Machine learning approach for low-dose CT imaging yields superior results

Machine learning has the potential to vastly advance medical imaging, particularly computerized tomography (CT) scanning, by reducing radiation exposure and improving image quality.

Those new research findings were just published in Nature Machine Intelligence by engineers at Rensselaer Polytechnic Institute and radiologists at Massachusetts General Hospital and Harvard Medical School.

According to the research team, the results published in this high-impact journal make a strong case for harnessing the power of artificial intelligence to improve low-dose CT scans.

"Radiation dose has been a significant issue for patients undergoing CT scans. Our machine learningtechnique is superior, or, at the very least, comparable, to the iterative techniques used in this study for enabling low-radiation dose CT," said Ge Wang, an endowed chair professor of biomedical engineering at Rensselaer, and a corresponding author on this paper. "It's a high-level conclusion that carries a powerful message. It's time for machine learning to rapidly take off and, hopefully, take over."

Low-dose CT imaging techniques have been a significant focus over the past several years in an effort to alleviate concerns about patient exposure to X-ray radiation associated with widely used CT scans. However, decreasing radiation can decrease image quality

To solve that, engineers worldwide have designed iterative reconstruction techniques to help sift through and remove interferences from CT images. The problem, Wang said, is that those algorithms sometimes remove useful information or falsely alter the image.

The team set out to address this persistent challenge using a machine learning framework. Specifically, they developed a dedicated deep neural network and compared their best results to the best of what three major commercial CT scanners could produce with iterative reconstruction techniques.

This work was performed in close collaboration with Dr. Mannudeep Kalra, a professor of radiology at Massachusetts General Hospital and Harvard Medical School, who was also a corresponding author on the paper.

The researchers were looking to determine how the performance of their deep learning approach compared to the selected representative algorithms currently being used clinically

Several radiologists from Massachusetts General Hospital and Harvard Medical School assessed all of the CT images. The deep learning algorithms developed by the Rensselaer team performed as well as, or better than, those current iterative techniques in an overwhelming majority of cases, Wang said.

Researchers found that their deep learning method is also much quicker, and allows the radiologists to fine-tune the images according to clinical requirements, Dr. Kalra said.

These positive results were realized without access to the original, or raw, data from all the CT scanners. Wang pointed out that if original CT data is made available, a more specialized deep learning algorithm should perform even better.

"This has radiologists in the loop," Wang said. "In other words, this means that we can integrate machine intelligence and human intelligence together in the deep learning framework, facilitating clinical translation."

He said that these results confirm that deep learning could help produce safer, more accurate CT images while also running more rapidly than iterative algorithms.

"Professor Wang's work is an excellent example of how advances in artificial intelligence, and machine and deep learning can improve biomedical tools and practices by addressing hard problems—in this case helping to provide high-quality CT images using a lower radiation dose.

Northwestern Medicine, Google AI System Outperforms Radiologists

A recent study by Northwestern Medicine and Google used artificial intelligence to improve detection of malignant lung cancer on low-dose CT scans.

The AI system provided an automated image evaluation tool to enhance the accuracy of early lung cancer diagnosis, resulting in earlier treatment.

The researchers compared the AI system performance to radiologists on low-dose CT scans of patients, some of whom had biopsy-confirmed cancer within a year. In most cases, the model performed as well as or better than the radiologists.

The AI system uses both the primary CT scan and a prior CT scan from the patient as input. Prior CT scans are useful in predicting lung cancer malignancy risk because the growth rate of suspicious lung nodules can suggest malignancy. The computer was trained using de-identified, biopsy-confirmed low-dose chest CT scans.

“This area of research is incredibly important, as lung cancer has the highest rate of mortality among all cancers, and there are many challenges in the way of broad adoption of lung cancer screening,” said Shravya Shetty, technical lead at Google.

More information: Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose CT image reconstruction, Nature Machine Intelligence (2019). DOI: 10.1038/s42256-019-0057-9 , https://www.nature.com/articles/s42256-019-0057-9

Journal information: Nature Machine Intelligence