Artificial intelligence enables low-dose CT scans, faster scan time


Science Highlights
June 10, 2019
Raymond A. MacDougall

Radiologists give high rating to low-dose CT processed with a novel deep-learning technique

A team of bioengineers at Rensselaer Polytechnic Institute (RPI), with funding from the National Institute of Biomedical Imaging and Bioengineering (NIBIB), have developed an artificial intelligence (AI) technique that uses image post-processing to rapidly convert low-dose computed tomography (CT) scans to images of superior quality, compared to low-dose scans that do not use the AI technique. CT has become a commonly prescribed imaging service in modern medicine, providing a non-invasive, detailed, and close-up view of internal anatomy and pathology. Low-dose CT minimizes x-ray radiation to a patient.

“This deep learning, hybrid, image-reconstruction technique integrates low-radiation dose CT images with emerging neural network methods and offers comparable images at much higher speed as those produced with iterative reconstruction methods,” said Behrouz Shabestari, Ph.D., director of the NIBIB program in Artificial Intelligence, Machine Learning, and Deep Learning. “Dr. Wang’s team has advanced deep learning techniques for tomographic imaging and pursued this research with NIH grant support to improve image quality and computational efficiency for low-dose dose CT.”

With its growing use, CT scanning contributes to 62% of the radiation dosage that people in the United States incur from all imaging modalities. While the risk of developing cancer from such radiation exposure is small, public concern has risen with the growing use of CT scans, making CT dose reduction a clinical goal. Medical imaging engineers are working to develop technologies that reduce radiation dose from CT without compromising its diagnostic performance.

CT scans are reconstructed from combinations of many X-rays taken from different angles. In their study published in the June 10, 2019, Nature Machine Intelligence, the team led by Ge Wang, Ph.D., Clark & Crossan Endowed Chair Professor in the RPI Department of Biomedical Engineering, and Mannudeep Kalra, M.D., associate professor of radiology at Harvard Medical School and radiologist at Massachusetts General Hospital, compared standard image reconstruction methods from commercial CT machines with a new method, called a modularized neural network. The new method is a type of AI that researchers refer to as machine learning, or deep learning.

Abdominal and chest CT images on chart

CT images of the chest and abdomen produced using, from left, low-dose AI reconstruction, low-dose conventional iterative reconstruction, and normal dose CT. The image quality of low-dose AI is comparable and, in some cases, superior to conventional low-dose CT scans and it is faster. (G Wang group, RPI/M Kalra group, MGH and Harvard)

The modularized neural network for CT image reconstruction progressively reduces data noise in a way that radiologists can interactively participate in the optimization of the reconstruction workflow. Each small increment of improved image quality can be evaluated by radiologists according to the medical diagnosis they want to make.

The researchers obtained low-dose CT scans of 60 patients; 30 which depicted abdominal anatomy and the other 30 that depicted chest anatomy. The scans represented three commercial CT scanner products, all that already use iterative image reconstruction algorithms—the conventional approach—to reduce image noise. The noise causes decreased image quality as a result of low radiation dose CT scanning. The iterative reconstruction approach refers to the repeated steps that medical imagers attempt towards generating the CT images consistent to some prior knowledge about imaging physics and image content. The researchers compared image reconstruction with currently used iterative methods and their novel deep neural network for image post-processing.

Three radiologists evaluated and scored images for two features: structural fidelity and image noise suppression. Structural fidelity is the ability of the image to accurately depict the anatomical structures in the field of view, which can be diminished by noise. Image noise shows up as random patterns on the image that detract from its clarity.

For abdominal imaging, the radiologists gave higher scores to images produced with the modularized neural network method on two of the three scanning devices and considered the images from the third device as of comparable quality with the iterative reconstruction method. For chest imaging, the experts found the image quality comparable between the two methods for all devices. Overall, the modularized neural network performed favorably or comparably relative to the iterative method when the radiologists evaluated structural fidelity and noise suppression.

The researchers add that their new method is much faster than the current commercial methods and that institutions with current CT scanners of various brands can utilize their technique to produce similar image results. Wang said that the study results confirm that deep learning could help to produce high quality CT images at lower dosages, and at the same time, this novel approach much more efficient than the iterative process, which is time consuming and subject to image noise artifacts.

The research was supported, in part, by a grant from NIBIB (EB017140) for research to develop systems for low-dose CT.

H Shan, A Padole, F Homayounieh, U Kruger, RD Khera, C Nitiwarangkul, MK Kalra and G Wang. Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose CT image reconstruction. Nature Machine Intelligence. June 10, 2019.