Machine-learning-based on multimodality radiomics analysis for the Preoperative Prediction for local relapse in osteosarcoma

doi:10.21203/rs.3.rs-3851721/v1

PURPOSE:

This study aimed to identify patients with local relapse (≤ 2 years) in osteosarcoma after surgical resection and make better clinical decisions by constructing a preoperative predictive model based on radiograph and multiparametric magnetic resonance imaging (MRI).

MATERIALS AND METHODS:

A retrospective study of 92 consecutive patients (training set, n = 61; testing set, n = 31) with extremity high-grade osteosarcoma were enrolled. The imaging features for each patient were extracted from radiograph, multiparametric MRI (T1WI, T2WI and T1WI-CE). In order to select features, three steps including minimal-redundancy-maximum-relevance (mRMR), least absolute shrinkage and selection operator (LASSO) regression and the random forest recursive feature elimination (RF-RFE) were performed. The classification performance was evaluated with four classifiers: extreme gradient boosting (XGB), logistic regression (LR), support vector machine (SVM) and random forest (RF). The receiver-operating characteristic curve (ROC) and the area under the curve (AUC) were used to evaluate the performance of the classifiers. DeLong’s test was utilized for comparing the AUCs.

RESULTS:

The performance (AUC, sensitivity, specificity, and accuracy) of four classifiers (RF, SVM, LR and XGB) using radiograph-MRI as image inputs were stable (all Hosmer–Lemeshow index > 0.05) with the fair to good prognosis efficacy. The RF classifier using radiograph-MRI features as training inputs exhibited better performance (AUC = 0.806, 0.868) than that using MRI-only (AUC = 0.774, 0.771) and radiograph-only (AUC = 0.613 and 0.627) in the training and testing sets (p < 0.05) while the other three classifiers showed no difference between MRI only and radiograph-MRI models.

CONCLUSION:

The tumoral radiograph and multiparametric MRI radiomics model can promisingly predict local relapse in extremity high-grade osteosarcoma. Our results highlighted the potential value of the tumoral radiomic model in osteosarcoma management.

Radiomic

Predictive value of tests

X-ray

Magnetic resonance imaging

Osteosarcoma

relapse

Osteosarcoma is the most prevalent primary malignant bone tumor, predominantly occurring in the pediatric and adolescent populations. [1] Despite the advancements in treatment strategies, such as surgery and chemotherapy, the prognosis for osteosarcoma patients remains challenging due to the high incidence of local relapse and distant metastasis.[2] Among the complications, local relapse following surgical resection is of significant concern. Hence, accurate preoperative prediction of relapse risk is critical in guiding treatment decisions and enhancing patient outcomes.

Traditional prognostic factors, such as tumor size, grade, and histological subtype, have limited predictive value in osteosarcoma. Recently, radiomic analysis, a technique that involves extracting quantitative features from medical imaging modalities, has emerged as a promising approach to overcome these limitations.[3–5] By analyzing texture, shape, and intensity characteristics of tumor images, radiomic analysis can unveil subtle patterns and relationships that may evade human observation. This enables the extraction of potentially valuable prognostic information and facilitates the development of predictive models.

In this study, we propose machine-learning-based approach that utilizes radiomic analysis of X-ray and Magnetic Resonance Imaging (MRI) images for preoperative prediction of local relapse in osteosarcoma patients. X-ray images provide valuable information regarding bone structure and the potential presence of metastasis, while MRI images offer detailed insights into tumor characteristics, including size, location, and vascularity. By amalgamating these imaging modalities and leveraging advanced machine learning techniques, our aim is to construct robust predictive model that accurately stratifies osteosarcoma patients based on their relapse risk.

By amalgamating these imaging modalities and leveraging advanced machine learning techniques, our aim is to construct a robust predictive model that accurately stratifies osteosarcoma patients based on their relapse risk. All patients will have X-ray and MRI images available, obtained prior to treatment initiation. Radiomic features will be extracted from these images employing state-of-the-art algorithms. Machine learning algorithms, such as support vector machines (SVM) or random forests (RF), will be trained using this dataset to build a predictive model for the local relapse risk. The performance of the model will be evaluated through cross-validation and compared to traditional clinical risk factors.

This study holds the potential to significantly advance the field of osteosarcoma prognostication and aid in personalized treatment planning. By integrating radiomic analysis with machine learning, we can harness the intrinsic characteristics of tumor images and uncover novel prognostic biomarkers. Ultimately, this could lead to improved long-term outcomes for osteosarcoma patients by enabling early identification of high-risk individuals who may benefit from more aggressive therapeutic interventions.

