Intelligent Detection and Grading Diagnosis of Fresh Rib Fractures Based on Deep Learning

doi:10.21203/rs.3.rs-5269042/v1

Download PDF

Research Article

Intelligent Detection and Grading Diagnosis of Fresh Rib Fractures Based on Deep Learning

https://doi.org/10.21203/rs.3.rs-5269042/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

Accurate detection and grading of fresh rib fractures are crucial for patient management but remain challenging due to the complexity of rib structures on CT images.

Methods

Chest CT images from 383 patients with rib fractures were retrospectively analyzed. The dataset was divided into a training set (n = 306) and an internal testing set (n = 77). An external testing set of 50 patients from the public RibFrac dataset was included. Fractures were classified into severe and non-severe categories. A modified YOLO-based deep learning model was developed for detection and grading. Performance was compared with thoracic surgeons using precision, recall, mAP50, and F1 score.

Results

The deep learning model showed excellent performance in diagnosing fresh rib fractures. For all fractures types in internal test set, the precision, recall, mAP50, and F1 score were 0.963, 0.934, 0.972, and 0.948, respectively. The model outperformed thoracic surgeons of varying experience levels (all p < 0.01).

Conclusion

The proposed deep learning model can automatically detect and grade fresh rib fractures with accuracy comparable to that of physicians. This model helps improve diagnostic accuracy, reduce physician workload, save medical resources, and strengthen health care in resource-limited areas.

Artificial Intelligence

Deep Learning

Fresh Rib Fracture

Detection

Grading Diagnosis

Rib fractures are the most common injuries in patients with blunt chest trauma [1; 2] and may be accompanied by severe complications such as pulmonary contusion, pneumonia, and hemopneumothorax [3; 4]. The number and displacement of rib fractures, as well as the presence of associated organ injuries, are correlated with mortality [5; 6]. Therefore, accurately identifying the number and severity of rib fractures is crucial.

With the increased use of chest multidetector computed tomography scans, the detection rate of rib fractures has significantly improved [7]. Compared to X-ray, CT offers greater contrast and resolution, providing a comprehensive view of the fracture site and enabling the detection of minor lesions and the assessment of other complications. However, the diversity and complexity of rib fracture shapes make it challenging to identify all fractures across hundreds of thin-slice CT images, often leading to missed diagnoses. The literature reports that 20.7% of initial chest CT scan reports misidentify the number or location of fractured ribs [8], and even after rib reconstruction, 20.9% of rib fractures are missed [9].

Recently, deep learning algorithms have been applied in the field of medical image processing, including image registration [10; 11], detection [12; 13], segmentation [14; 15] and disease prognosis [16; 17]. Some studies have reported good performance in detecting rib fractures [18; 19; 20; 21]. Azuma et al. [22] developed and validated a convolutional neural network (CNN) model for detecting rib fractures, showing that the CNN model could enhance the diagnostic ability of radiologists for any type of rib fracture. Zhou et al. [21] included clinical information in their CNN model, further improving diagnostic efficiency and reducing diagnosis time. However, the importance of fractures varies with their type. Chronic rib fractures usually do not require clinical intervention and do not affect treatment decisions or patient prognosis. In contrast, fresh rib fractures often exhibit severe displacement and may be accompanied by organ injuries, especially severely displaced and comminuted rib fractures, which are often accompanied by fatal complications such as severe pulmonary infection, persistent hemothorax, and massive hemoptysis[23]. Moreover, accurately grading rib fractures not only improves diagnostic precision but also plays a crucial role in determining the course of treatment. For instance, severe rib fractures with displacement exceeding the rib diameter often require surgical intervention, whereas non-severe fractures may be managed conservatively. Clinical guidelines suggest that fractures with significant displacement or multiple fractured ribs pose a higher risk for complications, necessitating timely intervention. To the best of our knowledge, there is a scarcity of studies focusing on the graded diagnosis of fresh rib fractures, with an emphasis on those that are severely displaced.

