A Bone to Pick
By Jeannette Sabatini
Vol. 23 No. 5 P. 18
When it comes to bone X-ray, recent advances give radiologists a break.
X-ray remains the MVP in the sport of diagnosing bone fracture, due to its ease of use, accessibility, and low exposure to patients, yet advances in technology and complementary AI software programs will soon up its game, making it possible to see more complicated fractures, prioritize images to be read, predict future fractures, and reveal bone health. Such advances will assist radiologists and all professionals who rely on X-ray to image bones, including those working in the hospital, emergency department (ED), urgent care, geriatric, orthopedic, and sports medicine settings, to name a few.
“X-ray is always good when looking for fractures, absolutely. Take two or three X-rays and you get the diagnosis,” says Ali Guermazi, MD, PhD, chief of radiology at VA Boston Healthcare System and a professor of radiology and medicine at Boston University School of Medicine. Integrate a user’s usual reading environment with AI software, and fractures that go unseen on X-ray are suddenly apparent and diagnoses can be made faster, explains the chief of radiology.
That experience includes Guermazi’s own AI validation studies as well as his involvement in the US clinical validation study for BoneView AI software, made by Gleamer, a French company. The BoneView algorithm, recently cleared for computer-assisted detection and diagnosis by the FDA, captures areas of suspected fracture within boxes that are visible on acquired X-rays. In the study, the application of AI to interpretations made by radiologists and nonradiologists lowered the false negative rate on X-rays by 29%, provided a 10.4% improvement of fracture detection sensitivity, and reduced reading time by 15%, the manufacturer notes. “This study shows that BoneView can improve sensitivity and specificity of readers, including radiologists and nonradiologists. It also can shorten the time of interpretation of the X-rays,” Guermazi says.
The AI software improved the specificity of fracture in numerous anatomical locations, including foot/ankle, knee/leg, hip/pelvis, hand/wrist, elbow/arm, shoulder/clavicle, rib cage, and thoracolumbar spine. “Detection is definitely a task where AI can complement the human eye, defeating cognitive biases and reducing medical errors, improving productivity, and prioritizing reads,” Guermazi says. “The radiologist is human. He sees a long day and may be tired. The fact that there is some kind of help is always appreciated.”
AI provides fast results, Guermazi notes. “AI software use can keep a radiologist from spending time going back and forth on a diagnosis, which can happen when they have been working long hours. That’s important,” he says. “Time is money, and, instead of having a patient sitting in the [ED] for hours waiting for the X-ray result, you can have a report coming very quickly and that will get the patient onto their next step in treatment and the doctor in the [ED] going right to the next patient, not wasting time and energy of their team.”
AI software can also let a radiologist know what images should be read first, the chief radiologist says. It can bring the most urgent cases to the top of a pile of hundreds of X-rays. “That’s extremely important,” Guermazi says, especially since radiology is seeing a daily increase in workload. “AI can help cope with this increased load by prioritizing reads and improving reading times. Whether you are in the [ED] or in the reading room of hospitals, academic or otherwise, one is always overwhelmed by X-ray. The fact that this software will put a hierarchy on the X-rays is really extremely important.”
The seven-month study of BoneView AI, completed in January 2021, involved a review of images acquired from multiple US centers, on instruments from a wide variety of manufacturers, Gleamer reports. Interpretations were made by radiologists and nonradiologists from Boston University School of Medicine, Stony Brook University Renaissance School of Medicine, and Massachusetts General Hospital-Harvard Medical School. In May, Gleamer announced that it is aligning with Aidoc, a provider of health care AI solutions with bases in New York and Israel, to integrate BoneView.
Radiologists are “excited” about the use of AI software to image bone, Guermazi says, anticipating that professionals will view it as a friend rather than a foe. “As it becomes clearer that it can beat the human eye at certain specific and repetitive or tedious tasks, AI will be viewed as a great add-on to heavy clinical workflow.”
