A fully automated rib fracture detection system on chest CT images and its impact on radiologist performance

Objective To compare rib fracture detection and classification by radiologists using CT images with and without a deep learning model. Materials and methods A total of 8529 chest CT images were collected from multiple hospitals for training the deep learning model. The test dataset included 300 chest CT images acquired using a single CT scanner. The rib fractures were marked in the bone window on each CT slice by experienced radiologists, and the ground truth included 861 rib fractures. We proposed a heterogeneous neural network for rib fracture detection and classification consisting of a cascaded feature pyramid network and a classification network. The deep learning-based model was evaluated based on the external testing data. The precision rate, recall rate, F1-score, and diagnostic time of two junior radiologists with and without the deep learning model were computed, and the Chi-square, one-way analysis of variance, and least significant difference tests were used to analyze the results. Results The use of the deep learning model increased detection recall and classification accuracy (0.922 and 0.863) compared with the radiologists alone (0.812 vs. 0.850). The radiologists achieved a higher precision rate, recall rate, and F1-score for fracture detection when using the deep learning model, at 0.943, 0.978, and 0.960, respectively. When using the deep learning model, the radiologist’s reading time was decreased from 158.3 ± 35.7 s to 42.3 ± 6.8 s. Conclusion Radiologists achieved the highest performance in diagnosing and classifying rib fractures on CT images when assisted by the deep learning model.