Simultaneous arteriole and venule segmentation with domain-specific loss function on a new public database.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Biomedical Image Analysis

Background:

• Accurate segmentation and classification of retinal arterioles and venules are crucial for diagnosing eye and systemic diseases.
• Challenges include complex vessel structures, varying illumination, and background differences.
• Current methods often use a two-step segmentation-classification approach, risking error propagation.

Purpose of the Study:

• To develop a novel method for simultaneous segmentation and classification of retinal arterioles and venules.
• To improve diagnostic capabilities for various diseases through enhanced retinal image analysis.
• To overcome limitations of existing segmentation-classification strategies.

Main Methods:

• Employed a fully convolutional network (FCN) for direct, simultaneous segmentation and classification.
• Configured the FCN to accept true color retinal images as input and produce multiple labels as output.
• Developed a domain-specific loss function to enhance performance.

Main Results:

• Achieved high sensitivity (0.944) and specificity (0.955) for overall vessel segmentation on the DRIVE dataset.
• Reported low misclassification rates: 10.3% for arterioles and 9.6% for venules.
• Outperformed state-of-the-art methods and avoided error propagation inherent in sequential approaches.

Conclusions:

• The proposed FCN method offers a robust and efficient solution for simultaneous retinal arteriole and venule segmentation and classification.
• Demonstrated strong performance on public and diverse datasets, indicating potential for clinical application.
• Holds significant promise for improving diagnostics and screening of eye and systemic diseases.