Studies from England are also promising. There, a group of researchers put AI’s sensitivity for detecting fractures at 91% to 92%, based on a review of 42 existing studies (37 X-ray and five CT) that compared the diagnostic performance in fracture detection between AI and clinicians. The study reports: “There were no statistically significant differences between clinician and AI performance.”
The review, conducted between January 2018 and July 2020, involved a systematic assessment and meta-analysis of peer-reviewed publications and literature. For internal validation test sets, the pooled sensitivity was 92% for AI and 91% for clinicians, and the pooled specificity was 91% for AI and 92% for clinicians. For external validation test sets, the pooled sensitivity was 91% for AI and 94% for clinicians, and the pooled specificity was 91% for AI and 94% for clinicians.
“We think that many clinical settings could benefit from the use of AI to diagnose fractures,” says study lead Rachel Kuo, MB BChir, MA (Cantab), MRCS, a National Institute for Health and Care Research academic clinical fellow in plastic and reconstructive surgery at Oxford University Hospitals National Health Service Foundation Trust in the United Kingdom. “Clinicians may benefit through having a ‘second reader’ to help them avoid error or to give them more confidence in their own diagnosis.”
Fracture-detection AI could reduce misdiagnosis at the point of care, according to Kuo. “This could be particularly helpful for difficult fracture types, such as scaphoid fractures, or in cases of polytrauma where there are multiple fractures.” The technology also shows potential for educational applications in fracture detection, she adds.
Another positive finding stemmed from the fact that the sources gathered came from various countries, including the United States, Europe, Southeast Asia, and Australia. “We were encouraged to see that AI appears to perform equally well on external validation—that is, data that is collected independently. This suggests that AI might be generalizable across populations, and so, for example, a fracture-detection AI developed in Europe might be useful in Australia. This could help patients by reducing rates of misdiagnosis and streamlining their treatment,” Kuo explains.
In addition to diagnosing fractures, AI software can predict them, according to studies spotlighted by RSNA. UK-based Ibex has developed AI software that reveals areas of poor bone health and osteoporosis, which represent areas that are susceptible to future fracture. The software, called Trueview, highlights areas of potential future fractures on a patient’s original X-ray.
According to results published in the British Medical Journal, Trueview proved as good a predictor of osteoporosis as the most commonly used technique for assessing osteoporosis, dual-energy X-ray absorptiometry (DXA). Trueview is designed to be added to a facility’s existing digital X-ray machine and can ready a report with highlighted areas at the same time that the initial X-ray is delivered.
Likewise, a research team out of Korea has factored AI into a model that can automatically diagnose osteoporosis from hip X-rays, according to an article in Radiology: Artificial Intelligence. The “deep-radiomic model” was developed from nearly 5,000 hip X-rays from 4,308 patients obtained over more than 10 years, according to the study. The team, led by Hee-Dong Chae, MD, from the department of radiology at Seoul National University Hospital in Korea, reports that the model could serve as a triage tool recommending DXA in patients with highly suspected osteoporosis.
More Work to Do
AI would surely benefit radiologists and nonradiologists alike, yet it may be some time before they see it within their workplace, Guermazi says. AI software has been used for over a year in Europe—notably in Spain, Italy, Portugal, France, Germany, and Switzerland—but its implementation in the United States still has some hurdles to jump. One of those pertains to patient privacy issues that come into play when the software is implemented on a workstation; the United States is more stringent with such policies than other countries. “We cannot use this software as standalone, and integration can take a bit of time. It’s always a challenge to implement something outside the workstation,” he says.
Both Guermazi and Kuo agree that AI software is in its infancy in the United States. “Ultimately, in a year from now, most of the hospitals in the US will be having this kind of software,” Guermazi says. From their studies, Kuo and her team gathered that further studies are needed. “We found that many studies had significant methodological flaws and shortfalls in transparent reporting,” she says. “It is important that any AI should be evaluated rigorously in reader studies and in prospective clinical trials to ensure patient safety.”
