Glaucoma detection: Binocular approach and clinical data in machine learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Glaucoma diagnosis traditionally relies on single-eye data.
• Automated diagnosis methods can be enhanced by integrating multi-modal data.
• Novel approaches are needed to improve early glaucoma detection accuracy.

Purpose of the Study:

• To develop and evaluate a multi-modal machine learning method for automated early glaucoma diagnosis.
• To investigate the benefits of using simultaneous data from both eyes (binocular mode) versus single-eye data (monocular mode).
• To compare the performance of direct fundus images against expert-segmented morphological data.

Main Methods:

• Utilized the PAPILA dataset including ocular fundus images, clinical data, and expert segmentations.
• Developed a binocular mode integrating data from both eyes, contrasting with a monocular baseline.
• Employed Gradient-Boosted Decision Trees (GBDT) and Convolutional Neural Networks (CNNs) including MobileNet, VGG16, ResNet-50, and Inception.
• Applied SHAP values and Deep Explainer for model interpretability.

Main Results:

• The binocular approach, combining morphological and clinical data, achieved an Area Under the Curve (AUC) of 0.796 (±0.003).
• CNN models using the binocular approach achieved an AUC of 0.764 (±0.005).
• Findings demonstrate improved model performance when considering data from both eyes.

Conclusions:

• Simultaneous use of ocular fundus images from both eyes and clinical data enhances automated glaucoma diagnosis.
• The proposed multi-modal machine learning method offers a promising avenue for early glaucoma detection.
• Integrating data from both eyes is a viable strategy to improve diagnostic accuracy in glaucoma screening.