Multimodal MRI Radiomics Features and Deep Learning Features in Predicting Lymphovascular Space Invasion in Endometrial Carcinoma

Multimodal MRI Radiomics Features and Deep Learning Features in Predicting Lymphovascular Space Invasion in Endometrial Carcinoma | 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 Multimodal MRI Radiomics Features and Deep Learning Features in Predicting Lymphovascular Space Invasion in Endometrial Carcinoma Yuan Tang, Feifei Shan, Wei Zhou, Kang Wang, Yuan Chen, Junqi Sun This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6281099/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 Background Lymphovascular space invasion (LVSI) is a key prognostic indicator in endometrial cancer, impacting disease progression, treatment strategies and overall survival of patients. The accurate identifying of LVSI of endometrial cancer before surgery is challenging due to certain limitations in radiological methods. Objective To develop an efficient approach for predicting LVSI in endometrial cancer before surgery using multimodal MRI Radiomics and deep learning, offering a crucial foundation for clinical treatment. Methods Two radiologists, unaware of pathology, manually outlined the region of interest (ROI) on preoperative DWI, T1WI + C, and T2WI images. The intersection of their ROIs was the final one. radiomics features were extracted from the ROI, and depth features from the largest - area ROI slice in MRI images. In the training set, 8 models were built via logistic regression after feature reduction. These models were tested against pathology, and the area under the receiver operating characteristic curve (AUC) was calculated. The best - performing imaging model was combined with a deep - learning model to form a hybrid, and its performance was evaluated.​ Results 308 patients were split 7:3 into a training set (n = 215) and a test set (n = 93). The LR model performed best in LVSI prediction. Combining clinical and radiomics with the LR algorithm enhanced performance. Deep - learning mixed - feature models had AUCs of 0.948 (training) and 0.704 (test), while mixed models had 0.953 (training) and 0.720 (test).​ Conclusion Deep - learning mixed - feature and mixed models are most effective for preoperative LVSI prediction in endometrial cancer, guiding clinical decision - making. Endometrial carcinoma lymphatic vascular space invasion MRI Radiomics Deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Endometrial carcinoma (EC) is one of the most common malignancies of the female reproductive system worldwide. In 2022, the number of new cases worldwide reached 420,242, accounting for 2.1% of the cancer incidence. It is more common in female cancers. From the overall trend of global cancer incidence, it is speculated that the number of cases may increase due to population growth, aging and lifestyle changes [1] . Lymphatic vascular space infiltration (LVSI) is an important pathological feature of EC, which is closely related to tumor stage, lymph node metastasis and patient prognosis [2,3] . Preoperative accurate prediction of LVSI is of great significance for formulating reasonable treatment plans and evaluating patient prognosis [4] .At present, surgical pathology is the "gold standard" for the diagnosis of LVSI, but this method is invasive and cannot provide a timely basis for preoperative clinical decision-making [6] . Magnetic resonance imaging (MRI) plays an important role in the preoperative evaluation of EC due to its high soft tissue resolution and multi-parameter imaging capabilities [5] . Multimodal MRI can not only clearly display the morphology, location and degree of tumor invasion, but also provide functional information reflecting tumor biological characteristics, providing a possibility for the prediction of LVSI [7,8] .Imageomics, as an emerging technology, can extract a large number of quantitative features from medical images with high throughput, mine hidden information in images, and achieve accurate analysis of tumors [9,10] . In EC research, imageomics has been used to predict tumor grading, myometrial infiltration depth, etc., but the prediction research of LVSI is still in the exploratory stage [11] . Some studies have shown that MRI-based radiomics models show certain potential in predicting LVSI [11,12] . However, the results of different studies vary, and the stability and reliability of the models still need to be improved.Deep learning has developed rapidly in the field of medical imaging, with powerful automatic feature extraction and classification capabilities [13] . By building deep neural networks, deep learning can learn complex feature representations from large amounts of image data, and has achieved remarkable results in disease diagnosis, prognosis assessment, etc. [13,14] . In EC research, deep learning has been applied to tumor recognition, staging, and molecular feature prediction [14,15] , but it is still less used in LVSI prediction.The purpose of this study is to explore the value of fusion of multimodal MRI radiomics features and deep learning features in predicting LVSI in patients with EC, and to provide new methods and ideas for preoperative assessment of LVSI by constructing and validating fusion models, so as to help optimize the treatment plan of patients with EC and improve the prognosis of patients. Materials and Methods 1. Research materials This study was a multicenter retrospective study and was approved by the ethics review committee because it was a retrospective study that exempted patients from informed consent. Data were taken from 456 patients with endometrial cancer diagnosed postoperatively at each center between January 2017 and June 2024. Enter the study flow chart below (Fig 1). 1.1 Inclusion criteria: Pathological diagnosis of endometrial carcinoma (including endometrioid carcinoma, mucinous carcinoma, serous carcinoma, clear cell carcinoma, undifferentiated carcinoma, excluding carcinosarcoma); With MMR immunohistochemical results; Complete preoperative pelvic MRI with complete clinical information. MRI showed that the maximum diameter of the tumor was > 0.5 cm. The expected survival period is ≥ 6 months. Age 20-80 years old, no serious heart, liver, kidney function abnormalities. 1.2 Exclusion criteria: Incomplete clinical or imaging data. Received radiotherapy and chemotherapy for endometrial carcinoma before enrollment. Imageomics and deep learning analysis are not possible due to artifacts in MRI images. The researchers determined that they were unfit to participate in the study. 1.3 Clinical data • Basic information: age, BMI (normal according to WHO standards: 18.5 - 24.9; overweight: 25 - 29.9; obesity: ≥ 30), menopausal status (defined as mail carriersopausal age ≥ 50 years), parity (Gravida/Para). Comorbidities: hypertension, diabetes mellitus (according to ICD-11 diagnostic criteria). Molecular markers: P53 (positive: ≥ 10% cell staining), Ki67 (positive: ≥ 20% cell staining), ER/PR (positive: ≥ 1% nuclear staining). FIGO staging: According to the 2023 FIGO guidelines, LVSI-positive individuals without lymph node metastases are upgraded to "stage IIIa (micrometastasis risk) ". 1.4 Pathological diagnosis of lymphatic vascular space invasion (LVSI) in endometrial carcinoma 1.4.1 LVSI pathological diagnostic criteria Morphological criteria: Tumor cells invade the vasculature space, with endothelial cell coating (confirmed by HE staining). Immunohistochemical assistance: CD31/D2-40 positive enhanced specificity; Priority was given to evaluating regions with high Ki67 expression (≥ 30%). Grading and quantification: Grading: Non-invasive/Focal (≤ 3 vessels)/Diffuse (> 3 vessels or extensive infiltration). Spatial distribution: Associated with depth of muscular infiltration, marked with pelvic/paraortic lymph node metastasis. 1.4.2 Pathological parameters of endometrial cancer cases: Histological subtype (WHO 2020 classification), grade (high/medium/low differentiation), depth of muscular infiltration (none/≤ 50%/> 50%), cervical/fallopian tube/ovarian infiltration status (yes/no), lymph node metastasis (pelvic/paraortic). Molecular markers: P53 (positive: ≥ 10% cell staining), Ki67 (positive: ≥ 20% cell staining), ER/PR (positive: ≥ 1% nuclear staining). 1.4.3 Quality control: Data de-identification processing, double-blind entry by two independent researchers, cross-checking, multiple interpolation method for missing values; 10% of slices are randomly selected every quarter for re-inspection, with an error rate of less than 5%; AI-assisted recognition (algorithm sensitivity > 90%). 1.5 Imaging data collection MRI sequence: Preoperative pelvic MRI raw DICOM data, including DWI, T1WI enhancement, and T2WI sequences. Image Archiving System (PACS): Matching images with clinicopathological information (e.g. FIGO staging, postoperative pathological results). Supplementary examination: In some cases, ultrasound or PET-CT were combined to verify the accuracy of MRI in assessing tumor extent and lymph node metastasis. 