An optimal way of subcutaneous vein detection using CNN: U-Net

Thong, Adrian Kheng Hoeng (2022) An optimal way of subcutaneous vein detection using CNN: U-Net. Final Year Project, UTAR.

Preview

PDF Download (3317Kb) | Preview

Abstract

Intravenous (IV) procedures are usually performed by nurses and medical practitioners through the patient’s veins. This is either to obtain blood samples or to administer drugs that cannot be taken orally. However, it has become an increasing concern since there are several factors affecting intravenous (IV) procedures to be done smoothly without hurting the patients repeatedly. This problem has been described as Peripheral Difficult Venous Access (PDVA) and performing such procedures repeatedly will heighten the patient’s anxiety level. Several state-of-art visualization and transformation techniques have been proposed for vein detection and finger vein detection for authentication purposes. However, deep learning has proved its ability and capability in performing automated tasks such as feature extraction, object recognition and detection. With considerations such as accuracy, automation, speed and being in real-time, we propose a deep learning solution (U-Net) to detect Near-Infrared (NIR) subcutaneous forearm veins. Results are then validated with groundtruths that are hand-plotted using Adobe Photoshop that has been verified by a certified radiologist. This report consists of an introduction part to the problem described, reviews of past related state-of-art works, proposing a deep learning framework and its variants as a feasible solution as well as performing performance analysis on the proposed framework.

Actions (login required)