Skip to main content
Insights into Imaging
Download PDF
  • Original Article
  • Open access
  • Published:

Determining rib fracture age from CT scans with a radiomics-based combined model: a multicenter retrospective study

Insights into Imaging volume 14, Article number: 214 (2023) Cite this article

Abstract

Objectives

We aimed to develop a combined model based on clinical and radiomic features to classify fracture age.

Methods

We included 1219 rib fractures from 239 patients from our center between March 2016 and September 2022. We created an external dataset using 120 rib fractures from 32 patients from another center between October 2019 and August 2023. According to tasks (fracture age between < 3 and ≥ 3 weeks, 3–12, and > 12 weeks), the internal dataset was randomly divided into training and internal test sets. A radiomic model was built using radiomic features. A combined model was constructed using clinical features and radiomic signatures by multivariate logistic regression, visualized as a nomogram. Internal and external test sets were used to validate model performance.

Results

For classifying fracture age between < 3 and ≥ 3 weeks, the combined model had higher areas under the curve (AUCs) than the radiomic model in the training set (0.915 vs 0.900, p = 0.009), internal test (0.897 vs 0.854, p < 0.001), and external test sets (0.881 vs 0.811, p = 0.003). For classifying fracture age between 3–12 and > 12 weeks, the combined model had higher AUCs than the radiomic model in the training model (0.848 vs 0.837, p = 0.12) and internal test sets (0.818 vs 0.793, p < 0.003). In the external test set, the AUC of the nomogram-assisted radiologist was 0.966.

Conclusion

The combined radiomic and clinical model showed good performance and has the potential to assist in the classification of rib fracture age. This will be beneficial for clinical practice and forensic decision-making.

Critical relevance statement

This study describes the development of a combined radiomic and clinical model with good performance in the classification of the age of rib fractures, with potential clinical and forensic applications.

Key points

• Complex factors make it difficult to determine the age of a fracture.

• Our model based on radiomic features performed well in classifying fracture age.

• Associating the radiomic features with clinical features improved the model’s performance.

Graphical Abstract

Introduction

Thoracic injuries are common among trauma patients [1]. Rib fractures are the most common form of thoracic trauma, serving as a marker of injury severity and a guide for legal decision-making [2, 3]. In the USA, 300,000 patients were diagnosed with rib fractures in 2004 [4], which increased to 350,000 in 2017 [5]. The risk of fracture increases significantly with age [6, 7]. Therefore, rib fracture cases are expected to increase with global aging [8]. The cause of rib fractures varies between age groups; they are usually caused by falls in older adults, traffic accidents in adults, and violent conflicts in adolescents [9, 10].

Accident compensation usually requires clarification of injury events leading to rib fractures [11]. In addition, the specific timing of rib fractures during accidents is important in legal cases. Providing the age of a rib fracture can clarify the association with a specific physical event and rule out irrelevant injuries, which is beneficial for settling compensation issues. Furthermore, the fracture healing state can be clinically assessed based on a patient’s radiological presentation and physical examination at different ages after the fracture [12]. However, this is influenced by multiple factors, including the patient’s age, nutritional status, sex, fracture location, and vascular injury, making it highly challenging to assess fracture union [13]. Combining imaging with fracture timing can improve surgeons’ confidence in assessing the healing state. Analyzing the fracture healing process can guide further treatment and care practices for patients with rib fractures [14].

Radiomics provides a new perspective for doctors by extracting quantitative features from medical images to reflect the heterogeneity of lesions that are unobservable by the naked eye. Its primary benefit lies in the quantitative and more objective interpretation of medical images, potentially overcoming the limitations of visual image assessment [15, 16]. To the best of our knowledge, no study has investigated whether a model derived from computed tomography (CT) imaging and clinical information can aid the diagnosis of fracture age. We aimed to evaluate the performance of a model built using radiomic and clinical data to diagnose fracture age.

Methods

Ethics consideration

This retrospective study was approved by the ethics committee of our hospital (No. 20220069), which waived the requirement for informed consent.

Study population of the internal dataset

We searched the Picture Archiving and Communication System of our hospital using the keyword phrase “chest trauma evaluation” to identify patients who visited our hospital for chest trauma evaluation between March 2016 and September 2022. The inclusion criteria comprised the following: (1) patient age > 18 years; (2) patients who experienced a traumatic event (including traffic accidents, falls, and fights involving blunt chest trauma); (3) imaging reports confirming at least one rib fracture; (4) slice thickness after image reconstruction < 1.5 mm; and (5) case reports documenting the precise time of fractures. The exclusion criteria comprised the following: (1) malignant and benign bone tumors involving the ribs; (2) metastatic diseases involving the ribs; (3) significant artifacts in CT images; and (4) surgical internal fixation of rib fractures. We obtained demographic features, including age and sex, from the patient hospital records. The scan parameters are described in the Supplementary Material. We identified 71 patients with rib fractures imaged < 3 weeks after trauma and matched them by age to ensure that the number of fractures at 3–12 weeks and > 12 weeks was approximately equal to that at < 3 weeks. These three frames for depicting fracture ages were sourced from previously published literature [17, 18]. For each patient, 12 bilateral ribs were included in the dataset.

