Feature article

AI Could Hold the Key to Identifying Pneumonia Via X-Ray

Pneumonia is a particularly difficult disease to diagnose because it could be caused by any number of pathogens that lead to a bacterial, fungal or viral infection in the lungs and it can be contracted almost anywhere, including in hospitals.1 

Chest X-rays have long been considered the best tool to detect any form of pneumonia, but it's not perfect. Pneumonia can appear similar to other conditions on a scan, and imaging cannot identify the infectious pathogen, making the diagnosis of pneumonia via X-ray a challenge. This is especially true when patients are experiencing multiple health problems simultaneously. 

Physicians must gather a complete picture of the patient's condition to make their best and most accurate diagnosis. Although a doctor can order a complete blood test in addition to the X-ray, it may be of limited value under these circumstances. The results could simply signal a pre-existing condition when there is an abnormality and not that the patient has contracted pneumonia. Indirect evidence and external symptoms like fevers and secretions must also be taken into account during the diagnostic process. 

“Most families think it’s just a lung infection and shouldn’t be hard to diagnose," said Dr. Dominique Piquette of the Sunnybrook Health Sciences Centre in Toronto. "But the evidence isn’t always clear-cut.”1,2

In order to address such issues when using X-rays, the Radiological Society of North America (RSNA) put together the second Pneumonia Detection Challenge, which began in August 2018. The machine learning data challenge invited teams to design and develop algorithms to recognize and localize pneumonia from chest X-rays.

“The expectation that artificial intelligence will soon provide valuable tools for radiology continues to grow,” said Dr. Luciano Prevedello, chair of the Machine Learning Steering Subcommittee of the RSNA Radiology Informatics Committee (RIC).3

AI-detection of pneumonia through radiology 

The progress that artificial intelligence (AI) has made with regard to radiology has indeed exceeded any and all expectations in terms of providing accurate, automated diagnoses, and, most cases, even outshined human healthcare professionals.4 

One prominent example includes the algorithm, ‘CheXNet,’ an artificial neural network designed to detect pneumonia from chest X-rays, at a performance rate greater than the average radiologist. The system was developed by Stanford University and tested against four practicing radiologists to diagnose 14 diseases and achieved “state-of-the-art results” on all of them.5

With advances like these, there is always the fear that technology would replace the skilled worker, and in this case, the radiologist. But leaders in the field don't see these innovations as replacements to skilled physicians.

“Radiologists have a critical role to play, even if they don’t have extensive experience with machine learning,”3 said Dr. Prevedello.

To tackle this delicate predicament, a paper, presented at the Society for Medical Imaging Informatics (SIIM) in Medicine’s Machine Intelligence in Medical Imaging conference, in September 2018, described a novel approach of using AI to enhance the diagnoses made by practitioners instead of substituting their abilities with a machine.4

Their innovative technology, Artificial Swarm Intelligence (ASI), is believed to be based on the biological phenomenon of swarm intelligence4 (i.e., a discipline, composed of many single bodies, such as ants in colonies or birds in flocks, that focuses on collective behaviors as a product of interactions between individuals themselves and their environment).6

Similarly, the researchers of the same study, led by CEO and founder of Unanimous AI, Louis Rosenberg, developed a method — “hive mind” — that uses ASI to combine and utilize the individual capacities of a small group of radiologists, together, in real time, to procure an optimal diagnosis or solution. It was found that this technique showed better results than individual doctors or algorithms detecting pneumonia by X-ray.4,7

It is interesting to note the pivotal role of the radiologists here — their extensive subject knowledge and clinical experience, the power of human insight, and their ability to pinpoint systematic errors, in case of any, among other responsibilities.3

Such studies show how quickly and correctly AI algorithms can be trained by employing new datasets. For instance, Stanford University’s CheXNet study3 used about 100,000 chest X-ray scans from the National Institutes of Health database to accurately learn and identify pneumonia better than trained medical imaging doctors, within a period of two months.8 Published in PLOS Medicine in 2018, the study used an innovative computer-based ‘heat-map’ tool on the chest X-rays was indicative of the area most likely to be pneumonia, which could possibly reduce chances of misdiagnoses and missed cases of the disease.5,9

What lies ahead for AI and diagnostics

Jeremy Irvin, a graduate student at Stanford University at the time the study was conducted, talked about the team’s efforts to continue this line of work by improving the medical algorithms in order to make it high-quality for detecting abnormalities. He added that the future of machine learning in healthcare has huge potential.9 In fact, it was reported in 2017 that algorithm-accuracy had grown from 75 to 95 percent in five years, exceeding the already fast rate at which these kinds of systems were being built.8

Despite being a hotly debated topic across disciplines with regard to the monopolization of jobs, artificial intelligence and machine learning has leveraged the solving of complex data analysis problems, optimization of practices, and the diagnosis of life-threatening diseases like pneumonia.  



  1. Pneumonia, NIH: National Heart, Lung and Blood Institute, https://www.nhlbi.nih.gov/health-topics/pneumonia. Accessed December 14, 2018.
  2. Why it can be hard to diagnosis pneumonia. YourHealthMatters. http://health.sunnybrook.ca/navigator/can-hard-diagnosis-pneumonia/. Accessed December 12, 2018.
  3. RSNA Announces Pneumonia Detection Machine Learning Challenge. RSNA News: Press Release. http://press.rsna.org/timssnet/media/pressreleases/PDF/pressreleasePDF.cfm?ID=2028. Accessed December 13, 2018.
  4. Artificial Swarm Intelligence employed to Amplify Diagnostic Accuracy in Radiology. IEMCON 2018 - 9th Annual Information Technology, Electronics, and Mobile Communication Conference. Accessed December 14, 2018.
  5. Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists. PLOS Medicine. https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1002686. Accessed December 26, 2018.
  6. Swarm intelligence. Scholarpedia. http://www.scholarpedia.org/article/Swarm_intelligence. Accessed December 14, 2018.
  7. Radiologists connect via AI ‘hives’ to enhance pneumonia diagnosis. HealthImaging. https://www.healthimaging.com/topics/artificial-intelligence/radiologists-connect-ai-enhance-pneumonia-diagnosis. Accessed December 14, 2018.
  8. Stanford trained AI to diagnose pneumonia better than a radiologist in just two months. Quartz, https://qz.com/1130687/stanford-trained-ai-to-diagnose-pneumonia-better-than-a-radiologist-in-just-two-months/. Accessed December 14, 2018.
  9. Algorithm better at diagnosing pneumonia than radiologists. Stanford Medicine. http://med.stanford.edu/news/all-news/2017/11/algorithm-can-diagnose-pneumonia-better-than-radiologists.html. Accessed December 14, 2018.