Enhancing the Automated Detection of Implantable Collamer Lens Vault Using Generative Adversarial Networks and...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate Implantable Collamer Lens (ICL) vault measurement is crucial for postoperative outcomes.
• Anterior segment optical coherence tomography (AS-OCT) is a key imaging modality for ICL assessment.
• Automating ICL vault estimation can improve efficiency and consistency in clinical practice.

Purpose of the Study:

• To evaluate the effectiveness of Generative Adversarial Networks (GANs) and synthetic data in enhancing a convolutional neural network (CNN) for automated ICL vault estimation.
• To improve the performance of CNN models for ICL vault measurement using AS-OCT images.

Main Methods:

• Retrospective analysis utilizing a deep learning framework with both real and synthetic AS-OCT images.
• Generation of approximately 100,000 synthetic ICL AS-OCT scans using GANs and image editing algorithms.
• Training a CNN on a dataset including real patient scans and synthetic images, followed by evaluation using metrics like MAPE, MAE, and RMSE.

Main Results:

• CNN trained solely on real images achieved MAPE of 15.31%, MAE of 44.68 µm, and RMSE of 63.3 µm.
• Inclusion of GAN-generated and edited synthetic images significantly improved performance: MAPE reduced to 8.09%, MAE to 24.83 µm, and RMSE to 32.26 µm.
• High correlation (R² = +0.98) between actual and predicted ICL vault distances was observed, with no significant difference in measured vs. predicted values (P = .58).

Conclusions:

• Integration of GAN-generated and synthetic images substantially enhances the accuracy of ICL vault estimation.
• GANs and synthetic data are effective in improving OCT image analysis, particularly in scenarios with limited real-world data.
• The developed model shows potential for assisting postoperative ICL evaluations and advancing OCT automation.