Finally, as shown in Fig. 1, based on our inclusion and exclusion criteria, 1219 rib fractures (< 3 weeks: n = 323; 3–12 weeks: n = 492; > 12 weeks: n = 404) from 239 patients (< 3 weeks: n = 71; 3–12 weeks: n = 90; > 12 weeks: n = 78) were included in our internal dataset. For classifying fracture age between < 3 and ≥ 3 weeks, 1219 fractures were randomly assigned to the training (n = 853) and internal test sets (n = 366). For classifying fracture age between 3–12 and > 12 weeks, 896 fractures were randomly assigned to the training (n = 627) and internal test sets (n = 269).

Fig. 1
figure 1

Flow chart of the recruitment pathway for the datasets used in this study

Full size image

Study population of the external dataset

To build a validation external dataset, we included 120 rib fractures from 32 patients (< 3 weeks: n = 16; ≥ 3 weeks: n = 16) who underwent chest trauma evaluation between October 2019 and August 2023, in another hospital in Shanghai according to the same inclusion and exclusion criteria to construct. This external dataset was used to confirm the efficacy of our model and assess the performance of radiologists in classifying rib fracture age based on CT images using the model. The scan parameters used for CT examination are provided in the Supplementary Material. Detailed clinical features are listed in Table S3.

Extraction of radiomic and clinical features

Referring to the diagnostic CT reports, a radiologist with 5 years of experience in chest diagnostics manually delineated the volume of interest (VOI) at the fractured ends of the rib using the 3D Slicer software (https://www.slicer.org) (Figure S1). Subsequently, another radiologist with 20 years of experience in chest diagnostics confirmed the VOIs. Next, we used the PyRadiomics package in Python to extract radiomic features from the VOIs [19]. For each VOI, we extracted 107 radiomic features. The detailed outlining process and radiomic features are shown in the Supplementary Material.

The clinical features included demographic and CT image features. The clinical features were as follows: (1) pleural inflammation, (2) 1–4 rib fractures, (3) lateral fractures, (4) bone fragments, (5) multiple fractures of the same rib, (6) cartilage junction fractures, (7) periosteal callus information, (8) intramedullary callus information, and (9) sex.

Selection of radiomic and clinical features

We hypothesized that meaningful features might differ when classifying fracture age between < 3 weeks, ≥ 3 weeks, 3–12 weeks, and > 12 weeks. Therefore, we selected the features separately according to these time frames.

All radiomic features were normalized using z-score normalization before feature selection. To minimize the impact of dimensionality, the selection of features was conducted in three steps using the training set. First, we calculated the p value for each feature using the t-test and selected features with p values < 0.05 for further consideration. Spearman’s correlation coefficients were calculated for the selected features to avoid the underlying severe linear dependence. When the value was < 0.9, we determined this to mean that there was no correlation between the selected features. Lastly, we used the least absolute shrinkage and selection operator (LASSO) analysis to determine the most useful radiomic features. Clinical features were selected using univariate and multivariate logistic regression analyses.

Development of radiomic and combined models

We used two feature sets instead of one for the precise classification of the fracture age because of the difference in features for fracture age between < 3 weeks, ≥ 3 weeks, 3–12 weeks, and > 12 weeks. Regarding both levels of classification, we developed radiomic and combined models.

The base classifier acts as a feature encoder and has a significant impact on classification [20]. We compared the performance of five common machine learning (ML) algorithms, support vector machine (SVM), K-nearest neighbor, random forest, decision tree, and eXtreme Gradient Boosting to determine the most suitable classifier for fracture age assessment. Hyperparameters were optimized using grid search and tenfold cross-validation. Only the hyperparameters listed in Table S4 were adjusted, while all other hyperparameters were maintained at the default values specified by models. The filtered radiomic features were incorporated into the classifier and trained within the training set to obtain the radiomic model.

Using one of the two levels of binary classification (< 3 weeks or 3–12 weeks) as the reference standard, coded as 0 in the training of the classifier, we obtained the predicted probability of each fracture in the remaining two groups as the radiomic signature. Lastly, we developed the combined model based on the radiomic signature and clinical features using multivariate logistic regression.

