Chester the AI Radiology Assistant’ is a free online tool that allows you to assess chest X-rays online within a few seconds. It ensures the security of private medical while predicting the chances of developing 14 diseases.
Though it is still rudimentary, Chester can process the upload of users with output diagnostic predictions for various diseases like cardiomegaly, atelectasis, infiltration, effusion, masses, pneumonia, nodules, consolidation, pneumothorax, emphysema, edema, pleural thickening, hernia, and fibrosis with as much as 80% accuracy. A red and green sliding scale points out the diagnostic probability for all these conditions, ranging from a healthy state to high-risk conditions.
Joseph Paul Cohen together with his colleagues from the Montreal Institute for Learning Algorithms made a debut of Chester in a paper that was published in 2020. In the paper, they wrote that they want to develop a system to scale the minimum cost of computation while preserving diagnostic accuracy and privacy. Anyone having access to the web and a browser may use this tool, including smartphones. It supplements the opinion of a professional but does not replace the same.
Cohen and team wrote, “Deep learning has shown promise to augment radiologists and improve the standard of care globally. Two key issues complicating the deployment of these systems include patient privacy as well as global population scaling.”
This team developed the tool, Chester, with the help of the CheXnet DenseNet-121 model. It is the same train-validation-test split just as Rajpukar’s 2017 paper on this subject. Technology facilitated Chester to further analyze the reports of chest scans with the web-based, locally running system.
The inference of this tool is designed to make it simple. It contains three key components – detection out of distribution, disease prediction, and explanation of prediction. After uploading the X-day, the Chester tool takes approximately 12 seconds to load models, another 1.3 seconds for computing relevant graphs, and 17 seconds more to compute the gradients explaining these predictions.
There is a separate function allowing patients to view the heat map of images influencing the prognosis of Chester. At any given point of time, they can view a heat map out-of-distribution where images varied from the training distribution of the team. A bright heatmap indicates too much difference of image from the training dataset of Chester for the tool to predict accurately.
The developers added, “We are equipped to prevent any dissimilar image from processing against the training data to prevent chances of errors in our predictions.”
Cohen and team said that they built Chester to help the medical fraternity so that researchers are able to “experiment with it to figure out how it can be useful.” Their aims include:
- Showing the strength of the open datasets and supporting the creation of more unrestricted public dataset projects
- Establish lower bound care, as the team pinpointed that radiologists need not be ‘no worse’ compared to Chester
- Design a model, which can be copied to global scale healthcare solutions without any untenable costs
- Highlighting the fact that patient privacy is preserved using a unique web-delivery system
Cohen and his co-authors also wrote, “This tool can be used as an assistant and as a teaching tool. The system has the necessary design to process locally. As a result, it ensures privacy of patients while enabling scaling from 1 to a million users with insignificant overheads. It is hopeful that this could prompt radiologists to offer us a feedback and help us to advance the tool as well as adapt to the changing needs.”
sepStream® gives great importance to technology and brings in sophisticated diagnostic tools to deliver quality results.