Human expert grading versus automated quantification of fluid volumes in nAMD, DME and BRVO

This study compared an automated deep learning algorithm with certified human graders from the Vienna Reading Center (VRC) in identifying intra- (IRF) and subretinal fluid (SRF) in OCT scans of patients treated for neovascular age-related macular degeneration (nAMD), diabetic macular edema (DME) and branch retinal vein occlusion (BRVO). Multicenter clinical trial data from the VRC imaging database was used for this post hoc analysis. OCT scans were analyzed using a validated algorithm (RetInSight, Vienna, Austria) to compute IRF and SRF volumes. These fluid volumes were compared to fluid presence graded by trained and experienced graders of the VRC. 6898 OCT scans were analyzed for fluid volumes and presence of IRF and SRF. For nAMD/DME /BRVO in the central millimeter: the overall concordance for the detection of IRF and SRF between the algorithm and manual grading reached an AUC of 0.94/0.92/0.98 and 0.89/0.95/0.92, respectively. This deep learning approach showed a high concordance with human expert grading for detection of IRF and SRF and provides precise volumetric information across different retinal fluid-associated diseases. Thus, automated fluid quantification is a feasible tool for standardized treatment decision support and disease monitoring in clinical practice at the highest human expert level.