Nomogram building

In order to simplify the combined model into an easy-to-understand tool, a nomogram was utilized for constructing a simplified graphical display. The sum of nomogram points was calculated based on clinical features and radiomic signature. For the convenience of calculations, instantly deployable online calculators were developed.

Statistical analysis

Statistical analysis was performed using SPSS 26.0 (IBM) and R Studio (ver. 4.3.1). Clinical features were measured using chi-squared or Fisher’s exact-probability testing as all were categorical variables. The clinical features with p values of < 0.05 in both univariate and multivariate logistic regression analyses were included in the combined model. Differences in radiomic features were assessed using the t-test or the Mann–Whitney U test as they were continuous variables. The statistical significance level was set at p values of < 0.05.

Results

Selection of clinical features

The clinical features were compared (Table S1 and S2), and we observed no significant differences between the clinical features of the training and internal test sets (p = 0.142–0.988). Results of univariate and multivariate logistic regression analyses for clinical features associated with the classification of fracture age are presented in Table 1. For classifying fracture age between < 3 and ≥ 3 weeks, six independent clinical features were selected, including sex, fractures of ribs 1–4, pleural inflammation, lateral fractures, intramedullary callus formation, and periosteal callus formation. For classifying fracture age between 3–12 and > 12 weeks, three independent clinical features were selected, including pleural inflammation, intramedullary callus formation, and multiple fractures of the same rib.

Table 1 Results of univariate and multivariate logistic regression analyses for classification of fracture age
Full size table

Development and validation of the radiomic model

Using the < 3 weeks as the reference standard, 17 radiomic features were obtained using the t-test, Spearman’s correlation analysis, and LASSO regression analysis, including 3 shape, 4 intensity, and 10 texture features (Figure S2).

Using the 3–12 weeks as the reference standard, 10 radiomic features were obtained using the t-test, Spearman’s correlation analysis, and LASSO regression analysis, including 3 shape, 2 intensity, and 5 texture features (Figure S3).

SVM performed the best on fracture age assessment (Table S5); therefore, we chose it as the base classifier for the radiomic model and obtained the radiomic signature. Regarding the classification of the fracture age between < 3 weeks and ≥ 3 weeks, using < 3 weeks as the reference standard, the performance of the radiomic model showed an AUC of 0.897 (95% confidence interval [CI]: 0.875–0.925) and 0.854 (95% CI: 0.804–0.903) for the training and internal test sets, respectively. Regarding the classification of the fracture age between 3–12 weeks and > 12 weeks, using 3–12 weeks as the reference standard, the performance of the radiomic model showed an AUC of 0.837 (95% CI: 0.805–0.868) and 0.793 (95% CI: 0.738–0.847) for the training and internal test sets, respectively.

Development and visualization of the combined model

For classifying fracture age between < 3 and ≥ 3 weeks, using the selected clinical features and radiomic signature, a combined model was developed and visualized in the form of a nomogram (Fig. 2). The online version of the nomogram could be used at https://myr234r.shinyapps.io/DynNomapp3/, and its user interface is illustrated in Fig. 4a. The AUC values of the combined model were 0.914 (95% CI = 0.892–0.935) and 0.889 (95% CI = 0.846–0.932) for the training and internal test sets, respectively. The calibration curves revealed good predictive accuracy between the actual probability and predicted probability (Fig. 2).

Fig. 2
figure 2

The combined model for classification between < 3 weeks and ≥ 3 weeks. The calibration curves of this model in the training and internal test sets were obtained by resampling 1000 times. The dotted line indicates the ideal ability, and the solid line represents the real ability of the model. The combined performed better when the solid line was closer to the dotted line

Full size image

For classifying fracture age between 3–12 and > 12 weeks, using the selected clinical features and radiomic signature, a combined model was developed and visualized in the form of a nomogram (Fig. 3). An online version of the nomogram could be used at https://myr234r.shinyapps.io/DynNomapp12/ and its user interface is illustrated in Fig. 4b. The AUC values of the combined model were 0.848 (95% CI: 0.818–0.878) and 0.818 (95% CI: 0.767–0.869) in the training and internal test sets, respectively. The calibration curves of the nomogram showed acceptable agreement between prediction and actual observation (Figs. 2 and 3).

