Retinal Vessel Extraction via Assisted Multi-Channel Feature Map and U-Net.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Early detection of retinal vessels is critical for preventing vision loss from retinopathies (e.g., glaucoma, diabetes).
• Existing methods struggle with thin vessel segmentation and discrimination in complex retinal regions.
• Red color and morphological variations in fundus images pose challenges for current vessel detection techniques.

Purpose of the Study:

• To develop an improved approach for accurate retinal vessel segmentation in fundus images.
• To combine traditional template-matching with deep learning (DL) for enhanced vessel detection.
• To improve the segmentation of thin vessels, vital for early diagnosis of retinopathies.

Main Methods:

• A novel hybrid method integrating Cauchy matched filter responses with a U-shaped fully convolutional neural network (U-net).
• The Cauchy matched filter response replaces the noisy red channel of fundus images for preprocessing.
• End-to-end pixel segmentation into vessel and background classes using the U-net architecture on image patches.

Main Results:

• The proposed method achieved an average extraction accuracy of 0.9640 on the DRIVE dataset.
• Evaluation metrics including Accuracy, Precision, Sensitivity, and Specificity were measured.
• The hybrid approach demonstrated effectiveness in segmenting retinal vessels.

Conclusions:

• The novel hybrid method effectively enhances retinal vessel detection, particularly for thin vessels.
• Combining template matching and deep learning offers a promising solution for improving retinopathy diagnosis.
• The U-net model, with Cauchy matched filter preprocessing, shows high accuracy in vessel segmentation.