2.1 Patient Selection

Informed consent was waived for this retrospective study, and approval was obtained from our institutional review board. The inclusion criteria were as follows: (i) osteosarcoma had not been treated surgically or medically prior to diagnosis; (ii) pre-treatment, each patient underwent radiograph and multiparametric MRI, including T1WI, T2WI, and CE-T1WI; (iii) diagnoses of osteosarcoma were confirmed through surgical resection and histopathological analysis; (iv) follow-up data for at least 2 years post-surgery were available. The exclusion criteria were as follows: (i) biopsy and locoregional therapy were conducted prior to the MRI and X-rays; (ii) poor image quality due to metallic artifacts or motion artifacts; (iii) missing images or relevant parameters; (iv) patients who did not undergo surgical resection of the primary tumor; (v) patients diagnosed with secondary osteosarcoma. A total of 92 patients who met these criteria were enrolled in the study. Subsequently, they were randomly divided into two groups: a training set consisting of 62 patients and a test group consisting of 30 patients. The training set was utilized for dimensionality reduction and modeling, while the test group was used for model validation.

2.2 X-ray and MR imaging acquisition

We conducted anteroposterior radiograph examination of lesions using GE Revolution XR/d and Philip Digital Diagnost X-ray equipment. All magnetic resonance imaging (MRI) scans were conducted using 1.5 or 3.0-T field strength magnets from Siemens Magnetom Aera 1.5, United Imaging uMR780 3.0, or GE Signa Excite and HDx 3.0. The parametric MR imaging consisted of axial T1-weighted images (TR: 457-798ms, TE: 6.7-13.2ms), T2-weighted images (TR: 3262-7904ms, TE: 83.0-106.3ms), and contrast-enhanced T1-weighted images (TR: 457-798ms, TE: 6.7-13.2ms). Gadolinium contrast agent was administered intravenously through the cubital vein at an injection rate of 2.5 ml/s and a dose of 0.1 mmol/kg body weight. For easier analysis of the patients' imaging datasets, we obtained Picture Archiving Communication System (PACS) images from our institutions and acquired Digital Imaging and Communications in Medicine (DICOM) images through download.

2.3 ROI segmentation

The region of interest (ROI) was segmented utilizing ITK-SNAP (version 3.8.0, http://www.itksnap.org/) layer by layer in the radiograph and MRI images by two radiologists (with 10 years and 15 years of imaging diagnostic experience, respectively). Every radiologist independently segmented the lesions without knowledge of the other's segmentation. Any discrepancies or disagreement in the specific image segmentation will be addressed through consensus meetings. Manual drawing was used to outline the soft tissue mass and bone destruction within the radiograph regions of interest (ROIs). During the MRI program, manual delineation was performed to outline any detectable tumors, necrosis, cyst degenerations, hemorrhages, periosteal reactions, and peritumoral edema.

2.4 Images Feature Extraction and Selection

Image preprocessing and feature extraction were conducted using the open-source package PyRadiomics 3.0 within Python software (version 3.7.6; http://www.radiomics.io/pyradiomics.html). To assess interobserver agreement, the same radiologist re-segmented 30 randomly selected radiograph and MR images after a two-month interval. Intra-rater reliability was evaluated by calculating intraclass correlation coefficients (ICCs). ICC values were categorized as poor (less than 0.5), moderate (0.5 to 0.75), good (0.75 to 0.9), or excellent (greater than 0.9). Features with ICC values above 0.75 were selected for inclusion during the feature selection process. [6]

The strategy for selecting features was as follows: As a first step, we ranked features based on mRMR (minimum-redundancy-maximum-relevance). The top 15 features were selected for the subsequent classifiers. Top features were selected based only on features from the training images. Second，the least absolute shrinkage and selection operator (LASSO) method on all features to grossly was applied to select those features with discriminative ability. The LASSO method restrains some feature coefficients to zero by adjusting the parameter λ. Third, the random forest recursive feature elimination (RF-RFE) was used to select the final feature subset based on calculating the area under the receiver operating characteristic curve (ROC-AUC) by using five fold cross-validation. All the radiomics features were standardized before feature selection. One-Hot encoding was used to encode categorical features.

