Objective:

To compare the accuracy of an AI model and experienced radiologists in estimating pleural effusion volume on chest radiography.

Approach:

• Study Design: Retrospective observational study of 88 adult patients who underwent chest radiography and computed tomography (CT) on the same day.
• Participants: Patients with CT-confirmed pleural effusion of at least 20 mL in 1 hemithorax.
• Methods: Two radiologists estimated pleural fluid volume on radiographs while blinded to CT measurements. An AI model analyzed the same radiographs.
• Evaluation: Agreement between CT-derived volumes and estimates from radiologists and AI was assessed, along with estimation error and identification of clinically significant effusions.

Key Findings:

• Moderate agreement with CT across all three approaches (radiologists and AI).
• Mean estimation errors were 300 mL for radiologist 1, 250 mL for radiologist 2, and 264 mL for AI.
• Positive bias in estimates: 103 mL for radiologist 1, 146 mL for radiologist 2, and 57 mL for AI.
• Area under the curve (AUC) for identifying pleural effusions of at least 300 mL was 0.84 for radiologist 1, 0.90 for radiologist 2, and 0.85 for AI.
• AI showed high within-session consistency but lower between-session consistency.

Interpretation:

The wide limits of agreement observed across all evaluators in this retrospective cohort confirm that chest radiography provides only a coarse approximation of pleural fluid volume, irrespective of the interpreter.

Limitations:

• Retrospective design and single institution study limit generalizability.
• Cohort selection enriched for positive cases, introducing spectrum bias.
• CT volumetry used as reference standard rather than actual drained fluid volume.
• Only one AI model evaluated, which may change with software updates.

Conclusion:

The findings indicate that both AI and radiologists have similar estimation capabilities for pleural effusion volume, but substantial variability exists.

Sources:

• European Radiology Experimental