The AI software was able to diagnose 75 percent of
cases of collapsed lung

New AI system identifies more pneumothorax cases in X-rays

October 18, 2019
by John R. Fischer, Senior Reporter
A new software program that utilizes artificial intelligence to assess chest X-ray images could enable greater accuracy for identifying more cases of collapsed lungs (pneumothorax).

Researchers at the University of Waterloo were able to determine 75 percent of cases of collapsed lungs with their developed system, a more than 25 percent increase from the average rate of less then 50 percent for medical specialists using chest X-rays.

"We are investigating a multitude of venues for breaking through the complexity barrier of diagnostic ambiguity. Our partners at UHN are in process of providing us with high-quality image data 'labeled' by the expert radiologists (labeled meaning that the location of the pneumothorax is highlighted)," Hamid Tizhoosh, a professor of systems design engineering and director of the Laboratory for Knowledge Inference in Medical Image Analysis (KIMIA Lab) at the university, told HCB News. "By combining the strength of unsupervised (search) and supervised (detection) methods we will be able to reduce the number of missed cases."

Pneumothorax occurs when air gets between the chest wall and the outside of a lung, causing pain and symptoms such as shortness of breath. If undiagnosed and not treated, it can cause the lung to suddenly deteriorate and can lead to death. Experienced radiologists can typically identify serious cases of pneumothorax on X-ray scans. Minor ones, however, are challenging to spot, leading to misdiagnoses that place up to 50 percent of patients at risk.

The software identifies the conditions by comparing the findings of the patient to a database of more than 550,000 X-rays, including 30,000 cases of collapsed lung. If a majority of similar X-rays from the database show a diagnosis of collapsed lung, the AI system recommends it as the patient’s diagnosis.

The team at Waterloo is working to increase the accuracy of the technology to 90 percent, as part of a project with the University Health network (UHN), a healthcare and medical research organization consisting of a number of Toronto-based hospitals. Once it has, the system will be integrated with another software program called Coral Review, a quality assurance tool used at UHN-affiliated hospitals to allow physicians to review medical imaging diagnoses made by peers and offer second opinions.

Waterloo and UHN see the AI system acting as a "computational second opinion" in Coral. If successful at the UHN affiliates, access to it would then be extended to other providers now using the Coral Review system.

In addition to second opinions, the system could be used to prioritize X-rays in busy work environments to reduce treatment delays that place patients at risk. There are also plans to apply the core AI search system to the assessment of other conditions.

"This specific search technology is somewhat customized for chest X-ray images. Hence, it may be used for other chest conditions like pneumonia and any other abnormality visible in X-rays," said Tizhoosh. "The core search technology we have developed at Kimia Lab is actually agnostic to image modality. Presently, we are closely working with other partners, such as Huron Digital Pathology in St. Jacobs, Ontario, to commercialize it for whole slide images of biopsy samples."

The project with UHN is supported by the Vector Institute, a not-for-profit corporation dedicated to advancing AI.