2. MRI image acquisition of endometrial cancer In this study, 3.0T Siemens MAGNETOM Vida, GE750W, GE Primier Magnetic Resonance Imager was used for examination. MRI scan sequences including T2WI, DWI and T1WI + C were used in this study. Fast spin echo (FSE) sequence or steady state free precession (SSFP) sequence were used in T2WI, single excitation plane echo imaging (SS-EPI) sequence was used in DWI, and fast spin echo (FSE) sequence was used in T1WI + C. MRI image scanning parameters are shown in Table 1. 3. MRI image segmentation Two radiologists experienced in the diagnosis of uterine tumors used ITK-SNAP 3.8.0 to perform semi-automatic ROI delineation of endometrial cancer MRI images on DWI, T1WI + C and T2WI images without knowing the pathological results. First, ensure that the data is complete and the sequence is correct and import it into the software. In the software, the window width and window level adjustment tools are used to optimize the image display, and the endometrium and lesion areas are clearly presented. Select the appropriate delineation tool in the toolbar. For example, when the boundary is clear, you can try the magic wand tool to quickly select, and if the boundary is blurred, use the hand-drawn tool to describe it in detail. The ROI contains the largest facet of the tumor tissue, avoiding surrounding normal tissue, necrotic tissue, and blood. If there are differences in the sketcher, it will be determined by a senior radiologist with more than 20 years of experience to ensure accuracy and reproducibility. 4. LVSI imaging and deep learning construction process for endometrial cancer 4.1 LVSI Imaging Process for Endometrial Cancer The research method of this study used ITK-SNAP to manually plot ROI on DWI, T1WI + C, and T2WI images, and used Onekey AI to extract radiomics features from MRI images in the radiomics library. Subsequently, depth features were obtained by cropping the largest area slice of the region of interest (ROI) from the MRI image. In this study, transfer learning techniques were used to further enrich these features with the help of a pre-trained ResNet101 model. To identify the most robust, non-redundant, and predictive features, correlation filtering and Lasso regression methods were employed in this study. Finally, radiomics feature labels and corresponding radiomics column plots were constructed on an independent validation cohort. The flowchart is as follows (Fig 2) 4.2 radiomics feature extraction Features are divided into three main categories: geometric features, intensity features, and texture features. Geometry features cover the three-dimensional shape properties of a tumor, providing a spatial representation of its structure. Intensity features capture first-order statistical properties of the intensity of endosomal elements in a tumor, helping to understand its internal uniformity or variability. Texture features reflect more complex patterns, depict second-order and higher-order spatial distributions of intensity, and encompass complex spatial relationships between voxels. These texture features are extracted from MRI images by pyradiomics library (http://pyradiomics.readthedocs.io), such as grey release co-occurrence matrix (GLCM), grey release stroke length matrix (GLRLM), grey release size region matrix (GLSZM) and neighborhood grey release difference matrix (NGTDM). A total of 1403 image omics features were extracted in this study: 36 geometric features, 208 intensity features and 1272 texture features. 4.3 Imageomics feature screening and feature selection In this study, all extracted features are normalized using Z-score normalization and evaluated for their statistical significance by a t-test or Mann-Whitney U-test. Only features with p-values below 0.05 are retained. To mitigate the collinearity problem, this study uses Pearson correlation coefficients to check for correlations between features, excluding one feature in pairs with correlation coefficients higher than 0.9. The feature set is further optimized by Lasso regression within the 10-fold cross-validation framework to determine the optimal regularization parameter lambda, thus effectively reducing the feature set to the most predictive and informative features. The ROI curve was used to evaluate the performance of 10 models such as LR, SVM, KNN, and RandomForest on the training dataset (lab-train) and test set (lab-test), including Accuracy (accuracy), AUC, 95% CI (confidence interval), Sensitivity (sensitivity), and Specificity (specificity). Select a better model to evaluate the performance of radiomics. 4.4 deep learning process of LVSI for endometrial cancer 4.4.1 Data preparation ROI Cropping: This study selects slices that present the greatest region of interest (ROI) as representative images. Data Enhancement: The approach of this study involves normalizing the intensity distribution by Z-score normalization of the RGB channels of the images. These normalized images are then used as inputs to the model. During the training phase, this study implements real-time data enhancement strategies including random cropping, horizontal flipping, and vertical flipping. For test images, only normalization processing is performed. 4.4.2 Training Model Transfer learning: This study assesses ResNet101 and CNN - based models. ResNet101, a deep residual network, overcomes layer - related issues in traditional neural networks with its residual connections. These connections enable direct information flow between layers, allowing for training of very deep networks. ResNet101 performs well in image classification, object detection, and semantic segmentation, advancing computer vision research. Hyperparameters: Transfer learning was used to adapt the model to different patient groups. The model was initialized with ImageNet pre - trained weights. A cosine annealing learning rate strategy, ηt = η mini+21( η maxi− η mini)(1+cos( Ti T cur π ))， η mini=0 The minimum learning rate is set， η maxi=0.01 The maximum learning rate is set， Ti =30 Represents the number of epochs during iterative training。Other important hyperparameters include using random layer descent (SGD) as the optimizer and using softmax cross entropy as the loss function. 4.4.3 Feature Label Model Construction Imageomics Feature Models : Using LASSO for feature selection, this study built imageomics risk models with algorithms like logistic regression, SVMs, and random forest. It compared model performances and explored feature fusion from multi - modal imaging to enhance prediction accuracy. Deep Learning Feature Models : CNN - computed output probabilities served as deep learning feature labels. Deep Learning radiomics Feature Model : The CNN model was the best in the test set. Features from its penultimate layer were extracted, PCA - reduced to 512 dimensions, and then combined with radiomics features via a pre - fusion algorithm to form DLR feature labels, followed by feature selection and model building. Combinatorial Feature Model : After early feature fusion and manual stitching, the study conducted feature selection and model construction similar to imageomics. Clinical Relevance Enhancement : Univariate and stepwise multivariate analyses identified key clinical features. These, combined with deep learning model predictions, formed an LR linear model and a combined feature label, which was visualized as a line diagram. 4.4.4 Statistical analysis: Data analytics were done on Onekey AI platform (v4.9.1) using Python 3.7.12. Specific library versions were used: Statsmodels 0.13.2 for stats, PyRadiomics 3.0.1 for imageomics, Scikit - learn 1.0.2 for machine learning, and a PyTorch 1.11.0 - based deep learning framework with CUDA and cuDNN. Python and statsmodels were for analysis, scikit - learn for machine learning model, and deep learning trained on NVIDIA 4090 GPUs with MONAI and PyTorch. The Shapiro - Wilk test checked clinical feature normality. Appropriate tests were used for different variable types. A p - value > 0.05 in Table 2 ensured unbiased groupings. In the test cohort, ROC curves evaluated model discriminative power, calibration curves with Hosmer - Lemeshow test analyzed calibration, and DCA evaluated clinical utility. Results 1. Clinical characteristics of LVSI in endometrial carcinoma 1.1 Baseline characteristics of clinical alignment A total of 456 patients with multicenter intrauterine intimacy were included, and a total of 148 patients were excluded according to the exclusion criteria, namely: Incomplete clinical or imaging data (n = 79); Received radiotherapy and chemotherapy for endometrial cancer before enrollment (n = 42); The MRI image contained artifacts (n = 27). Finally 308 met our study criteria. LVSI+(n=69),LVSI-(n=239). The training (215 cases) and test (93 cases) sets of 308 endometrial cancer patients were well - balanced in core clinicopathological features. (Table 2) The following is a clinical feature correlation matrix showing the correlation between the four variables "Myometrial_invasion (myometrial infiltration) ", "Histopathologic_subtype (histopathological subtype) ", "Para_aortic_lymph_node_metastasis (paraaortic lymph node metastasis) " and "label" (Fig 3-a) . The diagonal in the correlation figure is 1 by definition. Non - diagonal values, ranging from 0.126 - 0.382, show weak positive correlations between variables. Myometrial_invasion and Histopathologic_subtype have the lowest correlation coefficient (0.126), indicating the weakest link. This visualization helps preliminarily analyze variable correlations for future research. 