Weighted loss for imbalanced glaucoma detection: Insights from visual explanations.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a primary cause of irreversible vision loss due to optic nerve damage.
• Early glaucoma detection is critical but hindered by imbalanced fundus image datasets.
• Class imbalance poses a significant challenge for machine learning models in medical diagnostics.

Purpose of the Study:

• To improve the performance and interpretability of Convolutional Neural Networks (CNNs) for glaucoma detection.
• To address the challenge of class imbalance in glaucoma fundus image datasets.
• To evaluate the effectiveness of a weighted loss function in enhancing CNN performance.

Main Methods:

• Applied a weighted loss function to CNNs for glaucoma detection.
• Utilized the standardized SMDG-19 dataset, comprising data from 19 public sources.
• Assessed performance using recall, F1-score, precision, accuracy, and AUC.
• Employed Grad-CAM for interpretability analysis.

Main Results:

• Recall improved by 44.75% (from 60.3% to 87.3%), and F1-score increased by 7.25% (from 66.5% to 71.4%).
• Area Under the Curve (AUC) rose by 3.21% (from 84.2% to 87.4%).
• Minor decreases in precision (-6.53%) and accuracy (-4.10%) were observed.
• Grad-CAM visualizations confirmed focus on relevant optic nerve head regions.

Conclusions:

• The weighted loss function effectively enhances CNN performance for glaucoma detection, particularly improving recall and AUC.
• The strategy improves model interpretability by focusing on clinically significant areas.
• This approach offers a promising solution for overcoming class imbalance in ophthalmic image analysis.