Multi-parametric MRI-based Peritumoral Radiomics for Stage IIA and IIB Classification of Cervical Cancer：A Multicenter Study

Multi-parametric MRI-based Peritumoral Radiomics for Stage IIA and IIB Classification of Cervical Cancer：A Multicenter Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Multi-parametric MRI-based Peritumoral Radiomics for Stage IIA and IIB Classification of Cervical Cancer：A Multicenter Study Ying Wang, Weixiao Liu, Yulian Chen, Fei Wang, Xiaoyun Liang, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4772065/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Objective: The aim of the study is to establish a multiparametric MRI (mpMRI)-based peritumoral radiomics nomogram for preoperatively predicting IIA and IIB classification of cervical Cancer preoperatively. Methods: 208 patients with histologically confirmed cervical cancer from three institutions were enrolled in this study. All the cases were randomly divided into the training cohort (n=145) and the validation cohort (n=63). The performance of the nomogram was assessed with respect to its calibration, discrimination, and clinical usefulness. The independent-sample t test and the Chi-squared test were conducted to assess the significance of clinical factors between the training cohort and the validation cohort. The Pearson correlation coefficient analysis and recursive feature elimination algorithm were adopted successively to obtain the well-representative features. Different classifiers were compared to develop the optimal radiomics signature across 5-fold cross validation. The calibration curves and decision curve analysis were conducted to evaluate the clinical utility of the optimal model. The radiomics model was constructed using logistic regression. Results: The peritumoral radiomics models were superior to the intratumoral radiomics models, regardless of single sequence model or fusion model (all P <0.001*). DWI-based peritumoral radiomics model performed best with the AUCs of 0.975 (0.965−0.983) and 0.899 (0.880−0.916) in the training and validation cohort, respectively. There was no significant difference between the validation AUCs of DWI-based and fusion peritumoral radiomics model (0.899 vs. 0.895, P =0.566). In addition, 3 pixel peritumoral regions of radiomic signatures have a much better discrimination performance in distinguishing IIA and IIB stage by comparing the 2,4,5 pixels extension surrounding the tumor. Conclusion: MRI-based radiomics model from peritumoral regions of cervical cancer outperformed radiologists for the preoperative diagnosis of IIA and IIB stage, which could provide a noninvasive and reliable way of individualized treatment plans for patients with cervical cancer. Cervical cancer Peritumoral radiomics Magnetic resonance imaging Stage classification Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Cervical cancer, despite being highly preventable, remains the fourth most common cancer and a leading cause of cancer-related deaths among women worldwide [ 1 ] . The Federation of Gynecology and Obstetrics (FIGO) staging system is used to classify the extent of the disease. In the case of early-stage cervical cancer (FIGO IA to IIA), the recommended first-line treatment is radical hysterectomy with pelvic lymphadenectomy. For locally advanced cervical cancer (FIGO ≥ IIB), the standard treatment is adjuvant concurrent chemoradiotherapy (CCRT) [ 2 ] . It is important to accurately determine the FIGO stage because misclassification can impact treatment recommendations. Furthermore, there is an increasing demand for more conservative surgical approaches in young patients with cervical carcinoma. Precise identification of FIGO staging, particularly distinguishing between IIA and IIB stages before surgery in patients with cervical cancer, which can help avoid unnecessary surgical interventions and optimize treatment planning. This approach can be beneficial in preserving fertility and maintaining quality of life for these patients. Accurate staging of cervical cancer is crucial for treatment planning, however, current methods have their limitations. Local staging is usually done through a clinical examination under anesthesia, which includes invasive endoscopic examination. Despite this, the subjective nature of this method can pose challenges, especially when assessing endocervical tumor growth and the extent of parametrial tissue involvement [ 3 ] . Preoperative evaluation of the stage of cervical cancer frequently includes magnetic resonance imaging (MRI) as an adjunct imaging modality [ 4 ] . Research has shown notable differences between clinical examination and MRI-based staging, especially in evaluating tumor size larger than 2 cm, parametrial spread, extension to the pelvic wall, vaginal involvement, and adjacent organ involvement. MRI's specificity is limited and it has a high false-positive rate in stage classification for cervical cancer [ 5 ][ 6 ] . In accordance with the 2018 FIGO staging system, pathological examination via biopsy is integrated into disease staging [ 4 ] . However, this technique is invasive and susceptible to sampling errors, which may obscure the tumor's overall heterogeneity. In light of these limitations, there is an imperative need for noninvasive diagnostic methods [ 7 ] . Radiomics has emerged as a promising tool in the field of medical imaging, enabling the extraction of quantitative features from images and the development of models to assist in the management of various aspects of cervical cancer [ 8 ][ 9 ] . Radiomics has the potential to predict lymphovascular space invasion (LVSI), lymph node metastasis (LNM), and disease-free survival (DFS) according to previous research [ 10 – 13 ] . Moreover, some previous studies have suggested the peritumoral radiomics also provide important predictive information on LVSI, LNM and DFS prediction in cervical cancer [ 12 , 14 – 15 ] . Currently, there is a lack of radiomics analysis based on magnetic resonance imaging (MRI) for accurately distinguishing between stage IIA and IIB cervical cancer. Our previous study indicated that the features from 5-pixel peritumoral regions improved the area under receiver operating characteristic area under curve (AUC) in predicting LNM in cervical cancer [ 15 ] .Therefore, the incorporation of peritumoral radiomics into these analyses has the potential to provide valuable predictive information and enhance the staging and treatment planning of cervical cancer patients. The aim of this retrospective study was to assess the effectiveness of multi-parametric MRI (mpMRI)-based tumoral and peritumoral radiomics in differentiating between stage IIA and stage IIB cervical cancer, and further promote its potential clinical applications. Methods Patients Institutional ethics review board approval was acquired for this study, and written informed consent was not required for this retrospective study. The inclusion criteria for patients were as follows: (1) pathologically confirmed stage IIA and IIB cervical cancer, (2) patients underwent pelvic MRI before the treatment, and (3) clinical data were available, such as age, FIGO stage, histologic characteristics, differentiation. Patients were excluded if their tumor lesions were not visible or if the MRI images had poor quality. Patients who had received pretreatment therapy or conization prior to MRI were also excluded. In total, 208 patients were enrolled from Hospital 1 (n=36, from January 2015 to December 2019), Hospital 2 (n = 105, from January 2011 to December 2017), and Hospital 3 (n = 67, from December 2012 to June 2018). Patients were randomly divided into the training and validation cohort by an approximative ratio of 2:1, and then, 145 and 63 patients were assigned to each cohort, respectively. The study flowchart is illustrated in Figure 1. Clinical Data Collection and Analysis Age, human papillomavirus (HPV), squamous cell carcinoma (SCC), carcinoembryonic antigen (CEA), neutrophil, lymphocyte, and pathological staging, were collected from the electronic medical records for patients pathologically confirmed cervical cancer. Due to the difference or lack of quantitative information collected by different hospitals, only three baseline factors, including age, SCC and CEA, were analyzed in this study. Univariate and multivariate logistic regression (LR) analyses were successively applied to these factors for the clinical model development. Image Acquisition and Preprocessing All patients underwent pelvic MRI scans including sagittal contrast-enhanced T1-weighted imaging(CE-T1WI), axial T2-weighted imaging (T2WI), and axial diffusion-weighted imaging (DWI). The obtained MRI images were stored in the picture archiving and communication system in a DICOM format. Details regarding the acquisition parameters and MRI retrieval procedure were described in Supplementary Table S1. A general preprocessing pipeline was applied on the images to reduce the intensity and geometric variations caused by different MRI scanners, acquisition parameters, and patients. N4 algorithm embedded in 3D slicer software (version 5.0.3, www.slicer.org) was used to correct the distortion from non-uniform bias magnetic field in the MRI scanners. After standardizing the grayscale value of the images, all image voxels were uniformly resampled to 1×1×1 mm 3 with cubic B-spline interpolation algorithm. Image Assessment Two radiologists (JF and YL.C, with 10 and 5 years of experience in gynecologic MR imaging, respectively) independently determined the clinical stage based on the mpMRI in the absence of pathological and surgical results. According to 2018 FIGO stage, if the tumor extends beyond the cervix with the cervical stromal ring disrupted, on T2WI and DWI images, which show hyperintense signal extending to the parametrium were considered as IIB stage [16] . One radiologist (YL.C) repeated this procedure a month later. We then assessed the intra- and inter-rater consistency of clinical stage as observed directly by radiologists from MR images, using intra-class correlation coefficient (ICC) and Kendall’s rank correlation coeffcient (KCC). The MR image assessment results used for subsequent analysis were reviewed by a senior radiologist (FW, with 13 years of experience) and corrected to get the final version. Volume of Interest Segmentation and Reproducibility Evaluation Two types of volume of interest (VOI), the tumor core and the peritumoral area (abbreviated as intratumoral VOI and peritumoral VOI, respectively, seen in Figure 1 ), were used for the radiomics analysis. Similarly, the two radiologists (YW and JF), blinded to the clinical information, independently manually delineated the tumor core by using the ITK-SNAP software (version 3.6.0, www.itksnap.org). The intratumoral VOIs of 50 patients randomly selected from the whole cohort were re-segmented over a 2-week interval by another radiologist (JF) with the same tool and environment settings. Correspondingly, the intra- and inter-rater reproducibility were measured by dice similarity coefficient (DSC) and ICC (based on extracted radiomics features). The VOIs with large differences were also corrected by the senior radiologist (FW). To obtained the peritumoral information, the intratumoral VOIs were dilated by a circular structural element with a radius size (number of a pixel) of 2, 3, 4, and 5. Peritumoral VOIs were defined by subtracting the intratumoral VOIs from the dilated area. Radiomics Feature Extraction Compliant with the Image Biomarker Standardization Initiative guidelines (https://theibsi.github.io/), a total of 1130 MRI-based radiomics features were extracted from each VOI by using the Pyradiomics package (version 3.0.1, https://pyradiomics.readthedocs.io/). The features characterizing shape, intensity, and texture patterns of the VOIs were obtained from the original and derived MR images using the Laplacian of gaussian and wavelet filter. The extraction parameters and feature types