Faster image processing to fight lung cancer

Featured Video Play Icon

By Kate McAlpine
Michigan Engineering

A new study at the University of Michigan seeks to make low-dose computed tomography (CT) scans a viable screening technique by speeding up the image reconstruction from half an hour or more to just five minutes. The advance could be particularly important for fighting lung cancers, as symptoms often appear too late for effective treatment.

In December 2013, the U.S. Preventative Services Task Force recommended lung cancer screens for everyone between 55 and 80 years old who has been a smoker within the past 15 years. Roughly 90 percent of cases are related to smoking, and the healthcare costs are approximately $12 billion per year in the U.S.

"Reducing the radiation is like setting a very short shutter speed on your camera." Jeffrey Fessler

Unfortunately, the CT scans that reliably identify tumors by creating 3D images of the lungs also expose the patient to an X-ray dose comparable to about five to eight months’ worth of natural background radiation.

“It’s known that a radiation dose can increase the risk of cancer, but nobody knows exactly how much,” said Jeffrey Fessler, a professor of electrical and computer engineering, who leads the project.

On average, U.S. residents get about 3.2 milliSieverts (mSv) of radiation per year from natural sources such as bananas, cosmic rays from space and radon that escapes from rocks such as granite. Adding another six-month’s-worth of radiation every few years is undesirable for most patients. For this reason, the National Institutes of Health (NIH) prioritized reducing the radiation dose associated with these scans, soliciting proposals to make low-dose CT scans practical. The advance would also benefit children and adolescents, who are thought to be more sensitive to radiation.

Fessler’s team is investigating methods to reduce the dose from around 2 mSv to between 0.24 and 0.4 mSv, or a couple months of natural radiation. The drawback is that the images taken at these low X-ray doses are not as crisp as higher-dose images – initially, at least.

Traditional image reconstruction, which does not utilize knowledge of how X-rays behave in CT scans, leads to a grainy picture of a patient's neck and shoulders. Courtesy of Jeff Fessler.

Traditional image reconstruction, which does not utilize knowledge of how X-rays behave in CT scans, leads to a grainy picture of a patient's neck and shoulders. Courtesy of Jeff Fessler.

“Reducing the radiation is like setting a very short shutter speed on your camera,” said Fessler. “You’d get a grainy picture, but you could use Photoshop to try to get a better image. The data processing needed for low-dose X-ray CT is far more complicated.”

At present, it takes half an hour to an hour to reconstruct the low-dose images for diagnosis. That is impractical when the scans themselves take just a few minutes. Now, the NIH has provided Fessler and his team $1.9 million to cut that processing time down to five minutes.

Their software improves 3D CT scan images by modeling how the X-rays and detectors behave based on physics principles. However, the physics models are complicated, and the images are enormous – equivalent to 150 megapixels – resulting in those long processing times. Fessler thinks that today’s computers are fast enough for shorter processing times, but they need to work smarter.

Collaborating with Thomas Wenisch, an associate professor of computer science and engineering, the team aims to speed up the processing by taking advantage of multicore computing. Most modern computers, tablets and even smartphones come with multiple processors that share computing tasks, allowing our electronics to run faster. But the algorithm currently used to process CT scans predates the wide availability of multicore processing, and it does not fully use the available computing power.

Image reconstruction that includes knowledge of how X-rays behave during CT scans results in a more nuanced picture of the soft tissue in this patient's neck and shoulders. Courtesy of Jeff Fessler.

Image reconstruction that includes knowledge of how X-rays behave during CT scans results in a more nuanced picture of the soft tissue in this patient's neck and shoulders. Courtesy of Jeff Fessler.

The team’s solution is to develop new algorithms that divide the data among the processors, allowing each to handle a certain region, and then stitch the image back together at the end.

The team will work from a data set gathered in a previous study, from volunteer patients who agreed to be scanned twice – once at 20 percent radiation dose and once at 80 percent radiation dose. These matched scans will allow the team to compare their upgraded images with higher quality images.

If successful, the work could help make low-dose CT scans the norm. This would also benefit patients who need CT scans for other reasons, such as diagnosing infections, locating clogged arteries and mapping complex fractures.

Low doses are particularly important for suspected cancer cases because they require multiple scans, Fessler explained. Doctors watch for suspicious tissue that grows between scans spaced months to a year or two apart. “In repeat scan situations, it’s crucial that the dose be very low,” he said.

Fessler’s team also includes many radiologists at U-M who will interpret and analyze the enhanced low-dose images. They include Heang-Ping Chan, a professor of radiology; Lubomir Hadjiiski, a research professor of radiology; Ella Kazerooni, a medical doctor and professor of radiology; Prachi Agarwal, a medical doctor and associate professor of radiology; and Mitchell Goodsitt, a professor of radiology and an adjunct professor of nuclear engineering and radiological sciences.

Fessler is also a professor of biomedical engineering and radiology.