Early Detection of Keratoconus Using Corneal Tomography Image Analysis by GPT-5.4

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Early detection of keratoconus (KC) is crucial for timely intervention and preventing vision loss.
• Corneal tomography provides detailed structural information for KC diagnosis.
• Large language models (LLMs) are emerging as potential tools for medical image analysis.

Purpose of the Study:

• To assess the diagnostic performance of a general-purpose multimodal LLM (GPT-5.4) in identifying early stages of keratoconus.
• To evaluate the limitations of GPT-5.4 when analyzing corneal tomography images for keratoconus detection.
• To determine the accuracy, sensitivity, and specificity of the LLM in classifying normal, subclinical KC, and forme fruste KC (ffKC) eyes.

Main Methods:

• A retrospective analysis of 66 eyes (15 subclinical KC, 18 ffKC, 33 normal controls) using standardized four-map Pentacam corneal tomography images.
• GPT-5.4 was prompted to analyze images only, without clinical or demographic data.
• Diagnostic performance metrics including sensitivity, specificity, accuracy, and ROC analysis were calculated.

Main Results:

• GPT-5.4 achieved high specificity (100%) but low sensitivity (30.3%) for early KC detection, with an overall accuracy of 65.2%.
• The model exhibited a strong bias towards classifying eyes as normal, leading to a high false-negative rate.
• Performance varied by stage: excellent for normal corneas (100%), moderate for subclinical KC (53.3%), and poor for ffKC (11.1%), with most ffKC cases misclassified as normal.

Conclusions:

• GPT-5.4 demonstrates high specificity but insufficient sensitivity for early keratoconus detection using image-only input.
• The current version of GPT-5.4 is not a reliable tool for diagnosing or screening keratoconus.
• Further research is needed to improve LLM performance for ophthalmic image analysis.