A Deep Learning Model Detects Glaucoma Based on an OCT Report, but Where Should the Clinician Look to Identify Glaucomatous Damage?

Purpose: To assess if the heat maps from a deep learning model (DLM) can guide clinicians to relevant locations on an optical coherence tomography (OCT) glaucoma report. Methods: Grad-CAM class activation ("heat") maps were produced by a DLM, with an OCT glaucoma report as input. The reports were generated for 183 eyes/individuals with optic neuropathy consistent with glaucoma, as indicated by arcuate defects ranging from subtle to advanced. The number of eyes in which a particular element of the report had the hottest region and the region that reached a criterion level were tabulated. A theoretical framework was employed to assess if these locations were clinically relevant. This framework employs two prototypical arcuate patterns of damage, one for each hemiretina. Results: In 86.3% of the 183 eyes, the hottest spot appeared in one of the four thickness or deviation maps that have been used for published OCT grading criteria. These same maps accounted for 81.2% of all the hot spots reaching a criterion level. While the hot spots typically overlapped the predicted prototype regions, they were often misleading. For example, 57% of the 183 eyes had the hottest spot, or one or more hot spots, located largely outside the prototypical regions. Conclusions: Although the DLM heat maps are flagging clinically meaningful elements of the report, in their existing form, they are not optimal for guiding the clinician to relevant locations of damage. Translational Relevance: The DLM can tag eyes with glaucomatous damage, and the prototypical patterns overlaid on the reports might aid the clinician in identifying and understanding this damage.