1.2 LVSI univariate and multivariate analysis of endometrial cancer This study conducted univariate and multivariate analyses of endometrial cancer. Univariate analysis showed that progesterone receptor, estrogen receptor, etc., were significantly associated with reduced risk, while para - aortic lymph node metastasis was the only risk factor. Multivariate analysis identified three independent prognostic indicators: histological subtype, deep myometrial infiltration, and para - aortic lymph node metastasis, with OR > 4, indicating high clinical value. The strong negative correlation of PR and ER aligns with hormone receptor typing, and Ki67 and P53 confirm the model's rationality. Metabolic factors and FIGO staging were not significant. The overall data constructed a highly operable model, offering a reliable basis for personalized diagnosis and treatment. 2. Comparison of LVSI radiomics and deep learning radiomics feature fusion weights for endometrial cancer In the radiomics feature model, wavelet_LHL_firstorder_Kurtosis_1 (coefficient 0.071) has the most prominent positive effect and significantly improves the label; log_sigma_5_0_mm_3D_firstorder_RootMean Squared_0 (coefficient -0.062) has the strongest negative effect and suppresses the label value. In addition, wavelet_HLH_ firstorder_Skewness_1 (0.0336), log_sigma_5_0_mm_3D_ngtdm_Busyness_2 (0.0247) and other features also have obvious positive contributions, reflecting the key role of image texture and grey release statistical characteristics on the model, (Fig 3-b). In the fusion coefficient analysis of radiomics and deep learning radiomics features, wavelet_LHL_firstorder-_ Kurtosis_1 (coefficient 0.0715) had the most prominent positive effect in the model, significantly improving the label value; log_sigma_5_0_mm_ 3D_firstorder_RootMeanSquared_0 (coefficient -0.0641) had the strongest negative effect and suppressed the label value. In addition, log_sigma_5_0_mm_3D_ngtdm_Busyness_2 (0.0304), wavelet_HLL_glszm_ SmallAreaHighGrayLevelEmphasis _0 (0.0228) and other features also made significant positive contributions, reflecting the key role of image texture and grey release statistical characteristics on the model. (Fig 3-c). 3. Endometrial cancer LVSI deep learning imaging model feature label results The performance of 8 models such as LR, SVM, KNN, and RandomForest on the training dataset (lab-train) and test set (lab-test) includes Accuracy (accuracy), AUC, 95% CI (confidence interval), Sensitivity (sensitivity), Specificity (specificity) and other indicators. The accuracy of the XGBoost training dataset reached 0.991, the AUC was 0.999, but the accuracy of the test set was 0.613; most models showed that the training dataset was better than the test set. As shown in Table 4, Fig 3-d, Fig3-e show the corresponding generated ROC curves. The ROC curve shows that in the training set, the performance of each model is significantly differentiated. The XGBoost model performs the best, with an AUC of 0.999 (95% CI 0.998 - 1.000), which is almost close to the ideal ROC curve, and the sensitivity and specificity are highly balanced. The SVM (AUC 0.985), LR (AUC 0.948) and other model curves are also significantly better than the diagonal, which fully verifies the strong classification ability of deep learning radiomics features in the training data, indicating that the model fits the training dataset data well (Fig 3-d). In the test set, although the overall performance was weakened, the LR model test set AUC reached 0.704 (95% CI 0.578 - 0.829), which still maintained good discrimination in independent data, providing data support for model selection; other models such as SVM (AUC 0.672) also showed basic generalization potential. (Fig 3-e). 4. Comparison of LVSI clinical characteristics, radiomics characteristics, deep learning radiomics characteristics, and deep learning fusion tags for endometrial cancer The results showed that the ROC curve of the training cohort: Clinic AUC: 0.799 (95% CI: 0.770 - 0.821); Imageomics Feature Label (Rad) AUC: 0.846 (95% CI: 0.827 - 0.865); deep learning feature label (DTL) AUC: 0.668 (95% CI: 0.642 - 0.695); deep learning radiomics feature label (DLR) AUC: 0.948 (95% CI: 0.938 - 0.957); combined feature label (Combine) AUC: 0.953 (95% CI: 0.943 - 0.962). (Table 5, Fig 3-f) Test cohort ROC curve: Clinic AUC: 0.675 (95% CI: 0.631 - 0.718); deep learning radiomics feature label (DLR) AUC: 0.704 (95% CI: 0.663 - 0.745); radiomics feature label (Rad) AUC: 0.687 (95% CI: 0.640 - 0.733); deep learning feature label (DTL) AUC: 0.654 (95% CI: 0.602 - 0.707); combined feature label (Combine) AUC: 0.720 (95% CI: 0.681 - 0.759). (Table 5, Fig 3- g). 4.1 ROC curve analysis of training dataset Training dataset: Representing Clinic, Rad, DTL, DLR, and Combined models, with their corresponding area under the curve (AUC) and 95% confidence interval (CI) next to them.Among them, Combine has the highest AUC value, 0.953, and the 95% CI is 0.943-0.962, indicating that it has the strongest ability to distinguish between positive and negative cases on the training dataset; followed by DLR, whose AUC is 0.948; and DTL, whose AUC is lower, 0.668. (Fig 3-f) 4.2 ROC curve analysis of test set The ROC curves of the five models or indicators for predicting endometrial cancer in the test set and the corresponding AUC values and 95% CI. On the test set, the combined AUC value was 0.720, and the 95% CI was 0.681-0.759, which was still relatively high among several indicators; the DLR's AUC was 0.704; the DTL's AUC was the lowest, 0.654. (Fig 3-g) 4.3 Calibration curves of different LVSI models for endometrial cancer The Hosmer-Lemeshow (HL) test is used to quantify the difference between the predicted probability and the observed results, generating calibration graphs (Fig 3-h and 3-i). 4.3.1 Training dataset calibration curve (Cohort train Calibration): The training dataset calibration curve shows that most models match the predicted probability with the actual positive frequency well. Among them, the Combined and DLR models perform best. In the prediction probability interval of 0.2-0.8, the actual positive frequency is highly consistent with the predicted probability. For example, when the prediction probability is 0.4, the actual frequency is close to 0.4; when the prediction probability is 0.6, the actual frequency is stable at 0.58-0.62. Although Clinic, Rad and other models have slight deviations, the overall trend is close to the ideal line. (Fig 3-h). 4.3.2 Test set calibration curve (Cohort test Calibration): In the test set calibration curve, the actual positive frequency of the Combined model is closer to the ideal line in the prediction probability range of 0.4-0.6, which shows the rationality of the evaluation of this risk segment; the DLR model also shows the actual positive frequency consistent with the prediction probability at some probability points (such as 0.3-0.4), reflecting a certain calibration potential.(Fig 3-i). 5. DeLong test This "Cohort test Delong" graph presents the comparisons among different models on the test set. The Combined model stands out with extremely small P - values when compared with most models, indicating significantly superior performance, and it can be used as the main applied model. The Clinic model has differences in performance from some models and can leverage its own advantages in combination with clinical practice.(Fig 3-j) 6. Clinical application of LVSI in endometrial cancer: decision curve analysis (DCA) Two DCA graphs show different models' performance in training and test sets(Fig 4-a, Fig 4-b). In the training set, for 0 - 0.2, models are close to the perfect calibration line with similar performance; for 0.2 - 0.8, Combined and DLR excel, offering high net benefits, and Clinic and DTL also work; for 0.8 - 1.0, models near the line again. In the test set, 0 - 0.2 shows deviation. For 0.2 - 0.6, Combined still has good net benefits in parts, though reduced compared to the training set. For 0.6 - 1.0, models fluctuate greatly, yet this indicates improvement directions, and generalization ability can be enhanced. 