are summarized in Supplementary Table S2. Z-score normalization was applied to the radiomics features for parameter standardization before radiomics analysis. Radiomics Feature Selection and Signature Construction Pearson correlation coefficient (PCC) analysis and recursive feature elimination (RFE) algorithm were adopted successively to obtain the well-representative features in the feature selection procedure. If the PCC of feature pairs was greater than 0.99, one of the features was randomly eliminated to obtain a feature subset with low redundancy. The RFE algorithm was further used for feature screening and combined with different classifiers to obtain the optimal radiomics signature. Random forest (RF), support vector machine (SVM), logistic regression (LR), naive Bayes (NB), Adaboost, and multilayer perceptron (MLP) classifiers were compared for constructing the radiomics signature. Hyperparameter optimization was performed using 5-fold cross-validation. It is important to note that radiomics feature selection and signature construction were based solely on data from the training cohort, while the validation cohort was used for evaluating and comparing the performance of the constructed signatures. Statistical Analysis All the statistical analyses in this study were performed with SPSS version 22.0 and Python version 3.7.6. The independent-sample t test and the Chi-squared test were conducted to assess the significance of clinical factors between the training cohort and the validation cohort. Receiver operating characteristic curve (ROC), AUC, sensitivity, specificity, and accuracy were used to evaluate the prediction performance of different signatures. The difference between the AUCs were statistically analyzed using the method of DeLong et al [17] with MedCalc version 20.026. Furthermore, the calibration curve and decision curve analysis were conducted to evaluate the clinical utility of the optimal signature. Note that a two-tailed P -value less than 0.05 was indicated statistically significant in this work. Results Patient Characteristics Among 208 patients with cervical cancer, 156 patients were in stage IIA and 52 patients were in stage IIB. The clinical characteristics of the training cohort (n=145, average age: 51.63±9.90) and the validation cohort (n=63, average age: 52.48±9.52) are presented in Table 1 . It details the clinical characteristics such as age, squamous cell antigen, CEA, histology, degree of cellular differentiation. A statistical difference was observed between the age in the training set ( P = 0.044) and the CEA in the validation set ( P = 0.007) for stages IIA and IIB. Table 1 Characteristics of cervical cancer patients in training and validation cohorts Characteristics Training cohort (n=145) Validation cohort (n=63) IIA (n=109) IIB (n=36) P-value IIA （n=47） IIB （n=16） P-value Age (years)mean ± SD 52.58± 9.64 48.75±10.27 0.044 52.82± 9.84 48.81 ±9.14 0.176 CEA (ng/mL） 6.25±13.57 60.07±309.1 0.304 4.68±4.56 21.64±21.78 0.007 SCC-Ag (ng/mL) 5.65±8.62 6.35±8.16 0.673 5.13±7.36 8.57±10.67 0.190 Histology（%） Squamous 98（89.91%） 33（91.67%） 0.655 44（93.62%） 15（93.75%） 1.000 Adenocarcinoma 11（10.09%） 2（5.56%） 2（4.26%） 1（6.25%） Degree of cellular differentiation, n (%) Low 26（23.85%） 13（36.11%） 0.355 15（31.91%） 3（18.75%） 0.253 Middle 72（66.06%） 20（55.56%） 29 （61.70%） 10（62.50%） High 11 （10.09%） 3 （8.33%） 3（6.38%） 3（18.75%） Abbreviations: CEA , carcinoembryonic antigen; SCC , squamous cell carcinoma; P* < 0.05 indicates a significant difference. Segmentation Reproducibility Evaluation For manual delineation of VOIs by two radiologists, the mean and median values of intra- and inter-rater DSC exceeded 0.90, indicating that the difference in delineation was relatively small in this work ( Figure 2B ). To quantitatively compare the variability of the radiomic-features family with respect to the segmentation, we analyzed the variation in the number of features under different ICC thresholds ( Figure 2A ). The results showed that the radiomics features also presented good reproducibility under the condition of small segmentation difference, which was applicable to both intra- and inter-rater ICC. Furthermore, 91.6% (1035/1130), 91.5% (1034/1130), 91.9% (1039/1130) of the features were identified as highly reproducibility with ICC≥0.75 of both intra- and inter-rater variability. Predictive Performance of Clinical model Age (odds ratio[OR], 0.962; 95% confidence interval[CI], 0.925−0.999; P = 0.047*), SCC (OR, 1.009; 95% CI, 0.967−1.053; P = 0.671), and CEA (OR, 1.010; 95% CI, 0.986−1.034; P = 0.407) were assessed by univariate LR analysis. Age and CEA was found significantly associated with the stage and then used to construct a clinical model. The model yielded AUC values of 0.611 (95% CI, 0.573−0.649) in the training cohort and 0.652 (95% CI, 0.616−0.687) in the validation cohort, respectively ( Table 2 ). Predictive Performance of Image Assessment by Radiologists The low intra- and inter-rater agreement of staging obtained by visual observation of MR images were presented between the two radiologists. The intra- and inter-rater ICC were 0.555 and 0.553, whereas intra- and inter-rater KCC were 0.556 and 0.560, respectively ( Figure 2C ). As seen in Table 2 , the AUCs of the predicted clinical stage after correction by the senior radiologist were 0.710 (95% CI: (0.680−0.739) and 0.757 (95% CI: 0.730−0.785) in the training and validation cohort, respectively. Table 2 Prediction performance of different models for the classification of stage IIA and IIB. Training cohort Validation cohort AUC (95% CI) P-value Sensitivity Specificity Accuracy AUC (95% CI) P-value Sensitivity Specificity Accuracy Clinical model 0.611 (0.573−0.649) <0.001* 0.417 0.761 0.676 0.652 (0.616−0.687) <0.001* 0.438 0.766 0.683 Image assessment 0.710 (0.680−0.739) Ref. 0.750 0.670 0.690 0.757 (0.730−0.785) Ref. 0.813 0.702 0.730 Intratumoral radiomics signature DWI 0.783 (0.751−0.815) 0.002* 0.583 0.835 0.772 0.707 (0.670−0.745) 0.028* 0.438 0.851 0.746 CE-T1WI 0.818 (0.788−0.848) <0.001* 0.750 0.716 0.724 0.771 (0.739−0.803) 0.424 0.625 0.702 0.683 T2WI 0.676 (0.640−0.712) 0.085 0.639 0.688 0.676 0.680 (0.643−0.716) <0.001* 0.625 0.596 0.603 Fusion 0.872 (0.846−0.899) <0.001* 0.861 0.780 0.800 0.815 (0.784−0.847) <0.001* 0.625 0.787 0.746 Peritumoral radiomics signature DWI 0.975 (0.966−0.984) <0.001* 0.917 0.972 0.959 0.899 (0.878−0.920) <0.001* 0.625 0.979 0.889 CE-T1WI 0.949 (0.935−0.962) <0.001* 0.833 0.908 0.890 0.895 (0.877−0.913) <0.001* 0.750 0.872 0.841 T2WI 0.933 (0.916−0.950) <0.001* 0.861 0.890 0.883 0.777 (0.748−0.806) 0.360 0.500 0.702 0.651 Fusion 0.981 (0.973−0.988) <0.001* 0.917 0.972 0.959 0.895 (0.874−0.915) <0.001* 0.625 0.979 0.889 Abbreviations: AUC , area under the curve; CI , confidence interval; DWI , diffusion weighted imaging; CE-T1WI , contrast-enhanced T1-weighted imaging; T2WI , T2-weighted imaging; P* < 0.05 indicates a significant difference. Predictive Perfor mance of Intratumoral Radiomics Signature After feature selection, 8, 16, and 3 predictive features were separately identified to constructed the optimal intratumoral radiomics signatures for DWI, CE-T1WI, and T2WI. Details of selected features are shown in Supplementary Table S3 . The optimal classifiers were LR, MLP, and LR, respectively. We found that the predictive performance of CE-T1WI-based radiomics signature is better than that of DWI and T2WI in both the training and validation cohorts ( Table 2 ; all P < 0.05). The fusion model derived from joint DWI, CE-T1WI, and T2WI signatures was considered to be the optimal intratumoral radiomics signature, superior to the signatures from single sequences in differentiating stages (all P < 0.01). The AUCs were 0.872 (95% CI: 0.846−0.899) and 0.815 (95% CI: 0.784−0.847) in the training and validation cohorts, respectively. Pred ictive Performance of Peritumoral Radiomics Signature The peritumoral VOIs were rebuilt using a structural element with a radius of 3, which showed noticeably better discriminatory performance than that from the surrounding 2, 4, and 5-pixel extensions ( Supplementary Figure S1 ). 14, 14 and 7 radiomics features, respectively, in combination with SVM, MLP and LR classifier, were used to develop the optimal DWI-, CE-T1WI-, and T2WI-based radiomics signatures ( Supplementary Table S4 ). As reflected in Table 2 , DWI-based signature performed best among all the sequences with the AUCs of 0.975 (95% CI: 0.966−0.984) and 0.899 (95% CI: 0.878−0.920) in the training and validation cohort, respectively. There was no significant difference between the validation AUCs of DWI-based and fusion peritumoral radiomics model (0.899 vs. 0.895, P = 0.566), implying important predictive information contained in DWI. Thus, DWI-based signature was identified as the optimal peritumoral radiomics signature. Comparison of the Various Predictive Models The peritumoral radiomics signatures were superior to the intratumoral radiomics signatures, regardless of single sequence or fusion analysis (all P < 0.001). DWI-based peritumoral radiomics signature performed best for both cohorts, significantly superior to the optimal intratumoral radiomics signature, clinical model, and image assessment by radiologists (all P < 0.001). Figure 3A-C show ROC curves of the models corresponding to three hospitals, confirming that the DWI-based peritumoral radiomics signature could offer the optimal performance for distinguishing IIA and IIB stage cervical cancer in data from different centers. In particular, the optimal model correctly predicted the staging of all cases at Hospital 1 and Hospital 2. Clinical Use of the Optimal Peritumoral Radiomics Signature The calibration curves of the optimal peritumoral radiomics signature ( Figure 3D and E ) demonstrated good agreement between prediction and actual observation in both cohorts (mean absolute error = 0.023 and 0.047, Hosmer-Lemeshow test P = 0.824 and 0.135, respectively). The DCA plotted in Figure 3F indicated the incremental clinical advantage of information from peritumoral region over information from other sources. Patients with cervical cancer could benefit more within the whole risk threshold range. Discussion Distinguishing IIA and IIB stageis essential because the therapeutic regimens and follow-up care are vastly different for each of these conditions. According to 2018 FIGO stage system for cervical cancer, IIA stage limited to the upper two-thirds of the vagina without parametrial involvement; while IIB with parametrial involvement but not up to the pelvic wall. So the parametrial involvement (PMI) is the main differentiation between IIA and IIB stages. MR images are used as a supplementary diagnostic tool in distinguishing between IIA and IIB based on tumor size, disruption of the hypointense cervical stromal ring on T2WI and DWI. It is important to note that MRI may occasionally overestimate parametrial invasion in large tumors (with an accuracy of 70%) due to peritumoral edema or inflammation caused by tumor compression. This factor should be considered when planning treatment [10] . The diagnostic value of MRI assessment by radiologists in our study is 0.757, which is similar to that of previous reports [11] [18] . As shown in Figure 4, patient with FIGO stage IIA cervical cancer has preserved stromal ring on MR image. In addition, patient with FIGO stage IIB has involved stromal disruption with fat or soft-tissue strands in the parametrium. In the corresponding histopathological examination (HE) images ( C, D, G, and F in Figure 4 ), the stage IIA cervical cancer tumor did not invade the paracervical tissues, whereas stage IIB tumor invaded. Zhao et al. conducted a study on the development of radiomics models employing the T2WI sequence, which demonstrated outstanding performance in differentiating between early stages of cervical cancer (stage IB vs. stage IIA). Using a SVM classifier, the model obtained impressive