2.5 Machine learning model

Finally, the final selected features were utilized for modelling with four well-known machine-learning classifiers: extreme gradient boosting (XGB), logistic regression (LR), support vector machine (SVM) and random forest (RF). All classifiers were implemented using the Scikit-Learn package (version 0.21.0) in python software, which provides an overall and friendly interface to access many machine-learning algorithms. The classifiers were trained or tuned in the training set to determine the best parameter configuration for the classifier model by using an grid search method with five fold cross validation. The performance of the predictive model was then evaluated in the testing set. The diagnostic performance of each model to differentiate LRs from NLRs was quantitatively evaluated by the area under the curve (AUC) of the receiver operating characteristic (ROC), accuracy, sensitivity and specificity. The calibration of the radiomics model was calculated by the Hosmer–Lemeshow test. DeLong’s test was utilized for comparisons of AUCs. A p value < 0.05 indicates a significant difference. All the above processes were implemented in Python (version 3.7.6), except DeLong’s test, which was implemented with MedCalc 19.8 software (MedCalc, Ostend, Belgium). Figure 1 depicts a flow diagram that illustrates the process of radiomics analysis.

2.5 Statistical analysis

The t-test was employed to compare continuous variables, whereas the chi-squared or Fisher's exact test was utilized for the classification of variables between groups. Performance of the model was evaluated using receiver operating characteristic curves (ROC) curves. We calculated the sensitivity, specificity, and AUC. The statistical significance of differences between ROC curves was determined using DeLong's method. R 3.5.1 and Python 3.5.6 were used to conduct statistical analyses. A 2-tailed p-value of 0.05 was considered significant.

3.1 Patient characteristics

A total of 92 osteosarcoma patients were included in this study (63 men and 39 women), including 33 LRs and 59 NLRs. In accordance with the number of patients, the patients were randomly divided into 7:3 training and testing sets. No significant difference between LRs and NLRs in age, gender, pathological type, location, tumor size, bone destruction type, ALP, LDH and metastasis in the entire data, the training and testing sets. Table 1-2 shows the clinical characteristics of the osteosarcoma patients.

3.2 Feature Selection and Radiomics Signature Construction

There were a total of 1502 features extracted from the T1WI, T2WI, CE-T1WI and X-ray images. As a result of the ICC test (ICC＞0.75), 1037, 1070, 1037, and 1181 features collected from T1WI, T2WI, CE-T1WI and radiograph were entered into the subsequent analysis. After data standardisation and dimensionality reduction by mRMR, LASSO and RFE-RF, 1 feature in radiograph, 11 features in MRI (4 features in T1WI, 2 features in T2WI, 5 features in T1WI-CE) were determined to construct the radiomics models. The most significant radiomics features of different models and the measure of feature importance of RF. (Table 3)

3.3 Performance of Models

The average performance (AUC, Sens, Spec, and ACC) of classifiers for different radiomic models are presented in Table 4, and ROC curves are shown in Figure. 2-4. The pairwise comparison of ROC curves indicated that the combination of X-ray and MRI demonstrated the best performance for these machine-learning methods, followed by the MRI and X-ray in the test set. In the test set, the efficiency of the four classifiers in the combination of X-ray and MRI was RF (AUC = 0.868), SVM (AUC = 0.768), LR (AUC = 0.759) and XGB (AUC = 0.682) in descending order, with no significant difference (p > 0.05). The p value of the four classifiers in the combination of X-ray and MRI in the Hosmer–Lemeshow test were more than 0.05, so the calibration of the four classifiers were reliable. There was no difference between MRI and the combination of X-ray and MRI for the XGB, LR, and SVM classifiers. The RF classifier in the combination of X-ray and MRI exhibited the better performance (AUC = 0.868) than MRI (AUC = 0.7705) and X-ray (AUC = 0.627) via the DeLong test (p < 0.05) in the test set. Figure 5 shows an example of two correctly classified lesions with different local relapse.

Regarding the clinical risk factors, the study found that sex, age, tumor location, tumor size, pathology, destruction, ALP, LDH, and metastasis at diagnosis did not significantly differ between the patients with local relapse (LRs) and those without local relapse (NLRs). These findings are consistent with previous research.[7] However, other studies have identified age (pediatric patients having an increased risk of relapse), tumor size (tumors larger than 8cm having a significantly higher risk of relapse), higher pre-treatment levels of LDH being associated with an increased risk of relapse, and the presence of metastasis at diagnosis being a strong indicator of a higher risk of relapse. [8–11] The discrepancies in the results may be attributed to factors such as inadequate sample size, geographical constraints, and the complex biological processes involved in osteosarcoma relapse, including the aberration of multiple genes and the influence of various risk factors. Radiomics, an emerging field in medical imaging, holds great potential in enhancing the diagnosis and treatment of various diseases, including osteosarcoma[12–15]. Its application in investigating the relapse of osteosarcoma can shed light on the intricate biological processes involved, helping to unravel the aberration of multiple genes and the influence of various risk factors. By utilizing radiomics, researchers can expand the sample size, overcome geographical constraints, and delve into the complexities of osteosarcoma relapse, providing valuable insights and paving the way for improved management strategies.