In our study, we developed a deep learning-based model for the intelligent detection of fresh rib fractures and graded diagnoses based on the severity of rib displacement, categorizing fractures into severe and non-severe. We conducted a comprehensive analysis of different subgroups of fracture severity and compared the model's performance with that of experienced thoracic surgeons.

Data sets and classification criteria

A retrospective collection of initial chest CT scans of patients diagnosed with rib fractures from January 2021 to April 2023 was conducted at the First Affiliated Hospital of Army Medical University (Hospital A) and Chongqing Dianjiang People's Hospital (Hospital B). Ethical approval for this retrospective study was obtained from the Ethics Committee of the First Affiliated Hospital of Army Medical University (Ethics Committee Name: Ethics Committee of the First Affiliated Hospital of Army Medical University, Ethics ID: KY2023062) and the Ethics Committee of Chongqing Dianjiang People's Hospital (Ethics Committee of Chongqing Dianjiang People's Hospital, Ethics ID: DYLL-LW-2023-03). Given the retrospective nature of the study, the requirement for informed consent was waived, and clinical trial number was not applicable. The inclusion criteria were patients aged ≥ 18 years with a history of trauma. Exclusion criteria included: 1. old or healed fractures; 2. artifacts; 3. bone tumors or bone destruction; 4. postoperative patients.

A total of 383 patients from the two centers were randomly divided into a training set (n = 306) and an internal testing set (n = 77). A rigorous randomization method was employed to minimize bias introduced by data partitioning and ensure the reliability and scientific validity of the study results. The external testing dataset comprised 50 CT scans of the public RibFrac[24], and radiologists re-annotated the CT scans of these 60 cases. The specific screening criteria are shown in Fig. 1.

In this study, fractures were classified into non-severe and severe based on the severity. Non-severe fractures were defined as: 1. fissure fractures with bone defects between the ends of the cortical bone without displacement; 2. cortical bone distortion or density changes; 3. fractures with angulation or displacement less than the rib diameter. Severe fractures were defined as: 1. fractures with displacement greater than the rib diameter; 2. fractures with observed bone fragments or comminuted fractures (Fig. 2).

CT acquisition

Patients were positioned supine, with their arms raised above their heads (for those unable to do so due to shoulder or upper limb injuries, arms were placed naturally at their sides). The scanning range extended from the thoracic inlet to the lung bases or the entire ribcage. The scan was performed during a single breath-hold following inhalation. Two different CT scanners were utilized for the examinations: Philips Brilliance 16 (Philips Medical Systems) and SOMATOM Definition (Siemens Healthineers). The tube voltage was set to 120 kV, with automatic modulation of the tube current. The slice interval ranged from 1 to 5 mm, and the slice thickness varied between 1 and 5 mm.

Image annotation

The CT slices in the training set were independently annotated by a thoracic surgeon with 5 years of experience and a radiologist with 5 years of experience. These two physicians used Makesense (https://www.makesense.ai/) to mark rib fractures with rectangular boxes and used different colors to indicate severity. In cases of uncertainty regarding fracture type, the two physicians discussed until a consensus was reached, or a third physician was involved to reach a final agreement. Annotations were made without access to any clinical information of the patients, and these annotations were used as the ground truth for training the automated detection algorithm.

The internal and external test datasets were annotated by a junior thoracic surgeon, a middle grade thoracic surgeon, and a senior thoracic surgeon, with the outcomes subsequently compared against the model's predictions.

Model architecture:

This study employs a deep learning model based on the You-Only-Look-Once (YOLO) [25] architecture, which is widely recognized for its effectiveness in real-time object detection. This approach allows the model to process entire images in one step, making it particularly suitable for tasks such as detecting rib fractures from CT scans (Fig. 3a).

Backbone Network

The backbone network is primarily responsible for extracting relevant features from the input CT images. It uses a series of CBS modules (composed of convolution layers, batch normalization, and SiLU activation functions) to effectively extract and normalize the feature maps. To further enhance feature extraction, we integrated C2f_EMA modules, which incorporate parallel convolutional layers and feature aggregation mechanisms. This setup helps capture more detailed and diverse features. At the final stage, a Spatial Pyramid Pooling in Feature (SPPF) module is utilized to pool information across different scales, enabling the model to capture global features essential for precise detection (Fig. 3b).

