AI improves diagnostic imaging at Canadian hospitals

There is increasing evidence that artificial intelligence (AI) is poised to make a major impact on the practice of medicine in Canada.
GE Healthcare’s Critical Care Suite, for example, an AI application that is designed to identify and help prioritize cases such as pneumothorax, was cleared by the U.S. Food and Drug Administration (FDA) in September.

“Normally, X-ray cases are read on a first in, first out basis and if a patient is number 50 on the list and there’s a critical pneumothorax, the radiologist may not get to it right away,” explained Daniel Zikovitz, Principal Solutions Architect with GE Healthcare. “With Critical Care Suite, if pneumothorax is suspected an alert is sent directly to the radiologist for review.” Critical Care Suite is currently not available in Canada.

Artificial intelligence is pervasive in other industries and works behind the scenes when we browse the Internet, shop online and use our smartphones, but its deployment in healthcare has been challenging.

Training an algorithm to identify a pneumothorax, for example, requires that it be exposed to a large volume of chest X-rays from multiple hospitals, which hasn’t been easy given the siloed nature of their storage, privacy concerns and regulatory controls.

The development of Critical Care Suite and the pneumothorax algorithm was made possible by the launch of GE Healthcare’s Edison platform. Edison allows seamless uploading and sharing of images from partnering hospitals and provides a common Web-based workspace on which radiologists from different organizations can curate and annotate the images, an essential prerequisite to training an algorithm.

Humber River Hospital in Toronto, one of four institutions to sign a data sharing agreement for the development of Critical Care Suite, provided 156,000 privacy-compliant chest X-rays and associated reports to GE Healthcare. Two U.S. institutions and one in India also collaborated on the project.

“Using natural language processing, GE was able to go through thousands and thousands of reports and sort out the ones that had the word pneumothorax in them,” said Marina MacPherson, Senior PACS Analyst at Humber River Hospital. “Once curated, the algorithm then had to learn what a pneumothorax looks like, so radiologists were contracted to go through the images and electronically outline the pneumothorax. “Eventually, we expect GE to move on to develop algorithms to identify other pathologies.”

Currently, Critical Care Suite works on GE’s Optima XR240amx portable X-ray machine.

At the McGill University Health Centre’s (MUHC) Augmented Intelligence and Precision Health Laboratory (AIPHL) in Montreal, Dr. Reza Forghani and his team are using GE’s Edison platform and machine learning technology to predict the spread of head and neck squamous cell carcinoma to the cervical lymph nodes.

“CT imaging doesn’t pick up some types of early spread well, so a lot of head and neck cancer patients get their lymph nodes taken out,” explained Dr. Forghani, Associate Professor, Department of Radiology, MUHC, and Director and Lead Investigator, AIPHL. There’s a value to that because of the survival benefit, he said, but it results in overtreatment and an unnecessary invasive procedure with potential for significant complications and morbidity for as many as 60 to 70 percent of patients.

“If validated in larger studies, this approach could be combined with a radiologist’s expert evaluation of nodes as a clinical assistant tool, potentially increasing diagnostic confidence in prediction of absent nodal metastases and reducing negative neck dissections,” concludes a research paper published in European Radiology.

McGill University Health Centre’s Augmented Intelligence and Precision Health Laboratory is beta testing GE’s Edison platform and operates a “very broad AI program that includes everything from precision medicine to streamlining processes,” said Dr. Forghani. Edison allows his team to work collaboratively with other institutions around the world and add to its capabilities.

“Images are great but there are a lot of other things in a patient’s chart, including waveforms, ECGs and molecular phenotypes,” he noted. “For example, there’s a lot of interest in liquid biopsies. People have tumours, they shed dead cells and, therefore, DNA. In the future we could combine the molecular information and the information from the scan to help us make predictions.

“These are complex problems,” he added. “They’re feasible, but it will take years until they’re developed and validated. In the meantime, there’s much lower hanging fruit to optimize workflow because, let’s be honest, there are limited resources and medicine costs too much, so we need to be more efficient and demonstrate some wins in the short term.

“More and more things are done with imaging, which is fine, but someone has to read those images and how we deal with volume is an issue.”

One example of a workflow tool is GE Healthcare’s AIRx solution for magnetic resonance brain scans, which uses artificial intelligence to set the angle and thickness of the slice, as well as the energy level.

Manually setting the protocol for a head and neck scan could take 45 minutes, said GE’s Daniel Zikovitz. “The autoprotocol on our AIRx can do it in 15 minutes or less. That saves time, but what’s even more important is that it avoids variance. If you’re doing a head scan and you’re measuring a brain tumour and you’ve given the patient some sort of treatment – it could be radiotherapy or a drug – you want to determine within a fairly tight margin if the tumour is growing or contracting.”

If the scans are performed using different protocols, determining the growth or contraction of the tumour is more difficult.

Another Edison-powered example of low-hanging fruit is GE’s X-ray Quality Application featuring Repeat Reject Analytics, which Humber River uses to track and report on X-ray reject rates. The reports produced by the tool identify the type of exam, the technician and the reason for the reject, empowering the department to take corrective action.

“Our reject rate was originally upwards of 8 percent. The last time we looked, we’re trending below 5 percent,” said Dolores Dimitropoulos, Manager of Humber River’s Medical Imaging Department.

AI is here and it’s exciting,” said MacPherson. “It’s going to change medicine, but in particular, it will change medical imaging.”

“If you think artificial intelligence doesn’t work, what you would do without your iPhone?” McGill’s Dr. Forghani asks skeptics. “It’s not by chance that when you go on Google, you get targeted advertisements, and it shouldn’t be a surprise that Amazon is so popular when you order something on Friday night and receive it on Saturday. We can use a little of that in healthcare.”