AI Adoption into Radiology Clinical Practice

By 

Artificial Intelligence (AI) applications are improving medical imaging through their positive impact on patient satisfaction, image quality, and workflow efficiency. Deep learning (DL) reconstruction is one AI technique that is being used to allow scan time reduction without compromising image quality, says Suzie Bash, MD, the medical director of neuroradiology at San Fernando Valley Interventional Radiology at RadNet. 

Radiologists can now alter their image protocols to accelerate image acquisition, then apply a DL solution to denoise and enhance the sharpness of the magnetic resonance (MR) image, even to the point of exceeding the image quality of standard of care. This can be done with a vendor-neutral solution, such as SubtleMR from Subtle Medical, or through the original equipment manufacturer’s tool, such as AIR Recon DL from GE Healthcare, Deep Resolve from Siemens, Advanced intelligent Clear-IQ Engine (AiCE) from Canon, SmartSpeed and the FDA-pending Precise Image from Philips, and IP-RAPID from Fuji. This process can reduce a routine brain MR scan slot from approximately 30 to 40 minutes down to 15 minutes.

“Deep learning allows us to enable faster scans and get superior perceived image quality, higher perceived signal-to-noise ratio, spatial resolution and contrast-to-noise ratio, as well as reduced artifacts and reduced dose,” Dr. Bash explained.

 

 

Proven benefits of artificial intelligence tools

In positron emission tomography (PET) studies, DL tools like SubtlePET from Subtle Medical can reduce scan time by 75%, which also helps reduce radiation dose. “You apply the deep learning to restore the quality to that accelerated scan, but at dramatically decreased radiation dose,” Dr. Bash said.

SubtleGAD, a work-in-progress, is expected to permit clinicians to reduce gadolinium dose by 90% and apply DL reconstruction to restore contrast enhancement. “This is very pertinent in these days where there's so much talk about gadolinium deposition,” she said.

AI tools are also playing a critical role in cancer screening. RadNet’s DeepHealth subsidiary is a leading developer of AI tools for mammography interpretation. It’s been shown to help detect breast cancer before expert radiologists can diagnose it. “This is going to revolutionize mammography and help save lives,” Dr. Bash said. “Clinical validation of this AI tool was demonstrated in a manuscript published in Nature Medicine, where the technology successfully flagged the breast cancer on mammograms at the one- and two-year point, and consistently outperformed expert breast fellowship-trained radiologists.”  RadNet is also beginning to use an AI solution for prostate cancer screening through its Quantib subsidiary and is developing an AI tool for lung cancer screening through its Aidence subsidiary.

Triage application AI solutions, such as offerings by Viz.ai and Aidoc, are also providing great value by expediting diagnosis and treatment of critical imaging findings. These AI tools can automatically flag critical findings on computed tomography (CT) exams and prioritize these studies to the top of the radiologist’s worklist, detecting findings such as large vessel occlusions (LVOs), intracranial hemorrhages, aneurysms, cervical spine fractures and pulmonary embolisms. Viz.ai’s solution will also securely transmit the CT images, demonstrating the LVO and perfusion map to the multidisciplinary stroke team through a HIPPA-compliant cellular phone application to prompt early intervention, resulting in 40% improved patient outcomes and shorter hospital stays.


Quantitative volumetric tools

Quantitative volumetric AI tools, such as NeuroQuant from Cortechs.ai, icobrain from icometrix, and QuantibND segment and quantify the volume of brain structures, then compare those volumes to a large age- and gender-matched normative database, which allows volumetric tracking to assess for rate of change over time. The software can assess different patterns of brain atrophy, such as those seen in Alzheimer’s disease, frontotemporal dementia, and dementia with Lewy bodies.

In patients with multiple sclerosis, quantitative neuroimaging assesses demyelinating disease burden and dynamic change over time, which can impact disease-modifying therapy. The software is also useful in evaluating patients with epilepsy or traumatic brain injury. Cortech.ai’s OnQ Neuro product can also segment and quantitate the volume of brain tumors and help assess for true tumor progression versus treatment-related change. These quantitative AI solutions enhance diagnostic accuracy, add clinical value, decrease reader subjectivity, and positively impact patient care.

“Quantitative volumetric MR improves diagnostic value, improves detection of disease activity, and helps eliminate report bias. It's easy to add on to routine brain imaging at negligible cost to acquisition speed, and it doesn't require any radiology post-processing. It's straightforward to interpret and can offer a referral advantage,” said Dr. Bash.

 

 

Pairing two AI tools

Dr. Bash said it’s possible to use more than one AI solution to obtain the benefits of both. To investigate this, they paired NeuroQuant with fast acquisition deep learning sequences (FAST-DL) on brain images of patients experiencing memory loss. They found that DL reconstruction allows 60% scan time reduction while maintaining high volumetric quantification accuracy. In a prospective, multicenter, multireader trial, researchers determined that FAST-DL was statistically superior to standard of care for perceived quality across every imaging feature reviewed, despite the significant scan time reduction.

“The DL outperformed the standard of care for image quality assessment every single time. There was no difference in quantitative volumetric biomarkers or in the clinical classification for the patients in either the standard of care or FAST-DL group, which indicates that the DL post-processing did not introduce any errors into the database,” Dr. Bash explained. “DL reconstruction allowed 60% scan time reduction while maintaining a high volumetric quantification accuracy and perceived superior image quality when compared to standard of care. These shorter scan times may boost the utilization of volumetric quantitative MRI in routine clinical practice.”

 

 

AI Integration and adoption

Dr. Bash said radiologists are critical to AI technology adoption. “With AI solutions like image reconstruction, it is about improving image quality, workflow efficiency and the patient experience,” she explained.

At RadNet, radiologists uniformly agreed that the fast, deep-learning sequence looked better than the standard-of-care image, which helped adoption of new technologies. “They became champions in spreading that excitement throughout the practice, which provided momentum for successful implementation,” she said.

Editor’s Note: Suzie Bash, MD, is the medical director of neuroradiology at San Fernando Valley Interventional Radiology at RadNet. She participated in an expert forum sponsored by Bracco Diagnostics Inc., the US subsidiary of Bracco Imaging S.p.A., and Applied Radiology. This article is based on her presentation and the group’s resulting discussion.

Back To Top

AI Adoption into Radiology Clinical Practice.  Appl Radiol. 

December 07, 2022


Copyright © Anderson Publishing 2023