metrics, including AUC of 0.907%, sensitivity of 0.778, specificity of 0.999, and accuracy of 0.867 [18] . It is important to note that this particular study only considered tumoral regions and did not include data from the peritumoral regions. In our study, we explored the intratumoral and peritumoral regions of cervical cancer in differentiating IIA and IIB stage, and the results revealed that the most predictable features could be obtained from the peritumoral regions with 3 pixel dilation distances in the DWI images. Our results showed that DWI images could provide good functional information in distinguish IIA and IIB stage. DWI is a functional imaging technique in association with water diffusion properties, the tumor shows a higher signal intensity and lower apparent diffusion coefficient value than normal cervical tissue. Previous studies showed that DWI can significantly improve diagnostic efficacy in stage, risk factors prediction [19] [20] , which can detect early pathological changes associated with changes in water content in tissues, not effected by edema and inflammation around the tumor. The connective tissue of the parametrium is mainly composed of fat, blood vessels, nerves, fibrous tissues and lymphatic vessels. Due to the lack of fascial restriction, direct infiltration and invasion of the tumor is more likely to occur in the tumor parametrium. Therefore, the immediate surrounding tumor microenvironment may offer unique radiomics signatures of stage classification. Considering the peritumoral microenvironment of lymphatics and blood vessels invasion varies with tumor size, unlike previously reported study [13] [21] , we selected pixel instead of fixed diameter, because the pixel is more objective and reliable, which vary with the size of the tumor itself. In this study, we compare the 2,3,4,5 pixels extension surrounding the tumor, the results demonstrated that 3 pixel peritumoral regions of radiomics signatures show the best discrimination performance in distinguishing IIA and IIB stage, which can be partly explained by that the fact that peritumoral region provides more information about parametrial infiltration [22] . The choice of robust and reproducible radiomics features should be emphasized for promoting subsequent applicability in clinical. Manual segmentation suffers from high intra- or inter-reader variability, and the value of the extracted radiomics features may differ due to variances in tumor segmentations uncertainty. Thus, multiple-segmentation by multiple radiologists is a method to provided reproducible and reliable radiomic features. It is important to identify whether the features extracted from the two types of semiautomatic segmentations capture the same tumor image properties as manual delineation. We applied ICC analysis to preliminarily identify stable features that were less sensitive to segmentation differences to improve the reproducibility of feature extraction. Similar to the previous study [23] , we used PCC analysis to remove some highly correlated features. Thus, a feature subset with lower dimensions was obtained, and then the final feature filtering was performed. In the model development procedure, we compared the performance of multiple classifiers and selected the best one based on the corresponding MRI sequence. This makes sense in a lot of research because it is difficult for one classifier to perform well in all analysis tasks [24-27] . The results of the optimal peritumoral radiomics signatures highlights the potential of texture feature which were difficult to decipher through simple observation but provided considerably richer information for distinguishing IIA and IIB stage cervical cancer. Limitation Our study has several limitations. First, our study was performed with a relatively small population. Further work is required to focus on validating the proposed model on a larger multicenter dataset, and external validation in different centers is still needed. Secondly, our study manually extracted three-dimensional texture features, which is subjective and time consuming. The use of automatic segmentation technology may offer assistance in this regard. Additionally, important clinical characteristics (e.g. HPV) that were correlated with stage in cervical cancer were not involved in this study. The integration of additional multi-source characteristics into the existing model to further develop a clinical-AI framework will require exploration in order to facilitate its clinical application. Conclusion MRI-based radiomics model from peritumoral regions outperformed radiologists for the preoperative diagnosis of IIA and IIB stage. A noninvasive and reliable supplementary approach was provided to precise preoperative staging, which can help optimize individualized treatment plans for patients with cervical cancer. Declarations Conflict of Interest The authors declare no potential conflicts of interest. Acknowledgments This study was supported by Guangdong Basic and Applied Basic Research Foundation (grant number: 2021A1515110763); Medical science and Technology Research Fund of Guangdong Province (grant number: A2022267); The Science and Technology Program of Guangzhou (grant number: 202201020057,2023A03J1038); The Clinical Frontier Technology Program of the First Affiliated Hospital of Jinan University, China (No. JNU1AF-CFTP-2022-a01201). Authors contributions Concept and design: J.F, X.Z; Acquisition, analysis, or interpretation of data: J.F., Y.W., YL.C.; Drafting of the manuscript: Y.W., J.F., WX. L; Statistical analysis: F.W.,YL.C.; Performing experiments: YW.,WX.L; Supervision and revision of the manuscript: J.F. X.Z. All authors read and approved the final manuscript. Ethics declaration: not applicable References Castle P E, Einstein M H, Sahasrabuddhe V V. Cervical cancer prevention and control in women living with human immunodeficiency virus. Ca 2021;71(6): 505–526. Cohen PA, Jhingran A, Oaknin A, et al. Cervical cancer. Lancet 2019;393(10167):169-182. Bourgioti C, Chatoupis K, Rodolakis A, et al. Incremental Prognostic Value of MRI in the Staging of Early Cervical Cancer: A Prospective Study and Review of the Literature. Clinical Imaging 2016;40(1): 72–78. Manganaro L, Lakhman Y, Bharwani N, et al. Correction to: Staging, recurrence and follow-up of uterine cervical cancer using MRI: Updated Guidelines of the European Society of Urogenital Radiology after revised FIGO staging 2018. Eur Radiol 2022;32(1):738. Dappa E, Elger T, Hasenburg A, et al. The value of advanced MRI techniques in the assessment of cervical cancer: a review. Insights into Imaging 2017;8(5): 471–481. Knoth J, Pötter R, Jürgenliemk-Schulz I M, et al. Clinical and Imaging Findings in Cervical Cancer and Their Impact on FIGO and TNM Staging – An Analysis from the EMBRACE Study. Gynecologic Oncology 2020;159(1): 136–141. Zhou Y, Gu H-L, Zhang X-L, et al. Multiparametric Magnetic Resonance Imaging-Derived Radiomics for the Prediction of Disease-Free Survival in Early-Stage Squamous Cervical Cancer. Eur Radiol 2022;32(4): 2540–2551. Gillies R J, Kinahan P E, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278(2): 563–577. Lambin P, Leijenaar R T H, Deist T M, et al. Radiomics: The Bridge between Medical Imaging and Personalized Medicine. Nature Reviews. Clinical Oncology 2017;14(12): 749–762.. Balcacer P, Shergill A, Litkouhi B. MRI of Cervical Cancer with a Surgical Perspective: Staging, Prognostic Implications and Pitfalls. Abdominal Radiology 2019;44(7): 2557–2571. Xiao M, Ma F, Li Y et al. Multiparametric MRI-based radiomics nomogram for predicting lymph node metastasis in early-stage cervical cancer. J Magn Reson Imaging 2020;52:885–896 . Li Z, Li H, Wang S et al. MR-based radiomics nomogram of cervical cancer in prediction of the lymph-vascular space invasion preoperatively. J Magn Reson Imaging 2019; 49:1420–1426. Wang T, Gao T, Guo H et al. Preoperative prediction of parametrial invasion in early-stage cervical cancer with MRI based radiomics nomogram. Eur Radiol 2020;30:3585–3593. Zhou Y, Gu HL, Zhang XL, Tian ZF, Xu XQ, Tang WW. Multiparametric magnetic resonance imaging‑derived radiomics for the prediction of disease‑free survival in early‑stage squamous cervical cancer. Eur Radiol 2022;32(4):2540-2551. Wu QX, Wang S, Zhang SX, et al. Development of a Deep Learning Model to Identify Lymph Node Metastasis on Magnetic Resonance Imaging in Patients With Cervical Cancer. JAMA Netw Open 2020;3(7): e2011625. Balcacer P, Shergill A, Litkouhi B. MRI of Cervical Cancer with a Surgical Perspective: Staging, Prognostic Implications and Pitfalls. Abdominal Radiology 2019;44(7): 2557–2571. Delong E R , Delong D M, Clarke-Pearson D L. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach.[J].Biometrics, 1988, 44(3):837. Zhao X, Wang X, Zhang B, et al. Classifying Early Stages of Cervical Cancer with MRI-Based Radiomics. Magnetic Resonance Imaging 2022;89: 70–76. Park J J, Kim C K, Park S Y, et al. Value of Diffusion-Weighted Imaging in Predicting Parametrial Invasion in Stage IA2-IIA Cervical Cancer. European Radiology 2014;24(5): 1081–1088. Woo S, Suh C H, Kim S Y, et al. Magnetic Resonance Imaging for Detection of Parametrial Invasion in Cervical Cancer: An Updated Systematic Review and Meta-Analysis of the Literature between 2012 and 2016. European Radiology 2018;28(2): 530–541. Chen H, Zhang X, Wang X, et al. MRI-Based Radiomics Signature for Pretreatment Prediction of Pathological Response to Neoadjuvant Chemotherapy in Osteosarcoma: A Multicenter Study. European Radiology 2021;31(10): 7913–7924. Sun C, Tian X, Liu Z, et al. Radiomic Analysis for Pretreatment Prediction of Response to Neoadjuvant Chemotherapy in Locally Advanced Cervical Cancer: A Multicentre Study. EBioMedicine 2019;46: 160–169. Ren J, Li Y, Yang J-J, et al. MRI-Based Radiomics Analysis Improves Preoperative Diagnostic Performance for the Depth of Stromal Invasion in Patients with Early Stage Cervical Cancer. Insights into Imaging 2022;13(1): 17. Chong G O, Park S-H, Park N J-Y,et al. Predicting Tumor Budding Status in Cervical Cancer Using MRI Radiomics: Linking Imaging Biomarkers to Histologic Characteristics. Cancers 2021;13(20): 5140. Yu W, Lu Y, Shou H, et al. A 5-Year Survival Status Prognosis of Nonmetastatic Cervical Cancer Patients through Machine Learning Algorithms. Cancer Medicine 2023;12(6): 6867–6876. Lu Y, Liu H, Liu Q, et al. CT-Based Radiomics with Various Classifiers for Histological Differentiation of Parotid Gland Tumors. Frontiers in Oncology 2023;13:1118351. Zhang B, He X, Ouyang F, et al. Radiomic Machine-Learning Classifiers for Prognostic Biomarkers of Advanced Nasopharyngeal Carcinoma. Cancer Letters 2017;403: 21–27. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterial2024.0510.