Fig. 3
figure 3

The combined model for classification between 3–12 weeks and > 12 weeks. The calibration curves of this model in the training and internal test sets were obtained by resampling 1000 times. The dotted line indicates the ideal ability, and the solid line represents the real ability of the model. The combined performed better when the solid line was closer to the dotted line

Full size image
Fig. 4
figure 4

Online dynamic nomograms for classifying rib fracture age between < 3 and ≥ 3 weeks (a) and between 3–12 weeks and > 12 weeks (b). The figure displays the probabilities (with a 95% confidence interval) determined by the combined model for rib fracture ages of ≥ 3 weeks and > 12 weeks. The specific predicted values are provided in the “Numerical Summary” tab

Full size image

Performance comparison between the radiomic and combined models

The DeLong test showed that the combined model performed better than the radiomic model for classification between < 3 weeks, ≥ 3 weeks, 3–12 weeks, and > 12 weeks (Table 2).

Table 2 Performance comparison of the radiomic and combined models
Full size table

For classifying fracture age between < 3 weeks and ≥ 3 weeks, the receiver operating characteristic (ROC) curves of the radiomic and combined models are shown in Fig. 5. In the training set, the combined model (AUC = 0.915, 95% CI = 0.893–0.936) outperformed the radiomic model (AUC = 0.900, 95% CI = 0.875–0.925; p = 0.009). Similarly, in the internal test set, the combined model (AUC = 0.889, 95% CI = 0.846–0.932) outperformed the radiomic model (AUC = 0.854, 95% CI = 0.804–0.903; p < 0.001).

Fig. 5
figure 5

Receiver operating characteristic (ROC) analysis revealed that for the classification between < 3 and ≥ 3 weeks (a, b), as well as between 3–12 and > 12 weeks (c, d), the performance of the combined model (red line) surpassed that of the radiomics model (blue line) in both the training and internal test sets

Full size image

For classifying the fracture age between 3–12 weeks and > 12 weeks, the ROC curves of the radiomic and combined models are shown in Fig. 5. In the training set, the AUC of the combined model (0.848, 95% CI = 0.818–0.878) was slightly higher than that of the radiomic model (0.837, 95% CI = 0.805–0.868; p = 0.12). However, in the internal test set, the combined model (AUC = 0.818, 95% CI = 0.767–0.869) performed significantly better than the radiomic model (AUC = 0.793, 95% CI = 0.738–0.847; p = 0.003).

Performance in the external test set

In the external test set, we reached the same conclusion as that in the internal dataset (Table 3). We evaluated a radiologist’s (engaged in musculoskeletal imaging diagnosis for 15 years) performance in classifying rib fracture age aided by the combined model. For classifying fracture age between < 3 and ≥ 3 weeks, the AUC values for the radiomic model, combined model, and radiologist aided by the combined model were 0.811 (95% CI = 0.733–0.889), 0.881 (95% CI = 0.814–0.949), and 0.966 (95% CI = 0.935–0.996), respectively. This indicated that the combined model was better than the radiomic model, and the collaborative integration of the radiologist with the combined model further enhanced classification performance (Figure S4).

Table 3 Performance comparison of the radiomic model, combined model, and nomogram-assisted radiologist for classifying fracture age between < 3 and ≥ 3 weeks in the external test set
Full size table

Discussion

The results from the current study showed that our combined model, based on clinical and radiomic features, could accurately classify the age of rib fractures. The good performance of the radiologist assisted by the model demonstrated the potential feasibility of the model in practical applications, and it is expected to provide value for clinical and forensic decision-making.

After a fracture, the fracture region undergoes a series of characteristic events over time called the inflammatory, repair, and remodeling phases, all of which are influenced by many complex factors [21, 22]. Sex, 1–4 rib fractures, and lateral fractures impact fracture healing through estrogen levels, vascular injury, and respiration and body movements, respectively [14, 23,24,25,26]. Pleural inflammation and callus formation are considered important imaging features of the inflammatory and reparative phases of fracture healing, respectively [21, 27]. In late fracture healing, multiple fractures of the same rib end healing early by promoting bone resorption [28]. Therefore, the above clinical features were included in our combined model. Radiomic enables the quantification of heterogeneity of fracture regions, thereby compensating for the limitations of visual image assessment. The combination of all pertinent features (visually observable clinical features and visually difficult-to-identify radiomic features) in the combined model allows for a comprehensive assessment of fracture regions, providing radiologists with additional data-driven insights. Thus, aided by the combined model, radiologists achieve diagnoses by integrating subjective and objective interpretations of medical images, together with their individual clinical experience. This fusion of knowledge from both human expertise and data-driven analysis enhances the diagnostic decision-making process, ultimately improving diagnostic performance.

