Anterior segment biometric measurements explain misclassifications by a deep learning classifier for detecting...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning classifiers analyze anterior segment optical coherence tomography (AS-OCT) images for gonioscopic angle closure detection.
• Misclassifications by these AI tools necessitate identifying underlying biometric factors.

Purpose of the Study:

• To identify biometric parameters that explain misclassifications by a deep learning classifier for detecting gonioscopic angle closure in AS-OCT images.

Main Methods:

• AS-OCT images from the Chinese American Eye Study (CHES) were analyzed using a deep learning classifier.
• Biometric parameters were compared across true positive, true negative, false positive, and false negative predictions.
• Logistic regression models were used to differentiate prediction errors.

Main Results:

• Parameter measurements, including iris curvature (IC), lens vault (LV), and angle opening distance (AOD), differed significantly between prediction classes (p<0.001).
• False positives (FP) showed characteristics of both true positives (TP) and true negatives (TN) regarding anterior segment and angle parameters.
• Models for FP and false negative (FN) detection improved overall classifier accuracy from 84.8% to 89.0%.

Conclusions:

• Disagreement between anterior segment and angle parameters explains misclassifications in OCT-based deep learning for angle closure detection.
• These findings can guide improvements in AI classifier performance.
• The study highlights discrepancies between gonioscopic and AS-OCT definitions of angle closure.