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4772065","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":341817288,"identity":"9fbe4952-2c76-427e-84f0-a3cba56a5833","order_by":0,"name":"Ying Wang","email":"","orcid":"","institution":"The First Affiliated Hospital of Jinan University","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Wang","suffix":""},{"id":341817289,"identity":"5712cd20-ecd8-4370-988f-f12422f58454","order_by":1,"name":"Weixiao Liu","email":"","orcid":"","institution":"The First Affiliated Hospital of Jinan University","correspondingAuthor":false,"prefix":"","firstName":"Weixiao","middleName":"","lastName":"Liu","suffix":""},{"id":341817290,"identity":"b4772538-a2ae-408d-a316-bd0593eb700c","order_by":2,"name":"Yulian Chen","email":"","orcid":"","institution":"People's Hospital of YangJiang","correspondingAuthor":false,"prefix":"","firstName":"Yulian","middleName":"","lastName":"Chen","suffix":""},{"id":341817291,"identity":"c5772287-79ea-4ec6-8f6d-af6f475dd257","order_by":3,"name":"Fei Wang","email":"","orcid":"","institution":"The First Affiliated Hospital of Jinan University","correspondingAuthor":false,"prefix":"","firstName":"Fei","middleName":"","lastName":"Wang","suffix":""},{"id":341817292,"identity":"af165786-3d0e-4c59-873f-20345c445e71","order_by":4,"name":"Xiaoyun Liang","email":"","orcid":"","institution":"The First Affiliated Hospital of Jinan University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoyun","middleName":"","lastName":"Liang","suffix":""},{"id":341817293,"identity":"67cdea77-2365-4384-a3d4-de91148dcece","order_by":5,"name":"Xiao Zhang","email":"","orcid":"","institution":"The First Affiliated Hospital of Jinan University","correspondingAuthor":false,"prefix":"","firstName":"Xiao","middleName":"","lastName":"Zhang","suffix":""},{"id":341817294,"identity":"89de4229-0987-418c-9353-1facd3c6b245","order_by":6,"name":"Jin Fang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA10lEQVRIiWNgGAWjYBAC9gYgIWHAIMcIYjCwEaGF5wBEizGJWoAgEayDOC3sZw+/sCi4k948u8eA4UPZYQb+2Q0EtPDkpVlIGDzLbZxzxoBxxrnDDBJ3DuDXYs+QY2YgYXA4t3FGjgEzb9thBgOJBAK28L8Ba0lnBGn5S5QWiRzjB0AtCWAtjMRpeWMGDOTDho0z0goO9pxL55G4QdBhOcafJf4cljeckbzxwY8yazn+GQS0AAGbtASQNGxgYDgAMoOgeiBg/vgBSMoTo3QUjIJRMApGJgAAWxlAAQ67/gEAAAAASUVORK5CYII=","orcid":"","institution":"The First Affiliated Hospital of Jinan University","correspondingAuthor":true,"prefix":"","firstName":"Jin","middleName":"","lastName":"Fang","suffix":""}],"badges":[],"createdAt":"2024-07-20 08:29:04","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4772065/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4772065/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":63285524,"identity":"e64511fc-1f9a-46f3-b89a-581951e73b30","added_by":"auto","created_at":"2024-08-26 13:40:38","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":235530,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of the study.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4772065/v1/982f27b27bea9923af41a9ac.png"},{"id":63285526,"identity":"1e680cd6-8fa4-480c-ac48-bcfb2ddac9c4","added_by":"auto","created_at":"2024-08-26 13:40:39","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":66776,"visible":true,"origin":"","legend":"\u003cp\u003eReproducibility evaluation of segmentation and image assessment by radiologists. (\u003cstrong\u003eA\u003c/strong\u003e) variability of the radiomic-features family with respect to the segmentation under different ICC thresholds; (\u003cstrong\u003eB\u003c/strong\u003e) DSC from manual delineation of VOIs by two radiologists; (\u003cstrong\u003eC\u003c/strong\u003e) agreement of staging obtained by visual observation of MR images.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4772065/v1/8266ee7d7ca736ca9fd13e36.png"},{"id":63285523,"identity":"64fcbd1f-32a6-4d3e-8649-df9bc1f42b13","added_by":"auto","created_at":"2024-08-26 13:40:37","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":99874,"visible":true,"origin":"","legend":"\u003cp\u003eEvaluation curves of the optimal model obtained by various methods. ROC curves of four models for Hospital 1 (\u003cstrong\u003eA\u003c/strong\u003e), Hospital 2 (\u003cstrong\u003eB\u003c/strong\u003e), and Hospital 3 (\u003cstrong\u003eC\u003c/strong\u003e), verifying that the DWI-based peritumoral radiomics signature could offer the optimal performance for distinguishing IIA and IIB stage cervical cancer in data from different centers; calibration curves of the DWI-based peritumoral radiomics model for the training cohort (\u003cstrong\u003eD\u003c/strong\u003e) and validation cohort (\u003cstrong\u003eE\u003c/strong\u003e); decision curve analysis of the four models (\u003cstrong\u003eF\u003c/strong\u003e).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-4772065/v1/1aa994a4c53109759419dfac.png"},{"id":63285522,"identity":"d5a1791b-37d5-4984-93fc-7bb3daa565eb","added_by":"auto","created_at":"2024-08-26 13:40:37","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":417870,"visible":true,"origin":"","legend":"\u003cp\u003eMagnetic resonance imaging and pathological findings in patients with cervical cancer FIGO stage IIA and IIB.\u003cstrong\u003e (A-B)\u003c/strong\u003e Magnetic resonance images of patients with cervical cancer FIGO stage IIA, sagittal T2WI and DCE MR showed the red area of interest as the tumour area; \u003cstrong\u003e(C-D)\u003c/strong\u003e Cervical cancer tumour and paratesticular infiltration shown in HE.\u003cstrong\u003e (E-F)\u003c/strong\u003e Magnetic resonance images of a patient with cervical cancer FIGO stage IIB, with sagittal T2WI and DCE MR showed the red area of interest as the tumour area;\u003cstrong\u003e (H-G)\u003c/strong\u003e Pathological images of cervical cancer tumour and paratesticular infiltration in HE. Scale bars: 400 μm \u003cstrong\u003e(C, G)\u003c/strong\u003e, 200 μm \u003cstrong\u003e(D, H)\u003c/strong\u003e.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-4772065/v1/eb89bb5ec2485920a51243ae.png"},{"id":75201273,"identity":"e15f4ec6-3dee-473a-ac07-e4b0ddd098be","added_by":"auto","created_at":"2025-02-01 00:01:21","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1750177,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4772065/v1/db965cde-51f7-446f-81f4-e4b2c291001c.pdf"},{"id":63285525,"identity":"aab83755-e9e0-4f78-8296-47f9fa165729","added_by":"auto","created_at":"2024-08-26 13:40:39","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":136100,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial2024.0510.docx","url":"https://assets-eu.researchsquare.com/files/rs-4772065/v1/499f5b2fc14219063d62fb9b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multi-parametric MRI-based Peritumoral Radiomics for Stage IIA and IIB Classification of Cervical Cancer：A Multicenter Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eCervical cancer, despite being highly preventable, remains the fourth most common cancer and a leading cause of cancer-related deaths among women worldwide\u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e. The Federation of Gynecology and Obstetrics (FIGO) staging system is used to classify the extent of the disease. In the case of early-stage cervical cancer (FIGO IA to IIA), the recommended first-line treatment is radical hysterectomy with pelvic lymphadenectomy. For locally advanced cervical cancer (FIGO\u0026thinsp;\u0026ge;\u0026thinsp;IIB), the standard treatment is adjuvant concurrent chemoradiotherapy (CCRT) \u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e. It is important to accurately determine the FIGO stage because misclassification can impact treatment recommendations. Furthermore, there is an increasing demand for more conservative surgical approaches in young patients with cervical carcinoma. Precise identification of FIGO staging, particularly distinguishing between IIA and IIB stages before surgery in patients with cervical cancer, which can help avoid unnecessary surgical interventions and optimize treatment planning. This approach can be beneficial in preserving fertility and maintaining quality of life for these patients.\u003c/p\u003e \u003cp\u003eAccurate staging of cervical cancer is crucial for treatment planning, however, current methods have their limitations. Local staging is usually done through a clinical examination under anesthesia, which includes invasive endoscopic examination. Despite this, the subjective nature of this method can pose challenges, especially when assessing endocervical tumor growth and the extent of parametrial tissue involvement\u003csup\u003e[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/sup\u003e. Preoperative evaluation of the stage of cervical cancer frequently includes magnetic resonance imaging (MRI) as an adjunct imaging modality\u003csup\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003e. Research has shown notable differences between clinical examination and MRI-based staging, especially in evaluating tumor size larger than 2 cm, parametrial spread, extension to the pelvic wall, vaginal involvement, and adjacent organ involvement. MRI's specificity is limited and it has a high false-positive rate in stage classification for cervical cancer \u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e][\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e. In accordance with the 2018 FIGO staging system, pathological examination via biopsy is integrated into disease staging\u003csup\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003e. However, this technique is invasive and susceptible to sampling errors, which may obscure the tumor's overall heterogeneity. In light of these limitations, there is an imperative need for noninvasive diagnostic methods\u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eRadiomics has emerged as a promising tool in the field of medical imaging, enabling the extraction of quantitative features from images and the development of models to assist in the management of various aspects of cervical cancer\u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e][\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e. Radiomics has the potential to predict lymphovascular space invasion (LVSI), lymph node metastasis (LNM), and disease-free survival (DFS) according to previous research\u003csup\u003e[\u003cspan additionalcitationids=\"CR11 CR12\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e. Moreover, some previous studies have suggested the peritumoral radiomics also provide important predictive information on LVSI, LNM and DFS prediction in cervical cancer\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eCurrently, there is a lack of radiomics analysis based on magnetic resonance imaging (MRI) for accurately distinguishing between stage IIA and IIB cervical cancer. Our previous study indicated that the features from 5-pixel peritumoral regions improved the area under receiver operating characteristic area under curve (AUC) in predicting LNM in cervical cancer\u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e.Therefore, the incorporation of peritumoral radiomics into these analyses has the potential to provide valuable predictive information and enhance the staging and treatment planning of cervical cancer patients.\u003c/p\u003e \u003cp\u003eThe aim of this retrospective study was to assess the effectiveness of multi-parametric MRI (mpMRI)-based tumoral and peritumoral radiomics in differentiating between stage IIA and stage IIB cervical cancer, and further promote its potential clinical applications.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003ePatients\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInstitutional ethics review board approval was acquired for this study, and written informed consent was not required for this retrospective study. The inclusion criteria for patients were as follows: (1) pathologically confirmed stage IIA and IIB cervical cancer, (2) patients underwent pelvic MRI before the treatment, and (3) clinical data were available, such as age, FIGO stage, histologic characteristics, differentiation. Patients were excluded if their tumor lesions were not visible or if the MRI images had poor quality. Patients who had received pretreatment therapy or conization prior to MRI were also excluded. In total, 208 patients were enrolled from Hospital 1 (n=36, from January 2015 to December 2019), Hospital 2 (n = 105, from January 2011 to December 2017), and Hospital 3 (n = 67, from December 2012 to June 2018). Patients were randomly divided into the training and validation cohort by an approximative ratio of 2:1, and then, 145 and 63 patients were assigned to each cohort, respectively. The study flowchart is illustrated in \u003cstrong\u003eFigure 1.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Data Collection and Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAge, human papillomavirus (HPV), squamous cell carcinoma (SCC), carcinoembryonic antigen (CEA), neutrophil, lymphocyte, and pathological staging, were collected from the electronic medical records for patients pathologically confirmed cervical cancer. Due to the difference or lack of quantitative information collected by different hospitals, only three baseline factors, including age, SCC and CEA, were analyzed in this study. Univariate and multivariate logistic regression (LR) analyses were successively applied to these factors for the clinical model development.