Purpose: To compare the agreement between artificial intelligence (AI)-based classifiers and clinical experts in categorizing normal cornea from ectatic conditions.
Methods: Prospective diagnostic test study at Noor Eye Hospital. Two hundred twelve eyes of 212 patients were categorized into three groups of 92 normal, 52 subclinical keratoconus (SKCN), and 68 KCN eyes based on clinical findings by 3 independent expert examiners. All cases were then categorized using four different classifiers: Pentacam Belin/Ambrosio enhanced ectasia total deviation value (BADD) and Topographic Keratoconus Classification (TKC), Sirius Phoenix, and OPD-Scan III Corneal Navigator. The performance of classifiers and their agreement with expert opinion were investigated using the sensitivity, specificity, and Kappa index (κ).
Results: For detecting SKCN, Phoenix had the highest agreement with the clinical diagnosis (sensitivity, specificity, and κ of 84.62%, 90.0%, and 0.70, respectively) followed by BADD (55.56%, 86.08%, 0.42), TKC (26.92%, 97.50%, 0.30), and Corneal Navigator (30.77%, 93.75%, 0.29). For KCN diagnosis, the highest agreement with expert opinion was seen for Phoenix (80.02%, 96.60%, 0.79), BADD (95.59%, 85.42%, 0.75), TKC (95.59%, 84.03%, 0.73), and Corneal Navigator (67.65%, 96.45%, 0.68). Analysis of different classifiers showed that Phoenix had the highest accuracy for differentiating KCN (91.24%) and SKCN (88.68%) compared to other classifiers.
Conclusions: Although AI-based classifiers, especially Sirius Phoenix, can be very helpful in detecting early keratoconus, they cannot replace clinical experts' opinions, particularly for decision-making before refractive surgery. Albeit, there may be concerns about the accuracy of clinical experts as well.
Keywords: Keratoconus; artificial intelligence; corneal topography; subclinical keratoconus.