Deep learning-based image reconstruction (DLR) has been a hot topic in CT over the past five years, as researchers and vendors have continuously demonstrated the technology's potential to improve on legacy and filtered back-projection (FBP) reconstruction methods.
Owing primarily to its noise reduction capabilities, DLR has repeatedly been shown to sharply lower radiation dose while maintaining image quality. Accordingly, commercial and research activity has accelerated.
DLR is also demonstrating its value particularly in cardiac CT applications.
Deep learning-based image reconstruction (DLR) has been a hot topic in CT over the past five years, as researchers and vendors have continuously demonstrated the technology's potential to improve on legacy and filtered back-projection (FBP) reconstruction methods.
Owing primarily to its noise reduction capabilities, DLR has repeatedly been shown to sharply lower radiation dose while maintaining image quality. Accordingly, commercial and research activity has accelerated.
DLR is also demonstrating its value particularly in cardiac CT applications.
Overall, although further research is still needed, "DLR for CT is actively being improved as more vendor options are being developed and current DLR options are being enhanced with second generation algorithms being released," wrote Samuel Brady, PhD, in the department of radiology at Cincinnati Children’s Hospital Medical Center and the University of Cincinnati, for a 2023 article in the British Journal of Radiology.
Brady went on to suggest that DLR algorithms may eventually preserve noise texture similar to that of FBP and object boundary sharpness at dose levels currently used in conjunction with iterative reconstruction (IR).
History of CT image recon algorithms
For added historical perspective, radiologists at Stanford University, University of Wisconsin-Madison, and Leiden University Medical Center in the Netherlands early last year compiled a simple history of CT image reconstruction algorithms. FBP began 40 years ago. Then came model-based iterative reconstruction (MBIR); however, noise texture and reconstruction time were seen as drawbacks.
As of January 2023, hybrid iterative reconstruction (HIR) -- a blend of FBP and MBIR -- was the state-of-the-art image reconstruction technique, according to Lennart Koetzier; Martin Willemink, PhD; and colleagues, who wrote about the technical principles and clinical prospects for the RSNA journal Radiology. Around 2018, however, development efforts for DLR began to pick up.
It is generally understood that DL algorithms for cardiac CT images have been designed and trained to reconstruct low-quality raw data into quality images. The foundation of some DLR algorithms is a convolutional neural network (CNN). Supervised learning allows for an image with a noise texture that more closely resembles a typical FBP image, while retaining the noise-reduction capabilities of iterative reconstruction methods and shortening reconstruction times, according to the authors.
In their review, Willemink et al noted that DLR yields improved image quality compared with FBP and HBIR, as well as the potential of between 30% and 71% lower radiation dose compared with HIR. Thanks to its better noise reduction, DLR also maintains diagnostic image quality, according to the authors. What's more, they also concluded that deep learning-based metal artifact reduction could be more accurate than current techniques.
Looking back, looking forward
Two years ago Tim Leiner, MD, PhD, a radiologist and professor of cardiovascular radiology at Mayo Clinic, joined a distinguished group of experts at the National Heart, Lung, Blood Institute (NHLBI)'s workshop on AI in cardiovascular imaging.
Tim Leiner, MD, PhD.
Image acquisition and reconstruction were already capturing a significant share of algorithm development then with points of interest geared toward optimally managing image artifacts, critically evaluating the trade-offs of improving signal-to-noise versus losing spatial/temporal resolution, evaluating performance, and using AI-enabled acquisition in commercial products and considering what defines clinical viability.
Today, however, "image reconstruction is probably one of the areas furthest along," Leiner told AuntMinnie.com. "People are learning how to work with it, but I think most people are positively impressed with what you can do with functions in the equipment."
Additionally, Mayo Clinic is expanding its research into cardiovascular AI with the goal of developing a suite of algorithms that experts there believe will be clinically useful, Leiner said. The program addresses the heart and vasculature and complements Mayo Clinic's ongoing work in cardiac AI and AI coronary analysis.
"We're seeing now is that cardiac CT is a fantastic test to rule out the presence of flow-limiting coronary disease," Leiner said, adding that the future is not about single algorithms.
"It's about being able to orchestrate different sets of algorithms or sequential algorithms into something that makes sense from a clinical point of view," Leiner explained. "Plaque quantification, if you go five or six years ahead, you're going to see algorithms that can denoise the data, that can give you some aspect of image quality, that can identify motion artifacts, that can also quantify plaque.
"But also look at myocardial enhancement as a proxy for perfusion, cardiac chamber size, and then all of this will go into a quantitative report that will give you a holistic view of what's going on with the heart instead of just this one tiny aspect," Leiner concluded. "I think that's a future development. Ensembling this into a logical, complete package, if you will, is going to be the next frontier."
New quality measure for 2025
Efforts to advance low-dose CT and sufficient image quality are important, especially now that the U.S. Centers for Medicare and Medicaid Services (CMS) has called attention to monitoring the performance of diagnostic CT to discourage unnecessarily high radiation doses (See CMS Clinical Quality Measure 494, last updated June 3) while maintaining image quality.
The new quality measure, which was added to the Medicare Merit-Based Incentive Payment System (MIPS) program's Diagnostic Radiology measure set for 2025, suggests that radiologists track, optimize, and lower CT radiation doses to reduce potential cancer risks related to radiation dose in CT imaging. The Lown Institute called Measure 494 a victory for high-value care. While the quality measure has sparked concern as to its possible proprietary underpinnings and ambiguities, it's a measure CT imaging providers should still be watching.
AI can be incorporated into all steps of the cardiac imaging workflow, including scan ordering, patient scheduling, automated protocol generation, image acquisition and analysis, worklist prioritization and urgent results, report generation, and report communication, according to the authors of a 2023 overview of AI in cardiac CT that was compiled at the University of Louisville in Kentucky.
As CT image reconstruction AI finds its most effective use, radiologists, radiology administrators, and medical physicists will have choices to make that may involve commercial cardiac CT AI tools for evaluating heart and vascular flow dynamics, coronary plaque quantification, and other potential cardiovascular risk markers.