Quantitative estimation of pleural effusion volume on chest radiography: comparison of radiologists and a general-purpose multimodal AI model using CT as the reference standard.
Authors
Affiliations (3)
Affiliations (3)
- Department of Medical Imaging, Motol and Homolka University Hospital and Second Faculty of Medicine, Charles University, Prague, Czech Republic.
- Department of Pneumology, Motol and Homolka University Hospital and Second Faculty of Medicine, Charles University, Prague, Czech Republic.
- Department of Medical Imaging, Motol and Homolka University Hospital and Second Faculty of Medicine, Charles University, Prague, Czech Republic. [email protected].
Abstract
Chest radiography provides limited precision of pleural effusion volume estimates. This study compares volume estimates from experienced radiologists and a multimodal artificial intelligence (AI) model, using computed tomography (CT)-derived volumetry as the reference standard. In this retrospective observational study, same-day chest CT and radiographs from adult patients with CT-confirmed pleural effusion (≥ 20 mL) were evaluated. Manually segmented pleural effusion volumes from CT served as the reference standard. Two board-certified radiologists independently estimated left and right pleural fluid volume on radiographs. A general-purpose multimodal AI model analyzed each radiograph using a standardized prompt and generated volume estimates for each hemithorax. Eighty-eight patients (72.0 ± 14.4 years (mean ± standard deviation); 56% male) were included, and 176 hemithoraces were analyzed. Median CT-derived pleural effusion volume was 389 mL (interquartile range 188-703 mL). Correlation with CT volumes was moderate for both radiologists (intraclass correlation coefficient 0.64 and 0.74) and the AI model (0.66). Mean absolute error was 300 mL (95% confidence interval 252 to 356 mL) for radiologist #1, 250 mL (216 to 285 mL) for radiologist #2, and 264 mL (227 to 299 mL) for AI. For detection of effusions ≥ 300 mL, AUC values were 0.84 (radiologist #1), 0.90 (radiologist #2), and 0.85 (AI). Quantitative estimation of pleural effusion volume on chest radiography demonstrates only moderate agreement with CT, regardless of whether the interpretation is performed by experienced radiologists or a general multimodal AI model. This limitation appears intrinsic to the imaging modality. Question Can a general-purpose multimodal AI model estimate pleural effusion volume on chest radiographs with accuracy comparable to experienced radiologists, using CT volumetry as reference? Findings Across 88 patients (176 hemithoraces), radiologists and AI showed comparable moderate agreement with CT (mean absolute errors 250-300 mL; intraclass correlation coefficient 0.64-0.74); chest radiography is intrinsically imprecise. Relevance statement Chest radiography supports only coarse pleural effusion volume estimation regardless of interpreter; however, a general-purpose multimodal AI model achieved accuracy comparable to that of experienced radiologists.