Deep Learning-Based Diagnosis of Corneal Condition by Using Raw Optical Coherence Tomography Data

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Keratoconus (KC) is a progressive corneal ectasia impacting vision, necessitating early diagnosis for effective treatment.
• Raw medical data offer consistency across software versions, unlike preprocessed data which can be affected by updates.
• Early and stage-related diagnosis of KC is crucial for preserving visual acuity.

Purpose of the Study:

• To differentiate between healthy and keratoconus (KC) corneas using raw optical coherence tomography (OCT) data.
• To evaluate the efficacy of modified convolutional neural networks (CNNs) in analyzing raw OCT images for KC detection.

Main Methods:

• Utilized 2737 eye examinations from Casia2 anterior-segment OCT, classified by ophthalmologists into 'normal', 'ectasia', or 'other disease'.
• Employed modified DenseNet121, EfficientNet-B0, MobileNetV3-Large, and ResNet18 CNN models for image analysis.
• Focused on analyzing raw OCT data without preprocessing to maintain data integrity.

Main Results:

• Modified EfficientNet-B0 achieved the highest overall accuracy at 92.86%.
• CNN models demonstrated high performance across metrics, including macro-averaged sensitivity, specificity, Positive Predictive Value, and F1 scores.
• The models achieved accuracies ranging from 89.68% to 92.86%.

Conclusions:

• Convolutional neural networks can effectively diagnose keratoconus using raw optical coherence tomography data.
• Raw OCT data represent a viable and consistent alternative to preprocessed data for clinical analysis in ophthalmology.
• This approach holds significant potential for improving the early diagnosis of KC and guiding timely treatment decisions.