A residual connection enabled deep neural network model for optic disk and optic cup segmentation for glaucoma...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early glaucoma diagnosis is crucial for preventing vision loss and requires accurate cup-to-disk ratio (CDR) estimation.
• Current automatic CDR computation methods often lack accuracy and exhibit high complexity, hindering their use in diagnostic systems.
• Existing deep learning models for glaucoma diagnosis are computationally intensive, requiring significant training and testing time due to numerous parameters.

Purpose of the Study:

• To propose a novel Residual Connection Deep Neural Network (RC-DNN) for efficient and accurate joint optic disk (OD) and optic cup (OC) segmentation.
• To address the limitations of existing methods by reducing model complexity and improving diagnostic accuracy for glaucoma screening.

Main Methods:

• Developed a Residual Connection Deep Neural Network (RC-DNN) utilizing non-identity residual connectivity for joint OC and OD detection.
• The model employs efficient residual connectivity to enable simultaneous segmentation, mitigate vanishing gradients, and facilitate segmentation with fewer layers.
• Trained and evaluated the RC-DNN model on the RIM-ONE and DRISHTI-GS datasets.

Main Results:

• The RC-DNN model achieved high performance in OC segmentation across both datasets.
• Key metrics include Dice coefficients of 92.62% (DRISHTI-GS) and 86.52% (RIM-ONE), Jaccard coefficients of 86.87% (DRISHTI-GS) and 77.54% (RIM-ONE), sensitivity of 94.21% (DRISHTI-GS) and 95.36% (RIM-ONE), specificity of 99.83% (DRISHTI-GS) and 99.639% (RIM-ONE), and balanced accuracy of 94.2% (DRISHTI-GS) and 98.9% (RIM-ONE).
• Demonstrated significant performance enhancement for joint OC and OD segmentation with reduced computational complexity.

Conclusions:

• The developed RC-DNN model offers improved efficacy, robustness, and reliability for glaucoma diagnosis.
• The model's reduced complexity and high accuracy make it suitable for deployment in population-scale glaucoma screening programs.
• Highlights the potential of efficient deep learning architectures for advancing ophthalmic diagnostics.