|
| Reference | Dataset | Method used | Evaluation metrics | Research challenges |
|
| [23] | In-house dataset. No information about the dataset is mentioned in the paper | Rib region extract method and spatial coherence convolutional neural network | Accuracy, recall, and speed | There were limited comparative experiments and the potential of CNN networks was not fully researched |
| [24] | 1,079 patients and 25,054 2D annotations from 3 different hospitals, slice thicknesses range from 1 to 5 mm | Faster R-CNN | Precision recall, and -score | This work only used 2D CNN, and no 3D information was combined. The precision and recall of this work were not particularly high |
| [26] | A total of 7,473 annotated traumatic rib fractures from 900 patients from a single center, slice thicknesses range from 1 to 1.25 mm | Sliding widow mechanism and a modified U-Net called FracNet | Free response receiver-operating characteristic (FROC) analysis | This article only carried out a single-center study, and the landscape of deep neural networks was not fully explored |
| [28] | 8,529 chest CT images and 33,828 annotations, slice thickness of CT images was 0.625 mm | Rib fracture detection pipeline consisting of five stages: rib segmentation, vertebra detection, rib labeling, rib fracture detection, and rib fracture classification. VRB-Net for rib fracture detection | Recall, precision, and score | The ground truth for detection and classification may include incorrect cases caused by incorrect annotation |
|