Neck Network

The Neck network is designed to refine and combine the feature maps extracted by the Backbone. Through upsampling and merging features from different layers, the Neck helps in capturing information across various resolutions. The C2f_EMA modules are once again employed to enhance the fused features, allowing the model to better detect objects at different scales. This multi-scale fusion ensures that small and large fractures are both captured accurately.

Head Network

The Head network generates the final detection results. It processes the refined feature maps and produces bounding boxes and classification predictions for rib fractures. This layer leverages features from multiple levels of the network, allowing it to accurately localize and categorize fractures. This design ensures that the model performs well even in challenging cases, such as detecting subtle or complex fractures.

Model evaluation

To assess the model's effectiveness, we used metrics that are suitable for object detection tasks. These include mean Average Precision (mAP), precision, recall, and the F1-score. Unlike traditional classification metrics, these provide a more comprehensive evaluation of the model's ability to accurately detect and localize rib fractures. We used an Intersection over Union (IoU) threshold of 0.5 to define correct detections, meaning that the predicted and actual bounding boxes had to overlap by at least 50% to be considered accurate (Fig. 4). This threshold aligns with clinical practices, where predictions with sufficient overlap are deemed useful for guiding diagnosis and treatment decisions.

Statistical analysis

This study employed SPSS (version 26.0; IBM, NY for statistical analysis, with a two-tailed p-value threshold of < 0.05. Normality of continuous variables was assessed using the Kolmogorov-Smirnov test. Normally distributed data were presented as mean ± standard deviation; otherwise, M (P₂₅-P₇₅) were utilized. Categorical variables were expressed as frequencies and percentages.

For normally distributed data, two-sample t-tests were applied, and non-normally distributed data were analyzed using the Mann-Whitney U test. Categorical variables were compared using Chi-squared or Fisher's exact test. The z-test was used to compare whether there was a statistical difference between the system and the thoracic surgeons.

Patient Characteristics.

This study included data from 383 patients with rib fractures from two centers, comprising a total of 4673 annotations. Among these annotations, 3,043 were for non-severe fractures and 1,630 for severe fractures. Of the 383 patients, 130 were from Hospital A and 253 from Hospital B. We randomly selected 306 patients for the training set and 77 patients for the internal testing set. Our results showed no statistically significant differences in gender (P = 0.256) and age (P = 0.766) between the training and testing sets (Table 1). Additionally, there were no significant differences in age (P = 0.565) between patients from the two centers, and hospital A had a significantly higher proportion of male patients compared to hospital B (P < 0.01) as detailed in Table 1.

Table 1

Patient’s radiologic and clinical information in two hospitals.
Characteristic	Hospital A	Hospital B	P value
No. of patients, n (%)	140	243
Age, years, M (Q1, Q3)	57 (50 ~ 67)	56 (49 ~ 68)	0.565
Gender, male, n (%)	113 (80.7%)	153 (63.0%)	< 0.01
No. of fracture annotations (total)	1503	3170
Annotations of non-severe fracture	971	2072
Annotations of severe fracture	532	1098
Hospital A: Southwest hospital; hospital B: Dianjiang People's Hospital of Chongqing; No: Number; M (Q1, Q3): Median (First Quartile, Third Quartile); CT: Computed Tomography.

The training set consisted of 306 patients, with a total of 3,646 annotations, including 2,455 non-severe fractures and 1,191 severe fractures. The internal testing set included 77 patients, with a total of 1,027 annotations, comprising 685 non-severe and 342 severe fractures. Additionally, we incorporated 50 patients from the public RibFrac dataset, with a total of 821 annotations, including 660 non-severe fractures and 161 severe fractures (Table 2).