Radiomics are useful for detecting changes in bone microarchitecture due to injury and aging [29,30,31,32]. Several studies [33,34,35] utilized radiomic features from T2-weighted images to achieve early diagnosis of osteoarthritis. Lin et al. [36] developed a radiomic and clinical features-based model to predict the prognosis of osteoarthritis, with an AUC of 0.83 (95% CI = 0.70–0.96) in the validation set. Recent studies [37,38,39,40,41,42,43,44,45] also demonstrated the ability of radiomics in osteoporosis detection and fracture prediction. Apart from its applications in diagnosing bone disorders, radiomics research is starting to emerge in the field of forensic medicine. Giorgio et al. [46] demonstrated, based on a limited sample (n = 4), that CT-based radiomics could be associated with time of death. Subsequently, Klontzas et al. [47] applied CT-based radiomics to predict postmortem interval, yielding promising results in this exploratory study (AUC = 0.75, 95%CI = 0.584–0.916). A previous forensic study [48] indicated an association between CT-based quantitative parameters and the age of rib fractures. However, the limited sample size (n = 9) and the finite number of extracted features (n = 5) were insufficient for conducting ML analysis. The aforementioned radiomics research in forensic medicine predominantly focused on postmortem forensic investigations, whereas this study pertained to trauma assessment for legal proceedings. Our study was based on principles similar to those of the aforementioned studies, namely, that radiomics, which reflect changes in bone microstructure during the bone healing process, can be used for fracture age classification, and that the addition of clinical features can further enhance model performance.

A strength of this study is the development of a single radiomic signature that summarizes various radiomic features, providing a useful tool for clinical practice. In addition, the combined model was visualized as a nomogram, providing readers with more tangible interpretation of each factor’s impact on the classification. We have made our combined model public on an open-access website in the form of online calculator to simplify the calculations. However, some limitations exist in our study. First, this was a retrospective study. A prospective study is needed to further validate our model. Second, older participants were included in this study (median age, 57 years; interquartile range, 50–63 years) and the model’s applicability to younger patients is unknown. Younger patients should be included in future studies to improve the robustness of the model. Lastly, in the absence of data > 12 weeks from the external center, we did not perform external validation for the model in classifying fracture age between 3–12 and > 12 weeks. Future studies could be designed to overcome these limitations.

Conclusions

Our study established a combined model for rib fracture age classification based on CT images. The combined model and the model-assisted radiologist achieved good performance in classifying rib fracture age. This model has the potential to influence clinical practice and forensic decision-making.

Availability of data and materials

All the data will be shared upon reasonable request by the corresponding author.

Abbreviations

AUC:

Area under the curve

CI:

Confidence interval

LASSO:

Least absolute shrinkage and selection operator

OR:

Odds ratio

ROC:

Receiver operating characteristic

SVM:

Support vector machine

VOI:

Volume of interest

References

  1. Head W, Kumar N, Thomas C, Leon S, Dieffenbaugher S, Eriksson E (2021) Are rib fractures stable? An analysis of progressive rib fracture offset in the acute trauma setting. J Trauma Acute Care Surg 91:917–922

    Article  PubMed  Google Scholar 

  2. O’Donovan S, van den Heuvel C, Baldock M, Humphries MA, Byard RW (2022) Fatal blunt chest trauma: an evaluation of rib fracture patterns and age. Int J Legal Med 136:1351–1357

    Article  PubMed  PubMed Central  Google Scholar 

  3. Yu L, Baumann BM, Raja AS et al (2020) Blunt traumatic aortic injury in the pan-scan era. Acad Emerg Med 27:291–296

    Article  PubMed  Google Scholar 

  4. Lafferty PM, Anavian J, Will RE, Cole PA (2011) Operative treatment of chest wall injuries: indications, technique, and outcomes. J Bone Joint Surg Am 93:97–110

    Article  PubMed  Google Scholar 

  5. Pieracci FM, Majercik S, Ali-Osman F et al (2017) Consensus statement: surgical stabilization of rib fractures rib fracture colloquium clinical practice guidelines. Injury 48:307–321

    Article  PubMed  Google Scholar 

  6. Ensrud KE (2013) Epidemiology of fracture risk with advancing age. J Gerontol A Biol Sci Med Sci 68:1236–1242

    Article  PubMed  Google Scholar 

  7. Bonicelli A, Kranioti EF, Xhemali B, Arnold E, Zioupos P (2022) Assessing bone maturity: compositional and mechanical properties of rib cortical bone at different ages. Bone 155:116265

    Article  PubMed  CAS  Google Scholar 

  8. He Z, Zhang D, Xiao H et al (2019) The ideal methods for the management of rib fractures. J Thorac Dis 11:S1078–s1089

    Article  PubMed  PubMed Central  Google Scholar 

  9. Coary R, Skerritt C, Carey A, Rudd S, Shipway D (2020) New horizons in rib fracture management in the older adult. Age Ageing 49:161–167