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eImage Acquisition and Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll patients underwent pelvic MRI scans including sagittal contrast-enhanced T1-weighted imaging(CE-T1WI), axial T2-weighted imaging (T2WI), and axial diffusion-weighted imaging (DWI). The obtained MRI images were stored in the picture archiving and communication system in a DICOM format.\u0026nbsp;Details regarding the acquisition parameters and MRI retrieval procedure were described in Supplementary Table S1.\u003c/p\u003e\n\u003cp\u003eA general preprocessing pipeline was applied on the images to reduce the intensity and geometric variations caused by different MRI scanners,\u0026nbsp;acquisition parameters, and patients. N4 algorithm embedded in 3D slicer software (version 5.0.3, www.slicer.org) was used to correct the distortion from non-uniform bias magnetic field in the MRI scanners. After standardizing the grayscale value of the images, all image voxels were uniformly resampled to 1\u0026times;1\u0026times;1 mm\u003csup\u003e3\u003c/sup\u003e with cubic B-spline interpolation algorithm.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eImage Assessment\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTwo radiologists (JF and YL.C, with 10 and 5 years of experience in gynecologic MR imaging, respectively)\u0026nbsp;independently determined the clinical stage based on the mpMRI in the absence of pathological and surgical results. According to 2018 FIGO stage, if the tumor extends beyond the cervix with the cervical stromal ring disrupted, on T2WI and DWI images, which show hyperintense signal extending to the parametrium were considered as IIB stage\u003csup\u003e[16]\u003c/sup\u003e .\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOne radiologist (YL.C) repeated this procedure a month later. We then assessed the intra- and inter-rater consistency of clinical stage as observed directly by radiologists from MR images, using intra-class correlation coefficient (ICC) and Kendall\u0026rsquo;s rank\u0026nbsp;correlation\u0026nbsp;coeffcient (KCC). The MR image assessment results used for subsequent analysis were reviewed by a senior radiologist (FW, with 13 years of experience) and corrected to get the final version.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVolume of Interest Segmentation and Reproducibility Evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTwo types of volume of interest (VOI), the tumor core and the peritumoral area (abbreviated as intratumoral VOI and peritumoral VOI, respectively, seen in \u003cstrong\u003eFigure 1\u003c/strong\u003e), were used for the radiomics analysis. Similarly, the two radiologists (YW and JF), blinded to the clinical information, independently manually delineated the tumor core by using the ITK-SNAP software (version 3.6.0, www.itksnap.org). The intratumoral VOIs of 50 patients randomly selected from the whole cohort were re-segmented over a 2-week interval by another radiologist (JF) with the same tool and environment settings. Correspondingly, the intra- and inter-rater reproducibility were measured by dice similarity coefficient (DSC) and ICC (based on extracted radiomics features). The VOIs with large differences were also corrected by the senior radiologist (FW). To obtained the peritumoral information, the intratumoral VOIs were dilated by a circular structural element with a radius size (number of a pixel) of 2, 3, 4, and 5. Peritumoral VOIs were defined by subtracting the intratumoral VOIs from the dilated area.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRadiomics Feature Extraction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCompliant with the Image Biomarker Standardization Initiative guidelines (https://theibsi.github.io/), a total of 1130 MRI-based radiomics features were extracted from each VOI by using the Pyradiomics package (version 3.0.1, https://pyradiomics.readthedocs.io/). The features characterizing shape, intensity, and texture patterns of the VOIs were obtained from the original and derived MR images using the Laplacian of gaussian and wavelet filter. The extraction parameters and feature types are summarized\u0026nbsp;in Supplementary Table S2. Z-score normalization was applied to the radiomics features for parameter standardization before radiomics analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRadiomics Feature Selection and Signature Construction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePearson correlation coefficient (PCC) analysis and recursive feature elimination (RFE) algorithm were adopted successively to obtain the well-representative features in the feature selection procedure. If the PCC of feature pairs was greater than 0.99, one of the features was randomly eliminated to obtain a feature subset with low redundancy. The RFE algorithm was further used for feature screening and combined with different classifiers to obtain the optimal radiomics signature.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRandom forest (RF), support vector machine (SVM), logistic regression (LR), naive Bayes (NB), Adaboost, and multilayer perceptron (MLP) classifiers were compared for constructing the radiomics signature. Hyperparameter optimization was performed using 5-fold cross-validation. It is important to note that radiomics feature selection and signature construction were based solely on data from the training cohort, while the validation cohort was used for evaluating and comparing the performance of the constructed signatures.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll the statistical analyses in this study were performed with SPSS version 22.0 and Python version 3.7.6. The independent-sample t test and the Chi-squared test were conducted to assess the significance of clinical factors between the training cohort and the validation cohort. Receiver operating characteristic curve (ROC), AUC, sensitivity, specificity, and accuracy were used to evaluate the prediction performance of different signatures. The difference between the AUCs were statistically analyzed using the method of DeLong et al \u003csup\u003e[17]\u003c/sup\u003e with MedCalc version 20.026. Furthermore, the\u0026nbsp;calibration curve and decision curve analysis were conducted to evaluate the clinical utility of the optimal signature. Note that\u0026nbsp;a two-tailed \u003cem\u003eP\u003c/em\u003e-value less than 0.05 was indicated statistically significant in this work.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003ePatient Characteristics\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAmong 208 patients with cervical cancer, 156 patients were in stage IIA and 52 patients were in stage IIB. The clinical characteristics of the training cohort (n=145, average age: 51.63\u0026plusmn;9.90) and the validation cohort (n=63, average age: 52.48\u0026plusmn;9.52) are presented in \u003cstrong\u003eTable 1\u003c/strong\u003e. It details the clinical characteristics such as age, squamous cell antigen, CEA, histology, degree of cellular differentiation. A statistical difference was observed between the age in the training set (\u003cem\u003eP\u003c/em\u003e = 0.044) and the CEA in the validation set (\u003cem\u003eP\u003c/em\u003e = 0.007) for stages IIA and IIB.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u0026nbsp;\u003c/strong\u003eCharacteristics of cervical cancer patients in training and validation cohorts\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eCharacteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"35.38732394366197%\" colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eTraining cohort\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(n=145)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"30.809859154929576%\" colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eValidation cohort\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;(n=63)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.26890756302521%\" valign=\"top\"\u003e\n \u003cp\u003eIIA\u003c/p\u003e\n \u003cp\u003e(n=109)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"19.95798319327731%\" valign=\"top\"\u003e\n \u003cp\u003eIIB\u003c/p\u003e\n \u003cp\u003e(n=36)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"10.504201680672269%\" valign=\"top\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.80672268907563%\" valign=\"top\"\u003e\n \u003cp\u003eIIA\u003c/p\u003e\n \u003cp\u003e（n=47）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"19.95798319327731%\" valign=\"top\"\u003e\n \u003cp\u003eIIB\u003c/p\u003e\n \u003cp\u003e（n=16）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"10.504201680672269%\" valign=\"top\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eAge \u0026nbsp;(years)mean \u0026plusmn; SD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e52.58\u0026plusmn;\u0026nbsp;9.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e48.75\u0026plusmn;10.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.044\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e52.82\u0026plusmn;\u0026nbsp;9.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e48.81 \u0026plusmn;9.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.176\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eCEA (ng/mL）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e6.25\u0026plusmn;13.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e60.07\u0026plusmn;309.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.304\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e4.68\u0026plusmn;4.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e21.64\u0026plusmn;21.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eSCC-Ag (ng/mL)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e5.65\u0026plusmn;8.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e6.35\u0026plusmn;8.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.673\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e5.13\u0026plusmn;7.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e8.57\u0026plusmn;10.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.190\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eHistology（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eSquamous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e98（89.91%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e33（91.67%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.655\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e44（93.62%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e15（93.75%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e1.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eAdenocarcinoma\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e11（10.09%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e2（5.56%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e2（4.26%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e1（6.25%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eDegree of cellular differentiation, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eLow\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e26（23.85%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e13（36.11%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.355\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e15（31.91%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e3（18.75%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e0.253\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eMiddle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e72（66.06%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e20（55.56%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e29\u0026nbsp;（61.70%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e10（62.50%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.197183098591548%\" valign=\"top\"\u003e\n \u003cp\u003eHigh\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.661971830985916%\" valign=\"top\"\u003e\n \u003cp\u003e11\u0026nbsp;（10.09%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e3\u0026nbsp;（8.33%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.084507042253522%\" valign=\"top\"\u003e\n \u003cp\u003e3（6.38%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"16.725352112676056%\" valign=\"top\"\u003e\n \u003cp\u003e3（18.75%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.80281690140845%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations:\u003c/strong\u003e \u003cem\u003eCEA\u003c/em\u003e, carcinoembryonic antigen; \u003cem\u003eSCC\u003c/em\u003e, squamous cell carcinoma; P* \u0026lt; 0.05 indicates a significant difference.