Table 2

Patient’s radiologic and clinical information in training, internal and external testing datasets.
Characteristic	Training data set	Internal Testing data set	External Testing data set	P value
No. of patients, n (%)	306	77	50	-
Age, years, M (Q1, Q3)	57 (48 ~ 69)	55 (53 ~ 62)	-	0.256
Gender, male, n (%)	202 (66.0%)	59 (76.6%)	-	0.766
No. of fracture annotations (total)	3646	1027	821	-
Annotations of non-severe fracture	2455	685	660	-
Annotations of Severe fracture	1191	342	161	-
Hospital A: Southwest hospital; hospital B: Dianjiang People's Hospital of Chongqing; No: Number; M (Q1, Q3): Median (First Quartile, Third Quartile); CT: Computed Tomography.

Performance of the models.

In the internal test set, the overall precision for all fracture types was 0.963, recall was 0.934, mAP50 was 0.972, and F1 score was 0.948. For external test dataset, precision was 0.880, recall was 0.796, mAP50 was 0.857, and F1 score was 0.836 (Table 3).

Table 3

Model performance in two multicenter testing sets
Fracture type	Precision	Recall	mAP50	F1 score
Internal Testing data set	0.963	0.934	0.972	0.948
Severe Fracture	0.957	0.909	0.964	0.932
Non-Severe Fracture	0.968	0.959	0.981	0.963
External Testing data set	0.880	0.796	0.857	0.836
Severe Fracture	0.867	0.774	0.843	0.818
Non-Severe Fracture	0.892	0.819	0.872	0.854
mAP 50: mean Average Precision at 50; Hospital A: Southwest hospital; hospital B: Dianjiang County People's Hospital of Chongqing.

We visualized the model's performance using the Precision-Recall (PR) curve. Figure 5a illustrates the PR curves for the internal test set, where the AUPRC values are 0.964 for Non-severe fractures, 0.981 for Severe fractures, and 0.972 for all classes combined. Figure 5b shows the PR curves for the external test set, with AUPRC values of 0.843 for Non-severe fractures, 0.872 for Severe fractures, and 0.857 for all classes combined. These results demonstrate that our model exhibits excellent performance on the external test set, indicating good generalizability. Figure 6 presents examples of rib fractures successfully identified by our model. Additionally, a confusion matrix has been created to illustrate the model's performance on the internal and external testing set (Fig. 7a and Fig. 7b).

Comparison of the performance between our model and thoracic surgeons.

Table 4, Fig. 8a and Fig. 8b offer a comparative analysis of precision, recall, and F1 score for three thoracic surgeons of different experience levels, contrasted with our model’s performance. For non-severe fractures, Thoracic Surgeon 1 recorded a precision of 0.651, a recall of 0.670, and an F1 score of 0.660. In comparison, Thoracic Surgeon 2 achieved a precision of 0.669, a recall of 0.736, and an F1 score of 0.701, while Thoracic Surgeon 3 attained a precision of 0.802, a recall of 0.932, and an F1 score of 0.862. When evaluating severe fractures, Thoracic Surgeon 1's precision was 0.606, recall was 0.695, and F1 score was 0.648; Thoracic Surgeon 2 demonstrated a precision of 0.793, recall of 0.774, and F1 score of 0.783; and Thoracic Surgeon 3 showed a precision of 0.922, recall of 0.842, and F1 score of 0.880. For all classes, Thoracic Surgeon 1 achieved a precision of 0.635, recall of 0.678, and an F1 score of 0.656. Thoracic Surgeon 2 had a precision of 0.713, recall of 0.751, and an F1 score of 0.731, whereas Thoracic Surgeon 3's metrics were a precision of 0.845, recall of 0.895, and an F1 score of 0.869. These findings underscore that our deep learning model consistently outperforms thoracic surgeons with varying experience levels across all evaluated metrics, highlighting its significant promise in advancing rib fracture detection.

