Reverse-engineering the diagnostic process: An indirect finding-driven deep learning model for pelvic active bleeding...

Contrast-enhanced computed tomography (CT) is useful for detecting active bleeding in patients with pelvic trauma. However, early detection of active bleeding from CT images is challenging due to the extensive image volume, which can delay intervention. Although the development of accurate and rapid detection support systems for pelvic active bleeding is highly anticipated, a significant data imbalance between bleeding and non-bleeding images in real-world hospital settings poses a significant challenge to clinically applicable models. We reverse-engineered the physician's diagnostic process and implemented a curriculum-based deep learning algorithm using indirect findings to prevent overfitting caused by the imbalance. In this multicenter study, we developed and implemented a multistage deep learning algorithm for the detection of active bleeding using enhanced CT of pelvic trauma patients. From April 18, 2008 to December 31, 2023, we collected 258,057 slices of whole-body CT images of 2178 patients for blunt trauma with pelvic injuries from five emergency centers. Two types of classification-based deep learning models were developed and evaluated: anatomical structure extraction (ASE) and active bleeding detection (ABD) models. An algorithm using these models was implemented, and its accuracy in detecting active bleeding in the pelvis and the inference speed were evaluated. For validation data, the ASE and ABD models recognized the pelvis and active bleeding with areas under the receiver operating characteristic curve (AUCs) (accuracies) of 0.999 (99.4%) and 0.920 (83.0%), respectively. For test data, the pipeline achieved an AUC of 0.901 (accuracy: 77.5%) and an average inference time of 1.93 s. In the observer performance study, the algorithm outperformed residents and demonstrated no difference with that of board-certified specialists. Our curriculum-based deep learning algorithm achieves a sufficiently high accuracy and inference speed for practical use. Our training approach using indirect findings offers a viable solution for developing imaging AI models even with an imbalanced dataset.