Deep Learning-Based Automated Classification of Multi-Categorical Abnormalities From Optical Coherence Tomography...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Optical coherence tomography (OCT) is crucial for diagnosing retinal diseases.
• Accurate categorization of OCT images is essential for effective treatment.
• Current diagnostic methods can be time-consuming and require expert interpretation.

Purpose of the Study:

• To develop an intelligent system using deep learning for automated OCT image categorization.
• To evaluate the system's performance in differentiating various retinal conditions from OCT scans.
• To compare the system's diagnostic accuracy against human experts.

Main Methods:

• A dataset of 25,134 OCT images was utilized.
• One hundred one-layer convolutional neural networks (ResNet) were trained for categorization.
• 10-fold cross-validation was employed for algorithm optimization and performance evaluation.
• Key performance metrics included Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, and kappa values.

Main Results:

• The deep learning system achieved an AUC of 0.984 and an accuracy of 0.959.
• Specific disease detection accuracies included macular hole (0.978), cystoid macular edema (0.848), epiretinal membrane (0.957), and serous macular detachment (0.947).
• The system's kappa value (0.929) surpassed those of two independent human experts (0.882 and 0.889).

Conclusions:

• The developed deep learning system demonstrates high accuracy in automatically detecting and differentiating various OCT images.
• The system's performance is comparable to or better than that of human experts.
• This AI tool has significant potential to enhance the efficiency and impact of retinal disease diagnostics in clinical practice.