Among the studies Kuo suggests are those that would evaluate the impact of integrating fracture-detection AI in clinical practice. The studies included in her team’s review only tested clinicians in an artificial study setting, asking clinicians to review X-rays selected by the researchers. Only one study evaluated their algorithm in prospective clinical practice. “Our review did not identify any AI in routine clinical practice for fracture diagnosis,” she says. “There is a gap between research and implementation, and research evaluating the impact of integrating fracture-detection AI in clinical practice is urgently required.”
More qualitative work should assess clinician and patient attitudes toward AI as a diagnostic adjunct, compared to an independently functioning AI. “Currently, most research groups prioritize the development of ‘human-in-the-loop’ AI, in which clinicians use AI results to make decisions about patient care,” she notes. “Interestingly, only a minority of studies evaluated the performance of clinicians with AI assistance.”
The true cost benefits of fracture-detection AI also should be evaluated. None of the studies reviewed by Kuo’s team integrated an evaluation of cost. “Accurate AI may reduce direct costs to hospitals through streamlining patient pathways and making more efficient use of radiologist time, and indirect costs to patients who may need more time off work to attend hospital appointments,” she says. “However, this has not yet been investigated for fracture detection.”
A Sporting Chance
Overall, X-ray is not sitting back when it comes to bone imaging and is showing its worth in different settings, such as sports medicine. This past spring, the Women’s Tennis Association (WTA) put DXA in the game to help assess the bone health of athletes participating in the BNP Paribas Open in Indian Wells, California. “DXA provided information on player bone health by measuring the bone mineral density in the spine and hip. This information was compared to the normal bone density range of the general population,” explains Kathleen Stroia, MS, PT, ATC, senior vice president of sport sciences and medicine at the WTA. The results were used to guide recommendations for the players.
DXA technology utilizes low-dose X-ray to assess bone density and body composition, according to Joe Joyce, human performance specialist at Hologic, which manufactures the Horizon DXA system used on site at the event. During scanning, he explains, the system’s C-arm passes over the supine subject while the pulsing X-ray generator emits a true linear fan beam captured by the GADOX detectors, delivering real-time pixel-by-pixel calibration through bone and tissue equivalents.
Clinically, DXA is being used for lumbar spine and proximal femur studies, bone mineral density measurements of the entire skeleton, atypical femur fracture assessment, and many other clinical applications, Joyce notes. It served to screen player health at the event, yet Stroia saw its potential for assessing injured patients. “The technology was used as a screening tool to establish baseline values of bone mineral density in WTA athletes at the peak of their physical performance,” she says. “It could potentially also be used to provide predictive models for injury risk factors.”
During the event, Hologic’s team had the chance to work alongside health professionals within the field of sports, reflecting the benefits advanced X-ray technology has outside of the clinical setting. “Our radiology technologists worked alongside the WTA physicians, athletic trainers, and sports scientist to provide DXA scans for assessing body composition and bone density that helped tailor training and nutritional regimens,” Joyce says.
Stroia adds: “This was the first time our physicians and sport dietitians had the benefit of bone health and body composition measurements to make more complete and specific dietary, health, and lifestyle intervention plans. This data provides our team with a more complete health picture of our players.”
Overall, any advance in X-ray technology serves to complement the radiologist, rather than the other way around. Only a clinician can consider all the factors related to a patient when it comes to making a diagnosis and determining the next step for a patient, Guermazi says. Providing an example, he explains that AI is unable to distinguish bone abnormalities that are caused by cancer, whereas a radiologist has the ability to consider all of the factors and move the patient on for cancer therapy. “The software proposes the diagnosis, and the radiologist will decide if it is true or not. It is the signature of the radiologist, not the software,” he says.
— Jeannette Sabatini is a freelance writer based in Malvern, Pennsylvania.