Table 4

Comparison of diagnostic performance between thoracic surgeons and our model.
	Thoracic Surgeon 1			Thoracic Surgeon 2			Thoracic Surgeon 3
Fracture type	Precision	Recall	F1 score	Precision	Recall	F1 score	Precision	Recall	F1 score
Non-severe fracture	0.8019	0.9320	0.8621	0.6685	0.7361	0.7007	0.6512	0.6698	0.6604
Severe fracture	0.9218	0.8416	0.8799	0.7925	0.7741	0.7832	0.6061	0.6950	0.6475
Total	0.8447	0.8946	0.8689	0.7126	0.7506	0.7311	0.6352	0.6782	0.6560

Chest CT scans are often the preferred imaging modality for patients with chest trauma because they can identify many injuries that may be missed by chest X-rays, including pulmonary contusions, hemothorax, pneumothorax, and rib fractures. Moreover, rib fractures are considered indicators of severe trauma [26]. Traditional detection methods require physicians to meticulously evaluate the entire CT scan, which is a time-consuming and error-prone process, particularly for less experienced radiologists or thoracic surgeons. Additionally, the number and displacement of rib fractures are related to the follow-up treatment plan[27; 28]. Therefore, thoracic surgeons should prioritize different types of rib fractures. Chronic rib fractures, characterized by the presence of a mature callus or an invisible fracture line, do not require clinical intervention and do not affect treatment decisions or patient prognosis, making their detection unnecessary. In contrast, timely identification and localization of fresh rib fractures are crucial, especially in emergency settings, as severely displaced rib fractures may be associated with internal organ injuries, significantly impacting patient outcomes[29].

To address this issue, we developed a deep learning-based intelligent diagnostic model for the classification and detection of fresh rib fractures, which was validated on both internal multicenter datasets and external public datasets. The model outperformed experienced thoracic surgeons and demonstrated exceptional performance.

Recently, Yao et al.[30] developed a deep learning-based rib fracture detection system that achieved high-performance detection and diagnosis of rib fractures on chest CT images, significantly reducing physician workload and minimizing misdiagnoses. However, this study only addressed the binary classification of fracture presence. Zhou et al.[31] developed a convolutional neural network (CNN) model that classified fractures into fresh, healing, and old fractures but did not perform a graded diagnosis for precise fracture diagnosis. Zhou et al.[21] reported the use of clinical information in their CNN model, which similarly improved diagnostic efficiency and reduced diagnosis time. Xiong et al.[32] found that the performance of radiologists on night shifts was inferior to that on day shifts, and the use of a deep learning-based computer-aided diagnosis (CAD) system for rib fractures helped night shift radiologists achieve performance levels comparable to their daytime performance. Unlike previous studies, our research focused on fresh rib fractures, as their rapid localization and diagnosis are crucial components of intelligent diagnosis and treatment of acute chest trauma. Chronic rib fractures do not affect treatment decisions or patient outcomes, making the focus on fresh rib fractures more aligned with real clinical scenarios. Our model can also intelligently grade fractures based on their severity, aiding in treatment decisions and prognosis evaluation for posttraumatic rib fractures.

Our model's performance is also comparable to that of physicians, making it suitable as the "first reader." This approach can help improve diagnostic accuracy, reduce diagnosis time, and reduce the workload of physicians; additionally, it can aid in building medical resources in under resourced areas.

In our model improvements, we optimized the Backbone, Neck, and Head networks to enhance feature extraction, fusion, and detection capabilities. The Backbone, designed to extract deep features from medical images, alternates between CBS modules (Convolution, Batch Normalization, and SiLU activation) and C2f_EMA modules, which include convolution layers and parallel Bottleneck_EMA branches for richer feature representation. It concludes with the SPPF module, which combines multiple Maxpool and Concat operations to capture critical global information. In the Neck, we focused on refining and consolidating multi-scale features using an Upsample operation and the C2f_EMA module to merge and process features across different levels, improving detection accuracy and robustness. Finally, the Head network, composed of additional C2f_EMA modules, integrates and processes refined features before passing them to the Detect layer, where bounding boxes and class predictions are generated. This comprehensive multi-level integration allows our model to achieve high-precision detection in medical images.

