英文简介Deep Learning Colour Fundus Image Analysis for Cataract Grading

Why did we carry out this work?

· Current methods to triage non-urgent ophthalmic referrals are subjective and dependent on the quality of patient information provided in referral letters.

· Cataract diagnosis and grading require a slit lamp biomicroscope and proper training.   

· A deep learning system can serve as a clinical aid to 1) triage new patient referrals for non-urgent specialist ophthalmic care and 2) diagnose and grade cataracts through the objective analysis of retina photographs alone.

How did we do this?

· The first phase was training, tuning and internal validation of an automated deep learning system to triage of ophthalmic referrals and grade cataract severity using an existing dataset of 43,274 retina photographs.

· The second phase was a prospective comparison study for external validation of the deep learning system, recruited 2680 eyes from 1290 Chinese subjects, aged 50 and above, who were referred to the Ophthalmology specialist outpatient clinic at Grantham Hospital for blurring of vision for three or more months.

What did we find?

· Our deep learning system achieved high accuracy in the external validation test and moderate-to-substantial agreement with ophthalmologist assessment for triaging new non-urgent referrals and grading of cataracts.

· However, the deep learning model also generated 21.1% more early referrals than necessary