    Article  PubMed  Google Scholar 

  10. Chua MT, Pan DST, Lee MZ et al (2022) Epidemiology and outcomes of older trauma patients in Singapore: a multicentre study. Injury 53:3149–3155

    Article  PubMed  Google Scholar 

  11. Lau G, Gabbe BJ, Giummarra MJ (2021) Patterns and predictors of personal responsibility attributions after major trauma. Injury 52:2886–2895

    Article  PubMed  Google Scholar 

  12. Morshed S (2014) Current options for determining fracture union. Adv Med 2014:708574

    Article  PubMed  PubMed Central  Google Scholar 

  13. Zura R, Xu ZJ, Della Rocca GJ, Mehta S, Steen RG (2017) When is a fracture not “fresh”? Aligning reimbursement with patient outcome after treatment with low-intensity pulsed ultrasound. J Orthop Trauma 31:248–251

    Article  PubMed  Google Scholar 

  14. Fisher JS, Kazam JJ, Fufa D, Bartolotta RJ (2019) Radiologic evaluation of fracture healing. Skeletal Radiol 48:349–361

    Article  PubMed  Google Scholar 

  15. Gillies RJ, Kinahan PE, Hricak H (2016) Radiomics: images are more than pictures, they are data. Radiology 278:563–577

    Article  PubMed  Google Scholar 

  16. Abbasian Ardakani A, Bureau NJ, Ciaccio EJ, Acharya UR (2022) Interpretation of radiomics features-a pictorial review. Comput Methods Programs Biomed 215:106609

    Article  PubMed  Google Scholar 

  17. Wu X, Jiang Y (2015) Old fracture. Zhonghua Wai Ke Za Zhi 53:460–463

    PubMed  Google Scholar 

  18. Arredondo-Gómez E (2007) Treatment of traumatic clavicular pseudoarthrosis with the Hunec Colchero nail. Acta Ortop Mex 21:63–68

    PubMed  Google Scholar 

  19. van Griethuysen JJM, Fedorov A, Parmar C et al (2017) Computational radiomics system to decode the radiographic phenotype. Cancer Res 77:e104–e107

    Article  PubMed  PubMed Central  Google Scholar 

  20. Deist TM, Dankers F, Valdes G et al (2018) Machine learning algorithms for outcome prediction in (chemo)radiotherapy: an empirical comparison of classifiers. Med Phys 45:3449–3459

    Article  PubMed  PubMed Central  Google Scholar 

  21. Claes L, Recknagel S, Ignatius A (2012) Fracture healing under healthy and inflammatory conditions. Nat Rev Rheumatol 8:133–143

    Article  PubMed  CAS  Google Scholar 

  22. Palanisamy P, Alam M, Li S, Chow SKH, Zheng YP (2022) Low-intensity pulsed ultrasound stimulation for bone fractures healing: a review. J Ultrasound Med 41:547–563

    Article  PubMed  Google Scholar 

  23. Talbot BS, Gange CP Jr, Chaturvedi A, Klionsky N, Hobbs SK, Chaturvedi A (2017) Traumatic rib injury: patterns, imaging pitfalls, complications, and treatment. Radiographics 37:628–651

    Article  PubMed  Google Scholar 

  24. Claes L (2011) Biomechanical principles and mechanobiologic aspects of flexible and locked plating. J Orthop Trauma 25(Suppl 1):S4–7

    Article  PubMed  Google Scholar 

  25. Rizzoli R, Reginster JY, Boonen S et al (2011) Adverse reactions and drug-drug interactions in the management of women with postmenopausal osteoporosis. Calcif Tissue Int 89:91–104

    Article  PubMed  PubMed Central  CAS  Google Scholar 

  26. Fischer V, Haffner-Luntzer M (2022) Interaction between bone and immune cells: implications for postmenopausal osteoporosis. Semin Cell Dev Biol 123:14–21

    Article  PubMed  Google Scholar 

  27. Kolar P, Schmidt-Bleek K, Schell H et al (2010) The early fracture hematoma and its potential role in fracture healing. Tissue Eng Part B Rev 16:427–434

    Article  PubMed  Google Scholar 

  28. Loi F, Córdova LA, Pajarinen J, Lin TH, Yao Z, Goodman SB (2016) Inflammation, fracture and bone repair. Bone 86:119–130

    Article  PubMed  PubMed Central  CAS  Google Scholar 

  29. Le Corroller T, Halgrin J, Pithioux M, Guenoun D, Chabrand P, Champsaur P (2012) Combination of texture analysis and bone mineral density improves the prediction of fracture load in human femurs. Osteoporos Int 23:163–169

