Iterative Reconstruction: Ready for Its Close-up?
Enhanced image clarity, less noise, and half the radiation exposure for patients: sound good? It’s true. It’s iterative reconstruction, and it’s a mounting wave in multidetector CT technology. There’s only one catch—it’s time-consuming unless you happen to have the proprietary software of the single vendor that has come to market with a workflow-ready product at this point. 
W. Dennis Foley, MD, FACR W. Dennis Foley, MD, FACR, chief of digital imaging and a professor of radiology at the Medical College of Wisconsin in Milwaukee, has been championing iterative reconstruction since putting the technique to the test in a pilot demonstration in 2008. Foley and his colleagues judged iterative reconstructions against the same exams reconstructed in the more common way, using filtered back projection (FBP). FBP, Foley explains, involves “point-source radiation from a focal spot on the x-ray tube passing through as a single, thin line through the center of a voxel that goes to the center of each detector cell.” That’s what supposed to happen, anyway, but as Foley puts it, “That’s not the reality.” Instead, the x-ray beam is not a thin line, and the detector receives radiation across its face at various angles to the voxel, Foley says, noting that this can occur randomly. When this happens, image noise is created that must be overpowered by increasing the radiation exposure, so FBP uses a higher radiation dose than iterative reconstruction. FBP Noise “There’s a direct relationship between radiation dose and noise,” Foley says. “Noise is a statistical variation in the attenuation values not reflecting the underlying anatomy. Noise means that you don’t see a clear edge or clear contrast between two adjacent tissues. It gets blurred. Noise is a model that blurs anatomic features of an image.” Noise can happen because of variations in detector sensitivity, or it can occur through variations in electronics, Foley says. With FBP, dosage strength is used to minimize noise, but a new focus on patients’ radiation exposure from CT has placed a premium on being able to lower the dose—and this is what iterative reconstruction can do, without the gain in noise associated with FBP. “The problem is that most medical radiation the population receives now comes from CT, and the utilization of CT has increased markedly in the past 10 years,” Foley says. “There is some concern about excessive radiation of the population increasing the incidence of cancer, even though the risk is very slight.” The Software Approach While FBP takes a signal-to-detector approach to image creation, iterative reconstruction uses complex software, algorithms, mathematics, and modeled anatomical inputs to create the image by passing over the exam data from the modality again and again. It is the thousands of repeated iterations—each time, filling in data—that give the technique its name. Iterative-reconstruction software is continuously refining and building, using models and nearby anatomy to guess at missing data, until the image becomes the completed exam that the radiologist interprets. “Iterative reconstruction is done by a statistical modeling process. Lowering dose is more important than getting noise down,” Foley says. Nonetheless, he adds, iterative reconstruction gives both lower dose and equal or better structural detail in the completed image. Iterative reconstruction doesn’t have to overpower noise: It eliminates noise by refashioning data. What would be noisy in FBP is less noisy with iterative reconstruction. “The noise in the reconstructed image is fundamentally determined by the noisiest projection data,” Foley says. “If you can smooth out the noise differential by statistical modeling in projection data, you can make an image that is much smoother that still demonstrates the anatomy with equal fidelity.” ASIR and MBIR The first iterative-reconstruction software on the market that is fast enough to match the workflow of FBP, Foley says, is Adaptive Statistical Iterative Reconstruction (ASIR) from GE Healthcare, Waukesha, Wisconsin. The ASIR software works with any GE CT750HD scanner. With ASIR, most, if not all, CT studies can be completed at a lower radiation dose. Foley notes that ASIR is particularly useful for younger adults or people having repeated exams. It is also useful in studies looking for renal stones, in Crohn disease, for CT angiography, and for abdominal aortic and iliac studies in patients with stent grafts—all studies where low dose is especially beneficial, Foley says. What makes ASIR fast enough for everyday use is that its modeling only focuses on smoothing out noisy data. A more complete technique called model-based iterative reconstruction (MBIR) focuses on noise reduction and also corrects for system optics, Foley says. It is MBIR that holds promise for even further dose reductions and improved image clarity. The problem with MBIR is that, as Foley puts it, “MBIR is computationally dense.” In 2008, Foley and his research team did a study comparing a set of MBIR exams with the same exams reconstructed using FBP. It was a small study, ranking only 29 exams, but the results were striking, Foley says. “We did a pilot study with patients with low-dose CT scans of the upper abdomen as part of a multiphase evaluation of liver lesions or renal stones. We took other patients with IV contrast-enhanced scans of the abdomen and the chest,” Foley says. The exams included 17 abdomen/pelvis scans, one with a contrast agent; three chest scans, one with a contrast agent, and nine