While there’s no doubt that AI-based solutions can play a valuable role in diagnostic imaging, many of the radiology and health system leaders we spoke with at the 2022 RSNA conference were are asking an important question: are today’s AI models the right solutions for improving MRI efficiency?

We heard that and similar questions especially about MRI fast acquisition solutions. The search for practical and cost-effective techniques has become increasingly urgent due to a confluence of factors including an aging U.S. population, clinical staffing shortages, physician attrition from burnout, and intense financial pressure on health systems. In fact, we heard from health systems and radiology practices that have working groups or steering committees focused specifically on improving MRI speed and image clarity to address:

• MRI backlogs
• Extended inpatient stays due to limited scanner availability
• The need for additional or longer scans to run more protocols
• Longer reads and turnaround times because of reduced image quality and data resulting from shorter scans.

Faster acquisition enables more scans to be completed in the same amount of time. That can substantially improve patient throughput, machine utilization, access to advanced imaging and health system revenue. That’s the goal of a variety of AI models using synthetic MRI data as well as our physics-based STAGE (STrategically Acquired Gradient Echo) software platform.

The Differences Between STAGE and AI

Both types of solutions can accelerate MRI acquisition times by about 30 percent while maintaining or improving image quality. But they take very different approaches.

AI models enable faster acquisition by subsampling data then reconstructing or enhancing images using algorithms trained with a variety of data sources. Some models look promising but radiologists and health system leaders have valid concerns including:

• The appropriateness and currency of training data
• Patient data privacy
• Model validation and ongoing surveillance
• Data drift, concept drift and related data science issues
• Procedure billing and software ROI

In a December 2022 blog article published by the American College of Radiology’s Data Science institute, co-authors Bryant Chang and Dr. Tessa Cook MD, examine the pros and cons of AI-based synthetic data in MRI and CT. Their article mirrors many of the concerns about AI models for that we heard at RSNA: “Synthetic data might also not be complex and nuanced enough to mimic real data. Artifacts in synthetic data could be unique as compared to real data and could occur more frequently. Synthetic images of certain modalities, such as MRI for neuroimaging, are still inferior to their real counterparts. The creation, curation, and utilization of synthetic data is still a relatively new field, with a host of complications that must be addressed.”

STAGE software instead applies proven MRI physics to accelerate acquisition of fully sampled 1.5T and 3T data while reducing image noise and improving signal-to-noise ratio in post-processing. STAGE also takes advantage of the quantitative nature of MRI to generate susceptibility weighted image (SWI) maps for better detection of disease biomarkers. It also produces standardized output regardless of scanner brand or model.

The difference between STAGE and AI-based techniques is analogous to the differences in JPEG and RAW image formats produced by digital cameras. JPEG images are generated with subsampled data. File sizes are smaller for easier sharing and storage but image resolution is reduced. You’re also limited in the ways you can process and enhance the image because the data were never acquired. However, there are AI models available to enhance resolution sufficiently for most casual applications.

RAW image files are generated using all the data that a camera’s sensor is set to acquire. File sizes are much larger and require higher capacity storage than JPEG files. Acquisition times are comparable to JPEG images and limited only the performance of the camera’s components. But the ability to work with fully sampled data using “digital darkroom” post-processing software gives photographers tremendous flexibility to refine images based on their preferences and goals.

Shifting From Technology to Clinical and Financial Impacts

At first, many RSNA attendees assumed that STAGE was AI-based. But once we explained its physics-based techniques, the conversations quickly shifted from technology to how STAGE improves MRI efficiency and health system finances. That included a review of STAGE testing on 6 different MRI systems at a busy U.S. health system with high neuroradiology volumes. Results showed that multi-contrast scans processed with the latest version of STAGE with CROWN (Constrained Reduction Of White Noise) technology averaged 30% faster than conventional protocols. The time savings created 3 to 4 new scanning slots per scanner per day. Projected revenue increases for each 1.5T system was more than $240,000/year. The added slots on each 3T system increased projected revenues by $320,000/year.

To be clear, SpinTech MRI is not against AI at all. Rather, our physics-based approach is better suited to address the complicated technical, clinical and financial requirements for improving MRI efficiency. It’s also important to note that STAGE can support AI model development with standardized data output from all MRI manufacturers and models. Radiologists also can use AI models to augment image interpretation and reporting workflows.

Our biggest takeaway from those discussions at RSNA is that the search for practical and cost-effective solutions for improving MRI efficiency is more likely to succeed if we start by asking if we are using technology in the right ways to solve today’s most pressing problems.

By: Karen Holzberger, President & CEO of SpinTech MRI