhong Chen²Zongyuan Ge²Qian Shao⁵Jie Peng^1*Jili Chen^6*Peiquan Zhao^1*

• ¹Department of Ophthalmology, Xin Hua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China
• ²Beijing Airdoc Technology Co., Ltd., Beijing, China
• ³Monash Medical AI Group, Monash University, Clayton, Australia
• ⁴Department of Ophthalmology, Shanghai Children’s Hospital, School of medicine, Shanghai Jiao Tong University, Shanghai, China
• ⁵Zhenjiang Ruikang Hospital, Zhenjiang, China
• ⁶Shibei Hospital Jingan District Shanghai, Shanghai, China
• ⁷Medical College of China Three Gorges University, Yichang, China

Purpose: To develop and evaluate a deep learning system for screening multiple abnormal findings including hemorrhages, drusen, hard exudates, cotton wool spots and retinal breaks using ultra-widefield fundus images.Methods: The system consisted of three modules: (I) quality assessment module, (II) artifact removal module and (III) lesion recognition module. In Module III, a heatmap was generated to highlight the lesion area. A total of 4521 UWF images were used for the training and internal validation of the DL system. The system was evaluated in 2 external validation datasets consisting of 344 images and 894 images from two other hospitals. The performance of the system in these 2 datasets was compared with or without Module II.Results: In both external validation datasets, the deep learning system made better performance when recognizing lesions on processed images after Module II than on original images without Module II. Module II-enhanced preprocessing improved Module III's five-lesion recognition performance by an average of 6.73% and 14.4% areas under the curves, 14.47% and 19.62% accuracy in the two external validations.Conclusions: Our system showed reliable performance for detecting MAF in real-world UWF images. For deep learning systems to recognize real-world images, the artifact removal module was indeed helpful.