Although an AI-based lung nodule evaluation tool used with low-dose CT (LDCT) helped radiologists find more clinically actionable nodules, it did not necessarily reduce interpretation time, according to a study published March 18 in the American Journal of Roentgenology.
The finding is surprising — and merits further research and AI refinement, wrote a team led by Eui Jin Hwang, MD, PhD, of Seoul National University College of Medicine in South Korea.
“[Our study] provides pragmatic evidence regarding the implications of AI-assisted LDCT interpretation,” the group noted, “[but] achievement of improved time efficiency will likely require further optimization of the AI tool’s integration into legacy system infrastructures.”
AI tools designed for automated lung nodule detection, classification, and measurement have shown promise in experimental settings, but evidence from actual clinical practice is limited, the authors wrote. They explained that prior research has relied on retrospective reader studies conducted outside of clinical practice – a framework which then raises questions about whether laboratory-observed efficiency benefits translate to real-world radiology workflows.
Hwang and colleagues conducted a study that included 911 asymptomatic adults who underwent LDCT as part of self-initiated general health checkups at Seoul National University Hospital between May and September 2025. The investigators randomized the participants into either an AI-assisted interpretation group (n=447) or a standard interpretation without AI group (n = 464).
For the intervention arm, the team used a commercially approved AI tool (Aview Lung Nodule CAD; Coreline Soft) that identified, classified, and measured nodules. Ten thoracic radiologists with one year to 25 years of postfellowship experience interpreted all the exams and reported only noncalcified nodules measuring 4 mm or larger. The study’s primary outcome was interpretation time per examination, while secondary outcomes included detection rates of Lung-RADS category 3 or 4 nodules, all nodules, and frequency of follow-up LDCT recommendations.
The investigators reported the following:
|
Performance of radiologist readers of LDCT exams with and without AI assistance |
|||
|
Measure |
No AI assistance |
AI assistance |
p-value |
| Interpretation time per exam |
172 seconds |
187 seconds |
0.23 |
| Frequency of follow-up LDCT recommendations |
7.4% |
15.3% |
0.04 |
|
Detection rate |
|||
| Lung-RADS-positive nodules |
10.3% |
16.9% |
0.03 |
| All nodules |
32.6% |
52.9% |
0.002 |
No individual in either group was diagnosed with lung cancer (median follow-up of 216 and 215 days in control and intervention groups, respectively), they wrote.
Hwang and colleagues suggested that the lack of time savings likely reflected “real-world constraints” on AI-PACS integration and the heightened scrutiny radiologists may apply when AI output directly influences patient management. They noted that, although AI increased detection of clinically meaningful nodules, it also produced more follow-up recommendations — prompting the question of downstream resource use in populations with low lung cancer prevalence.
Access the full study here.