Back to the list
Congress: ECR24
Poster Number: C-17087
Type: EPOS Radiologist (scientific)
DOI: 10.26044/ecr2024/C-17087
Authorblock: F. Souschek, P. Mildenberger, L. Müller, T. Jorg, M. C. Halfmann; Mainz/DE
Disclosures:
Fabio Souschek: Nothing to disclose
Peter Mildenberger: Nothing to disclose
Lukas Müller: Nothing to disclose
Tobias Jorg: Nothing to disclose
Moritz Christian Halfmann: Nothing to disclose
Keywords: Artificial Intelligence, Thorax, Plain radiographic studies, Comparative studies, Quality assurance
Purpose Artificial intelligence is a rapidly evolving tool in radiology that can significantly increase efficiency and decrease costs. However, the diagnostic accuracy of AI tools across all population subgroups is essential for clinical usage. Recent studies have raised doubts about the generalizability of artificial intelligence (AI) tools in clinical applications outside their narrow training corridors [1]. Therefore, this study aimed to compare the diagnostic performance of AI between supine chest X-rays of intensive care unit patients and posteroanterior (PA) radiographs.
Read more Methods and materials In this retrospective study, a commercially available AI tool analysed 151 supine chest X-rays from ICU patients and 70 PA view chest X-rays from patients at a tertiary care hospital (Figure 1). [fig 1] Reports from board-certified radiologists served as a reference standard. Diagnostic accuracies for detecting lung opacities, consolidation, pleural effusion, pulmonary edema, pneumothorax, and cardiomegaly were compared by means of sensitivity, specificity, positive predictive values (PPV), and negative predictive values (NPV).
Read more Results Despite generally lower diagnostic performance in supine chest radiographs compared to PA radiographs, sensitivity and specificity were moderate to high in both groups for detecting lung opacities (85% vs. 92%; 63% vs. 95%), consolidation (80% vs. 75%; 81% vs. 97%), and pleural effusion (77% vs. 87%; 82% vs. 96%, respectively).  Lower diagnostic accuracy was observed in detecting pulmonary edema (59% vs. 25%; 61% vs. 100%), pneumothorax (0 % vs. 20%; 100% vs. 100%), and cardiomegaly (21% vs. 46%; 75% vs....
Read more Conclusion While diagnostic accuracy remained moderate to high for detecting opacities, consolidation, and pleural effusion in supine chest X-rays of ICU patients, the performance in detecting pneumothorax and cardiomegaly falls short of clinical applicability. Additionally, supine chest X-rays showed markedly lower positive and negative predictive values with a broader range, even though the prevalence of target findings was considerably higher in ICU patients. This highlights the importance of rigorous evaluation of AI tools prior to their application in new clinical settings.
Read more References 1. Huisman M, Hannink G. The AI Generalization Gap: One Size Does Not Fit All. Radiol Artif Intell. 2023 Aug 30;5(5):e230246. doi: 10.1148/ryai.230246. PMID: 37795134; PMCID: PMC10546357.
Read more
GALLERY