Densely connected U-Net retinal vessel segmentation algorithm based on multi-scale feature convolution extraction.

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Accurate segmentation of retinal blood vessels is crucial for diagnosing ophthalmic diseases.
• Existing algorithms may have limitations in capturing complex vessel structures.

Purpose of the Study:

• To improve the accuracy of retinal blood vessel segmentation.
• To develop an algorithm that enhances feature extraction and utilization for better segmentation performance.

Main Methods:

• A novel U-Net structure incorporating parallel fusion and serial embedding of multiscale features was designed.
• Inception multiscale convolution and dense blocks were utilized for feature extraction.
• Dense connection structures replaced classical convolution blocks in the U-Net encoder for multiscale feature extraction.

Main Results:

• The proposed methods demonstrated high performance on standard datasets (DRIVE and CHASE_DB1).
• Metrics including sensitivity, specificity, accuracy, and AUC were evaluated for both parallel fusion and serial embedding approaches.
• Specific results showed sensitivity ranging from 0.7854 to 0.8113 and accuracy up to 0.9574.

Conclusions:

• Multiscale feature detection and dense connections significantly enhance retinal blood vessel segmentation.
• The proposed algorithm outperforms the standard U-Net and other mainstream segmentation methods.
• This advancement holds promise for improved automatic diagnosis of ophthalmic conditions.