Fig. 2. Comparison of model with clinicians.

Our model identified three biomarkers that were annotated by clinicians. We present ROC (left column) and precision-recall (right column) curves for SLIVER-net and the baseline 3d CNN model along with individual annotator performance. For subretinal drusenoid deposits, SLIVER-net appears to outperform retina fellows in terms of both AUC and precision-recall, while the reverse is true for hyporeflective drusen cores.