Does BMI affect AI's accuracy when assessing CT scans?
New research is offering insight into how body mass index affects CT image quality, and whether artificial intelligence-enabled tools are more vulnerable to inaccuracies when analyzing images from certain patients.
A higher BMI can decrease scan quality, which increases the likelihood of interpretation errors for both humans and AI. The performance of AI, however, is only as effective as the data on which it was trained, and many datasets lack diversity in terms of patient BMI and body habitus.
Published in the European Journal of Radiology, new findings compare how BMI impacts the performance of both human and AI readers at interpreting low-dose CT (LDCT) cancer screenings.
“Participants with higher BMI have increased X-ray beam attenuation, and decreased image quality by more noise, particularly in low-dose CT. This may therefore decrease AI’s detection performance,” Nikos Sourlos, with the department of radiology at University Medical Center of Groningen in the Netherlands, and colleagues explained. “Prior studies have focused on the effect of BMI on human reader’s performance, although literature is limited to ultra-LDCT. These studies showed conflicting results, with some showing that BMI influences nodule detection performance and others showing no effect.”
For their work, researchers analyzed chest LDCTs from the Lifelines cohort, comparing the findings from participants with the highest BMI (top 1.5%) to those with the lowest (1.5% as well). An AI software and human reader assessed the scans for the presence of lung nodules; their performance was later evaluated by two chest radiologists who compared readers’ sensitivity and false positive rates per scan.
Each patient group contained a total of 176 participants. In the high BMI group, 131 nodules were identified, while 136 were spotted in the low BMI cohort. The AI detected more nodular findings than the human reader, but 154 nodules were spotted by both. The performance of both readers was similar regardless of patient BMI, with both yielding slightly lower sensitivity in patients with higher BMI. AI reads, however, resulted in higher rates of false positives per scan, more so in the low BMI group.
Though AI consistently kept pace with radiologists, the group suggested that its yield could be improved even further with additional training.
“One implication from our findings is the recommendation that in the development of a benchmark dataset intended for validating AI software for lung nodule detection, inclusion of both low and high BMI cases is important,” the authors advised. “It is also advisable to evaluate the performance of various commercial software packages in lung nodule detection in a diverse range of BMI participants to compare their performance, as was done in the work of Leeuwen et al. for chest radiographs.”
Read more about the findings here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.