abdominal scans, one with a contrast agent. The FBP images and the iterative reconstructions were compared side by side. The exams were ranked by image quality using axial-, sagittal-, and coronal-plane displays, Foley says. The images were ranked from minus 5 to plus 5. “We looked at the usual image-quality factors: spatial resolution, noise, high and low contrast, uniformity parameters, and artifact level, such as streaks and beam hardening,” Foley says. “Iterative reconstruction was perceived superior in spatial resolution, noise suppression, and both high-contrast and low-contrast detail,” he adds. For unknown reasons—perhaps the way that the algorithms handle the data—the iterative-reconstruction studies were even more striking in the sagittal and coronal displays than in the axial displays. “Increasingly, radiologists are looking at coronal-plane displays in addition to axial ones, and it’s going to be an important issue to get good image quality in all three planes,” Foley notes. Drawbacks Iterative reconstruction is not a perfect technique. In the Foley group’s pilot study, iterative reconstruction created more artifacts than FBP. He adds that in the year since the study was done, most of the artifact bugs have been worked out, however. There are other tradeoffs with iterative reconstruction. As noted in the study, if it is used with the same radiation dose as FBP, iterative reconstruction gives superior noise reduction. If it is used at a lower dose, iterative reconstruction gives image quality equal to that of a higher-dose FBP, but the intensity of the iterative-reconstruction modeling process itself is another variable that can be controlled, and often needs to be, Foley says. That’s because when the iterative-reconstruction modeling process is tuned to its full capability, the images produced can look overly smoothed. “There are varying levels of the algorithm that are applied, and when you get to 100%, the image looks very smooth—waxy,” Foley says. “It’s a problem, so you use 40% to 50%.” There are concerns, he adds, that with very waxy reconstructions in iterative reconstruction, small lesions in solid organs or some instances of stenosis might not show up, but not enough research has been done to confirm this. Foley says, “Experience suggests 40% to 50% application produces good image quality at half the radiation dose.” Even at a lower capacity, the iterative-reconstruction images do take getting used to, Foley says. “What radiologists have assumed to be normal noise is not the normal noise on iterative reconstruction,” he notes. “It takes time to reset the expectations of what normal is.” Foley says that as speed in MBIR is improved, clinical trials will be needed to compare MBIR with ASIR. “Radiation dosage with ASIR can drop to 50% of FBP and retain the same diagnostic image quality. The expectation is that with MBIR, you should be able to go further, but that hasn’t been studied,” he says.George Wiley is a contributing writer for ImagingBiz.com.
W. Dennis Foley, MD, FACR W. Dennis Foley, MD, FACR, chief of digital imaging and a professor of radiology at the Medical College of Wisconsin in Milwaukee, has been championing iterative reconstruction since putting the technique to the test in a pilot demonstration in 2008. Foley and his colleagues judged iterative reconstructions against the same exams reconstructed in the more common way, using filtered back projection (FBP). FBP, Foley explains, involves “point-source radiation from a focal spot on the x-ray tube passing through as a single, thin line through the center of a voxel that goes to the center of each detector cell.” That’s what supposed to happen, anyway, but as Foley puts it, “That’s not the reality.” Instead, the x-ray beam is not a thin line, and the detector receives radiation across its face at various angles to the voxel, Foley says, noting that this can occur randomly. When this happens, image noise is created that must be overpowered by increasing the radiation exposure, so FBP uses a higher radiation dose than iterative reconstruction. FBP Noise “There’s a direct relationship between radiation dose and noise,” Foley says. “Noise is a statistical variation in the attenuation values not reflecting the underlying anatomy. Noise means that you don’t see a clear edge or clear contrast between two adjacent tissues. It gets blurred. Noise is a model that blurs anatomic features of an image.” Noise can happen because of variations in detector sensitivity, or it can occur through variations in electronics, Foley says. With FBP, dosage strength is used to minimize noise, but a new focus on patients’ radiation exposure from CT has placed a premium on being able to lower the dose—and this is what iterative reconstruction can do, without the gain in noise associated with FBP. “The problem is that most medical radiation the population receives now comes from CT, and the utilization of CT has increased markedly in the past 10 years,” Foley says. “There is some concern about excessive radiation of the population increasing the incidence of cancer, even though the risk is very slight.” The Software Approach While FBP takes a signal-to-detector approach to image creation, iterative reconstruction uses complex software, algorithms, mathematics, and modeled anatomical inputs to create the image by passing over the exam data from the modality again and again. It is the thousands of repeated iterations—each time, filling in data—that give the technique its name. Iterative-reconstruction software is continuously refining and building, using