Development and Validation of a Deep Learning System for Multiple Ocular Fundus Diseases Using Retinal Images

Development and Validation of a Deep Learning System for Multiple Ocular Fundus Diseases Using Retinal Images: a Multi-center Prospective Study

Retinal images can reflect both fundus and systemic conditions (diabetes and cardiovascular disease) and firstly to be used for medical artificial intelligence (AI) algorithm training due to its advantages of clinical significance and easy to obtain. Here, the investigators developed a single network model that can mine the characteristics among multiple fundus diseases, which was trained by plenty of fundus images with one or several disease labels (if they have) in each of them. The model performance was compared with those of both native and international ophthalmologists. The model was further tested by datasets with different camera types and validated by three external datasets prospectively collected from the clinical sites where the model would be applied.

Inclusion Criteria:

• The quality of fundus images should clinical acceptable. More than 80% of the fundus image area including four main regions (optic disk, macular, upper and lower retinal vessel archs) are easy to read and discriminate.

Exclusion Criteria:

• Images with light leakage (>30% of area), spots from lens flares or stains, and overexposure were excluded from further analysis.