7. Endometrial cancer LVSI nomogram The Nomogram uses Myometrial_invasion (myometrial infiltration), Histopathologic_subtype (histopathological subtype), Para_aortic_lymph_node_metastasis (paraaortic lymph node metastasis) and DLR as variables to assign corresponding Points scores to each characteristic and different states; by accumulating Total Points, it is finally mapped to the Risk axis to realize the visual prediction of risk probability (Fig 5). Discussion This study constructed multimodal MRI imaging, deep learning, and feature extraction fusion models for predicting lymphovascular space invasion (LVSI) in endometrial cancer. The LR model showed the best performance in LVSI prediction on the training dataset and test set. Combining clinical and radiomics features using the LR algorithm in column plots had excellent performance. The AUC of the deep learning hybrid feature model on the training dataset and test set and the AUC of the hybrid model on the training dataset and test set are consistent with previous studies showing that multimodal data fusion can capture information on tumor heterogeneity [15] . Compared with Liu et al.'s study [17] , although the training performance of our combined model was slightly highe, the test set AUC in our study was lower. The possible reason is that deep learning in our study can mine deep associations in multimodal data, but it has high requirements for data homogeneity, and data collection bias and sample heterogeneity in the test set challenge its generalization ability. We explored the ResNet101 network to enhance the performance of traditional CNN models and compared different models. Feature extraction and selection in this study ensured feature effectiveness. The good performance of the independent LR model may be related to its complex data processing ability. radiomics lineup diagrams are useful for clinics, yet the study has limitations such as a small sample size affecting generalization and subjectivity in ROI sketching causing errors. Radiomicsfeature selection is crucial. Texture features (e.g., texture complexity, contrast, uniformity) can reflect tumor tissue heterogeneity and cell structure complexity related to LVSI [18, 19] . These features are closely related to tumor aggressiveness, such as high contrast texture suggesting internal structural disorder of tumor, and long-run emphasis reflecting abnormal cell arrangement [20] . Shape features (e.g., tumor area, perimeter) can reflect tumor growth and invasiveness [21] . We selected various Radiomicsfeatures including first - order statistical, texture, and shape features, which can reflect tumor microstructures and biological characteristics from different angles [22] . First - order statistical features mainly describe the distribution of image grey release values, such as mean, standard deviation, skewness, and kurtosis. These features can reflect the density and uniformity of tumor tissue, and may have a certain relationship with the occurrence of LVSI [23] . Some features like angular second - order moment in texture features and maximum diameter in shape features are significantly correlated with LVSI [24, 25] . Deep learning models, through attention mechanisms like GAM modules, focus on tumor margins and vascular structures, identifying microscopic infiltrations [26] . Our innovation is the "complementary feature fusion" strategy, combining radiomics interpretability features with deep learning abstract features. The experimental results show that the fusion model has a slightly higher test - set AUC (0.720) than the individual radiomics model (0.687) and the deep learning feature fusion model (0.704), possibly due to the complementarity of the two types of features [27, 28] . Research limitations and prospects This study has the following limitations This study, though insightful, has areas for improvement.Study data—clinical, pathological, and imaging—show heterogeneity. Tumor - related imaging data and subjectivity in LVSI evaluation of pathological sections may affect finding precision and reliability [27 - 29] .The model has good predictive performance, yet its biological mechanism is unclear. Further study with gene expression or immunohistochemical data is needed [28 - 30] .The multimodal MRI radiomics - deep - learning model is complex, requiring professional skills for operation. This complexity could impede its clinical use. Conclusion The fusion model of multimodal MRI radiomics and deep learning features has shown good advantages in EC-LVSI prediction, providing a new technical path for precise risk stratification before surgery. Declarations Author Contribution Y. T. participated in the study design, data collection and manuscript writing. F.F. S. conducted a critical review of the manuscript. Y. T.,F.F. S. authors have contributed equally to this work .W. Z. participated in the data collection and data analytics. K. W. assisted in the collection and collation of references. Y. C. participated in data collection and typesetting. J.Q. S. (corresponding author) supervised the entire research project, participated in the study design, case collection and was responsible for the final approval of the manuscript. All authors read and approved the final version of the manuscript. References Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin . 2024;74(3):229-263. doi:10.3322/caac.21834 Bosse T, Peters EE, Creutzberg CL, et al. Substantial lymph-vascular space invasion (LVSI) is a significant risk factor for recurrence in endometrial cancer--A pooled analysis of PORTEC 1 and 2 trials. Eur J Cancer . 2015;51(13):1742-1750. doi:10.1016/j.ejca.2015.05.015 Guntupalli SR, Zighelboim I, Kizer NT, et al. Lymphovascular space invasion is an independent risk factor for nodal disease and poor outcomes in endometrioid endometrial cancer. Gynecol Oncol . 2012;124(1):31-35. doi:10.1016/j.ygyno.2011.09.017 Liu D, Huang J, Zhang Y, et al. Multimodal MRI-based radiomics models for the preoperative prediction of lymphovascular space invasion of endometrial carcinoma. BMC Med Imaging . 2024;24(1):252. Published 2024 Sep 20. doi:10.1186/s12880-024-01430-1 Ironi G, Mapelli P, Bergamini A, et al. Hybrid PET/MRI in Staging Endometrial Cancer: Diagnostic and Predictive Value in a Prospective Cohort. Clin Nucl Med . 2022;47(3):e221-e229. doi:10.1097/RLU.0000000000004064 Saleh GA, Abdelrazek R, Hassan A, Hamdy O, Tantawy MSI. Diagnostic utility of apparent diffusion coefficient in preoperative assessment of endometrial cancer: are we ready for the 2023 FIGO staging?. BMC Med Imaging . 2024;24(1):226. Published 2024 Aug 28. doi:10.1186/s12880-024-01391-5 Zhang Q, Ouyang H, Ye F, et al. Multiple mathematical models of diffusion-weighted imaging for endometrial cancer characterization: Correlation with prognosis-related risk factors. Eur J Radiol . 2020;130:109102.doi:10.1016/j.ejrad.2020.109102. Li S, Wang Y, Sun Y, et al. Both intra- and peri-tumoral radiomics signatures can be used to predict lymphatic vascular space invasion and lymphatic metastasis positive status from endometrial cancer MR imaging. Abdom Radiol (NY) . 2024;49(11):4140-4150. doi:10.1007/s00261-024-04432-3 Yan B, Jia Y, Li Z, et al. Preoperative prediction of lymphovascular space invasion in endometrioid adenocarcinoma: an MRI-based radiomics nomogram with consideration of the peritumoral region. Acta Radiol . 2023;64(9):2636-2645. doi:10.1177/02841851231181681 Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer . 2012;48(4):441-446. doi:10.1016/j.ejca.2011.11.036 Ma C, Zhao Y, Song Q, et al. Multi-parametric MRI-based radiomics for preoperative prediction of multiple biological characteristics in endometrial cancer. Front Oncol . 2023;13:1280022. Published 2023 Dec 15. doi:10.3389/fonc.2023.1280022 Jin X, Yan R, Li Z, et al. Evaluation of Amide Proton Transfer-Weighted Imaging for Risk Factors in Stage I Endometrial Cancer: A Comparison With Diffusion-Weighted Imaging and Diffusion Kurtosis Imaging. Front Oncol . 2022;12:876120. Published 2022 Apr 14. doi:10.3389/fonc.2022.876120 Wang Y, Liu W, Lu Y, et al. Fully Automated Identification of Lymph Node Metastases and Lymphovascular Invasion in Endometrial Cancer From Multi-Parametric MRI by Deep Learning. J Magn Reson Imaging . 2024;60(6):2730-2742. doi:10.1002/jmri.29344 Jiang X, Zhai W, Song J, et al. Associations between MRI radiomic phenotypes and clinical outcomes in endometrial cancer: Implications for preoperative risk stratification. Magn Reson Imaging . 2025;117:110298. doi:10.1016/j.mri.2024.110298 Zhang Y, Chen S, Wang Y, et al. Deep learning-based methods for classification of microsatellite instability in endometrial cancer from HE-stained pathological images. J Cancer Res Clin Oncol . 2023;149(11):8877-8888. doi:10.1007/s00432-023-04838-4 Zhang Y, Chen S, Wang Y, et al. Deep learning-based methods for classification of microsatellite instability in endometrial cancer from HE-stained pathological images. J Cancer Res Clin Oncol . 2023;149(11):8877-8888. doi:10.1007/s00432-023-04838-4 Liu D, Huang J, Zhang Y, et al. Multimodal MRI-based radiomics models for the preoperative prediction of lymphovascular space invasion of endometrial carcinoma. BMC Med Imaging . 