Abstract
Purpose :
To evaluate the performance of an automated deep learning algorithm in identifying retinal fluids in OCT scans of patients treated for neovascular age-related macular degeneration (nAMD).
Methods :
OCT scans acquired with routinely-used devices - Cirrus-Zeiss Meditec, Spectralis-Heidelberg Engineering and Topcon, were analyzed using a previously described deep learning algorithm in order to derive intraretinal fluid (IRF) and subretinal fluid (SRF) volumes, and hence fluid presence, in the 1, 3 and 6mm areas within the macula. These volumes were compared to expert grading of fluid status (presence/absence) from a centralized reading center using the area under the curve (AUC) of receiver operating characteristic as a measure of concordance.
Results :
19,034 OCT scans were analyzed from monthly visits over 2 years in HARRIER. In the central 1 mm, the concordance for the detection of IRF and SRF between the algorithm and grading reached an AUC of 0.93 and 0.87, respectively. In the central 6 mm, the concordance for both IRF and SRF had an AUC of 0.90. Regarding IRF in the central 1mm, 18% of scans graded with no fluid were found to have fluid by the algorithm, while 8% of scans graded to have IRF had no fluid identified by the algorithm. Regarding SRF in the central 1mm, 23% of scans graded with no fluid were found to have fluid by the algorithm, while 15% of scans graded to have SRF had no fluid identified by the algorithm. Across the fluid types, the highest concordance was achieved on scans from Spectralis (AUC=0.92), while Topcon and Cirrus had a slightly lower concordance (AUC=0.89 and 0.86). Results from HAWK were analysed in the same manner.
Conclusions :
Identification of IRF and SRF from deep learning-based segmentation of retinal fluid showed high concordance with expert grading. Furthermore, automated quantification of nAMD features across the entire macula, beyond retinal thickness measurements, illustrates the potential of such approaches to objectively analyze OCT scans in order to support the treatment decision in nAMD patients in clinical practice.
This is a 2020 ARVO Annual Meeting abstract.