Externally validated deep learning model for pulmonary embolism on CTPA: high accuracy overall but poor for subsegmental-only clots

An externally validated deep learning model for pulmonary embolism on CTPA achieved sensitivity 0.80, specificity 0.97, and AUROC 0.94 for any PE. No central PE was missed, but 100/375 peripheral PEs (26.7%) went undetected. Subsegmental-only PE detection was near chance (AUROC…

• The model, a second-place algorithm from the RSNA 2020 challenge, was tested on 1,038 CTPA scans from 2015-2022.
• For subsegmental-only PEs, sensitivity was 0.94 but specificity only 0.22, suggesting the model was not reliable for this subtype, likely because it was not explicitly trained for it.
• Performance was slightly better for right-sided than left-sided PEs (AUROC 0.95 vs 0.92, p < 0.05).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.