In this study, we used four machine-learning-based radiomics models based on radiograph and MRI to predict local relapse (LR) in extremity high-grade osteosarcoma. X-Ray imaging is routinely used for diagnosis and evaluation of osteosarcoma. In the past, most of the radiomics studies on bone tumors focused on CT and MRI, but in recent years, machine-learning-based radiomic analysis has emerged as a powerful tool for extracting quantitative features from X-Ray images and using these features for predicting clinical outcomes in bone tumors. Several studies have demonstrated the potential of radiograph radiomic analysis in predicting for the differentiation of benign and malignant bone tumors[16], efficacy of newadjuvant chemotherapy[17] and lung metastasis[18] in osteosarcoma patients. These studies have shown that radiomic features extracted from radiograph images can be used as predictors of local relapse. This study describes the use of preoperative radiograph radiomics to distinguish between LRs and NLRs. The AUC of the model for each classifier (XGB, LR, SVM and RF) in the test set were 0.611, 0.541, 0.723 and 0.627, respectively. The AUC obtained from radiomics based on radiograph were not very high. The explanation behind this is that radiograph imaging may have inherent limitations, such as low spatial resolution or limited ability to capture subtle variations in tissue characteristics. These limitations can affect the ability of radiomics models to accurately predict outcomes, resulting in a lower AUC. Furthermore, we discovered that among the features extracted from the radiograph images, only one firstorder feature (logarithm_firstorder_Median) was selected. This feature described the distribution of pixel or voxel intensities, providing additional quantitative information within the ROI. However, when compared to other studies, the feature selection from the radiograph we made may not fully capture the intratumoral heterogeneity of osteosarcoma.

In our study, the AUC using MRI-based radiomics of the model for each classifier (XGB, LR, SVM and RF) in the testing set were 0.609, 0.759, 0.727 and 0.705, respectively. The AUC obtained through the utilization of MRI-based radiomics, specifically employing the LR, SVM, and RF models, exhibited a slight superiority over the utilization of radiograph in the prediction of local relapse (LR) in osteosarcoma. Our analysis involved the examination of image data from a cohort of 92 osteosarcoma patients, and through the application of machine learning techniques, we were able to establish a strong correlation between pre-therapy MR radiomics features and tumor heterogeneity. Subsequently, we identified eleven statistically significant high-level features from the MRI images that displayed a substantial association with local relapse. It is worth noting that a previous study has already demonstrated that the radiomic features significantly linked to relapse in osteosarcoma are derived from CE-T1WI[19]. Therefore, our study, along with previous research findings, provides evidence to endorse the inclusion of CE-T1WI in radiomic investigations pertaining to osteosarcoma. Among the radiomic attributes derived from MRI, nine of them are associated with texture features acquired through wavelet transformation. The wavelet transformation technique assesses signal resolution across different temporal, spatial, and frequency scales, thereby providing insights into the organization of tissue microstructure. The extraction of wavelet features from DCE-MRI has demonstrated its efficacy in capturing the heterogeneity observed in breast cancer[20] and osteosarcoma[21].

While both X-Ray and MRI radiomic analyses have demonstrated potential in predicting local relapse in osteosarcoma, the integration of these two modalities has the potential to further enhance prediction accuracy. By amalgamating the complementary information derived from radiograph and MRI images, machine-learning algorithms can acquire more robust predictive models. Recent studies have documented the successful integration of radiograph and MRI radiomic features for preoperative prognostication of osteosarcoma outcomes, encompassing local relapse. The diagnostic performance of the combined features of radiograph and MRI in this investigation achieved a slightly higher AUCs compared to that of the individual sequence. The AUC using radiograph and MRI-based radiomics of the model for each classifier (XGB, LR, SVM and RF) in the testing set were 0.681, 0.759, 0.768 and 0.868, respectively. Previous and current research indicates that combining multiple models proves to be more beneficial, particularly in cases where the performance of a single model is lacking.[22, 23]

RF is effective at capturing non-linear relationships between radiomics features and the target variable. [24] It performs well in complex and non-linear scenarios and can handle high-dimensional data with a limited number of samples. RF also provides a measure of feature importance, allowing for the identification of significant radiomics features that contribute to the prediction and enhancing model interpretability. [25] These advantages make RF particularly suitable for small-sample radiomics studies with limited samples and complex relationships between features and outcomes. The RF had best predictive efficiency (AUCs: training set 0.999 vs. test set 0.868) of four machine learning models in our study. It was well consistent with the obtained results and previous literature.[26–28]

