A near infrared image of forearm subcutaneous vein extraction using U-Net

Abstract

Machine learning is in demand for acquiring important perceptions from big data or producing advanced revolutionary technologies and helps most the human tasks effortlessly. Healthcare is one of the industries that receive benefits from it. In the medical industry, venipuncture is one of the most crucial procedures, and locating the patient’s vein is the challenge faced by clinicians. The difficulty leads to multiple trials of venipuncture and causing harm such as bleeding, bruising, damaging surrounding cells, and other effects on the patient. If the case is worst, the patient might have to go to central venous access. Near-Infrared (NIR) has some strong properties such as non-invasive technique, low cost, and small size for the implementation locating the forearm subcutaneous vein; thus, it is a popular method among researchers. The technique has a weakness in that it requires image processing for the enhancement and the vein is more visible and located. This paper is approaching Deep Learning to automatically extract the forearm subcutaneous vein from the NIR image using two architectures: the standard convolutional neural network (CNN) with U-Net architecture and Residual U-Net architecture. The purpose of using two types of architecture is to compare the result and will use the highest accuracy method for the forearm subcutaneous vein extraction. The research found that the U-Net architecture with common CNN has results dice score of 0.6995 while deep residual architecture results 0.7599. It proves that the deep residual architecture has a better extraction than the common CNN block. This project is expected to expand to extract the live video of the forearm subcutaneous vein in future.