Comparison of a human versus deep learning-based evaluation of fluid change in real-world OCT images of neovascular AMD

Abstract

Purpose : To compare the clinical evaluation of intra- and subretinal fluid (IRF, SRF) presence and volume change between eye care professionals and a deep learning-based algorithm in real-world OCT images of patients with neovascular age-related macular degeneration (nAMD).

Methods : Patients were diagnosed with nAMD and treated at our retina department between 2007 and 2018 and are included in the Vienna Imaging Biomarker Eye Study (VIBES) registry. Five retinologists (RET), three ophthalmology residents (RES), three ophthalmologists working in private practice (PRIVO), three orthoptists (ORTH) and three certified readers from the Vienna Reading Center (VRC) subjectively graded the presence of IRF/SRF at two consecutive visits and the level of change (increase/no-change/decrease). There were 28 to 120 days between two visits and anti-VEGF injections were given at any timepoint between the two visits. For the presence of IRF/SRF, the majority vote of RET was used as ground truth and compared to the majority vote of each of the other four professional groups as well as to an algorithm for automated fluid quantification (fluid presence ≥5nl). For the comparison of fluid change (change ≥ ±5nl), the same algorithm served as the objective ground truth and was compared to all five professional groups.

Results : Spectralis OCT volumes of 248 visits (=124 visit pairs) were included in our analysis. Compared to RET, the agreement on IRF presence was 93% (RES), 93% (PRIVO), 92% (ORTH), 88% (VRC), 89% (algorithm); agreement on SRF presence was 92% (RES), 89% (PRIVO), 89% (ORTH), 94% (VRC), 90% (algorithm). Compared to the algorithm, agreement on IRF change was 84% (RET), 87% (RES), 85% (PRIVO), 80% (ORTH), 87% (VRC); agreement on SRF change was 87% (RET), 87% (RES), 85% (PRIVO), 83% (ORTH), 88% (VRC).

Conclusions : The consensus of human experts from different professional background appears to be consistent in respect to overall fluid evaluation. However, in quantitative terms, artificial intelligence allows a precise assessment of fluid change over time and thus a management of macular edema that is better or comparable to different human eye-care professionals.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.