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSegmentation Reproducibility Evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor manual delineation of VOIs by two radiologists, the mean and median values of intra- and inter-rater DSC exceeded 0.90, indicating that the difference in delineation was relatively small in this work (\u003cstrong\u003eFigure 2B\u003c/strong\u003e). To quantitatively compare the variability of the radiomic-features family with respect to the segmentation, we analyzed the variation in the number of features under different ICC thresholds (\u003cstrong\u003eFigure 2A\u003c/strong\u003e). The results showed that the radiomics features also presented good reproducibility under the condition of small segmentation difference, which was applicable to both intra- and inter-rater ICC. Furthermore, 91.6% (1035/1130), 91.5% (1034/1130), 91.9% (1039/1130) of the features were identified as highly reproducibility with ICC\u0026ge;0.75 of both intra- and inter-rater variability.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictive Performance of Clinical model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAge (odds ratio[OR], 0.962; 95% confidence interval[CI], 0.925\u0026minus;0.999; \u003cem\u003eP\u0026nbsp;\u003c/em\u003e= 0.047*), SCC (OR, 1.009; 95% CI, 0.967\u0026minus;1.053; \u003cem\u003eP\u003c/em\u003e = 0.671), and CEA (OR, 1.010; 95% CI, 0.986\u0026minus;1.034; \u003cem\u003eP\u003c/em\u003e = 0.407) were assessed by univariate LR analysis. Age and CEA was found significantly associated with the stage and then used to construct a clinical model. The model yielded AUC values of 0.611 (95% CI, 0.573\u0026minus;0.649) in the training cohort and 0.652 (95% CI, 0.616\u0026minus;0.687) in the validation cohort, respectively (\u003cstrong\u003eTable 2\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictive Performance of\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eImage Assessment by Radiologists\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe low intra- and inter-rater agreement\u0026nbsp;of staging obtained by visual observation of MR images\u0026nbsp;were presented between the two radiologists. The intra- and inter-rater ICC were 0.555 and 0.553, whereas intra- and inter-rater KCC were 0.556 and 0.560, respectively\u0026nbsp;(\u003cstrong\u003eFigure 2C\u003c/strong\u003e). As seen in \u003cstrong\u003eTable 2\u003c/strong\u003e, the AUCs of the predicted clinical stage after correction by the senior radiologist were 0.710 (95% CI: (0.680\u0026minus;0.739) and 0.757 (95% CI: 0.730\u0026minus;0.785) in the training and validation cohort, respectively.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u0026nbsp;\u003c/strong\u003ePrediction performance of different models for the classification of stage IIA and IIB.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"598\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.863105175292155%\" colspan=\"2\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"41.235392320534224%\" colspan=\"5\"\u003e\n \u003cp\u003eTraining cohort\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"40.90150250417362%\" colspan=\"5\"\u003e\n \u003cp\u003eValidation cohort\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.863105175292155%\" colspan=\"2\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.021702838063439%\"\u003e\n \u003cp\u003eAUC (95% CI)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"6.510851419031719%\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.178631051752921%\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.345575959933222%\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.021702838063439%\"\u003e\n \u003cp\u003eAUC (95% CI)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.863105175292155%\" colspan=\"2\"\u003e\n \u003cp\u003eClinical\u0026nbsp;model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.021702838063439%\"\u003e\n \u003cp\u003e0.611\u0026nbsp;(0.573\u0026minus;0.649)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"6.510851419031719%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.178631051752921%\"\u003e\n \u003cp\u003e0.417\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.345575959933222%\"\u003e\n \u003cp\u003e0.761\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.676\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.021702838063439%\"\u003e\n \u003cp\u003e0.652\u0026nbsp;(0.616\u0026minus;0.687)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.438\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.766\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.683\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.863105175292155%\" colspan=\"2\"\u003e\n \u003cp\u003eImage\u0026nbsp;assessment\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.021702838063439%\"\u003e\n \u003cp\u003e0.710\u0026nbsp;(0.680\u0026minus;0.739)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"6.510851419031719%\"\u003e\n \u003cp\u003eRef.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.178631051752921%\"\u003e\n \u003cp\u003e0.750\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.345575959933222%\"\u003e\n \u003cp\u003e0.670\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.690\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.021702838063439%\"\u003e\n \u003cp\u003e0.757\u0026nbsp;(0.730\u0026minus;0.785)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003eRef.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.813\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.702\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.011686143572621%\"\u003e\n \u003cp\u003e0.730\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"10.367892976588628%\" rowspan=\"4\"\u003e\n \u003cp\u003eIntratumoral\u0026nbsp;radiomics\u0026nbsp;signature\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.357859531772576%\"\u003e\n \u003cp\u003eDWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.043478260869565%\"\u003e\n \u003cp\u003e0.783\u0026nbsp;(0.751\u0026minus;0.815)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"6.521739130434782%\"\u003e\n \u003cp\u003e0.002*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.190635451505017%\"\u003e\n \u003cp\u003e0.583\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.357859531772576%\"\u003e\n \u003cp\u003e0.835\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.772\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.043478260869565%\"\u003e\n \u003cp\u003e0.707\u0026nbsp;(0.670\u0026minus;0.745)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.028*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.438\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.851\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.746\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003eCE-T1WI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.818\u0026nbsp;(0.788\u0026minus;0.848)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.276119402985074%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.022388059701493%\"\u003e\n \u003cp\u003e0.750\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003e0.716\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.724\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.771\u0026nbsp;(0.739\u0026minus;0.803)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.424\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.702\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.683\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003eT2WI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.676\u0026nbsp;(0.640\u0026minus;0.712)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.276119402985074%\"\u003e\n \u003cp\u003e0.085\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.022388059701493%\"\u003e\n \u003cp\u003e0.639\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003e0.688\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.676\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.680\u0026nbsp;(0.643\u0026minus;0.716)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.596\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.603\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003eFusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.872\u0026nbsp;(0.846\u0026minus;0.899)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.276119402985074%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.022388059701493%\"\u003e\n \u003cp\u003e0.861\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003e0.780\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.800\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.815\u0026nbsp;(0.784\u0026minus;0.847)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.787\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.746\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"10.367892976588628%\" rowspan=\"4\"\u003e\n \u003cp\u003ePeritumoral\u0026nbsp;radiomics\u0026nbsp;signature\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.357859531772576%\"\u003e\n \u003cp\u003eDWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.043478260869565%\"\u003e\n \u003cp\u003e0.975\u0026nbsp;(0.966\u0026minus;0.984)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"6.521739130434782%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.190635451505017%\"\u003e\n \u003cp\u003e0.917\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.357859531772576%\"\u003e\n \u003cp\u003e0.972\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.959\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.043478260869565%\"\u003e\n \u003cp\u003e0.899\u0026nbsp;(0.878\u0026minus;0.920)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.979\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.023411371237458%\"\u003e\n \u003cp\u003e0.889\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003eCE-T1WI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.949\u0026nbsp;(0.935\u0026minus;0.962)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.276119402985074%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.022388059701493%\"\u003e\n \u003cp\u003e0.833\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003e0.908\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.890\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.895\u0026nbsp;(0.877\u0026minus;0.913)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.750\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.872\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.841\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003eT2WI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.933\u0026nbsp;(0.916\u0026minus;0.950)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.276119402985074%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.022388059701493%\"\u003e\n \u003cp\u003e0.861\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003e0.890\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.883\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.777\u0026nbsp;(0.748\u0026minus;0.806)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.360\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.702\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.651\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003eFusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.981\u0026nbsp;(0.973\u0026minus;0.988)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.276119402985074%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.022388059701493%\"\u003e\n \u003cp\u003e0.917\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.208955223880597%\"\u003e\n \u003cp\u003e0.972\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.959\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.552238805970148%\"\u003e\n \u003cp\u003e0.895\u0026nbsp;(0.874\u0026minus;0.915)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e\u0026lt;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.979\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"7.835820895522388%\"\u003e\n \u003cp\u003e0.889\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations:\u003c/strong\u003e \u003cem\u003eAUC\u003c/em\u003e, area under the curve; \u003cem\u003eCI\u003c/em\u003e, confidence interval; \u003cem\u003eDWI\u003c/em\u003e, diffusion weighted imaging; \u003cem\u003eCE-T1WI\u003c/em\u003e, contrast-enhanced T1-weighted\u0026nbsp;imaging; \u003cem\u003eT2WI\u003c/em\u003e, T2-weighted\u0026nbsp;imaging;\u0026nbsp;P* \u0026lt; 0.05 indicates a significant difference.