Andrew Garnett, Joyann Todd, Owen Grant, Joel Joseph & Girendra Persaud
Published: October 4, 2021 • Book of Abstracts – Student Research, Volume 2 [Forthcoming]
Andrew Garnett, Joyann Todd ✉️ School of Allied Health. College of Medical Sciences. University of Guyana – Turkeyen Campus. Greater Georgetown, Guyana.
Owen Grant, Joel Joseph, Girendra Persaud Department of Computer Science. Faculty of Natural Sciences. The University of Guyana – Turkeyen Campus. Greater Georgetown, Guyana.
In medical imaging practice, technical errors on diagnostic Chest X-rays can result in an image being rejected and repeated. This research proposed the use of Computer Vision, specifically Convolutional Neural Networks as an automatic alternative to manual quality inspection to identify technical errors and decrease the rate of erroneously validated images. The approach was to create, train, and test a Convolutional Neural Network to identify valid and invalid Chest X-rays using a machine learning library. The first testing set consisted of 450 digital X-rays retrieved from local Picture Archiving and Communication Systems (PACSs). These were manually evaluated and placed into valid and invalid categories by a team of 11 radiologists, along with the reasons for their classifications. The same testing set was analysed by the Convolutional Neural Network, which outputted whether the image was valid or invalid along with a heat-map highlighting erroneous areas. A second testing set retrieved from an online data source that consisted of 450 images pre-labelled by NIH based radiologists was also analysed by the Convolutional Neural Network. The results of these comparisons were used to statistically evaluate and determine the rate of accuracy of the Convolutional Neural Network by calculating the model’s F1 score. Accuracy of the heat-map output in recognising invalid parts of an X-ray was also tested through comparison with the reasons listed on the radiologist evaluation sheet. When compared to radiologist-evaluated images, the Convolutional Neural Network was able to surpass the baseline percentage accuracy for medical imaging technologists of 90%. Heat-map visualisation was also successfully implemented, showing that the Convolutional Neural Network had shortcomings in establishing the degree of rotation necessary for an image to be invalid as well as in its ability to disregard radiologic labels and pathology in technical evaluations. This research has shown that automating the process of quality inspection can increase the accuracy and consistency of diagnostic Chest X-rays.
Keywords: Quality inspection, Convolutional Neural Network, chest X-ray, medical images