models and nearby anatomy to guess at missing data, until the image becomes the completed exam that the radiologist interprets. “Iterative reconstruction is done by a statistical modeling process. Lowering dose is more important than getting noise down,” Foley says. Nonetheless, he adds, iterative reconstruction gives both lower dose and equal or better structural detail in the completed image. Iterative reconstruction doesn’t have to overpower noise: It eliminates noise by refashioning data. What would be noisy in FBP is less noisy with iterative reconstruction. “The noise in the reconstructed image is fundamentally determined by the noisiest projection data,” Foley says. “If you can smooth out the noise differential by statistical modeling in projection data, you can make an image that is much smoother that still demonstrates the anatomy with equal fidelity.” ASIR and MBIR The first iterative-reconstruction software on the market that is fast enough to match the workflow of FBP, Foley says, is Adaptive Statistical Iterative Reconstruction (ASIR) from GE Healthcare, Waukesha, Wisconsin. The ASIR software works with any GE CT750HD scanner. With ASIR, most, if not all, CT studies can be completed at a lower radiation dose. Foley notes that ASIR is particularly useful for younger adults or people having repeated exams. It is also useful in studies looking for renal stones, in Crohn disease, for CT angiography, and for abdominal aortic and iliac studies in patients with stent grafts—all studies where low dose is especially beneficial, Foley says. What makes ASIR fast enough for everyday use is that its modeling only focuses on smoothing out noisy data. A more complete technique called model-based iterative reconstruction (MBIR) focuses on noise reduction and also corrects for system optics, Foley says. It is MBIR that holds promise for even further dose reductions and improved image clarity. The problem with MBIR is that, as Foley puts it, “MBIR is computationally dense.” In 2008, Foley and his research team did a study comparing a set of MBIR exams with the same exams reconstructed using FBP. It was a small study, ranking only 29 exams, but the results were striking, Foley says. “We did a pilot study with patients with low-dose CT scans of the upper abdomen as part of a multiphase evaluation of liver lesions or renal stones. We took other patients with IV contrast-enhanced scans of the abdomen and the chest,” Foley says. The exams included 17 abdomen/pelvis scans, one with a contrast agent; three chest scans, one with a contrast agent, and nine abdominal scans, one with a contrast agent. The FBP images and the iterative reconstructions were compared side by side. The exams were ranked by image quality using axial-, sagittal-, and coronal-plane displays, Foley says. The images were ranked from minus 5 to plus 5. “We looked at the usual image-quality factors: spatial resolution, noise, high and low contrast, uniformity parameters, and artifact level, such as streaks and beam hardening,” Foley says. “Iterative reconstruction was perceived superior in spatial resolution, noise suppression, and both high-contrast and low-contrast detail,” he adds. For unknown reasons—perhaps the way that the algorithms handle the data—the iterative-reconstruction studies were even more striking in the sagittal and coronal displays than in the axial displays. “Increasingly, radiologists are looking at coronal-plane displays in addition to axial ones, and it’s going to be an important issue to get good image quality in all three planes,” Foley notes. Drawbacks Iterative reconstruction is not a perfect technique. In the Foley group’s pilot study, iterative reconstruction created more artifacts than FBP. He adds that in the year since the study was done, most of the artifact bugs have been worked out, however. There are other tradeoffs with iterative reconstruction. As noted in the study, if it is used with the same radiation dose as FBP, iterative reconstruction gives superior noise reduction. If it is used at a lower dose, iterative reconstruction gives image quality equal to that of a higher-dose FBP, but the intensity of the iterative-reconstruction modeling process itself is another variable that can be controlled, and often needs to be, Foley says. That’s because when the iterative-reconstruction modeling process is tuned to its full capability, the images produced can look overly smoothed. “There are varying levels of the algorithm that are applied, and when you get to 100%, the image looks very smooth—waxy,” Foley says. “It’s a problem, so you use 40% to 50%.” There are concerns, he adds, that with very waxy reconstructions in iterative reconstruction, small lesions in solid organs or some instances of stenosis might not show up, but not enough research has been done to confirm this. Foley says, “Experience suggests 40% to 50% application produces good image quality at half the radiation dose.” Even at a lower capacity, the iterative-reconstruction images do take getting used to, Foley says. “What radiologists have assumed to be normal noise is not the normal noise on iterative reconstruction,” he notes. “It takes time to reset the expectations of what normal is.” Foley says that as speed in MBIR is improved, clinical trials will be needed to compare MBIR with ASIR. “Radiation dosage with ASIR can drop to 50% of FBP and retain the same diagnostic image quality. The expectation is that with MBIR, you should be able to go further, but that hasn’t been studied,” he says.George Wiley is a contributing writer for ImagingBiz.com.