labeled Intraveneous (IV) infusion container images

We generated an IV fluid-specific dataset to maximize the accuracy of the measurement. We developed our system as a smartphone application, utilizing the internal camera for the nurses or patients. Thus, users should be able to capture the surface of the fluid in the container by adjusting the smartphone's position or angle to reveal the front view of the container. Thus, we collected the front view of the IV fluid containers when generating the training dataset.

We captured the IV fluid container in video mode and extracted an image for every 15 frames. To simulate the actual operation, we changed the capturing angle by moving from the front of the container to the sides, closer, away from the container. We captured the images at a nursing practice lab maintained by the Department of Nursing at Konkuk University Glocal Campus in South Korea. The lab was set up as a mock-up of a general hospital’s infirmary, including the actual IV infusion systems. This process resulted in 2,266 images. We used an iPhone 12 capable of capturing images at FHD (1920*1080) or higher quality. The phone was vertically held when capturing the images. Furthermore, we added 370 IV fluid container images to increase the dataset diversity by searching the web using Google. Thus, there were 2,636 images in the dataset at the end.

We used Labelme for image labeling. Labelme is an open-source image annotation tool for object recognition, tracking, and segmentation by enabling the labeling of detailed information about images [16]. We annotated the IV infusion container images as either Vessel or Liquid General as masking coordinates, then stored the outcomes in JSON files, which are input to the computer vision system.