Vein segmentation and visualization of upper and lower extremities using convolution neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Vascular Surgery

Background:

• Accurate vein localization is crucial for effective venipuncture, vascular surgeries, and Chronic Venous Disease (CVD) management.
• Current methods for vein visualization can be challenging, especially under unconstrained conditions.

Purpose of the Study:

• To develop a reliable real-time framework for venous localization, identification, and visualization using a deep learning (DL) self-parametrized Convolution Neural Network (CNN).
• To segment venous maps from lower and upper limb datasets acquired under unconstrained conditions using near-infrared (NIR) imaging.
• To assist vascular surgeons in procedures such as venipuncture, vascular surgeries, and CVD treatments.

Main Methods:

• A portable image acquisition setup was used to collect venous data from 72 subjects.
• A manually annotated dataset trained and compared conventional CNN architectures (ResNet, VGGNet) with a self-parameterized U-Net.
• The focus was on improving automated vein segmentation and visualization.

Main Results:

• The self-parameterized U-Net demonstrated superior performance in segmenting unconstrained datasets compared to conventional CNN models.
• Achieved a Dice score of 0.58 and 96.7% accuracy for real-time vein visualization.
• The model proved effective for real-time vein location under unconstrained conditions.

Conclusions:

• The self-parameterized U-Net shows significant potential for vein segmentation and visualization.
• It can potentially reduce risks associated with traditional venipuncture and CVD treatments.
• This advanced CNN architecture offers improved vascular assistance, enhancing patient care and treatment outcomes.