Multimodal-Based Non-Contact High Intraocular Pressure Detection Method

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma, a leading cause of irreversible blindness, necessitates accurate intraocular pressure (IOP) monitoring.
• Traditional IOP measurements are susceptible to corneal biomechanical variations, impacting diagnostic reliability.

Purpose of the Study:

• To develop a non-contact, deep learning-based method for detecting elevated IOP by integrating Scheimpflug images and corneal biomechanical data.
• To improve the accuracy of high IOP prediction and glaucoma screening.

Main Methods:

• A multi-modal deep learning framework utilizing CycleGAN for data augmentation, Swin Transformer for visual feature extraction, and Kolmogorov-Arnold Network (KAN) for data fusion.
• Integration of Scheimpflug imaging with clinical parameters for comprehensive data analysis.
• Dataset augmentation to address data scarcity and class imbalance.

Main Results:

• The proposed model achieved a diagnostic accuracy of 0.91 on a real-world dataset, surpassing traditional methods.
• Grad-CAM visualizations highlighted key anatomical features, including corneal thickness and anterior chamber depth, related to IOP fluctuations.

Conclusions:

• The study demonstrates the efficacy of a deep learning approach combining ocular imaging and biomechanics for accurate IOP assessment.
• Findings emphasize the significance of corneal structure in IOP regulation and suggest potential for non-invasive, biomechanics-informed IOP screening.