Improved fully convolutional neuron networks on small retinal vessel segmentation using local phase as attention.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal vessel segmentation is vital for diagnosing diseases like diabetes, glaucoma, and hypertension.
• Current segmentation methods often overlook small retinal vessels, which are sensitive indicators of circulatory health and early disease.
• Accurate segmentation of small vessels is essential for timely diagnosis and disease warning.

Purpose of the Study:

• To develop an advanced method for accurate retinal vessel segmentation, with a specific focus on small vessels.
• To improve the early diagnosis and monitoring of systemic diseases through enhanced retinal image analysis.
• To introduce a new evaluation metric for assessing small vessel segmentation performance.

Main Methods:

• Combined unsupervised methods: local phase congruency (LPC) and orientation scores (OS).
• Integrated LPC and OS into a U-Net-based deep learning network with attention mechanisms (UN-LPCOS).
• Proposed a new metric, sensitivity on a small ship (Se), to specifically evaluate small vessel segmentation.

Main Results:

• The proposed UN-LPCOS method demonstrated remarkable ability in identifying and segmenting small retinal vessels.
• Achieved outstanding segmentation performance on both overall vessel structures and small vessels.
• Validated effectiveness on the DRIVE dataset and data from The Maastricht Study.

Conclusions:

• The UN-LPCOS method significantly improves retinal vessel segmentation, particularly for small vessels.
• This technique holds great potential for early disease detection and monitoring through enhanced retinal image analysis.
• The proposed Se metric provides a valuable tool for evaluating small vessel segmentation performance.