However, our study had several limitations. One limitation of the current model is its inability to identify specific anatomical structures, such as which rib is fractured, which may reduce its utility in acute settings requiring rapid, precise localization. Additionally, as AI is increasingly integrated into clinical workflows, ethical concerns around data security, patient privacy, and clinician reliance on automated systems must be addressed. Our model has demonstrated robustness across both internal and external datasets. However, future studies should focus on validating the model in a variety of clinical settings, particularly in hospitals with differing patient demographics, trauma protocols, and imaging equipment to ensure broad applicability and our sample size could be further expanded. Finally, prospective studies remain relatively scarce. We aim to collect more data and conduct prospective studies to further validate and optimize the model.

In summary, we developed a deep learning-based intelligent detection model for the detection and grading of fresh rib fractures. The model demonstrated high fracture detection rates and localization accuracy. It is suitable as a "first reader" to assist physicians in quickly and accurately determining the condition of patients with rib fractures, reducing diagnostic time, improving diagnostic accuracy, and enhancing physician efficiency. Additionally, the model can help strengthen medical resources in under resourced areas, including rural and township regions.

Artificial Intelligence

Computed Tomography

CNN

Convolutional Neural Network

CAD

Computer-Aided Diagnosis

mAP

Mean Average Precision

ROC

AUC-Receiver Operating Characteristic-Area Under Curve

Precision-Recall

IoU

Intersection over Union

EMA

Exponential Moving Average

CBS

Convolution, Batch normalization, SiLU activation

SPPF

Spatial Pyramid Pooling in Feature

YOLO

You Only Look Once

AUPRC

Area Under the Precision-Recall Curve

Ethics approval: This study was performed in line with the principles of the Declaration of Helsinki. Ethical approval for this retrospective study was obtained from the Ethics Committee of the First Affiliated Hospital of Army Medical University (Ethics Committee Name: Ethics Committee of the First Affiliated Hospital of Army Medical University, Ethics ID: KY2023062) and the Ethics Committee of Chongqing Dianjiang People's Hospital (Ethics Committee of Chongqing Dianjiang People's Hospital, Ethics ID: DYLL-LW-2023-03).

Human Ethics and Consent to Participate declarations: Given the retrospective nature of the study, the requirement for informed consent was waived

Availability of data and materials: The datasets used and analysed during the current study are available from the corresponding author on reasonable request.

Competing Interests: The authors have no relevant financial or non-financial interests to disclose.

Funding: This work was supported by University Funded Science and Technology Innovation Capacity Improvement Project (No. 2019XYY14), Science-Health Joint Medical Scientific Research Project of Chongqing (No. 2022ZDXM018), Chongqing Science and Technology Talent Project (CQYC201905037), Chongqing key research and development program (CSTB2022TIAD-KPX0181) and Joint Project of Chongqing Health Commission and Science and Technology Bureau (No. 2024MSXM054).

Author contribution:

Conceptualization: Qingyuan Xu and Yi Yu
Data curation: Tongxin Li, Luya Shen and Yong Fu
Formal analysis: Mingyi Liao, Junliang Che, Shulei Wu and Jie Liu
Funding acquisition: Wei Wu, Yong Fu and Yi Wu
Visualization: Tongxin Li, Mingyi Liao, Fanghong Zhang and Ping He
Writing-original draft: Tongxin Li
Writing-review & editing: All authors

