AI support platform saves time in routine radiology practices

By integrating an artificial intelligence (AI) support platform into routine radiology practice, radiologists will save about an hour a day interpreting chest CT scans compared to reading the scans without it, a randomized study suggests.

dr. Joseph Schoepfa

“Radiology is one of the main potential uses of AI simply because we are digitally based, something that a computer can easily read and analyze,” said Joseph Schoepf, MD, of the Medical University of South Carolina, Charleston, South Carolina. Medscape Medical News

“So radiology is indeed one of the first areas that I think will benefit a lot from AI,” he said. He noted that this is the first study to assess the impact of an AI support platform on breast CT interpretation times in a real clinical setting.

The study was recently published online in the American Journal of Roentgenology (AJR).

Training on the AI ​​platform

Prior to the start date of the study, three cardiothoracic radiologists received at least 30 days of training on how to use the AI ​​platform. All scans were performed and interpreted as part of real world clinical practice. The final sample consisted of 390 chest CT scans performed on 390 patients; 195 of those scans were interpreted in the AI-assisted arm and 195 were interpreted in the non-AI 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] poor,” the authors report.

For reader 1, the time spent reading a scan was 289 seconds using AI, versus 344 seconds without AI assistance (p < .001). For reader 2, the time spent reading a scan using AI was 449 seconds, versus 649 seconds without AI assistance (p < .001). For the third reader, the time spent reading a scan with AI assistance was 281 seconds, versus 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 to 412 seconds in the non-AI 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 note.

The same difference in favor of the AI-assisted arm was seen in contrast-enhanced scans and in non-contrast-enhanced scans; on 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 percentage reduction in interpretation times

scans Average difference in favor of the AI-assisted arm Percent reduction in interpreting times in favor of the AI-assisted arm p where the
Contrast Enhanced Scans 83 seconds 20% <.001
Scans without contrast 104 seconds 24.2% <.001
Negative scans 84 seconds 26.2% =.001
Positive scans with no new findings 117 seconds 25.7% -.001
Positive scans with new findings 92 seconds 20.4% <.001

CT scans doubled

The researchers note that from 2000 to 2016, the number of CT scans of the chest for adults in the United States more than doubled. “However, the number of practicing radiologists has not kept pace with the growth in imaging use; the resulting mismatch leads to increased workload per practicing radiologist and subsequent increases in burnout,” they argue. New solutions are needed to reduce the burden of repetitive tasks to ease the workload of these practitioners.

When asked whether the learning curve could make the adoption of AI support platforms a challenge in less expert centers, Schoepf said it was easy to learn. “The learning curve was actually quite short because the output of the particular system we’re using (AI-Rad Companion; Siemens Healthineers) is pretty much a one-pager and it’s pretty intuitive,” he explained.

The software provides automated image analysis, quantification and visualization of structures on CT scans and includes cardiac, lung 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 the size and volume of the lesion, quantifies the diameter of the thoracic aorta and calculates coronary calcium volumes, in addition other functions.

“The most important thing I think AI is doing for radiology right now is that it relieves us of a lot of time-consuming quantification tasks,” says Schoepf.

For example, the coronary artery calcium score is currently performed manually by circling each calcification in 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.

Likewise, measuring the size of the aorta is normally a relatively arduous task because a radiologist must measure the dimensions of the aorta in three-dimensional space, which is both challenging and time-consuming. “The computer basically does all of that for us. And if we do, say, a surveillance scan to see if an aortic aneurysm has grown over time, the computer just makes the job a lot easier,” Schoepf said.

Even though radiologists look at the entire imaging study, they are still prone to “zooming in” on the question they need to answer: For example, does this patient have a pulmonary embolism, yes or no? Of course, the radiologist asked this question will focus on the pulmonary arteries to see if there’s any sign of a clot, but by becoming so focused on one organ system, it’s easy to miss things that are out of the immediate field of view. interest.

“The computer does a really good job of finding pathologies that we’re not necessarily looking for,” says Schoepf. “I think a lot of people will recognize the benefits of AI, and eventually, when people have embraced the idea of ​​integrating AI into their practice, we’ll see it be more widely adopted.”

A promise here

Asked by Medscape Medical News to comment on the findings, Mikael Hammer, MD, of Harvard Medical School, in Boston, Massachusetts, agreed with Schoepf that it is quite easy to apply AI in radiology compared to many other fields.

“I think there’s promise here, although a lot has been promised to us over the years, and we’re still waiting to see things come to fruition,” he said.

On the other hand, Hammer noted that it is not clear from the article why interpretations were faster using this software. Presumably, it’s because radiologists copied the measurements directly from the software rather than taking the measurements manually themselves, he said.

“I think we’re really looking forward to the next step in automation where the software can feed measurements right back into your report, so that has the potential to be even more efficient. When the computer does the measurements, you verify them on the one hand.” somehow, and then the measurements just automatically go into your report,” Hammer explained.

“So that’s what we look forward to happening at some point and that has the potential to be even more useful, and we look forward to further refinements of this,” he added.

The research has not received specific grants from funding agencies in the public, commercial or non-profit sectors. Schoepf and Hammer have not disclosed any relevant financial relationships.

Ben J Roentgenol. Published online June 8, 2022. Full text

For more news, follow Medscape on Facebook, TwitterInstagram and YouTube.

Leave a Comment