Checkbox Detection on Rwandan Perioperative Flowsheets using Convolutional Neural Network

Abstract

Millions of surgical operations are performed every year in African countries and the lack of digitization of data associated with them inhibit the ability to study the linkages of perioperative data with perioperative moralities [1]. Contrary to American operating rooms, where medical personnel are assisted by technologies that record and analyze patient vitals and other surgical data, low-income African operating rooms lack these resources and require their personnel to manually scribe this information onto paper flowsheets. In order to provide perioperative data to health care providers in Rwanda, the team designed and implemented image processing and machine learning techniques to automate checkbox detection for the digitization of surgical flowsheet data. A checkbox image is cropped based on its location with template matching and then processed through a trained convolutional neural network (CNN) to classify it as checked or unchecked. The template matching and CNN process were tested using 18 flowsheets. Of the 666 possible images, the template matching achieved an accuracy of 99.8%, and 96.7% of the cropped images were correctly classified using the CNN model.