VC-Net: Deep Volume-Composition Networks for Segmentation and Visualization of Highly Sparse and Noisy Image Data

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Extracting in-vivo microstructures, especially cerebrovasculature, is challenging due to data sparseness, noise, and complex vessel geometry.
• Accurate 3D visualization of microvascular networks is crucial for diagnosing vascular diseases but remains difficult with traditional methods.

Purpose of the Study:

• To present a novel visualization-guided computing paradigm for enhanced 3D exploration and microvascular structure extraction.
• To introduce VC-Net, an end-to-end deep learning method for robust 3D microvascular segmentation.

Main Methods:

• Developed VC-Net, a multi-stream convolutional neural network (CNN) framework.
• Embedded maximum intensity projection (MIP) image composition into the 3D volumetric image learning process.
• Created a joint volume-composition embedding space by integrating 3D volume and 2D MIP features.

Main Results:

• VC-Net demonstrated robust extraction of 3D microvascular structures, improving signal-to-noise ratio and handling geometric variability.
• The method effectively captured small/micro vessels and enhanced vessel connectivity compared to traditional and state-of-the-art deep learning approaches.
• Experimental results on public and patient datasets validated the method's superior performance in segmenting complex cerebrovasculature.

Conclusions:

• The proposed VC-Net method offers a significant advancement in 3D microvascular structure extraction and visualization.
• This approach facilitates accurate segmentation of sparse and complex vascular networks, with potential applications in MR arteriogram and venogram diagnosis.
• The synergistic integration of volume rendering and deep learning opens new avenues for 3D medical image analysis.