Background
Thoracic trauma is a common injury encountered in the emergency department and accounts for approximately 10%–15% of all trauma cases [
1]. The mortality rate globally ranges from 20 to 25% [
2]. Traumatic rib fracture, caused by a tremendous impact force on the chest wall, is the most common form of blunt thoracic injury and accounts for approximately 35% of all cases of thoracic traumas [
3]. Rib fractures are associated with significant morbidity and mortality, both of which increase as the number of fractured ribs increases [
4,
5]. Hence, rib fracture is an essential indicator of trauma severity. Accurately detecting rib fractures, compared to other injuries, can result in a higher treatment rate, avoid complications, and help solve medical-legal disputes such as traffic accidents and physical fighting [
6].
Chest radiography and computed tomography (CT) are two main methods for detecting rib fractures. Chest radiography is usually the initial imaging modality for rib fracture screening because the radiation dose, time and economic cost of CT are all higher than those of radiography [
7‐
9]. The American College of Radiology criteria for the evaluation of rib fractures recommend chest radiography with a posteroanterior view at four variant evaluations for suspected rib fractures in non-high-energy blunt trauma [
8]. Chest radiography is also a complementary examination for high-energy blunt trauma [
10]. However, research has shown that the overall incidence of rib fractures is probably higher than that previously recognised [
11]. A previous investigation reported that up to 50% of rib fractures may be missed on plain radiographs, which may lead to potential risks to patients [
12]. The detection of rib fractures on chest radiographs depends mostly on the reader’s experience, quality of the displayed images, and/or clinical scenario of chest radiograph scanning. Rib fracture detection is a time-consuming and demanding task for radiologists. Thus, a fast, easily available, and highly accurate method for rib fracture screening, which could be adopted to relieve radiologists and develop a cost-effective tool for clinical application, is urgently needed.
Artificial intelligence (AI) is widely used in the medical field, particularly in radiology. The deep learning algorithm of AI demonstrates good diagnostic accuracy and can be used to improve the quality and speed of image interpretation and increase the efficiency of physicians [
13‐
15]. Convolutional neural network (CNN) is an essential branch of deep learning. The multiple processing layers of CNN are more sensitive to image features and can enhance recognition accuracy [
16], which are commonly used AI techniques in medical imaging among radiology researchers [
17,
18]. Yamashita et al. [
16], divided the application of CNN into classification, segmentation, detection, and other applications [
16] such as lung nodule classification [
19], liver segmentation [
20] and breast cancer detection [
21]. CNN also demonstrates high feasibility and potential for fracture detection. Studies on lateral wrist fractures, proximal humerus fractures, thighbone fractures and orthopaedic trauma have shown promising results [
22‐
25].You Only Look Once v3 (YOLOv3) is a classic CNN algorithm with an excellent network structure, which is performed well in objection detection but seldom been used in the rib fractures detection.
For rib fracture detection, several research studies have been conducted, based on CNN. Jin et al. [
26], developed an automatic system, named FracNet, which is based on 3D UNet, to detect and segment rib fractures on CT images and achieved a detection sensitivity of 92.9% [
26]. Weikert et al. [
27] constructed a deep learning-based algorithm to detect acute and chronic rib fractures on trauma CT images, and it achieved good performance with a sensitivity of 87.4% [
27]. Yang et al. [
28] verified that the use of a deep learning system could be used to diagnose and classify rib fractures with better efficiency, faster speed, and similar results as those of radiologists’ readings [
28]. However, previous work has primarily focused on CT images, but few studies have verified the performance of the CNN model for detecting rib fractures by using radiography. Compared to CT, radiography is usually the first choice for diagnosing rib fractures in a clinical environment. Some object detection algorithms have been tested to detect fractures, based on radiography such as Faster RCNN [
29], Libra RCNN [
30], Dynamic RCNN [
31], Cascade RCNN [
32] and CCE-Net [
33]. While these methods were all be tested in single center dataset and without the application of YOLOv3. In this study, we applied YOLOv3 because YOLOv3 is a faster, relatively easier to comply, and more convenient framework than the CNN frameworks mentioned before [
34]. YOLOv3 has also been proven to have very good performance in multi-target detection. The traditional single detection method is more prone to miss detection for rib fractures because multiple rib fractures are more common for patients with rib fractures, instead single rib fractures are less common, so the multiple rib fracture detection model is very meaningful.
Thus, the aim of this study was to create a novel model for multiple rib fracture detection by using a CNN, based on multicenter and quality-normalised chest radiographs. The contributions of our study can be listed as follows. First, radiographs from four hospitals and external validation were collected, and the multicenter dataset could improve the robustness of the model. Second, image quality normalisation using the multiscale image contrast amplification (MUSICA) algorithm has been chosen, which is excellent in image enhancement and could ensure the model’s consistency but seldom applied in radiographs [
35]. Third, A CNN model was then constructed using the YOLOv3 algorithm, which is seldom used to detect rib fractures and achieved an outstanding detection rate in our study. Finally, the detection ability of the CNN model was compared with that of junior and senior radiologists and found the performance is better than the junior radiologists and similar to the senior radiologists. This study is the first to develop and verify the performance of a CNN model for detecting rib fractures on chest radiographs through normalised images from a multicenter dataset.
Discussion
In this study, we created a powerful CNN model for the detection of rib fractures by using chest radiographs. First, the quality of the input image was standardised. The CNN model was then trained to detect rib fractures with all lesions found. It showed promising results with high sensitivity and accuracy. Finally, a standardised model for rib fracture detection was developed. It outperformed the detection ability of senior and junior radiologists.
