DilUnet: A U-net based architecture for blood vessels segmentation

Area of Science:

• Medical imaging analysis
• Computer vision
• Ophthalmology

Background:

• Retinal image segmentation is crucial for detecting pathological disorders by analyzing retinal blood vessels.
• Early detection of vascular changes can prevent blindness and vision impairments.
• Existing segmentation methods require improved sensitivity and robustness.

Purpose of the Study:

• To propose an automatic retinal blood vessel segmentation method using an enhanced U-net architecture.
• To improve the accuracy and robustness of blood vessel segmentation for better clinical detection of eye diseases.

Main Methods:

• Developed an end-to-end U-net based framework incorporating preprocessing and data augmentation.
• Implemented multiscale input and multioutput modules with improved skip connections.
• Utilized dilated convolutions with varying rates for effective feature extraction.

Main Results:

• Achieved high accuracy (0.9680-0.9701) and Intersection over Union (0.7951-0.8698) on DRIVE, STARE, and CHASE datasets.
• Demonstrated superior sensitivity (0.8263-0.8837) compared to baseline methods.
• Ablation studies confirmed the contribution of each proposed module.

Conclusions:

• The proposed U-net based method outperforms original U-net and other state-of-the-art segmentation techniques.
• The enhanced architecture shows robustness to noise in retinal images.
• This method offers a more effective tool for automated retinal blood vessel segmentation.