Efficacy of a deep learning-based software for chest X-ray analysis in an emergency department.
Data
404 patients included in the study (74.5% of exams didn’t have any of the ChestView abnormalities)
295 had only CXRs
109 had CXRs + subsequent CT exam
Design
Retrospective single-center study
3 emergency physicians, 1 junior radiologist and 1 general radiologist independently reviewed the CXRs (with available clinical information)
Without AI → Washout ≥ 2 weeks → With AI
Ground truth = consensus between 2 experienced radiologists (with all available clinical and imaging information)
Results
Standalone performance [see study]
Readers
DL-assisted reading had a significantly higher combined sensitivity compared to unassisted reading
No significant change in the combined specificity between unassisted and assisted readings
Most important increase in sensitivity with DL assistance was obtained for nodules (38.1% → 67.6%)
Combined sensitivity was higher for bedside CXRs while the specificity and accuracy were significantly lower
