Feasibility of rib fracture detection in low-dose computed tomography images with a large, multicenter datasets-based model

Abstract

PurposeTo evaluate the feasibility of rib fracture detection in low-dose computed tomography (CT) images with a RetinaNet-based approach and to evaluate the potential of lowdose CT for rib fracture detection compared with regular-dose CT images. Materials and methodsThe RetinaNet-based deep learning model was trained using 7300 scans with 50,410 rib fractures that were used as internal training datasets from four multicenter. The external test datasets consisted of both regular-dose and low-dose chest-abdomen CT images of rib fractures; the MICCAI 2020 RibFrac Challenge Dataset was used as the public dataset. Radiologists' interpretations were used as reference standards. The performance of the model in rib fracture detection was compared with the radiologists’ interpretation. ResultsIn total, 728 traumatic rib fractures of 100 patients [60 men (60 %); mean age, 53.45 ± 11.19 (standard deviation (SD)); range, 18–77 years] were assessed in the external datasets. In these patients, the regular-dose group had a mean CT dose index volume (CTDIvol) of 7.18 mGy (SD: 2.22) and a mean dose length product (DLP) of 305.38 mGy cm (SD: 95.31); the low-dose group had a mean CTDIvol of 2.79 mGy (SD: 1.11) and a mean DLP of 131.52 mGy cm (SD: 55.58). The sensitivity of the RetinaNet-based model and that of the radiologists was 0.859 and 0.721 in the low-dose CT images and 0.886 and 0.794 in the regular-dose CT images, respectively. ConclusionsThese findings indicate that the RetinaNet-based model can detect rib fractures in low-dose CT images with a robust performance, indicating its feasibility in assisting radiologists with rib fracture diagnosis.