Indirect finding-driven deep learning model matches specialist accuracy for pelvic active bleeding on CT

A multistage deep learning model detected pelvic active bleeding on CT with an AUC of 0.901 and accuracy of 77.5%, matching specialists and outperforming residents (inference time 1.93 s).

• The anatomical structure extraction (ASE) and active bleeding detection (ABD) models achieved validation AUCs of 0.999 and 0.920, respectively.
• In an observer performance study, the algorithm's accuracy was non-inferior to board-certified specialists and significantly better than residents.
• The curriculum-based training using indirect findings addressed the significant data imbalance between bleeding and non-bleeding images across five emergency centers.

Automated summary

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