Early forays of AI into healthcare have typically followed a similar path.  Algorithms would be trained on vast quantities of medical data (images usually), and would then use its computational might to spot early signs of an illness faster than human doctors are capable of.

The latest project of this ilk comes from a Stanford team who have developed an algorithm, called ChexNet, that diagnoses pneumonia from chest X-rays.

“Interpreting X-ray images to diagnose pathologies like pneumonia is very challenging, and we know that there’s a lot of variability in the diagnoses radiologists arrive at,” the team say.

The algorithm was trained using the publicly available data held at the National Institutes of Health Clinical Center.  The database contains over 110,000 X-ray images that are each labeled with one of 14 possible pathologies.  The algorithm is capable of analyzing each iamge and returning a diagnosis for any of these pathologies.

Signs of progress

The work is interesting because it builds upon initial efforts by the Center to do this themselves.  The data and their own algorithm were released in September 2017, and the Stanford team have managed to significantly improve on its productivity in a very short space of time.

The team believe that their work is especially useful as spotting it early enough is very difficult from X-ray images.  They were able to rapidly improve upon the work done by the Center, to the extent that within a week they had developed an algorithm that could accurately diagnose 10 of the pathologies, with the full 14 delivered in just over a month.  Each of the diagnoses was more accurate than current state-of-the-art tests.

The interpretation of X-rays are a crucial part of the diagnoses of many diseases.  Currently however, that analysis largely relies on the skills of the radiologist, but as talented as they are, they are nonetheless fallible.

“The motivation behind this work is to have a deep learning model to aid in the interpretation task that could overcome the intrinsic limitations of human perception and bias, and reduce errors,” the authors say.

After a month of tweaking and improvements, the system was consistently able to outperform a number of experienced radiologists in detecting pneumonia.

X-ray heat maps

The team have also developed a digital tool that is capable of producing a ‘heat map’ style image from the X-ray.  The ‘temperature’ of the image correlates with the areas most likely to represent pneumonia.

They hope that the tool will significantly reduce the number of pneumonia cases that go undiagnozed and speed up attempts to treat those with the condition.

“We plan to continue building and improving upon medical algorithms that can automatically detect abnormalities and we hope to make high-quality, anonymized medical datasets publicly available for others to work on similar problems,” they explain.  “There is massive potential for machine learning to improve the current health care system, and we want to continue to be at the forefront of innovation in the field.”