Purpose: Body mass index (BMI) can influence image quality in low dose computed tomography (LDCT) through higher image noise levels. We evaluated whether BMI affects lung nodule detection by artificial intelligence (AI) software and a human reader.
Method: The study utilized chest LDCT scans from the Lifelines cohort. We included 1.5 % participants at highest BMI (mean = 39.8, sd = 3.0), and 1.5 % at lowest BMI (mean = 18.7, sd = 0.9). Nodule detection was performed by AI software and by a trained human reader (HR). Two chest radiologists reviewed detection discrepancies, with disagreements resolved by an expert radiologist. Sensitivity and false positives per scan (FP/scan) were compared between BMI groups, for AI versus HR.
Results: There were 176 participants in both groups, with 131 nodules in high BMI, and 136 in low BMI. AI detected 356 nodular findings and HR 251, including 154 nodules found by both. AI's sensitivity was 0.75 (95 % confidence interval 0.66-0.82) in high BMI, and 0.80 (0.72-0.86) in low BMI groups (p = 0.37). FP/scan was 0.30 and 0.55 in high and low BMI, respectively (p = 0.005). HR's sensitivity was 0.76 (0.68-0.83) in high BMI, and 0.84 (0.76-0.89) in low BMI groups (p = 0.17), with FP/scan of 0.05 and 0.16, respectively (p = 0.09). In both BMI groups, AI had more FP/scan than the human reader (p < 0.001).
Conclusions: Sensitivity for lung nodule detection in LDCT was not significantly different for high versus low BMI, either for AI or human reader. Compared to the human reader, AI had higher FP/scan in both BMI groups.
Keywords: Body Mass Index; Detection; Pulmonary nodules; Software validation; Tomography (x-ray computed