2024;24(1):252. Published 2024 Sep 20. doi:10.1186/s12880-024-01430-1 Yan BC, Li Y, Ma FH, et al. Radiologists with MRI-based radiomics aids to predict the pelvic lymph node metastasis in endometrial cancer: a multicenter study. Eur Radiol . 2021;31(1):411-422. doi:10.1007/s00330-020-07099-8 Mao W, Chen C, Gao H, Xiong L, Lin Y. A deep learning-based automatic staging method for early endometrial cancer on MRI images. Front Physiol . 2022;13:974245. Published 2022 Aug 30. doi:10.3389/fphys.2022.974245 Liu XF, Yan BC, Li Y, Ma FH, Qiang JW. Radiomics feature as a preoperative predictive of lymphovascular invasion in early-stage endometrial cancer: A multicenter study. Front Oncol . 2022;12:966529. Published 2022 Aug 18. doi:10.3389/fonc.2022.966529 Bonatti M, Pedrinolla B, Cybulski AJ, et al. Prediction of histological grade of endometrial cancer by means of MRI. Eur J Radiol . 2018;103:44-50. doi:10.1016/j.ejrad.2018.04.008 Liu XF, Yan BC, Li Y, Ma FH, Qiang JW. Radiomics Nomogram in Assisting Lymphadenectomy Decisions by Predicting Lymph Node Metastasis in Early-Stage Endometrial Cancer. Front Oncol . 2022;12:894918. Published 2022 May 31. doi:10.3389/fonc.2022.894918 Feng M, Zhao Y, Chen J, et al. A deep learning model for lymph node metastasis prediction based on digital histopathological images of primary endometrial cancer. Quant Imaging Med Surg . 2023;13(3):1899-1913. doi:10.21037/qims-22-220 Coada CA, Santoro M, Zybin V, et al. A Radiomic-Based Machine Learning Model Predicts Endometrial Cancer Recurrence Using Preoperative CT Radiomic Features: A Pilot Study. Cancers (Basel) . 2023;15(18):4534. Published 2023 Sep 13. doi:10.3390/cancers15184534 Meng X, Zhang X, Tian S, et al. Evaluation of lymphovascular space invasion in endometrial carcinoma by APTw and mDixon-Quant. Acta Radiol . 2024;65(11):1440-1446. doi:10.1177/02841851241277339 Hodneland E, Dybvik JA, Wagner-Larsen KS, et al. Automated segmentation of endometrial cancer on MR images using deep learning. Sci Rep. 2021;11(1):179. Published 2021 Jan 8. doi:10.1038/s41598-020-80068-9 Takahashi Y, Sone K, Noda K, et al. Automated system for diagnosing endometrial cancer by adopting deep-learning technology in hysteroscopy. PLoS One . 2021;16(3):e0248526. Published 2021 Mar 31. doi:10.1371/journal.pone.0248526 Dobrzycka B, Terlikowska KM, Kowalczuk O, Niklinski J, Kinalski M, Terlikowski SJ. Prognosis of Stage I Endometrial Cancer According to the FIGO 2023 Classification Taking into Account Molecular Changes. Cancers (Basel) . 2024;16(2):390. Published 2024 Jan 17. doi:10.3390/cancers16020390 Bosse T, Peters EE, Creutzberg CL, et al. Substantial lymph-vascular space invasion (LVSI) is a significant risk factor for recurrence in endometrial cancer--A pooled analysis of PORTEC 1 and 2 trials. Eur J Cancer . 2015;51(13):1742-1750. doi:10.1016/j.ejca.2015.05.015 Fasmer KE, Hodneland E, Dybvik JA, et al. Whole-Volume Tumor MRI Radiomics for Prognostic Modeling in Endometrial Cancer. J Magn Reson Imaging . 2021;53(3):928-937. doi:10.1002/jmri.27444 Tables Tables 1 to 5 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1.docx Table2.docx Table3.docx Table4.docx Table5.docx Listofacronyms.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-6281099","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":444557736,"identity":"96a628d4-2d63-4049-94c6-f7851d3a8d65","order_by":0,"name":"Yuan Tang","email":"","orcid":"","institution":"Department of Radiology, Affiliated Yuebei People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yuan","middleName":"","lastName":"Tang","suffix":""},{"id":444557737,"identity":"d654f71d-7aed-459d-9310-377e059e5ec1","order_by":1,"name":"Feifei Shan","email":"","orcid":"","institution":"Ultrasound Diagnosis Department, Affiliated Yuebei People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Feifei","middleName":"","lastName":"Shan","suffix":""},{"id":444557740,"identity":"0430ed32-7f76-4521-976c-097f86634a98","order_by":2,"name":"Wei Zhou","email":"","orcid":"","institution":"Department of Radiology, Affiliated Yuebei People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Zhou","suffix":""},{"id":444557746,"identity":"12c3ce2c-7140-45d9-ab28-977bf55e7938","order_by":3,"name":"Kang Wang","email":"","orcid":"","institution":"Department of Radiology, Shaoguan First People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Kang","middleName":"","lastName":"Wang","suffix":""},{"id":444557750,"identity":"8de21751-67e3-422d-967f-4a88fa2789a7","order_by":4,"name":"Yuan Chen","email":"","orcid":"","institution":"Department of Gynecology, Jiangmen Central Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yuan","middleName":"","lastName":"Chen","suffix":""},{"id":444557752,"identity":"da489be6-adf1-46c9-b18a-4efae620f438","order_by":5,"name":"Junqi Sun","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/0lEQVRIiWNgGAWjYJACAyDmYWNvP/zgQ4WNHAMzsVr4eM6kGc44k2ZMlBYwkJNIMJDmbDmc2EBIpfyM3APFvDusZdgYEhKMGRvS0ue38x78wFBjE43TUTfyEox5z6TzsDEcPPC4cIdN7obDfMkSDMfScnFZZyCRY2DM23aYh42xIcF45pm03A3MPAYSjA2HcWqRnwHTwsxgIA1kpMs38xj/wKeF4QZMCxtESwLDYR4zvLYYnHljYDi3DegXHh5wIBtuAGqxSMDjF/n2HDODt23W9vLzn4OjUl6+/4zxjQ81NrgdxsDABoxK9OhLwK0cBJgfYGoZBaNgFIyCUYAEADvaVMNRuYqhAAAAAElFTkSuQmCC","orcid":"","institution":"Department of Radiology, Affiliated Yuebei People's Hospital","correspondingAuthor":true,"prefix":"","firstName":"Junqi","middleName":"","lastName":"Sun","suffix":""}],"badges":[],"createdAt":"2025-03-22 03:08:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6281099/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6281099/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":81022666,"identity":"f06ee0a7-2e45-432a-a59e-565ce3304df8","added_by":"auto","created_at":"2025-04-21 10:03:00","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":91851,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version\u003c/p\u003e","description":"","filename":"Fig1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/f97ee52ca1f07a13cbde9f3f.jpg"},{"id":81022152,"identity":"40df5dd8-6c9b-40cd-848e-6712e43617fb","added_by":"auto","created_at":"2025-04-21 09:55:00","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":410369,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version\u003c/p\u003e","description":"","filename":"Fig2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/95fc29c5ff05f6757bf77737.jpg"},{"id":81022667,"identity":"c41aa7f1-aac8-4050-a418-903053f141a4","added_by":"auto","created_at":"2025-04-21 10:03:01","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1336645,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version\u003c/p\u003e","description":"","filename":"Fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/08d4b972fb05a0ccd4100e18.jpg"},{"id":81024040,"identity":"72b3fca3-1d04-4be5-9a4f-a6294d06618a","added_by":"auto","created_at":"2025-04-21 10:11:01","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":410603,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version\u003c/p\u003e","description":"","filename":"Fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/2e563617718ae918e031974d.jpg"},{"id":81024038,"identity":"89130f75-650f-4beb-a5db-e4366a561bb2","added_by":"auto","created_at":"2025-04-21 10:11:01","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":58456,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version\u003c/p\u003e","description":"","filename":"Fig5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/d37596e479040224c3304f8c.jpg"},{"id":81767556,"identity":"9e4056c9-d01a-4294-87e0-669c42869523","added_by":"auto","created_at":"2025-05-01 13:16:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3536509,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/db69426e-267d-4843-9923-69e48a95cc8a.pdf"},{"id":81022149,"identity":"a495ceff-93f2-4e6b-a379-82d323f23d10","added_by":"auto","created_at":"2025-04-21 09:55:00","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":17665,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/18186d385d79400bb0b445e3.docx"},{"id":81022665,"identity":"1c4fad03-8276-4e61-8937-f10509e63632","added_by":"auto","created_at":"2025-04-21 10:03:00","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":20201,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.docx","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/f3c5f2be1a09b1782a05852e.docx"},{"id":81022154,"identity":"313bb7e5-f1f1-4dc7-ae83-ebfc60a35c84","added_by":"auto","created_at":"2025-04-21 09:55:00","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":20269,"visible":true,"origin":"","legend":"","description":"","filename":"Table3.docx","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/222d9ea654166ab72346cff5.docx"},{"id":81022155,"identity":"e03ae93a-8864-4095-8957-2a8db8af5fc7","added_by":"auto","created_at":"2025-04-21 09:55:00","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":19868,"visible":true,"origin":"","legend":"","description":"","filename":"Table4.docx","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/9c4f7f55bbcd757e0fdabde5.docx"},{"id":81022160,"identity":"db319acf-0c2e-4072-adee-f918a446f713","added_by":"auto","created_at":"2025-04-21 09:55:01","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":18613,"visible":true,"origin":"","legend":"","description":"","filename":"Table5.docx","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/bc502dd236e9f218e1e32d5e.docx"},{"id":81022674,"identity":"99a5e0e2-3094-40cc-8be6-9089f3966ca1","added_by":"auto","created_at":"2025-04-21 10:03:01","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":18371,"visible":true,"origin":"","legend":"","description":"","filename":"Listofacronyms.docx","url":"https://assets-eu.researchsquare.com/files/rs-6281099/v1/c25748f3b890655f6be241de.