    Article  PubMed  Google Scholar 

  30. Xuan A, Chen H, Chen T et al (2023) The application of machine learning in early diagnosis of osteoarthritis: a narrative review. Ther Adv Musculoskelet Dis 15:1759720x231158198

    Article  PubMed  PubMed Central  Google Scholar 

  31. Fritz B, Yi PH, Kijowski R, Fritz J (2023) Radiomics and deep learning for disease detection in musculoskeletal radiology: an overview of novel MRI- and CT-based approaches. Invest Radiol 58:3–13

    Article  PubMed  Google Scholar 

  32. Levi R, Garoli F, Battaglia M et al (2023) CT-based radiomics can identify physiological modifications of bone structure related to subjects’ age and sex. Radiol Med 128:744–754

    Article  PubMed  Google Scholar 

  33. Peuna A, Hekkala J, Haapea M et al (2018) Variable angle gray level co-occurrence matrix analysis of T(2) relaxation time maps reveals degenerative changes of cartilage in knee osteoarthritis: Oulu knee osteoarthritis study. J Magn Reson Imaging 47:1316–1327

    Article  PubMed  Google Scholar 

  34. Xie Y, Dan Y, Tao H et al (2021) Radiomics feature analysis of cartilage and subchondral bone in differentiating knees predisposed to posttraumatic osteoarthritis after anterior cruciate ligament reconstruction from healthy knees. Biomed Res Int 2021:4351499

    Article  PubMed  PubMed Central  Google Scholar 

  35. Hirvasniemi J, Klein S, Bierma-Zeinstra S, Vernooij MW, Schiphof D, Oei EHG (2021) A machine learning approach to distinguish between knees without and with osteoarthritis using MRI-based radiomic features from tibial bone. Eur Radiol 31:8513–8521

    Article  PubMed  PubMed Central  Google Scholar 

  36. Lin T, Peng S, Lu S et al (2023) Prediction of knee pain improvement over two years for knee osteoarthritis using a dynamic nomogram based on MRI-derived radiomics: a proof-of-concept study. Osteoarthritis Cartilage 31:267–278

    Article  PubMed  CAS  Google Scholar 

  37. Farzi M, Pozo JM, McCloskey E et al (2022) Quantitating age-related BMD textural variation from DXA region-free-analysis: a study of hip fracture prediction in three cohorts. J Bone Miner Res 37:1679–1688

    Article  PubMed  CAS  Google Scholar 

  38. Hong N, Park H, Kim CO et al (2021) Bone radiomics score derived from DXA hip images enhances hip fracture prediction in older women. J Bone Miner Res 36:1708–1716

    Article  PubMed  Google Scholar 

  39. Rastegar S, Vaziri M, Qasempour Y et al (2020) Radiomics for classification of bone mineral loss: a machine learning study. Diagn Interv Imaging 101:599–610

    Article  PubMed  CAS  Google Scholar 

  40. Huang CB, Hu JS, Tan K, Zhang W, Xu TH, Yang L (2022) Application of machine learning model to predict osteoporosis based on abdominal computed tomography images of the psoas muscle: a retrospective study. BMC Geriatr 22:796

    Article  PubMed  PubMed Central  Google Scholar 

  41. Dai H, Wang Y, Fu R et al (2023) Radiomics and stacking regression model for measuring bone mineral density using abdominal computed tomography. Acta Radiol 64:228–236

    Article  PubMed  Google Scholar 

  42. Zhang J, Liu J, Liang Z et al (2023) Differentiation of acute and chronic vertebral compression fractures using conventional CT based on deep transfer learning features and hand-crafted radiomics features. BMC Musculoskelet Disord 24:165

    Article  PubMed  PubMed Central  Google Scholar 

  43. Biamonte E, Levi R, Carrone F et al (2022) Artificial intelligence-based radiomics on computed tomography of lumbar spine in subjects with fragility vertebral fractures. J Endocrinol Invest 45:2007–2017

    Article  PubMed  CAS  Google Scholar 

  44. Zhang H, Yuan G, Wang C et al (2023) Differentiation of benign versus malignant indistinguishable vertebral compression fractures by different machine learning with MRI-based radiomic features. Eur Radiol 33:5069–5076

    Article  PubMed  Google Scholar 

  45. Chen YC, Li YT, Kuo PC et al (2023) Automatic segmentation and radiomic texture analysis for osteoporosis screening using chest low-dose computed tomography. Eur Radiol 33:5097–5106

    Article  PubMed  Google Scholar 

  46. De-Giorgio F, Ciasca G, Fecondo G et al (2022) Post mortem computed tomography meets radiomics: a case series on fractal analysis of post mortem changes in the brain. Int J Legal Med 136:719–727

    Article  PubMed  PubMed Central  Google Scholar 

  47. Klontzas ME, Leventis D, Spanakis K, Karantanas AH, Kranioti EF (2023) Post-mortem CT radiomics for the prediction of time since death. Eur Radiol 33(11):8387–8395.

