Clinical Interpretable Deep Learning Model for Glaucoma Diagnosis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Clinical interpretability of Convolutional Neural Networks (ConvNets) for disease diagnosis remains a significant challenge.
• Accurate and transparent glaucoma diagnosis is crucial for timely intervention and patient outcomes.

Purpose of the Study:

• To propose a novel clinical interpretable ConvNet architecture for accurate glaucoma diagnosis.
• To enhance the transparency of deep learning models by highlighting distinct diagnostic regions in fundus images.
• To introduce a new feature aggregation scheme (M-LAP) for improved glaucoma diagnosis performance.

Main Methods:

• Developed a novel interpretable ConvNet architecture for glaucoma diagnosis.
• Proposed the Multi-scale Local Attention Pooling (M-LAP) scheme for feature aggregation.
• Modeled the correspondence between binary diagnosis and spatial pixels to generate glaucoma activation maps.
• Utilized fundus images from the ORIGA dataset for experimentation.

Main Results:

• Achieved the highest Area Under the Curve (AUC) of 0.88 for glaucoma diagnosis, outperforming state-of-the-art methods.
• Demonstrated effective optic disc segmentation with a Dice score of 0.9.
• Successfully localized disease foci using generated evidence maps, showcasing clinical interpretability.

Conclusions:

• The proposed interpretable ConvNet architecture offers accurate glaucoma diagnosis with enhanced transparency.
• The M-LAP scheme and glaucoma activation maps effectively bridge the gap between semantic diagnosis and precise localization.
• This work represents a significant advancement in providing clinically interpretable deep learning for glaucoma detection.