Integrating an artificial intelligence (AI) support platform into routine radiology practice saves radiologists about an hour a day interpreting chest CT scans compared to reading the scans without it, a randomized study suggests.
“Radiology is one of the first possible applications of AI simply because we are digitally based, which is something that a computer can easily read and analyze,” Joseph Schoepf, MD, from the Medical University of South Carolina, Charleston, South Carolina, told Medscape Medical News.
“So radiology is indeed one of the first fields that I think will benefit greatly from AI,” he said. He noted that this is the first study to assess the impact of an AI support platform on chest CT interpretation times in a real-world clinical setting.
The study was recently published online in the American Journal of Roentgenology (AJR).
Training on the AI Platform
Prior to the study start date, three cardiothoracic radiologists received at least 30 days of training on how to use the AI platform. All the scans were performed and interpreted as part of real-world clinical practice. The final sample consisted of 390 chest CT scans performed for 390 patients; 195 of those scans were interpreted in the AI-assisted arm, and 195 were interpreted in the non-AI-assisted arm. Each reader interpreted 65 scans in each arm.
A total of 190 scans were performed without IV contrast material; 200 were performed with IV contrast material. “For each reader, the mean interpretation time was significantly shorter in the AI-assisted than in the non-AI [assisted] arm,” the authors report.
For reader 1, the time spent reading a scan was 289 seconds with the help of AI, vs 344 seconds without AI help (P < .001). For reader 2, the time spent reading a scan with the help of AI was 449 seconds, vs 649 seconds without AI help (P < .001). For the third reader, the time spent reading a scan with AI help was 281 seconds, vs 348 seconds without AI assistance (P = .01). From pooling data from the three readers, the mean interpretation time was 328 seconds in the AI-assisted arm, compared with 412 seconds in the non-AI-assisted arm (P < .001); the mean difference in interpretation time was 93 seconds (95% CI, 63 – 123 seconds).
This corresponds to a 22.1% reduction in time to interpret a scan in favor of the AI-assisted arm (95% CI, 14.9% – 29.2%; P < .001), the authors point out.
The same difference in favor of the AI-assisted arm was seen in contrast-enhanced scans and in non–contrast-enhanced scans; in negative scans and positive scans; and in positive scans without new findings and in those with new findings (Table 1).
Table 1. Mean Difference and Percent Reduction in Interpretation Times
|scans||Mean difference in favor of the AI-assisted arm||Percent reduction in interpretation times in favor of the AI-assisted arm||P value|
|Contrast-enhanced scans||83 sec||20%||<.001|
|Non–contrast-enhanced scans||104 sec||24.2%||<.001|
|Negative scans||84 sec||26.2%||=.001|
|Positive scans without new findings||117 sec||25.7%||-.001|
|Positive scans with new findings||92 seconds||20.4%||<.001|
CT Scans Doubled
The investigators note that from 2000 to 2016, the number of chest CT scans carried out for adults in the United States more than doubled. “However, the number of practicing radiologists has not kept pace with the growth of imaging utilization; the resulting mismatch is leading to increased workloads per practicing radiologist and subsequent increases in burnout,” they state. Novel solutions to reduce the burden of repetitive tasks are called for to lighten the workload of these practitioners.
Asked whether the learning curve might make the adoption of AI support platforms challenging in less expert centers, Schoepf said that it was easy to learn. “The learning curve was actually pretty short because the output of the particular system we use (AI-Rad Companion; Siemens Healthineers) is pretty much a one-pager and it’s pretty intuitive,” he elaborated.
The software provides automated image analysis, quantification, and visualization of structures on CT scans and includes cardiac, pulmonary, and musculoskeletal modules. With the use of these modules, the application detects and segments lung lesions, provides the number and location of lesions as well as measurements of lesion size and volume, quantifies thoracic aortic diameters, and calculates coronary calcium volumes, in addition to other functions.
“The number one thing I believe that AI does right now for radiology is that it relieves us of a lot of quantification tasks that take a lot of time,” said Schoepf.
For example, coronary artery calcium scoring is currently performed manually by circling every calcification within the coronary artery tree, which takes time. “The computer basically just spits out a number that saves a lot of radiology hours to quantify,” Schoepf added.
Similarly, measuring the size of the aorta is normally a relatively arduous task in that a radiologist has to measure dimensions of the aorta in a 3-dimensional space, which is both challenging and time consuming. “The computer basically does all that for us. And if we perform a surveillance scan, for example, to see if an aortic aneurysm has grown over time, the computer just makes the task a whole lot easier,” Schoepf said.
Although radiologists do look at the entire imaging study, they are still prone to “zoom in” on the question they are being asked to answer: for instance, does this patient have a pulmonary embolism, yes or no? Naturally, the radiologist who is asked this question is going to focus in on the pulmonary arteries to see whether there is any sign of a clot, but by becoming so focused on one organ system, it’s easy to miss things that are outside of the immediate field of interest.
“The computer does a very good job in finding pathologies we are not necessarily looking for in the first place,” said Schoepf. “I think a lot of people will recognize the benefits of AI, and in time, when people have embraced the idea of integrating AI into their practice, we will see much more widespread adoption of it.”
Some Promise Here
Asked by Medscape Medical News to comment on the findings, Mikael Hammer, MD, from Harvard Medical School, in Boston, Massachusetts, agreed with Schoepf that it’s fairly easy to apply AI in radiology compared to many other fields.
“I think there is a promise here, although we’ve been promised a lot over the years, and we are still waiting to see things come to fruition,” he said.
On the other hand, Hammer noted that it’s not clear from the article exactly why interpretations were faster by using this software. Presumably it’s because radiologists were copying the measurements from the software directly instead of manually doing the measurements themselves, he said.
“I think what we are really looking forward to is the next step in automation, where the software can directly input measurements back into your report, so that has the potential to be even more efficient. When the computer does the measurements, you verify them in some way, and then the measurements just go automatically into your report,” Hammer elaborated.
“So that’s what we are looking forward to happening at some point and that has the potential to be even more helpful, and we are looking forward to continued refinements of this,” he added.
The research did not receive any specific grant from funding agencies in the public, commercial or not-for-profit sectors. Schoepf and Hammer have disclosed no relevant financial relationships.
Am J Roentgenol. Published online June 8, 2022. Full-text