Detecting glaucoma from multi-modal data using probabilistic deep learning.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma diagnosis relies on expert interpretation of fundus photographs and visual field tests.
• Deep learning models show promise in automating diagnostic tasks.
• Probabilistic models offer an advantage by quantifying diagnostic uncertainty.

Purpose of the Study:

• To evaluate the accuracy of probabilistic deep learning models in distinguishing normal eyes from those with glaucoma.
• To compare the performance of models using fundus photographs, visual fields, and combined data.
• To assess the utility of probabilistic models in providing diagnostic confidence levels.

Main Methods:

• Developed three probabilistic deep convolutional neural network (CNN) models using fundus photographs and visual field data.
• Trained and tested models on a dataset of 1,655 eyes and validated on an independent set of 196 eyes.
• Compared the performance of probabilistic models against deterministic CNN models.

Main Results:

• Probabilistic models achieved high accuracy, with Area Under the Curve (AUC) values up to 0.98 for combined modalities on the validation dataset.
• Combined fundus photograph and visual field models demonstrated superior performance, especially in early glaucoma detection (AUC 0.91).
• Misclassified eyes showed higher uncertainty, and probabilistic models provided a quantifiable certainty level.

Conclusions:

• Probabilistic deep learning models accurately detect glaucoma using multi-modal data.
• Combining fundus photographs and visual fields enhances diagnostic accuracy.
• Probabilistic models offer superior decision-making support by quantifying diagnostic certainty.