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictive Perfor\u003c/strong\u003e\u003cstrong\u003emance of Intratumoral Radiomics Signature\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter feature selection, 8, 16, and 3 predictive features were separately identified to constructed the optimal intratumoral radiomics signatures for DWI, CE-T1WI, and T2WI. Details of selected features are shown in\u0026nbsp;\u003cstrong\u003eSupplementary Table S3\u003c/strong\u003e. The optimal classifiers were LR, MLP, and LR, respectively. We found that the predictive performance of CE-T1WI-based radiomics signature is better than that of DWI and T2WI in both the training and validation cohorts (\u003cstrong\u003eTable 2\u003c/strong\u003e; all \u003cem\u003eP\u0026nbsp;\u003c/em\u003e\u0026lt; 0.05). The fusion model derived from joint DWI, CE-T1WI, and T2WI signatures was considered to be the optimal intratumoral radiomics signature, superior to the signatures from single sequences in differentiating stages (all \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.01). The AUCs were 0.872 (95% CI: 0.846\u0026minus;0.899) and 0.815 (95% CI: 0.784\u0026minus;0.847) in the training and validation cohorts, respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePred\u003c/strong\u003e\u003cstrong\u003eictive Performance of Peritumoral Radiomics Signature\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe peritumoral VOIs were rebuilt using a structural element with a radius of 3, which showed noticeably better discriminatory performance than that from the surrounding 2, 4, and 5-pixel extensions (\u003cstrong\u003eSupplementary Figure S1\u003c/strong\u003e).\u0026nbsp;14, 14 and 7 radiomics features, respectively, in combination with SVM, MLP and LR classifier, were used to develop the optimal DWI-, CE-T1WI-, and T2WI-based radiomics signatures (\u003cstrong\u003eSupplementary Table S4\u003c/strong\u003e). As reflected in \u003cstrong\u003eTable 2\u003c/strong\u003e, DWI-based signature performed best among all the sequences with the AUCs of 0.975 (95% CI: 0.966\u0026minus;0.984) and 0.899 (95% CI: 0.878\u0026minus;0.920) in the training and validation cohort, respectively. There was no significant difference between the validation AUCs of DWI-based and fusion peritumoral radiomics model (0.899 vs. 0.895, \u003cem\u003eP\u003c/em\u003e = 0.566), implying important predictive information contained in DWI. Thus, DWI-based signature was identified as the optimal\u0026nbsp;peritumoral\u0026nbsp;radiomics signature.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eComparison of the Various Predictive Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe peritumoral radiomics signatures were superior to the intratumoral radiomics signatures, regardless of single sequence or fusion analysis (all \u003cem\u003eP\u0026nbsp;\u003c/em\u003e\u0026lt; 0.001). DWI-based\u0026nbsp;peritumoral\u0026nbsp;radiomics signature performed best for both cohorts, significantly superior to the optimal intratumoral radiomics signature, clinical model, and image assessment by radiologists (all \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001). \u003cstrong\u003eFigure 3A-C\u003c/strong\u003e show ROC curves of the models corresponding to three hospitals, confirming that the DWI-based\u0026nbsp;peritumoral\u0026nbsp;radiomics signature could offer the optimal performance for\u0026nbsp;distinguishing IIA and IIB stage cervical cancer\u0026nbsp;in data from different centers. In particular, the optimal model correctly predicted the staging of all cases at Hospital 1 and Hospital 2.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Use of the Optimal Peritumoral Radiomics Signature\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe calibration curves of the optimal\u0026nbsp;peritumoral\u0026nbsp;radiomics signature (\u003cstrong\u003eFigure 3D and E\u003c/strong\u003e) demonstrated good agreement between prediction and actual observation in both cohorts (mean absolute error = 0.023 and 0.047, Hosmer-Lemeshow test \u003cem\u003eP\u003c/em\u003e = 0.824 and 0.135, respectively). The DCA plotted in \u003cstrong\u003eFigure 3F\u0026nbsp;\u003c/strong\u003eindicated the incremental clinical advantage of information from peritumoral region over information from other sources. Patients with cervical cancer could benefit more within the whole risk threshold range.\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eDistinguishing\u0026nbsp;IIA and IIB stageis essential because the therapeutic regimens and follow-up care are vastly different for each of these conditions. According to 2018 FIGO stage system for cervical cancer, IIA stage limited to the upper two-thirds of the vagina without parametrial involvement; while IIB with parametrial involvement but not up to the pelvic wall. So the parametrial involvement (PMI) is the main differentiation between IIA and IIB stages.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMR images are used as a supplementary diagnostic tool in distinguishing between IIA and IIB based on tumor size, disruption of the hypointense cervical stromal ring on T2WI and DWI. It is important to note that MRI may occasionally overestimate parametrial invasion in large tumors (with an accuracy of 70%) due to peritumoral edema or inflammation caused by tumor compression. This factor should be considered when planning treatment \u003csup\u003e[10]\u003c/sup\u003e. The diagnostic value of MRI assessment by radiologists in our study is 0.757, which is similar to that of previous reports \u003csup\u003e[11]\u003c/sup\u003e\u003csup\u003e\u0026nbsp;[18]\u003c/sup\u003e. As shown in \u003cstrong\u003eFigure 4,\u003c/strong\u003e patient with FIGO stage IIA cervical cancer has preserved stromal ring on MR image. In addition, patient with FIGO stage IIB has involved stromal disruption with fat or soft-tissue strands in the parametrium. In the corresponding histopathological examination (HE) images (\u003cstrong\u003eC, D, G, and F in Figure 4\u003c/strong\u003e), the stage IIA cervical cancer tumor did not invade the paracervical tissues, whereas stage IIB tumor invaded.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eZhao\u003csup\u003e\u0026nbsp;\u003c/sup\u003eet al. conducted a study on the development of radiomics models employing the T2WI sequence, which demonstrated outstanding performance in differentiating between early stages of cervical cancer (stage IB vs. stage IIA). Using a SVM classifier, the model obtained impressive metrics, including AUC of 0.907%, sensitivity of 0.778, specificity of 0.999, and accuracy of 0.867\u003csup\u003e[18]\u003c/sup\u003e. It is important to note that this particular study only considered tumoral regions and did not include data from the peritumoral regions. In our study, we explored the intratumoral and peritumoral regions of cervical cancer in differentiating IIA and IIB stage, and the results revealed that the most predictable features could be obtained from the peritumoral regions with 3 pixel dilation distances in the DWI images.\u003c/p\u003e\n\u003cp\u003eOur results showed that DWI\u0026nbsp;images could provide good functional information in distinguish IIA and IIB stage.\u0026nbsp;DWI is a functional imaging technique in association with water diffusion properties, the tumor shows a higher signal intensity and lower apparent diffusion coefficient value than normal cervical tissue.\u0026nbsp;Previous studies showed that DWI can significantly improve diagnostic efficacy in stage, risk factors prediction\u0026nbsp;\u003csup\u003e[19]\u003c/sup\u003e\u003csup\u003e[20]\u003c/sup\u003e,\u0026nbsp;which can detect early pathological changes associated with changes in water content in tissues, not effected by edema and inflammation around the tumor.\u003c/p\u003e\n\u003cp\u003eThe connective tissue of the parametrium is mainly composed of fat, blood vessels, nerves, fibrous tissues and lymphatic vessels. Due to the lack of fascial restriction, direct infiltration and invasion of the tumor is more likely to occur in the tumor parametrium. Therefore,\u0026nbsp;the immediate surrounding tumor microenvironment may offer unique radiomics signatures of stage classification.\u0026nbsp;Considering the peritumoral microenvironment of lymphatics and blood vessels invasion varies with tumor size, unlike previously reported study\u003csup\u003e[13]\u003c/sup\u003e\u003csup\u003e\u0026nbsp;[21]\u003c/sup\u003e, we selected\u0026nbsp;pixel instead of fixed diameter, because the pixel is more objective and reliable, which vary with the size of the tumor itself.\u0026nbsp;In this study, we\u0026nbsp;compare the 2,3,4,5 pixels extension surrounding the tumor,\u0026nbsp;the results demonstrated that 3 pixel peritumoral regions of\u0026nbsp;radiomics signatures show the best discrimination performance in distinguishing IIA and IIB stage, which can be partly explained by that the fact that peritumoral region provides more information about parametrial infiltration\u0026nbsp;\u003csup\u003e[22]\u003c/sup\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe choice of robust and reproducible radiomics features should be emphasized for promoting subsequent applicability in clinical. Manual segmentation suffers from high intra- or inter-reader variability, and the value of the extracted radiomics features may differ due to variances in tumor segmentations uncertainty. Thus, multiple-segmentation by multiple radiologists is a method to provided reproducible and reliable radiomic features. It is important to identify whether the features extracted from the two types of semiautomatic segmentations capture the same tumor image properties as manual delineation. We applied ICC analysis to preliminarily identify stable features that were less sensitive to segmentation differences to improve the reproducibility of feature extraction. Similar to the previous study\u003csup\u003e[23]\u003c/sup\u003e, we used PCC analysis to remove\u0026nbsp;some highly correlated features. Thus, a feature subset with lower dimensions was obtained, and then the final feature filtering was performed. In the model development procedure, we compared the performance of multiple classifiers and selected the best one based on the corresponding MRI sequence. This makes sense in a lot of research because it is difficult for one classifier to perform well in all analysis tasks\u003csup\u003e[24-27]\u003c/sup\u003e. The results of the optimal peritumoral radiomics signatures highlights the potential of texture feature which were difficult to decipher through simple observation but provided considerably richer information for distinguishing IIA and IIB stage cervical cancer.