Deep learning has advanced significantly with new algorithms and optimised network structures, and these greatly contributed to the current study. Kim et al. [
22] used X-ray-based AI to detect carpal fractures. The model showed sensitivity, specificity, and AUC values of 90%, 88% and 0.954, respectively [
22]. Chung et al. [
23] used a deep learning model to detect proximal humeral fractures, and the sensitivity, specificity, and AUC were 99%, 97% and 0.97, respectively [
23]. AI showed promising results in fracture detection in the aforementioned two studies, as well as in this study. Three reasons could explain these results. The first reason is quality normalisation, as discussed in the preceding paragraph. The second reason is the application of the innovative network YOLOv3, which combines YOLOv2, Darknet-19 and other new residual networks. Compared to ResNet-152 and ResNet-101, YOLOv3 has better training speed and accuracy [
37], further expanding its use. The third reason is that k-means was used to count the fracture marker box in the labelled sample. And result of the loss values of training and testing sets also proved the model performance. We could find the loss values of training and testing sets are hardly overlapping at the end of training. The reason is that the samples for training and testing sets were randomly obtained, and imbalance existed. While with the number of iterations increased, the curves tended to be more consistent, and the final training results showed the effectiveness of our detection algorithm.
A series of subgraphs was trained to locate multiple foci and were free of the hand-engineered region, which is rarely used in rib fracture detection. By examining three different scale feature maps, the number and specific locations of rib fractures could be better detected. Signal features detected by handcrafted analysis were challenged by the CNN model with a sliding window [
38,
39]. Comparative testing showed that the sensitivity for detecting rib fractures was significantly higher with the CNN model than that by the junior radiologist and close to that by the senior radiologist at the fracture level. In addition, the precision of the model was slightly lower than that of the radiologists, although the model can still provide radiologists with specific locations for suspect fractures, thereby reducing the rate of lateral missed diagnosis.
The CNN is the most commonly used AI technique for medical imaging [
17,
18]. The CNN model is also becoming a popular constituent of medical diagnosis with respect to efficiency and to precision medicine. Studies [
3,
40] have shown that specific organ injuries are often correlated with a specific fractured rib [
3,
40]. The number of displaced rib fractures could also be a strong predictor for developing pulmonary complications [
41], which makes the detection of rib fractures important to prevent complications and help mitigate patient pain. This model used FROC to test its multilesion detection ability. The sensitivity was 91.3% when the false-positive rate of each case was set at 0.56. By comparison, only 49% of rib fractures are traditionally detected on the physical evaluation of radiographs [
42]. This result may expand the clinical value of chest radiographs and reduce the rate of recommendations for additional imaging (RAIs). Harvey et al. [
7] reported that the rate of RAIs have increased by as much as 200% since 1995. In particular [
7], in radiographic imaging of the chest, the increase is because of the low diagnostic accuracy of radiography. Some critics have implicated RAIs as a cause of the increased use of additional imaging and associated costs.
The process of image standardisation also includes some image enhancement techniques so that images from different devices can have the same image quality. An important factor is that image standardisation forms the basis for the performance of the CNN model. Nearly all radiology applications are highly dependent on radiographic image quality, especially when combined with AI. However, image quality standardisation has long presented a challenge and has affected the intelligent diagnostic development of radiography, ultrasonography, CT, and magnetic resonance imaging. Several methods have been proposed to solve this problem. As in the research by Li et al. [
43], several steps were performed for standardisation, including rescaling, downsizing, and transformation. Smoothing, normalisation, and resampling have also been performed in diabetic retinopathy research [
44]. However, most studies have focused on noise elimination or uniform size rather than on feature enhancement. In the current study, proper image enhancement to reduce the variability of different machine images was necessary, particularly because images were reviewed by different display systems.
Among various imaging modalities, chest radiography is the appropriate initial imaging modality for patients with rib fractures. CT may provide a more accurate diagnosis; however, it is usually only performed after diagnostic chest radiography [
8]. Missed diagnoses of rib fractures on chest radiographs may cause legal disputes, especially in traffic accidents and physical fights. It more importantly may lead to delayed treatment. Therefore, this study focused on chest radiography for the early detection of rib fractures. The developed CNN model has a wide range of real-life applications, including but not limited to the following four areas: (1) assisting clinicians in an initial diagnosis of images; (2) screening images for errors after clinicians have made a diagnosis; (3) detecting rib fractures in unlabelled images in subsequent research studies; and (4) serving other studies that have been linked to rib fractures.
Limitations
This study has few limitations, despite its promising results. First, the fractures on the radiographs were labelled according to the physician’s comprehensive diagnosis without gold standard modalities such as pathology. Second, only posteroanterior radiographs were obtained, and the lateral position of the rib was not considered. Third, radiography cannot consistently demonstrate fractures in the costal cartilage, which is an inherent problem that decreases the detection rate. Fourth, the model was trained and tested by using different image resolutions. The off-label use in clinical environment of the model deserves further research. Finally, because only 19,974 subgraphs (918 radiographs) were included to train the model, more radiographs should be enrolled in the training data to improve model efficacy. This study is only a preliminary attempt at using a CNN model to examine rib fractures, based on radiographs. The efficiency of CNN data models is expected to continue to improve with the advent of computer technology and big data.
Conclusions
In this study, we proposed a model for detecting multiple rib fractures, by using a CNN based on quality-normalised chest radiographs through data collection from multicentre, image quality normalisation, CNN model construction, the model’s performance validation and its comparison with radiologists. The CNN model showed high diagnostic efficiency, which indicated that CNN can improve the detection rate of rib fractures on chest radiographs, help reduce missed diagnoses, avoid medical accidents, and relieve radiologists’ workload. Implementing AI models in a clinic is the tendency of medical development. Our research implies the potential value of using CNN in rib fracture diagnoses. The detection ability requires further validation, although CNN is promising for medical diagnosis.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.