The study provides valuable insights into the potential of machine learning algorithms to predict local relapse in osteosarcoma patients. However, there are certain limitations in the study that need to be acknowledged. Firstly, the study is based on retrospective data analysis, which introduces inherent biases and limitations. The data used in the study might not represent the true population and could be subject to selection bias. The accuracy of the predictions generated by the machine learning model can only be as good as the quality and representativeness of the data used for training it. Secondly, the sample size in this study is relatively small. With only a limited number of patients included, the generalizability and reliability of the results may be compromised. A larger sample size would have provided a more robust and statistically meaningful analysis. Additionally, manual delineation of ROI can be time-consuming and challenging, especially with extensive datasets or intricate anatomical structures. This can slow down the research process and hinder efficiency. Moreover, manual delineation relies on the subjective judgment of the observer, leading to potential bias and limiting the reproducibility and generalizability of the findings. It is crucial to assess the accuracy and repeatability of VOI in future studies. For larger sample sizes, automated or semi-automated segmentation methods could be utilized to save time and streamline the process. Furthermore, the impact of external factors, such as patient demographics, lifestyle, and comorbidities, is not considered in this study. These factors could significantly influence the risk of local relapse in osteosarcoma patients. Incorporating such variables into the machine learning model could enhance its predictive accuracy and clinical utility. Overall, the study provides a promising foundation for the use of machine learning in predicting local relapse in osteosarcoma. However, these limitations should be considered when interpreting the findings, and further research with larger sample sizes and comprehensive predictive models is warranted.

Machine-learning-based multimodality radiomics analyses hold great potential for preoperative prediction of local relapse in osteosarcoma. The integration of X-Ray and MRI radiomic features can improve prediction accuracy, helping identify patients at high risk of local relapse and enabling tailored treatment approaches. Further research is needed to validate the findings and establish standardized protocols for radiomic analysis in clinical practice. Ultimately, machine-learning-based radiomics has the potential to improve patient outcomes and contribute to personalized medicine in osteosarcoma treatment.

ALP Alkaline phosphatase

AUC Area under the curve

CI Confidence intervals

ICC intraclass correlation coefficients

LASSO Least Absolute Shrinkage and Selection Operator

LRs local relapse

LDH lactate dehydrogenase.

LR logistic regression

NLRs non-local relapse

MRI magnetic resonance images

mRMR minimal-redundancy-maximum-relevance

RF random forest

RF-RFE random forest recursive feature elimination

ROC Receiver operating characteristic

ROI Regions of interest

SVM support vector machine

T1WI T1-weighted images

T2WI T2-weighted images

XGB extreme gradient boosting

Ethical Approval

This study was approved by the local institutional review board of The University of Hong Kong - Shenzhen Hospital, Shenzhen, China.

Consent to participate

Written informed consent was waived by the Institutional Review Board of The University of Hong Kong - Shenzhen Hospital, Shenzhen, China.

Consent for publication

Not applicable

Funding

This study has received funding by Shenzhen Science and Technology Program (No. JCYJ20230807113017035).

Availability of data and materials

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Conflict of interest

The authors declare no competing interests.

