Ultra-wide-field fundus photography and AI-based screening and referral for multiple ocular fundus diseases.

Abstract

To address the difficulty in comprehensive screening of fundus diseases, we develop three deep learning algorithms (DLAs) based on different algorithms (Swin Transformer and cross-domain collaborative learning [CdCL]) and imaging modalities (ultra-wide-field [UWF] images and the cropped posterior-pole-region [PPR] images) to identify 25 fundus conditions and provide referral suggestions: WARM (CdCL + UWF images), BASE (Swin Transformer + UWF images), and WARM-PPR (CdCL + PPR images). 59,475 UWF images are included to establish internal and external datasets. WARM shows the best performance on the internal test (area under the receiver operating characteristic curve [AUC] for screening = 0.915; AUC for referral = 0.911) and the external multi-center test (AUC for screening = 0.912; AUC for referral = 0.902). UWF images and the CdCL approach significantly enhance the DLA's ability to detect abnormalities in the peripheral retina. The WARM model shows promise as a reliable and accurate tool for comprehensive fundus screening on a large scale.