This user has not added any information to their profile yet.
An IV drip infusion kit containing a bottle which constitutes infusion fluid, a drip chamber and a tube connected to the venous catheter. The novelty of the system proposes the use of a camera to film the drip chamber. The video recorded by the camera is sent to a neural network (the system comprises of SIPEED MAIXDUINO KIT FOR RISC-V where the neural network is deployed). The neural network is trained in such a way that it is able to locate the position of the drop and more important its state, i.e. whether the drop has just begun to take shape or is instead well formed. The alternation of these two states can be easily exploited to count drops and therefore measure a flowrate. The neural network is trained on a larger and bad illumination condition data sets in order to acquire efficiency. However, for the implementation of the project a reasonably dim and focused light beam can be used. In the neural network, a squared W x W RGB image is given as input. The output of the network will be N x S x S x 2 tensor. The neural network has an architecture contains a core network which is a sequence of convolutional layers that use an increase in number of 5 x 5 filters and leaky rectified linear activation. This helps in increase the level of abstraction. Problems relative to framing can also be completely avoided by designing a specific support for the camera to be applied directly on the infusion bottle using a clip like support. The output from the neural network is used to control the servo to prevent the backflow. It is finally used to avoid unnecessary risks, while monitoring contagious patients in isolation.
The drop counting operation is straight forward, given the output of the neural network for each frame of the video acquired through the camera. The drop count is increased when the state of the drop changes from 1 to 0. The instantaneous flow rate (in drops per minute, gtt/mil) is measured in such a way. The dataset was used to train and validate the neural network, splitting it according to an 80/20 ratio (training by validation). At the end of training the accuracy in the prediction of the state of drop was 100% for the validation dataset.
It is used to avoid unnecessary risks, while monitoring contagious patients in isolation. It counts the drop virtually without any error. It does not interfere with the normal operation required for traditional IV infusion (simple installation process, no risk of contaminating IV fluid)