This prospective study compared the performance of retinal specialists and an artificial intelligence (AI) algorithm in detecting retinal fluid from spectral-domain OCT scans.
Study design
Using spectral-domain OCT scans from 2 common devices, researchers compared retinal fluid grades from human retinal specialists against those from the artificial intelligence algorithm (Notal OCT Analyzer; NOA). They evaluated 1,127 eyes from 651 patients enrolled in the AREDS2 10-year follow-on study. Study investigators graded each scan for the presence/absence of intraretinal and subretinal fluid. The main outcome was accuracy; sensitivity, specificity, precision and F1 scores were also assessed.
Outcomes
Retinal fluid was present in 32.8%. For detecting retinal fluid, the investigators had an accuracy of 0.805, sensitivity of 0.468 and specificity of 0.970. The AI algorithm had an accuracy of 0.851, sensitivity of 0.822 and specificity of 0.865.
Investigators detected intraretinal fluid with an accuracy of 0.815, sensitivity of 0.403 and specificity of 0.978 while NOA metrics were 0.877, 0.763 and 0.922, respectively. For detecting subretinal fluid, the investigator had an accuracy of 0.946, sensitivity of 0.583 and specificity of 0.973; the NOA metrics were 0.863, 0.940 and 0.857, respectively.
Limitations
The 10-year AREDS2 study was not representative of real-world practice.
Clinical significance
This study revealed that retinal specialists had imperfect accuracy and low sensitivity in detecting retinal fluid, particularly for intraretinal fluid and difficult cases (e.g., lower fluid volumes appearing on fewer B-scans). Overall, AI-based detection achieved a higher level of accuracy—demonstrating that it could assist physicians in detecting retinal fluid.