Design of a Low-Cost, Self-Supervised NIR Vein Visualization System for Cannulation Assistance

We present a $1000 handheld prototype that combines a coaxial 850 nm LED with an industrial near-infrared camera and a 132 k-parameter Tiny-U-Net trained self-supervisedly from Frangi pseudo-labels; the full pipeline runs at over 30 fps on a laptop CPU. In a ten-subject laboratory study the system raised the contrast-to-noise ratio from 0.40 ± 0.10 (ambient light) to 0.72 ± 0.05 (IR LED + U-Net), increased the vessel-pixel fraction from 12.0 ± 3.0 % to 28.0 ± 6.0 %, and boosted the largest-segment coverage from 25.0 ± 11.0 % to 60.0 ± 10.0 %, all with $p$ <. 001. These results show that self-supervised near-infrared imaging can deliver continuous, high-contrast vein maps without manual annotations, paving the way for affordable vein visualisation in resource-constrained settings.