Deep Learning Versus Corneal Tomography Features to Detect Subclinical Corneal Edema in Fuchs Endothelial Corneal Dystrophy

Purpose: To compare a deep learning model with corneal tomography features for detecting subclinical corneal edema in patients with Fuchs endothelial corneal dystrophy (FECD). Methods: We trained a deep learning model to detect corneal edema on 379 optical coherence tomography B-scans of normal and edematous corneas. 51 eyes of 32 patients with FECD were analyzed and compared with 100 eyes of 50 normal patients. For each eye, the cornea was scanned on the same day using 2 modalities of the same swept-source optical coherence tomography device (Anterion): corneal tomography maps and 6 high-resolution radial B-scans. The 6 radial B-scans were analyzed using our model from which an en face map of edema was reconstructed. The location exhibiting the highest probability of edema was derived from that map. Two corneal surgeons assessed the tomography maps and labeled the location of the supposed highest edema. This location was compared with our model's en face map. Results: According to tomography features, 64.7% of eyes presented subclinical edema. Our model and tomography features agreed in 80% of cases for the presence or absence of subclinical edema. The average distance between the location of maximal edema determined by human experts on tomography maps and that derived from our model's en face map was 1.91 ± 1.15 mm. Our model detected superior peripheral edema in a majority of eyes. Conclusions: Despite being based on different approaches, both methods agreed in the detection of subclinical edema in most cases. The location of detected edema was very similar in both methods. In cases where both methods disagree, our approach provides new objective results that might help the surgeon in making a decision in difficult cases.