Body mass index can affect image quality in low-dose chest computed tomography by increasing image noise, raising questions about whether lung nodule detection remains consistent across different patient groups. A recent analysis published in the European Journal of Radiology evaluated artificial intelligence software and a trained human reader using chest scans from a population imaging cohort. The comparison focused on participants at the highest and lowest ends of the body mass index range and assessed lung nodule detection sensitivity and false positives per scan. The results indicate no significant difference in sensitivity between high and low body mass index groups for either artificial intelligence or human reading, while artificial intelligence produced more false positives per scan than the human reader in both groups.
Detection Performance Across Body Mass Index Groups
The dataset came from a population imaging cohort that included more than 12,000 individuals aged 45 years and above who underwent chest low-dose computed tomography between 2017 and 2022. The comparison selected participants at the highest and lowest ends of the body mass index distribution. Cases with a high number of detected nodules were excluded.
The final analysis included equal-sized high and low body mass index groups, with 176 participants in each. Both groups had a similar total number of nodules, with slightly more nodules in the low body mass index group. Mean body mass index was around 40 in the high group and below 19 in the low group. The overall participant group was mostly female, and more than half were ever smokers.
Image noise was higher in scans from the high body mass index group. Computed tomography dose was also higher in this group. These differences reflect the imaging challenge associated with higher body mass index, while the detection results allow comparison of whether that challenge affected artificial intelligence and human reader performance.
Artificial Intelligence and Human Reader Results
Artificial intelligence detected more nodular findings than the human reader. A substantial number of nodules were detected by both approaches, while discrepant findings were assessed by chest radiologists. When the two radiologists disagreed on the nature of a finding, a third radiologist established the final categorisation.
Must read: Continuous Fields Improve LDCT Nodule Segmentation
Artificial intelligence sensitivity was slightly lower in the high body mass index group than in the low body mass index group, but the difference was not statistically significant. The same pattern appeared for the human reader, with somewhat lower sensitivity in the high body mass index group and no significant difference between groups.
False positives showed a clearer difference. Artificial intelligence produced more false positives per scan in the low body mass index group than in the high body mass index group. The human reader also had more false positives per scan in the low body mass index group, although the difference was not significant. In both body mass index groups, artificial intelligence produced more false positives per scan than the human reader.
Overall, artificial intelligence and the human reader had similar sensitivity for lung nodule detection across the body mass index groups. The main distinction was not sensitivity, but the higher false-positive burden associated with artificial intelligence.
Subgroup Findings and Validation Needs
Additional analyses examined detection after excluding typical perifissural nodules and bronchovascular lymph nodes. These findings are benign and do not require follow-up below the stated volume threshold. After their exclusion, artificial intelligence sensitivity increased in both body mass index groups and was highest in the low body mass index group. Human reader sensitivity changed less after these exclusions.
A further subanalysis focused on participants at the most extreme body mass index values. Artificial intelligence sensitivity remained similar between the highest and lowest body mass index groups. Human reader sensitivity was lower in the highest body mass index group than in the lowest group. Artificial intelligence still had more false positives per scan than the human reader in both groups.
Several limitations affect interpretation. Nodules detected by both artificial intelligence and the human reader were treated as true positives and were not reviewed by radiologists, so some false positives may have been included. Both approaches may also have missed nodules. Body mass index may not fully reflect tissue distribution around the thorax. The proportion of women was also higher in the selected body mass index groups than in the whole cohort, which may have influenced thoracic image quality and nodule detection.
Sensitivity for lung nodule detection in low-dose chest computed tomography was not significantly affected by low or high body mass index for artificial intelligence or the human reader. Artificial intelligence produced more false positives per scan than the human reader in both body mass index groups, with a higher false-positive rate in the low body mass index group. The results support inclusion of both low and high body mass index cases when building benchmark datasets for validating artificial intelligence software in lung nodule detection.
Source: European Journal of Radiology
Image Credit: iStock
References:
Sourlos N, van Tuinen M, Sidorenkov G et al. (2025) Does BMI influence AI and human reader lung nodule detection in low-dose chest CT? European Journal of Radiology, 193:112453.
