Does BMI influence AI and human reader lung nodule detection in low-dose chest CT?

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Body mass index (BMI) is known to increase image noise in low-dose computed tomography (LDCT), potentially impacting diagnostic accuracy.
• Evaluating the influence of BMI on lung nodule detection is crucial for understanding the performance of both human readers and AI algorithms.

Purpose of the Study:

• To assess the effect of high versus low body mass index (BMI) on lung nodule detection rates using artificial intelligence (AI) software and a human reader (HR) in low-dose computed tomography (LDCT) scans.
• To compare the sensitivity and false positive rates (FP/scan) between different BMI groups for both AI and human readers.

Main Methods:

• Chest LDCT scans from the Lifelines cohort were analyzed, focusing on participants in the highest (mean BMI 39.8) and lowest (mean BMI 18.7) BMI categories.
• Lung nodule detection was performed independently by AI software and a trained human reader.
• Discrepancies in detection were reviewed by two radiologists, with final disagreements resolved by an expert radiologist. Sensitivity and FP/scan were statistically compared between BMI groups for AI and HR.

Main Results:

• AI sensitivity was 0.75 in high BMI and 0.80 in low BMI groups (p=0.37), while HR sensitivity was 0.76 and 0.84, respectively (p=0.17).
• False positives per scan (FP/scan) for AI were 0.30 (high BMI) and 0.55 (low BMI) (p=0.005), and for HR were 0.05 (high BMI) and 0.16 (low BMI) (p=0.09).
• AI consistently demonstrated a higher FP/scan rate than the human reader in both BMI groups (p < 0.001).

Conclusions:

• Lung nodule detection sensitivity in LDCT is not significantly influenced by high versus low body mass index for either AI or human readers.
• Artificial intelligence software exhibited a higher rate of false positives per scan compared to human readers, irrespective of BMI.
• These findings suggest that while BMI does not impair detection sensitivity, AI's higher false positive rate warrants further investigation and potential optimization.