  48. Viero A, Biehler-Gomez L, Messina C et al (2022) Utility of micro-CT for dating post-cranial fractures of known post-traumatic ages through 3D measurements of the trabecular inner morphology. Sci Rep 12:10543

    Article  PubMed  PubMed Central  CAS  Google Scholar 

Download references

Funding

This work was supported by the Shanghai Key Lab of Forensic Medicine, Key Lab of Forensic Science, Ministry of Justice, China (Academy of Forensic Science) (grant number: KF202113); the Youth Medical Talents—Medical Imaging Practitioner Program (grant number: AB83030002019004); Science and Technology Planning Project of Shanghai Science and Technology Commission (grant number: 22Y11910700); the Health Commission of Shanghai (grant number: 2018ZHYL0103); the National Natural Science Foundation of China (grant number: 61976238); Shanghai “Rising Stars of Medical Talent” Youth Development Program “Outstanding Youth Medical Talents” (SHWJRS (2021)-99), Emerging Talent Program (XXRC2213), and Leading Talent Program (LJRC2202) of Huadong Hospital, Excellent Academic Leaders of Shanghai (2022XD042); Shanghai Municipal Health Commission Program (202240366); and Ministry of Science and Technology Program (2020YFC2008600).

Author information

Author notes

  1. Yilin Tang and Liang Jin contributed equally to this work.

Authors and Affiliations

  1. Radiology Department, Huadong Hospital Affiliated to Fudan University, No. 221 West Yan’an Road, Jing’an District, Shanghai, 200040, China

    Yilin Tang, Liang Jin, Zhuangxuan Ma, Dechun Li & Ming Li

  2. Radiology Department, Huashan Hospital Affiliated to Fudan University, Shanghai, China

    Liang Jin

  3. Radiology Department, Shanghai Electric Power Hospital, Shanghai, China

    Wenbin Ji

  4. Department of Geriatrics and Gerontology, Huadong Hospital Affiliated to Fudan University, No. 221 West Yan’an Road, Jing’an District, Shanghai, 200040, China

    Wei Hong

  5. Institute of Functional and Molecular Medical Imaging, No. 221 West Yan’an Road, Jing’an District, Shanghai, 200040, China

    Ming Li

Authors
  1. Yilin Tang
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Liang Jin
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Wenbin Ji
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Zhuangxuan Ma
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Dechun Li
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Wei Hong
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Ming Li
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

Conceptualization, Y.T., L.J., and M.L.; methodology, Y.T., L.J., W.J., and D.L.; software, Y.T.; formal analysis, Y.T., Z.M., and L.J.; investigation, L.J.; resources, L.J. and M.L.; data curation, L.J., Y.T., W.H., and D.L.; writing—original draft, L.J.; writing—review and editing, L.J. and M.L.; supervision, W.H. and M.L.

Corresponding authors

Correspondence to Liang Jin, Wei Hong or Ming Li.

Ethics declarations

Ethics approval and consent to participate

This retrospective study was approved by the ethics committee of Huadong Hospital affiliated to Fudan University (NO. 20220069), which waived the requirement for informed consent.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1: Supplementary methods. Table S1.

Baseline characteristics of the internal dataset for classification between <3 and ≥3 weeks. Table S2. Baseline characteristics of the internal dataset for classification between 3‒12 weeks and >12 weeks. Table S3. Baseline characteristics of the external dataset. Table S4. Optimal hyperparameter configuration list for machine learning algorithms. Table S5. Performance comparison of different models. Figure S1. Delineation of the volume of interest (VOI) for the rib fracture region. Figure S2. LASSO regression analysis procedure for the selection of radiomic features. Figure S3. LASSO regression analysis procedure for the selection of radiomic features. Figure S4. Receiver operating characteristic (ROC) comparison among radiomic model, combined model, and nomogram-assisted radiologist using the external test set.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Tang, Y., Jin, L., Ji, W. et al. Determining rib fracture age from CT scans with a radiomics-based combined model: a multicenter retrospective study. Insights Imaging 14, 214 (2023). https://doi.org/10.1186/s13244-023-01546-y

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13244-023-01546-y

Keywords