Lin FC-F, Li R-Y, Tung Y-W, Jeng K-C, Tsai SC-S. Morbidity, mortality, associated injuries, and management of traumatic rib fractures. J Chin Med Assoc. 2016;79:329–34.
Marini CP, Petrone P, Soto-Sánchez A, García-Santos E, Stoller C, Verde J. Predictors of mortality in patients with rib fractures. Eur J Trauma Emerg Surg. 2021;47:1527–34.
Miller LA. Chest wall, lung, and pleural space trauma. Radiol Clin North Am 44 (2006).
Talbot BS, Gange CP, Chaturvedi A, Klionsky N, Hobbs SK, Chaturvedi A. Traumatic Rib Injury: Patterns, Imaging Pitfalls, Complications, and Treatment. Radiographics. 2017;37:628–51.
Brasel KJ, Moore EE, Albrecht RA, deMoya M, Schreiber M, Karmy-Jones R, Rowell S, Namias N, Cohen M, Shatz DV. Biffl, Western Trauma Association Critical Decisions in Trauma: Management of rib fractures. J Trauma Acute Care Surg. 2017;82:200–3.
Lee RB, Bass SM, Morris JA, MacKenzie EJ. Three or more rib fractures as an indicator for transfer to a Level I trauma center: a population-based study. J Trauma. 1990;30:689–94.
Cherney AR, Richardson DM, Greenberg MR, Choo EK, McGregor AJ, Safdar B. Prevalence and Clinical Import of Thoracic Injury Identified by Chest Computed Tomography but Not Chest Radiography in Blunt Trauma: Multicenter Prospective Cohort Study. Ann Emerg Med. 2016;68:133–4.
Cho SH, Sung YM, Kim MS. Missed rib fractures on evaluation of initial chest CT for trauma patients: pattern analysis and diagnostic value of coronal multiplanar reconstruction images with multidetector row CT. Br J Radiol. 2012;85:e845–50.
Urbaneja A, De Verbizier J, Formery A-S, Tobon-Gomez C, Nace L, Blum A, Gondim PA, Teixeira. Automatic rib cage unfolding with CT cylindrical projection reformat in polytraumatized patients for rib fracture detection and characterization: Feasibility and clinical application. Eur J Radiol. 2019;110:121–7.
Nazib A, Fookes C, Perrin D. Towards Extreme-Resolution Image Registration with Deep Learning, 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), 2019, pp. 512–516.
Wang AQ, Yu EM, Dalca AV, Sabuncu MR. A robust and interpretable deep learning framework for multi-modal registration via keypoints. Med Image Anal. 2023;90:102962.
Kuo W, Hӓne C, Mukherjee P, Malik J, Yuh EL. Expert-level detection of acute intracranial hemorrhage on head computed tomography using deep learning. Proc Natl Acad Sci U S A. 2019;116:22737–45.
Liu F, Zhou Z, Samsonov A, Blankenbaker D, Larison W, Kanarek A, Lian K, Kambhampati S, Kijowski R. Deep Learning Approach for Evaluating Knee MR Images: Achieving High Diagnostic Performance for Cartilage Lesion Detection. Radiology. 2018;289:160–9.
Monteiro M, Newcombe VFJ, Mathieu F, Adatia K, Kamnitsas K, Ferrante E, Das T, Whitehouse D, Rueckert D, Menon DK, Glocker B. Multiclass semantic segmentation and quantification of traumatic brain injury lesions on head CT using deep learning: an algorithm development and multicentre validation study. Lancet Digit Health. 2020;2:e314–22.
Xue J, Wang B, Ming Y, Liu X, Jiang Z, Wang C, Liu X, Chen L, Qu J, Xu S, Tang X, Mao Y, Liu Y, Li D. Deep learning-based detection and segmentation-assisted management of brain metastases. Neuro Oncol. 2020;22:505–14.
Suk H-I, Wee C-Y, Lee S-W, Shen D. State-space model with deep learning for functional dynamics estimation in resting-state fMRI. NeuroImage. 2016;129:292–307.
Jiang Y, Zhang Z, Wang W, Huang W, Chen C, Xi S, Ahmad MU, Ren Y, Sang S, Xie J, Wang J-Y, Xiong W, Li T, Han Z, Yuan Q, Xu Y, Xing L, Poultsides GA, Li G, Li R. Biology-guided deep learning predicts prognosis and cancer immunotherapy response. Nat Commun. 2023;14:5135.
Jin L, Yang J, Kuang K, Ni B, Gao Y, Sun Y, Gao P, Ma W, Tan M, Kang H, Chen J, Li M. Deep-learning-assisted detection and segmentation of rib fractures from CT scans: Development and validation of FracNet. eBioMedicine 62 (2020).