Ma O, Cai W, Zender L, Dayaram T, Shen J, Herron A, et al. MMP13, Birc2 (cIAP1), and Birc3 (cIAP2), amplified on chromosome 9, collaborate with p53 deficiency in mouse osteosarcoma progression. Cancer research. 2009; 69(6):2559-2567.
Klein M, Siegal G. Osteosarcoma: anatomic and histologic variants. American journal of clinical pathology. 2006; 125(4):555-581.
Lambin P, Leijenaar R, Deist T, Peerlings J, de Jong E, van Timmeren J, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nature reviews Clinical oncology. 2017; 14(12):749-762.
van Griethuysen J, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer research. 2017; 77(21):e104-e107.
Aerts H, Velazquez E, Leijenaar R, Parmar C, Grossmann P, Carvalho S, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature communications 2014; 5:4006.
Koo T, Li M. A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med. 2016; 15(2):155-163.
Bacci G, Longhi A, Versari M, Mercuri M, Briccoli A, Picci P. Prognostic factors for osteosarcoma of the extremity treated with neoadjuvant chemotherapy: 15-year experience in 789 patients treated at a single institution. Cancer. 2006; 106(5):1154-1161.
Mirabello L, Troisi R, Savage S. Osteosarcoma incidence and survival rates from 1973 to 2004: data from the Surveillance, Epidemiology, and End Results Program. Cancer. 2009; 115(7):1531-1543.
Bielack S, Kempf-Bielack B, Delling G, Exner G, Flege S, Helmke K, et al. Prognostic Factors in High-Grade Osteosarcoma of theExtremities or Trunk: An Analysis of 1,702 Patients Treatedon Neoadjuvant Cooperative Osteosarcoma Study GroupProtocols. Journal of clinical oncology 2023; 41(27):4323-4337.
Zhai C, Gu K, Zhai X, Wang J, Zhang J. Prognostic Value of Serum Lactate Dehydrogenase in Patients with Nasopharyngeal Carcinoma: a Meta-Analysis. Clinical laboratory. 2017; 63(11):1777-1785.
Marina N, Smeland S, Bielack S, Bernstein M, Jovic G, Krailo M, et al. Comparison of MAPIE versus MAP in patients with a poor response to preoperative chemotherapy for newly diagnosed high-grade osteosarcoma (EURAMOS-1): an open-label, international, randomised controlled trial. The Lancet Oncology. 2016; 17(10):1396-1408.
Xu L, Yang P, Hu K, Wu Y, Xu-Welliver M, Wan Y, et al. Prediction of neoadjuvant chemotherapy response in high-grade osteosarcoma: added value of non-tumorous bone radiomics using CT images. Quantitative imaging in medicine surgery 2021; 11(4):1184-1195.
Pereira H, Leite Duarte M, Ribeiro Damasceno I, de Oliveira Moura Santos L, Nogueira-Barbosa M. Machine learning-based CT radiomics features for the prediction of pulmonary metastasis in osteosarcoma. The British journal of radiology. 2021; 94(1124):20201391.
Bouhamama A, Leporq B, Khaled W, Nemeth A, Brahmi M, Dufau J, et al. Prediction of Histologic Neoadjuvant Chemotherapy Response in Osteosarcoma Using Pretherapeutic MRI Radiomics. Radiology Imaging cancer. 2022; 4(5):e210107.
White L, Atinga A, Naraghi A, Lajkosz K, Wunder J, Ferguson P, et al. T2-weighted MRI radiomics in high-grade intramedullary osteosarcoma: predictive accuracy in assessing histologic response to chemotherapy, overall survival, and disease-free survival. Skeletal radiology. 2023; 52(3):553-564.
von Schacky C, Wilhelm N, Schäfer V, Leonhardt Y, Jung M, Jungmann P, et al. Development and evaluation of machine learning models based on X-ray radiomics for the classification and differentiation of malignant and benign bone tumors. Eur Radiol. 2022.
Luo Z, Li J, Liao Y, Huang W, Li Y, Shen X. Prediction of response to preoperative neoadjuvant chemotherapy in extremity high-grade osteosarcoma using X-ray and multiparametric MRI radiomics. Journal of X-ray science technology,2023; 31(3):611-626.
Yin P, Zhong J, Liu Y, Liu T, Sun C, Liu X, et al. Clinical-radiomics models based on plain X-rays for prediction of lung metastasis in patients with osteosarcoma. BMC medical imaging. 2023; 23(1):40.
Chen H, Liu J, Cheng Z, Lu X, Wang X, Lu M, et al. Development and external validation of an MRI-based radiomics nomogram for pretreatment prediction for early relapse in osteosarcoma: A retrospective multicenter study. European journal of radiology. 2020; 129:109066.
Mahrooghy M, Ashraf A, Daye D, McDonald E, Rosen M, Mies C, et al. Pharmacokinetic Tumor Heterogeneity as a Prognostic Biomarker for Classifying Breast Cancer Recurrence Risk. IEEE transactions on bio-medical engineering. 2015; 62(6):1585-1594.
Zhang L, Ge Y, Gao Q, Zhao F, Cheng T, Li H, et al. Machine Learning-Based Radiomics Nomogram With Dynamic Contrast-Enhanced MRI of the Osteosarcoma for Evaluation of Efficacy of Neoadjuvant Chemotherapy. Frontiers in oncology. 2021; 11:758921.
Zhang Y, He K, Guo Y, Liu X, Yang Q, Zhang C, et al. A Novel Multimodal Radiomics Model for Preoperative Prediction of Lymphovascular Invasion in Rectal Cancer. Frontiers in oncology. 2020; 10:457.
Dong F, Li Q, Jiang B, Zhu X, Zeng Q, Huang P, et al. Differentiation of supratentorial single brain metastasis and glioblastoma by using peri-enhancing oedema region-derived radiomic features and multiple classifiers. European radiology. 2020; 30(5):3015-3022.
L B. Random forests. Machine Learning. 2001; 45(1):5-32.
Schaduangrat N, Malik A, Nantasenamat C. ERpred: a web server for the prediction of subtype-specific estrogen receptor antagonists. PeerJ. 2021; 9:e11716.
Ye J, Fang P, Peng Z, Huang X, Xie J, Yin X. A radiomics-based interpretable model to predict the pathological grade of pancreatic neuroendocrine tumors. European radiology. 2023.
Wu C, Du X, Zhang Y, Zhu L, Chen J, Chen Y, et al. Five machine learning-based radiomics models for preoperative prediction of histological grade in hepatocellular carcinoma. Journal of cancer research. 2023.
Wang W, Fan Z, Zhen JJTJoimr. MRI radiomics-based evaluation of tuberculous and brucella spondylitis. 2023; 51(8):3000605231195156.

