Optical coherence tomography-based deep-learning model for detecting central serous chorioretinopathy.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Central serous chorioretinopathy (CSC) is a common retinal condition affecting the posterior pole.
• Accurate diagnosis and differentiation between acute and chronic CSC are crucial for effective management.
• Current diagnostic methods rely on expert interpretation of spectral domain optical coherence tomography (SD-OCT) images.

Purpose of the Study:

• To develop and evaluate a deep learning system for diagnosing CSC.
• To assess the model's ability to distinguish between acute and chronic CSC cases.
• To compare the diagnostic performance of the deep learning model against human experts and established AI models.

Main Methods:

• A convolutional neural network was trained using SD-OCT images from patients with CSC and a control group.
• The model's performance was evaluated using sensitivity, specificity, accuracy, and AUROC.
• Comparative analysis was performed against VGG-16, Resnet-50, and diagnoses from five ophthalmologists.

Main Results:

• For CSC diagnosis, the model achieved 93.8% accuracy, 90.0% sensitivity, and 99.1% specificity (AUROC 98.9%).
• For distinguishing acute from chronic CSC, the model achieved 97.6% accuracy, 100.0% sensitivity, and 92.6% specificity (AUROC 99.4%).
• The deep learning model's performance was comparable or superior to VGG-16, Resnet-50, and ophthalmologists' diagnoses.

Conclusions:

• The developed deep learning system demonstrates high accuracy in diagnosing CSC and differentiating between its acute and chronic forms.
• Automated AI algorithms show potential for independent diagnostic roles in CSC, complementing human expert capabilities.
• SD-OCT imaging combined with deep learning offers a powerful approach for objective and efficient CSC assessment.