Wang S, Wu D, Ye L, Chen Z, Zhan Y, Li Y. Assessment of automatic rib fracture detection on chest CT using a deep learning algorithm. Eur Radiol. 2022;33:1824–34.
Wu M, Chai Z, Qian G, Lin H, Wang Q, Wang L, Chen H. Development and Evaluation of a Deep Learning Algorithm for Rib Segmentation and Fracture Detection from Multicenter Chest CT Images. Radiology: Artificial Intelligence 3 (2021).
Zhou Q-Q, Tang W, Wang J, Hu Z-C, Xia Z-Y, Zhang R, Fan X, Yong W, Yin X, Zhang B, Zhang H. Automatic detection and classification of rib fractures based on patients’ CT images and clinical information via convolutional neural network. Eur Radiol. 2020;31:3815–25.
Azuma M, Nakada H, Takei M, Nakamura K, Katsuragawa S, Shinkawa N, Terada T, Masuda R, Hattori Y, Ide T, Kimura A, Shimomura M, Kawano M, Matsumura K, Meiri T, Ochiai H, Hirai T. Detection of acute rib fractures on CT images with convolutional neural networks: effect of location and type of fracture and reader’s experience. Emerg Radiol. 2021;29:317–28.
Park H-S, Ryu S-M, Cho S-J, Park S-M, Lim S-H. A treatment case of delayed aortic injury: the patient with posterior rib fracture. Korean J Thorac Cardiovasc Surg. 2014;47:406–8.
Jin L, Yang J, Kuang K, Ni B, Gao Y, Sun Y, Gao P, Ma W, Tan M, Kang H, Chen J, Li M. Deep-learning-assisted detection and segmentation of rib fractures from CT scans: Development and validation of FracNet. EBioMedicine. 2020;62:103106.
Redmon J, Divvala S, Girshick R, Farhadi A, You Only Look Once: Unified, Real-Time Object Detection, 2016 IEEE Conference on Computer Vision and, Recognition P. (CVPR), 2016, pp. 779–788.
Murphy CE, Raja AS, Baumann BM, Medak AJ, Langdorf MI, Nishijima DK, Hendey GW, Mower WR, Rodriguez RM. Rib Fracture Diagnosis in the Panscan Era. Ann Emerg Med. 2017;70:904–9.
Sirmali M, Türüt H, Topçu S, Gülhan E, Yazici U, Kaya S, Taştepe I. A comprehensive analysis of traumatic rib fractures: morbidity, mortality and management. Eur J Cardiothorac Surg. 2003;24:133–8.
Bugaev N, Breeze JL, Alhazmi M, Anbari HS, Arabian SS, Holewinski S, Rabinovici R. Magnitude of rib fracture displacement predicts opioid requirements. J Trauma Acute Care Surg. 2016;81:699–704.
Boyles AD, Taylor BC, Ferrel JR. Posterior rib fractures as a cause of delayed aortic injury: A case series and literature review. Injury Extra. 2013;44:43–5.
Yao L, Guan X, Song X, Tan Y, Wang C, Jin C, Chen M, Wang H, Zhang M. Rib fracture detection system based on deep learning. Sci Rep. 2021;11:23513.
Zhou QQ, Wang J, Tang W, Hu ZC, Xia ZY, Li XS, Zhang R, Yin X, Zhang B, Zhang H. Automatic Detection and Classification of Rib Fractures on Thoracic CT Using Convolutional Neural Network: Accuracy and Feasibility. Korean J Radiol. 2020;21:869–79.
Xiong S, Hu H, Liu S, Huang Y, Cheng J, Wan B. Improving diagnostic performance of rib fractures for the night shift in radiology department using a computer-aided diagnosis system based on deep learning: A clinical retrospective study. J Xray Sci Technol. 2023;31:265–76.

No competing interests reported.

Download PDF

Editorial decision: Revision requested
25 Oct, 2024
Editor assigned by journal
25 Oct, 2024
Submission checks completed at journal
24 Oct, 2024
First submitted to journal
15 Oct, 2024

You are reading this latest preprint version

Intelligent Detection and Grading Diagnosis of Fresh Rib Fractures Based on Deep Learning

Status:

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

Introduction

Methods

Data sets and classification criteria

CT acquisition

Image annotation

Model architecture:

Backbone Network

Neck Network

Head Network

Model evaluation

Statistical analysis

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1