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLimitation \u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOur study has several limitations. First, our study was performed with a relatively small population. Further work is required to focus on validating the proposed model on a larger multicenter dataset, and external validation in different centers is still needed. Secondly, our study manually extracted three-dimensional texture features, which is subjective and time consuming. The use of automatic segmentation technology may offer assistance in this regard. Additionally, important clinical characteristics (e.g. HPV) that were correlated with stage in cervical cancer were not involved in this study. The integration of additional multi-source characteristics into the existing model to further develop a clinical-AI framework will require exploration in order to facilitate its clinical application.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eMRI-based radiomics model from peritumoral regions outperformed radiologists for the preoperative diagnosis of IIA and IIB stage. A noninvasive and reliable supplementary approach was provided to precise preoperative staging, which can help optimize individualized treatment plans for patients with cervical cancer.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict of Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no potential conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by Guangdong Basic and Applied Basic Research Foundation (grant number: 2021A1515110763); Medical science and Technology Research Fund of Guangdong Province (grant number: A2022267); The Science and Technology Program of Guangzhou (grant number: 202201020057,2023A03J1038); The Clinical Frontier Technology Program of the First Affiliated Hospital of Jinan University, China (No. JNU1AF-CFTP-2022-a01201).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors contributions\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConcept and design: J.F, X.Z; Acquisition, analysis, or interpretation of data: J.F., Y.W., YL.C.; Drafting of the manuscript: Y.W., J.F., WX. L; Statistical analysis: F.W.,YL.C.; Performing experiments: YW.,WX.L; Supervision and revision of the manuscript: J.F. X.Z. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics declaration:\u0026nbsp;\u003c/strong\u003enot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCastle P E, Einstein M H, Sahasrabuddhe V V. Cervical cancer prevention and control in women living with human immunodeficiency virus. Ca 2021;71(6): 505\u0026ndash;526.\u003c/li\u003e\n\u003cli\u003eCohen PA, Jhingran A, Oaknin A, et al. Cervical cancer. Lancet 2019;393(10167):169-182.\u003c/li\u003e\n\u003cli\u003eBourgioti C, Chatoupis K, Rodolakis A, et al. Incremental Prognostic Value of MRI in the Staging of Early Cervical Cancer: A Prospective Study and Review of the Literature. Clinical Imaging 2016;40(1): 72\u0026ndash;78.\u003c/li\u003e\n\u003cli\u003eManganaro L, Lakhman Y, Bharwani N, et al. Correction to: Staging, recurrence and follow-up of uterine cervical cancer using MRI: Updated Guidelines of the European Society of Urogenital Radiology after revised FIGO staging 2018. Eur Radiol 2022;32(1):738.\u003c/li\u003e\n\u003cli\u003eDappa E, Elger T, Hasenburg A, et al. The value of advanced MRI techniques in the assessment of cervical cancer: a review. Insights into Imaging 2017;8(5): 471\u0026ndash;481. \u003c/li\u003e\n\u003cli\u003eKnoth J, P\u0026ouml;tter R, J\u0026uuml;rgenliemk-Schulz I M, et al. Clinical and Imaging Findings in Cervical Cancer and Their Impact on FIGO and TNM Staging \u0026ndash; An Analysis from the EMBRACE Study. Gynecologic Oncology 2020;159(1): 136\u0026ndash;141. \u003c/li\u003e\n\u003cli\u003eZhou Y, Gu H-L, Zhang X-L, et al. Multiparametric Magnetic Resonance Imaging-Derived Radiomics for the Prediction of Disease-Free Survival in Early-Stage Squamous Cervical Cancer. Eur Radiol 2022;32(4): 2540\u0026ndash;2551.\u003c/li\u003e\n\u003cli\u003eGillies R J, Kinahan P E, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278(2): 563\u0026ndash;577.\u003c/li\u003e\n\u003cli\u003eLambin P, Leijenaar R T H, Deist T M, et al. Radiomics: The Bridge between Medical Imaging and Personalized Medicine. Nature Reviews. Clinical Oncology 2017;14(12): 749\u0026ndash;762..\u003c/li\u003e\n\u003cli\u003eBalcacer P, Shergill A, Litkouhi B. MRI of Cervical Cancer with a Surgical Perspective: Staging, Prognostic Implications and Pitfalls. Abdominal Radiology 2019;44(7): 2557\u0026ndash;2571.\u003c/li\u003e\n\u003cli\u003eXiao M, Ma F, Li Y et al. Multiparametric MRI-based radiomics nomogram for predicting lymph node metastasis in early-stage cervical cancer. J Magn Reson Imaging 2020;52:885\u0026ndash;896 .\u003c/li\u003e\n\u003cli\u003eLi Z, Li H, Wang S et al. MR-based radiomics nomogram of cervical cancer in prediction of the lymph-vascular space invasion preoperatively. J Magn Reson Imaging 2019; 49:1420\u0026ndash;1426.\u003c/li\u003e\n\u003cli\u003eWang T, Gao T, Guo H et al. Preoperative prediction of parametrial invasion in early-stage cervical cancer with MRI based radiomics nomogram. Eur Radiol 2020;30:3585\u0026ndash;3593.\u003c/li\u003e\n\u003cli\u003eZhou Y, Gu HL, Zhang XL, Tian ZF, Xu XQ, Tang WW. Multiparametric magnetic resonance imaging‑derived radiomics for the prediction of disease‑free survival in early‑stage squamous cervical cancer. Eur Radiol 2022;32(4):2540-2551.\u003c/li\u003e\n\u003cli\u003eWu QX, Wang S, Zhang SX, et al. Development of a Deep Learning Model to Identify Lymph Node Metastasis on Magnetic Resonance Imaging in Patients With Cervical Cancer. JAMA Netw Open 2020;3(7): e2011625.\u003c/li\u003e\n\u003cli\u003eBalcacer P, Shergill A, Litkouhi B. MRI of Cervical Cancer with a Surgical Perspective: Staging, Prognostic Implications and Pitfalls. Abdominal Radiology 2019;44(7): 2557\u0026ndash;2571. \u003c/li\u003e\n\u003cli\u003eDelong E R , Delong D M, Clarke-Pearson D L. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach.[J].Biometrics, 1988, 44(3):837.\u003c/li\u003e\n\u003cli\u003eZhao X, Wang X, Zhang B, et al. Classifying Early Stages of Cervical Cancer with MRI-Based Radiomics. Magnetic Resonance Imaging 2022;89: 70\u0026ndash;76.\u003c/li\u003e\n\u003cli\u003ePark J J, Kim C K, Park S Y, et al. Value of Diffusion-Weighted Imaging in Predicting Parametrial Invasion in Stage IA2-IIA Cervical Cancer. European Radiology 2014;24(5): 1081\u0026ndash;1088.\u003c/li\u003e\n\u003cli\u003eWoo S, Suh C H, Kim S Y, et al. Magnetic Resonance Imaging for Detection of Parametrial Invasion in Cervical Cancer: An Updated Systematic Review and Meta-Analysis of the Literature between 2012 and 2016. European Radiology 2018;28(2): 530\u0026ndash;541.\u003c/li\u003e\n\u003cli\u003eChen H, Zhang X, Wang X, et al. MRI-Based Radiomics Signature for Pretreatment Prediction of Pathological Response to Neoadjuvant Chemotherapy in Osteosarcoma: A Multicenter Study. European Radiology 2021;31(10): 7913\u0026ndash;7924. \u003c/li\u003e\n\u003cli\u003eSun C, Tian X, Liu Z, et al. Radiomic Analysis for Pretreatment Prediction of Response to Neoadjuvant Chemotherapy in Locally Advanced Cervical Cancer: A Multicentre Study. EBioMedicine 2019;46: 160\u0026ndash;169. \u003c/li\u003e\n\u003cli\u003eRen J, Li Y, Yang J-J, et al. MRI-Based Radiomics Analysis Improves Preoperative Diagnostic Performance for the Depth of Stromal Invasion in Patients with Early Stage Cervical Cancer. Insights into Imaging 2022;13(1): 17.\u003c/li\u003e\n\u003cli\u003eChong G O, Park S-H, Park N J-Y,et al. Predicting Tumor Budding Status in Cervical Cancer Using MRI Radiomics: Linking Imaging Biomarkers to Histologic Characteristics. Cancers 2021;13(20): 5140. \u003c/li\u003e\n\u003cli\u003eYu W, Lu Y, Shou H, et al. A 5-Year Survival Status Prognosis of Nonmetastatic Cervical Cancer Patients through Machine Learning Algorithms. Cancer Medicine 2023;12(6): 6867\u0026ndash;6876. \u003c/li\u003e\n\u003cli\u003eLu Y, Liu H, Liu Q, et al. CT-Based Radiomics with Various Classifiers for Histological Differentiation of Parotid Gland Tumors. Frontiers in Oncology 2023;13:1118351.\u003c/li\u003e\n\u003cli\u003eZhang B, He X, Ouyang F, et al. Radiomic Machine-Learning Classifiers for Prognostic Biomarkers of Advanced Nasopharyngeal Carcinoma. Cancer Letters 2017;403: 21\u0026ndash;27. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Cervical cancer, Peritumoral radiomics, Magnetic resonance imaging, Stage classification","lastPublishedDoi":"10.21203/rs.3.rs-4772065/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4772065/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjective: \u003c/strong\u003eThe aim of the study is to establish a multiparametric MRI (mpMRI)-based peritumoral radiomics nomogram for preoperatively predicting IIA and IIB classification of cervical Cancer preoperatively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003e208 patients with histologically confirmed cervical cancer from three institutions were enrolled in this study. All the cases were randomly divided into the training cohort (n=145) and the validation cohort (n=63). The performance of the nomogram was assessed with respect to its calibration, discrimination, and clinical usefulness. The independent-sample t test and the Chi-squared test were conducted to assess the significance of clinical factors between the training cohort and the validation cohort. The Pearson correlation coefficient analysis and recursive feature elimination algorithm were adopted successively to obtain the well-representative features. Different classifiers were compared to develop the optimal radiomics signature across 5-fold cross validation. The calibration curves and decision curve analysis were conducted to evaluate the clinical utility of the optimal model. The radiomics model was constructed using logistic regression.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eThe peritumoral radiomics models were superior to the intratumoral radiomics models, regardless of single sequence model or fusion model (all \u003cem\u003eP\u003c/em\u003e \u0026lt;0.001*). DWI-based peritumoral radiomics model performed best with the AUCs of 0.975 (0.965−0.983) and 0.899 (0.880−0.916) in the training and validation cohort, respectively. There was no significant difference between the validation AUCs of DWI-based and fusion peritumoral radiomics model (0.899 vs. 0.895, \u003cem\u003eP\u003c/em\u003e=0.566). In addition, 3 pixel peritumoral regions of radiomic signatures have a much better discrimination performance in distinguishing IIA and IIB stage by comparing the 2,4,5 pixels extension surrounding the tumor.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eMRI-based radiomics model from peritumoral regions of cervical cancer outperformed radiologists for the preoperative diagnosis of IIA and IIB stage, which could provide a noninvasive and reliable way of individualized treatment plans for patients with cervical cancer.\u003c/p\u003e","manuscriptTitle":"Multi-parametric MRI-based Peritumoral Radiomics for Stage IIA and IIB Classification of Cervical Cancer：A Multicenter Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-08-26 13:40:30","doi":"10.21203/rs.3.rs-4772065/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"50e77077-1f4c-47f2-85a2-f737ac6c4523","owner":[],"postedDate":"August 26th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-01-31T23:53:14+00:00","versionOfRecord":[],"versionCreatedAt":"2024-08-26 13:40:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4772065","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4772065","identity":"rs-4772065","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}