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multimodal MRI Radiomics Features and Deep Learning Features in Predicting Lymphovascular Space Invasion in Endometrial Carcinoma","fulltext":[{"header":"Introduction","content":"\u003cp\u003eEndometrial carcinoma (EC) is one of the most common malignancies of the female reproductive system worldwide. In 2022, the number of new cases worldwide reached 420,242, accounting for 2.1% of the cancer incidence. It is more common in female cancers. From the overall trend of global cancer incidence, it is speculated that the number of cases may increase due to population growth, aging and lifestyle changes\u003csup\u003e[1]\u003c/sup\u003e. Lymphatic vascular space infiltration (LVSI) is an important pathological feature of EC, which is closely related to tumor stage, lymph node metastasis and patient prognosis\u003csup\u003e[2,3]\u003c/sup\u003e. Preoperative accurate prediction of LVSI is of great significance for formulating reasonable treatment plans and evaluating patient prognosis\u003csup\u003e[4]\u003c/sup\u003e.At present, surgical pathology is the \"gold standard\" for the diagnosis of LVSI, but this method is invasive and cannot provide a timely basis for preoperative clinical decision-making\u003csup\u003e[6]\u003c/sup\u003e. Magnetic resonance imaging (MRI) plays an important role in the preoperative evaluation of EC due to its high soft tissue resolution and multi-parameter imaging capabilities\u003csup\u003e[5]\u003c/sup\u003e. Multimodal MRI can not only clearly display the morphology, location and degree of tumor invasion, but also provide functional information reflecting tumor biological characteristics, providing a possibility for the prediction of LVSI\u003csup\u003e[7,8]\u003c/sup\u003e.Imageomics, as an emerging technology, can extract a large number of quantitative features from medical images with high throughput, mine hidden information in images, and achieve accurate analysis of tumors\u003csup\u003e[9,10]\u003c/sup\u003e. In EC research, imageomics has been used to predict tumor grading, myometrial infiltration depth, etc., but the prediction research of LVSI is still in the exploratory stage\u003csup\u003e[11]\u003c/sup\u003e. Some studies have shown that MRI-based radiomics models show certain potential in predicting LVSI\u003csup\u003e[11,12]\u003c/sup\u003e. However, the results of different studies vary, and the stability and reliability of the models still need to be improved.Deep learning has developed rapidly in the field of medical imaging, with powerful automatic feature extraction and classification capabilities\u003csup\u003e[13]\u003c/sup\u003e. By building deep neural networks, deep learning can learn complex feature representations from large amounts of image data, and has achieved remarkable results in disease diagnosis, prognosis assessment, etc. \u003csup\u003e[13,14]\u003c/sup\u003e. In EC research, deep learning has been applied to tumor recognition, staging, and molecular feature prediction\u003csup\u003e[14,15]\u003c/sup\u003e, but it is still less used in LVSI prediction.The purpose of this study is to explore the value of fusion of multimodal MRI radiomics features and deep learning features in predicting LVSI in patients with EC, and to provide new methods and ideas for preoperative assessment of LVSI by constructing and validating fusion models, so as to help optimize the treatment plan of patients with EC and improve the prognosis of patients.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e1. Research materials\u003c/p\u003e\n\u003cp\u003eThis study was a multicenter retrospective study and was approved by the ethics review committee because it was a retrospective study that exempted patients from informed consent. Data were taken from 456 patients with endometrial cancer diagnosed postoperatively at each center between January 2017 and June 2024. Enter the study flow chart below (Fig 1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.1 Inclusion criteria:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePathological diagnosis of endometrial carcinoma (including endometrioid carcinoma, mucinous carcinoma, serous carcinoma, clear cell carcinoma, undifferentiated carcinoma, excluding carcinosarcoma);\u003c/p\u003e\n\u003cp\u003eWith MMR immunohistochemical results;\u003c/p\u003e\n\u003cp\u003eComplete preoperative pelvic MRI with complete clinical information.\u003c/p\u003e\n\u003cp\u003eMRI showed that the maximum diameter of the tumor was \u0026gt; 0.5 cm.\u003c/p\u003e\n\u003cp\u003eThe expected survival period is ≥ 6 months.\u003c/p\u003e\n\u003cp\u003eAge 20-80 years old, no serious heart, liver, kidney function abnormalities.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.2 Exclusion criteria:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIncomplete clinical or imaging data.\u003c/p\u003e\n\u003cp\u003eReceived radiotherapy and chemotherapy for endometrial carcinoma before enrollment.\u003c/p\u003e\n\u003cp\u003eImageomics and deep learning analysis are not possible due to artifacts in MRI images.\u003c/p\u003e\n\u003cp\u003eThe researchers determined that they were unfit to participate in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.3 Clinical data\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e• Basic information: age, BMI (normal according to WHO standards: 18.5 - 24.9; overweight: 25 - 29.9; obesity: ≥ 30), menopausal status (defined as mail carriersopausal age ≥ 50 years), parity (Gravida/Para).\u003c/p\u003e\n\u003cp\u003eComorbidities: hypertension, diabetes mellitus (according to ICD-11 diagnostic criteria).\u003c/p\u003e\n\u003cp\u003eMolecular markers: P53 (positive: ≥ 10% cell staining), Ki67 (positive: ≥ 20% cell staining), ER/PR (positive: ≥ 1% nuclear staining).\u003c/p\u003e\n\u003cp\u003eFIGO staging: According to the 2023 FIGO guidelines, LVSI-positive individuals without lymph node metastases are upgraded to \"stage IIIa (micrometastasis risk) \".\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.4 Pathological diagnosis of lymphatic vascular space invasion (LVSI) in endometrial carcinoma\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.4.1 LVSI pathological diagnostic criteria\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMorphological criteria: Tumor cells invade the vasculature space, with endothelial cell coating (confirmed by HE staining). Immunohistochemical assistance: CD31/D2-40 positive enhanced specificity; Priority was given to evaluating regions with high Ki67 expression (≥ 30%). Grading and quantification: Grading: Non-invasive/Focal (≤ 3 vessels)/Diffuse (\u0026gt; 3 vessels or extensive infiltration). Spatial distribution: Associated with depth of muscular infiltration, marked with pelvic/paraortic lymph node metastasis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.4.2 Pathological parameters of endometrial cancer cases:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHistological subtype (WHO 2020 classification), grade (high/medium/low differentiation), depth of muscular infiltration (none/≤ 50%/\u0026gt; 50%), cervical/fallopian tube/ovarian infiltration status (yes/no), lymph node metastasis (pelvic/paraortic).\u003c/p\u003e\n\u003cp\u003eMolecular markers: P53 (positive: ≥ 10% cell staining), Ki67 (positive: ≥ 20% cell staining), ER/PR (positive: ≥ 1% nuclear staining).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.4.3 Quality control:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData de-identification processing, double-blind entry by two independent researchers, cross-checking, multiple interpolation method for missing values; 10% of slices are randomly selected every quarter for re-inspection, with an error rate of less than 5%; AI-assisted recognition (algorithm sensitivity \u0026gt; 90%).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.5 Imaging data collection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMRI sequence: Preoperative pelvic MRI raw DICOM data, including DWI, T1WI enhancement, and T2WI sequences.\u003c/p\u003e\n\u003cp\u003eImage Archiving System (PACS): Matching images with clinicopathological information (e.g. FIGO staging, postoperative pathological results).