Artificial vascular image generation using blood vessel texture maps.

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Traditional methods for blood vessel identification in digital images require extensive pixel-wise annotations.
• Manual annotation of entire vessel trees is time-consuming and costly.
• Existing pre-training methods often focus on vessel geometry, with appearance modeled by general statistics or separate generative networks.

Purpose of the Study:

• To develop a novel methodology for generating artificial blood vessel images with realistic textures.
• To reduce the manual annotation effort required for training neural networks in medical image analysis.
• To improve the performance of network pre-training for blood vessel segmentation.

Main Methods:

• Manually annotate borders of small vessel segments to create texture maps.
• Utilize piecewise linear transformation to apply texture maps to Bézier curves representing vessel geometries.
• Generate artificial images by combining vessel geometries and extracted textures.

Main Results:

• Achieved high Dice scores (0.87 on CORTEX, 0.74 on DRIVE) with minimal annotation (0.03% and 0.29% respectively).
• Performance closely approximates baseline scores obtained with full dataset annotation.
• Demonstrated effectiveness on both fluorescence microscopy and fundus photography datasets.

Conclusions:

• The proposed method effectively disentangles blood vessel geometry and texture.
• This approach significantly enhances network pre-training performance compared to existing methods.
• Offers a cost-effective solution for training accurate blood vessel identification models.