Evaluating the Implementation of Deep Learning in LibreHealth Radiology on Chest X-Rays

Abstract

Respiratory diseases are the dominant cause of deaths worldwide. In the US, the number of deaths due to chronic lung infections (mostly pneumonia and tuberculosis), lung cancer and chronic obstructive pulmonary disease has increased. Timely and accurate diagnosis of the disease is highly imperative to diminish the deaths. Chest X-ray is a vital diagnostic tool used for diagnosing lung diseases. Delay in X-Ray diagnosis is run-of-the-mill milieu and the reasons for the impediment are mostly because the X-ray reports are arduous to interpret, due to the complex visual contents of radiographs containing superimposed anatomical structures. A shortage of trained radiologists is another cause of increased workload and thus delay. We integrated CheXNet, a neural network algorithm into the LibreHealth Radiology Information System, which allows physicians to upload Chest X-rays and identify diagnosis probabilities. The uploaded images are evaluated from labels for 14 thoracic diseases. The turnaround time for each evaluation is about 30 seconds, which does not affect clinical workflow. A Python Flask application hosted web service is used to upload radiographs into a GPU server containing the algorithm. Thus, the use of this system is not limited to clients having their GPU server, but instead, we provide a web service. To evaluate the model, we randomly split the dataset into training (70%), validation (10%) and test (20%) sets. With over 86% accuracy and turnaround time under 30 seconds, the application demonstrates the feasibility of a web service for machine learning based diagnosis of 14-lung pathologies from Chest X-rays.