\u003c/p\u003e\n\u003cp\u003eSupplementary examination: In some cases, ultrasound or PET-CT were combined to verify the accuracy of MRI in assessing tumor extent and lymph node metastasis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. MRI image acquisition of endometrial cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, 3.0T Siemens MAGNETOM Vida, GE750W, GE Primier Magnetic Resonance Imager was used for examination. MRI scan sequences including T2WI, DWI and T1WI + C were used in this study. Fast spin echo (FSE) sequence or steady state free precession (SSFP) sequence were used in T2WI, single excitation plane echo imaging (SS-EPI) sequence was used in DWI, and fast spin echo (FSE) sequence was used in T1WI + C. MRI image scanning parameters are shown in Table 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. MRI image segmentation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTwo radiologists experienced in the diagnosis of uterine tumors used ITK-SNAP 3.8.0 to perform semi-automatic ROI delineation of endometrial cancer MRI images on DWI, T1WI + C and T2WI images without knowing the pathological results. First, ensure that the data is complete and the sequence is correct and import it into the software. In the software, the window width and window level adjustment tools are used to optimize the image display, and the endometrium and lesion areas are clearly presented. Select the appropriate delineation tool in the toolbar. For example, when the boundary is clear, you can try the magic wand tool to quickly select, and if the boundary is blurred, use the hand-drawn tool to describe it in detail. The ROI contains the largest facet of the tumor tissue, avoiding surrounding normal tissue, necrotic tissue, and blood. If there are differences in the sketcher, it will be determined by a senior radiologist with more than 20 years of experience to ensure accuracy and reproducibility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. LVSI imaging and deep learning construction process for endometrial cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1 LVSI Imaging Process for Endometrial Cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe research method of this study used ITK-SNAP to manually plot ROI on DWI, T1WI + C, and T2WI images, and used Onekey AI to extract radiomics features from MRI images in the radiomics library. Subsequently, depth features were obtained by cropping the largest area slice of the region of interest (ROI) from the MRI image. In this study, transfer learning techniques were used to further enrich these features with the help of a pre-trained ResNet101 model. To identify the most robust, non-redundant, and predictive features, correlation filtering and Lasso regression methods were employed in this study. Finally, radiomics feature labels and corresponding radiomics column plots were constructed on an independent validation cohort. The flowchart is as follows (Fig 2)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 radiomics feature extraction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFeatures are divided into three main categories: geometric features, intensity features, and texture features. Geometry features cover the three-dimensional shape properties of a tumor, providing a spatial representation of its structure. Intensity features capture first-order statistical properties of the intensity of endosomal elements in a tumor, helping to understand its internal uniformity or variability. Texture features reflect more complex patterns, depict second-order and higher-order spatial distributions of intensity, and encompass complex spatial relationships between voxels. These texture features are extracted from MRI images by pyradiomics library (http://pyradiomics.readthedocs.io), such as grey release co-occurrence matrix (GLCM), grey release stroke length matrix (GLRLM), grey release size region matrix (GLSZM) and neighborhood grey release difference matrix (NGTDM). A total of 1403 image omics features were extracted in this study: 36 geometric features, 208 intensity features and 1272 texture features.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Imageomics feature screening and feature selection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, all extracted features are normalized using Z-score normalization and evaluated for their statistical significance by a t-test or Mann-Whitney U-test. Only features with p-values below 0.05 are retained. To mitigate the collinearity problem, this study uses Pearson correlation coefficients to check for correlations between features, excluding one feature in pairs with correlation coefficients higher than 0.9. The feature set is further optimized by Lasso regression within the 10-fold cross-validation framework to determine the optimal regularization parameter lambda, thus effectively reducing the feature set to the most predictive and informative features. The ROI curve was used to evaluate the performance of 10 models such as LR, SVM, KNN, and RandomForest on the training dataset (lab-train) and test set (lab-test), including Accuracy (accuracy), AUC, 95% CI (confidence interval), Sensitivity (sensitivity), and Specificity (specificity). Select a better model to evaluate the performance of radiomics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 \u0026nbsp;deep learning process of LVSI for endometrial cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4.1 Data preparation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eROI Cropping: This study selects slices that present the greatest region of interest (ROI) as representative images.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eData Enhancement: The approach of this study involves normalizing the intensity distribution by Z-score normalization of the RGB channels of the images. These normalized images are then used as inputs to the model. During the training phase, this study implements real-time data enhancement strategies including random cropping, horizontal flipping, and vertical flipping. For test images, only normalization processing is performed.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4.2 Training Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTransfer learning: This study assesses ResNet101 and CNN - based models. ResNet101, a deep residual network, overcomes layer - related issues in traditional neural networks with its residual connections. These connections enable direct information flow between layers, allowing for training of very deep networks. ResNet101 performs well in image classification, object detection, and semantic segmentation, advancing computer vision research.\u003c/p\u003e\n\u003cp\u003eHyperparameters: Transfer learning was used to adapt the model to different patient groups. The model was initialized with ImageNet pre - trained weights. A cosine annealing learning rate strategy,\u003cem\u003eηt\u003c/em\u003e=\u003cem\u003eη\u003c/em\u003emini+21(\u003cem\u003eη\u003c/em\u003emaxi−\u003cem\u003eη\u003c/em\u003emini)(1+cos(\u003cem\u003eTi\u003c/em\u003e\u003cem\u003eT\u003c/em\u003ecur\u003cem\u003eπ\u003c/em\u003e))，\u003cem\u003eη\u003c/em\u003emini=0\u0026nbsp;The minimum learning rate is set，\u003cem\u003eη\u003c/em\u003emaxi=0.01\u0026nbsp;The maximum learning rate is set，\u003cem\u003eTi\u003c/em\u003e=30\u0026nbsp;Represents the number of epochs during iterative training。Other important hyperparameters include using random layer descent (SGD) as the optimizer and using softmax cross entropy as the loss function.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4.3 Feature Label Model Construction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eImageomics Feature Models\u003c/strong\u003e: Using LASSO for feature selection, this study built imageomics risk models with algorithms like logistic regression, SVMs, and random forest. It compared model performances and explored feature fusion from multi - modal imaging to enhance prediction accuracy.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeep Learning Feature Models\u003c/strong\u003e: CNN - computed output probabilities served as deep learning feature labels.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeep Learning radiomics Feature Model\u003c/strong\u003e: The CNN model was the best in the test set. Features from its penultimate layer were extracted, PCA - reduced to 512 dimensions, and then combined with radiomics features via a pre - fusion algorithm to form DLR feature labels, followed by feature selection and model building.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCombinatorial Feature Model\u003c/strong\u003e: After early feature fusion and manual stitching, the study conducted feature selection and model construction similar to imageomics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Relevance Enhancement\u003c/strong\u003e: Univariate and stepwise multivariate analyses identified key clinical features. These, combined with deep learning model predictions, formed an LR linear model and a combined feature label, which was visualized as a line diagram.