Interpretable longitudinal glaucoma visual field estimation deep learning system from fundus images and clinical...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Visual field (VF) testing is crucial for glaucoma assessment but is time-consuming and requires patient cooperation.
• Color fundus photographs (CFPs) offer an accessible alternative for monitoring disease progression.

Purpose of the Study:

• To develop a multi-modal longitudinal estimation deep learning (MLEDL) system.
• To predict current and future visual field (VF) status using CFPs and clinical text.
• To enhance the efficiency and reliability of glaucoma patient monitoring.

Main Methods:

• Developed an MLEDL system utilizing cross-sectional and longitudinal datasets.
• Trained and validated the model on extensive patient records (1598 cross-sectional, 3278 longitudinal, 446 external test records).
• Employed VF grading methods for clinical validation and generated heatmaps to visualize disease-related damage.

Main Results:

• The MLEDL system demonstrated accurate VF prediction capabilities.
• Pointwise mean absolute errors across models ranged from 3.098 to 4.131 dB.
• Heatmaps effectively illustrated the correlation between fundus damage and vision loss, confirming clinical reliability.

Conclusions:

• The MLEDL system effectively predicts visual fields from CFPs and clinical narratives.
• This technology offers a potential tool for long-term glaucoma assessment and improves clinical efficiency.
• Facilitates non-invasive, reliable monitoring of glaucoma progression over time.