Table 1 Clinical characteristics of 92 cases of osteosarcoma

Characteristic	All (n=92)	NLRs (n=59)	LRs (n-33)	P value
Gender:				1.000
Female	36 (39.1%)	23 (29.3%)	13 (39.4%)
Male	56 (60.9%)	36 (70.7%)	20 (60.6%)
Age (median [IQR])	15.0 [12.00, 20.25]	15.00 [13.00, 19.00]	14.00 [12.00, 23.00]	0.874
Pathology:				0.595
Osteoblastic	19 (20.7%)	11 (18.6%)	8 (24.2%)
Others	73 (79.3%)	48 (81.4%)	25 (75.8%)
Location:				0.387
Femur	55 (59.8%)	38 (64.4%)	17 (51.5%)
Tibia	21 (22.8%)	11 (18.6%)	10 (30.3%)
Others	16 (17.4%)	10 (16.9%)	6 (18.2%)
Destrution:				0.270
Mix	44 (47.8%)	25 (42.4%)	19 (57.5%)
Osteolytic	40 (43.5%)	29 (49.2%)	11 (33.3%)
Osteoblastic	8 (8.7%)	5 (8.5%)	3 (9.1%)
Size (median [IQR])	6.90 [5.77, 8.03]	6.90 [6.05, 8.05]	6.60 [5.70, 7.80]	0.748
Metastases:				1.000
No	66（71.7%）	23 (39.0%)	13 (39.4%)
Yes	26（28.3%）	36 (61.0%)	20 (60.6%)
ALP (median [IQR])	351.25 [185.50, 503.30]	361.20 [183.10, 643.25]	319.00 [215.40, 461.20]	0.594
LDH (median [IQR])	221.15 [180.85, 302.72]	222.10 [179.95, 291.45]	220.20 [188.10, 353.50]	0.317

LRs local relapse, NLRs non-local relapse, ALP alkaline phosphatase, LDH lactate dehydrogenase.

Table 2 Clinical characteristics of 102 cases of osteosarcoma in training and testing set

Characteristic	Training set (n = 61)			Testing set (n = 31)
Characteristic	NLRs (n = 39)	LRs (n = 22)	p value	NLRs (n = 20)	LRs (n = 11)	p value
Gender			1.000			0.458
Female	12 (30.8%)	9 (40.9%)		17 (85.0%)	6 (54.5%)
Male	27 (69.2%)	13(59.1%)		3 (15.0%)	5 (45.5%)
Age (median [IQR])	15.00 [13.00, 18.50]	15.00 [12.00, 22.00]	0.775	13.50 [12.00, 19.25]	12.00 [11.00, 26.00]	0.885
Pathology			0.342			0.631
Osteoblastic	7 (179%)	7 (31.8%)		4 (20.0%)	1 (9.1%)
Others	32 (82.1%)	15 (68.2%)		16 (80.0%)	10 (90.9%)
Location			0.285			0.882
Femur	26 (66.7%)	11 (50.0%)		12 (60.0%)	6 (54.5%)
Tibia	7 (17.9%)	8 (36.4%)		4 (20.0%)	2 (18.2%)
Others	6 (15.4%)	3 (13.6%)		4 (20.0%)	3 (27.3%)
Destruction			0.611			0.414
Mix	15 (38.5%)	11 (50.0%)		10 (50.0%)	8 (72.7%)
Osteolytic	21 (53.8%)	9 (40.9%)		8 (40.0%)	2 (18.2%)
Osteoblastic	3 (7.70%)	2 (9.1%)		2 (10.0%)	1 (9.1%)
Size (median [IQR])	7.20 [6.40, 8.60]	6.75 [5.85, 8.02]	0.471	6.40 [5.25, 7.03]	6.30 [5.25, 7.60]	0.695
Metastases:			0.778			0.095
No	28(71.8%)	15(68.2%)		17 (85.0%)	6 (54.5%)
Yes	11(28.2%)	7(31.8%)		3 (15.0%)	5 (45.5%)
ALP (median [IQR])	352.90 [189.75, 537.30]	279.40 [184.93, 423.75]	0.224	382.25 [142.98, 805.42]	411.00 [303.90, 524.90]	0.536
LDH (median [IQR])	222.10 [179.95, 267.50]	218.85 [178.75, 342.42]	0.471	231.60 [182.80, 370.15]	267.90 [199.45, 413.70]	0.483

LRs local relapse, NLRs non-local relapse, ALP alkaline phosphatase, LDH lactate dehydrogenase.

