Background: Emphysema influences the appearance of lung tissue in chest computed tomography (CT) scans. We evaluate the effect of emphysema presence on lung nodule detection by artificial intelligence (AI) software and by a human reader.
Methods: Patients were selected from a population-based cohort with low-dose chest CT, with non-emphysema patients matching in nodule size to emphysema findings. AI results for nodular findings sized 30-100mm3 and 101-300mm3 were compared to human reading, and two expert radiologists blindly reviewed discrepancies in detected findings. Sensitivity and false positives (FPs)/scan of AI and human reader were compared for emphysema and non-emphysema groups.
Results: 39 participants with at least moderate emphysema and 82 without emphysema were included (N=121, mean age 61±8 years, 48% men). AI and human reader detected 196 and 206 nodular findings, respectively, yielding 109 concordant nodules and 184 discrepancies (including 118 true nodules). For AI, sensitivity of nodule detection was 0.68 (95% CI 0.57-0.77) in emphysema vs 0.71 (95% CI 0.62-0.78) in non-emphysema groups. FPs/scan were 0.51 and 0.22, respectively (p=0.028). Human reader had a sensitivity of 0.76 (95% CI 0.65-0.84) in emphysema and 0.80 (95%CI 0.72-0.86) in non-emphysema group. FPs/scan were 0.15 and 0.27 respectively (p=0.23). Sensitivity of AI was not significantly different from the reader, but FPs/scan were higher for AI than the reader in emphysema (p=0.02), particularly for 101-300mm3 sized nodules.
Conclusion: Compared to non-emphysema, emphysema had more FPs/scan for AI, but not reader. Emphysema did not impact sensitivity of lung nodule detection for either AI software or reader.
Keywords: Pulmonary nodules, AI software validation, emphysema, bias, chest CT