A Deep Learning Model for Rib Fracture Classification Based on CT Sequence Images

This study presents a deep learning model for rib fracture classification based on computed tomography (CT) sequence images. We retrospectively collected chest CT image data of 571 rib fractures in blunt chest injuries patients with rib fractures admitted to Tianjin Hospital in China from 2021 to 2022, and provided high-quality labeling of lesion locations and fracture types. A two-stage rapid diagnostic model for rib fractures was developed. Firstly, an object detection network based on CSPDarknet53 is designed to achieve rapid detection and localization of fracture lesion areas. Then, a SE-EfficientNet-B0 network is constructed to classify the fracture types of the single rib containing fracture lesions into five categories based on 3D reconstructed images. The lightweight fracture rapid diagnosis model designed in this paper has a fracture recognition recall rate of 93.5 %, an accuracy rate of 87.9 % and AP@ 0.5 of 0.856. The accuracy rate of classification for misaligned/non misaligned fractures is over 70 %, and the accuracy and recall rates for three newly occurring fractures(misaligned fractures, incomplete fractures, and healing fractures) are <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 0. 6 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 8. 0 \%}$</tex>, respectively.