Table 3 The measure of feature importance of RF

Model	No.	Features	Importance
X-ray	1	X-ray_logarithm_firstorder_Median	0.115654537
MRI-T1WI	4	T1WI_exponential_glrlm_ShortRunEmphasis	0.118829617
		T1WI_wavelet-HHL_glcm_ClusterShade	0.090243216
		T1WI_lbp-3D-k_gldm_DependenceVariance	0.060639051
		T1WI_wavelet-HLL_firstorder_TotalEnergy	0.136371946
MR-T2WI	2	T2WI_wavelet-LLL_firstorder_InterquartileRange	0.048598005
		T2WI_wavelet-LLH_firstorder_Skewness	0.096621067
MR-T1WI+CE	5	T1WI+CE_wavelet-LLH_gldm_LargeDependenceEmphasis	0.068312636
		T1WI+CE_wavelet-LHH_firstorder_Skewness	0.07982036
		T1WI+CE_wavelet-LLH_glrlm_RunEntropy	0.091828419
		T1WI+CE_wavelet-LLH_glszm_GrayLevelVariance	0.044466254
		T1WI+CE_wavelet-LHL_glcm_ClusterShade	0.048614892

Table 4 Predictive performances of four machine learning methods in the training and testing sets

	Machine learning	Model	AUC (95%CI)	Sensitivity	Specifcity	Accuracy
Training set	RF	X-ray	0.895(0.816-0.973)	0.591	0.949	0.820
		MRI	0.999(0.996-1.000)	0.864	1.000	0.951
		X-ray+MRI	0.999(0.996-1.000)	0.955	1.000	0.984
	XGB	X-ray	0.866(0.769-0.961)	0.591	0.974	0.836
		MRI	1.000(1.000-1.000)	1.000	1.000	1.000
		X-ray+MRI	1.000(1.000-1.000)	1.000	1.000	1.000
	LR	X-ray	0.621(0.476-0.767)	0.227	0.974	0.705
		MRI	0.927(0.854-0.999)	0.840	0.846	0.852
		X-ray+MRI	0.941(0.872-1.000)	0.864	0.897	0.885
	SVM	X-ray	0.379(0.233-0.524)	0.000	1.000	0.639
		MRI	0.945(0.883-1.000)	0.864	0.974	0.934
		X-ray+MRI	0.958(0.904-1.000)	0.818	0.974	0.918
Testing set	RF	X-ray	0.627(0.427-0.828)	0.636	0.700	0.677
		MRI	0.705(0.490-0.919)	0.545	0.900	0.774
		X-ray+MRI	0.868(0.742-0.995)	0.909	0.750	0.806
	XGB	X-ray	0.611(0.401-0.821)	0.636	0.600	0.613
		MRI	0.609(0.396-0.822)	1.000	0.300	0.548
		X-ray+MRI	0.682(0.482-0.881)	0.818	0.500	0.613
	LR	X-ray	0.540(0.313-0.769)	0.727	0.550	0.613
		MRI	0.759(0.583-0.934)	0.727	0.700	0.710
		X-ray+MRI	0.759(0.585-0.935)	0.818	0.650	0.710
	SVM	X-ray	0.459(0.231-0.687)	0.182	1.000	0.710
		MRI	0.759(0.571-0.947)	0.818	0.750	0.774
		X-ray+MRI	0.768(0.579-0.957)	0.727	0.800	0.774

AUC, area under the summary receiver operating characteristic curve; CI, confidence interval. XGB, extreme gradient boosting, LR, logistic regression, SVM, support vector machine, RF, random forest

No competing interests reported.

Machine-learning-based on multimodality radiomics analysis for the Preoperative Prediction for local relapse in osteosarcoma

Status:

Version 1

Abstract

PURPOSE:

MATERIALS AND METHODS:

RESULTS:

CONCLUSION:

Figures

1. Introduction

2. Materials and methods

3. Results

4. Discussion

5. Conclusions

Abbreviations

Declarations

References

Tables

Additional Declarations

Status:

Version 1