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4.4 Statistical analysis:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData analytics were done on Onekey AI platform (v4.9.1) using Python 3.7.12. Specific library versions were used: Statsmodels 0.13.2 for stats, PyRadiomics 3.0.1 for imageomics, Scikit - learn 1.0.2 for machine learning, and a PyTorch 1.11.0 - based deep learning framework with CUDA and cuDNN. Python and statsmodels were for analysis, scikit - learn for machine learning model, and deep learning trained on NVIDIA 4090 GPUs with MONAI and PyTorch.\u003c/p\u003e\n\u003cp\u003eThe Shapiro - Wilk test checked clinical feature normality. Appropriate tests were used for different variable types. A p - value \u0026gt; 0.05 in Table 2 ensured unbiased groupings. In the test cohort, ROC curves evaluated model discriminative power, calibration curves with Hosmer - Lemeshow test analyzed calibration, and DCA evaluated clinical utility.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e1. Clinical characteristics of LVSI in endometrial carcinoma\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.1 Baseline characteristics of clinical alignment\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 456 patients with multicenter intrauterine intimacy were included, and a total of 148 patients were excluded according to the exclusion criteria, namely: Incomplete clinical or imaging data (n = 79); Received radiotherapy and chemotherapy for endometrial cancer before enrollment (n = 42); The MRI image contained artifacts (n = 27). Finally 308 met our study criteria. LVSI+(n=69),LVSI-(n=239).\u003c/p\u003e\n\u003cp\u003eThe training (215 cases) and test (93 cases) sets of 308 endometrial cancer patients were well - balanced in core clinicopathological features. (Table 2)\u003c/p\u003e\n\u003cp\u003eThe following is a clinical feature correlation matrix showing the correlation between the four variables \"Myometrial_invasion (myometrial infiltration) \", \"Histopathologic_subtype (histopathological subtype) \", \"Para_aortic_lymph_node_metastasis (paraaortic lymph node metastasis) \" and \"label\" (Fig 3-a) .\u003c/p\u003e\n\u003cp\u003eThe diagonal in the correlation figure is 1 by definition. Non - diagonal values, ranging from 0.126 - 0.382, show weak positive correlations between variables. Myometrial_invasion and Histopathologic_subtype have the lowest correlation coefficient (0.126), indicating the weakest link. This visualization helps preliminarily analyze variable correlations for future research.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.2 LVSI univariate and multivariate analysis of endometrial cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study conducted univariate and multivariate analyses of endometrial cancer. Univariate analysis showed that progesterone receptor, estrogen receptor, etc., were significantly associated with reduced risk, while para - aortic lymph node metastasis was the only risk factor. Multivariate analysis identified three independent prognostic indicators: histological subtype, deep myometrial infiltration, and para - aortic lymph node metastasis, with OR \u0026gt; 4, indicating high clinical value. The strong negative correlation of PR and ER aligns with hormone receptor typing, and Ki67 and P53 confirm the model's rationality. Metabolic factors and FIGO staging were not significant. The overall data constructed a highly operable model, offering a reliable basis for personalized diagnosis and treatment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. Comparison of LVSI radiomics and deep learning radiomics feature fusion weights for endometrial cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the radiomics feature model, wavelet_LHL_firstorder_Kurtosis_1 (coefficient 0.071) has the most prominent positive effect and significantly improves the label; log_sigma_5_0_mm_3D_firstorder_RootMean Squared_0 (coefficient -0.062) has the strongest negative effect and suppresses the label value. In addition, wavelet_HLH_ firstorder_Skewness_1 (0.0336), log_sigma_5_0_mm_3D_ngtdm_Busyness_2 (0.0247) and other features also have obvious positive contributions, reflecting the key role of image texture and grey release statistical characteristics on the model, (Fig 3-b). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn the fusion coefficient analysis of radiomics and deep learning radiomics features, wavelet_LHL_firstorder-_ Kurtosis_1 (coefficient 0.0715) had the most prominent positive effect in the model, significantly improving the label value; log_sigma_5_0_mm_ 3D_firstorder_RootMeanSquared_0 (coefficient -0.0641) had the strongest negative effect and suppressed the label value. In addition, log_sigma_5_0_mm_3D_ngtdm_Busyness_2 (0.0304), wavelet_HLL_glszm_ SmallAreaHighGrayLevelEmphasis _0 (0.0228) and other features also made significant positive contributions, reflecting the key role of image texture and grey release statistical characteristics on the model. (Fig 3-c).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. Endometrial cancer LVSI deep learning imaging model feature label results\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe performance of 8 models such as LR, SVM, KNN, and RandomForest on the training dataset (lab-train) and test set (lab-test) includes Accuracy (accuracy), AUC, 95% CI (confidence interval), Sensitivity (sensitivity), Specificity (specificity) and other indicators. The accuracy of the XGBoost training dataset reached 0.991, the AUC was 0.999, but the accuracy of the test set was 0.613; most models showed that the training dataset was better than the test set. As shown in Table 4, Fig 3-d, Fig3-e show the corresponding generated ROC curves. \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe ROC curve shows that in the training set, the performance of each model is significantly differentiated. The XGBoost model performs the best, with an AUC of 0.999 (95% CI 0.998 - 1.000), which is almost close to the ideal ROC curve, and the sensitivity and specificity are highly balanced. The SVM (AUC 0.985), LR (AUC 0.948) and other model curves are also significantly better than the diagonal, which fully verifies the strong classification ability of deep learning radiomics features in the training data, indicating that the model fits the training dataset data well (Fig 3-d).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn the test set, although the overall performance was weakened, the LR model test set AUC reached 0.704 (95% CI 0.578 - 0.829), which still maintained good discrimination in independent data, providing data support for model selection; other models such as SVM (AUC 0.672) also showed basic generalization potential. (Fig 3-e).\u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. Comparison of LVSI clinical characteristics, radiomics characteristics, deep learning radiomics characteristics, and deep learning fusion tags for endometrial cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe results showed that the ROC curve of the training cohort: Clinic AUC: 0.799 (95% CI: 0.770 - 0.821); Imageomics Feature Label (Rad) AUC: 0.846 (95% CI: 0.827 - 0.865); deep learning feature label (DTL) AUC: 0.668 (95% CI: 0.642 - 0.695); deep learning radiomics feature label (DLR) AUC: 0.948 (95% CI: 0.938 - 0.957); combined feature label (Combine) AUC: 0.953 (95% CI: 0.943 - 0.962). (Table 5, Fig 3-f)\u003c/p\u003e\n\u003cp\u003eTest cohort ROC curve: Clinic AUC: 0.675 (95% CI: 0.631 - 0.718); deep learning radiomics feature label (DLR) AUC: 0.704 (95% CI: 0.663 - 0.745); radiomics feature label (Rad) AUC: 0.687 (95% CI: 0.640 - 0.733); deep learning feature label (DTL) AUC: 0.654 (95% CI: 0.602 - 0.707); combined feature label (Combine) AUC: 0.720 (95% CI: 0.681 - 0.759). (Table 5, Fig 3- g). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1 ROC curve analysis of training dataset\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTraining dataset: Representing Clinic, Rad, DTL, DLR, and Combined models, with their corresponding area under the curve (AUC) and 95% confidence interval (CI) next to them.Among them, Combine has the highest AUC value, 0.953, and the 95% CI is 0.943-0.962, indicating that it has the strongest ability to distinguish between positive and negative cases on the training dataset; followed by DLR, whose AUC is 0.948; and DTL, whose AUC is lower, 0.668. (Fig 3-f)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 ROC curve analysis of test set\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe ROC curves of the five models or indicators for predicting endometrial cancer in the test set and the corresponding AUC values and 95% CI. On the test set, the combined AUC value was 0.720, and the 95% CI was 0.681-0.759, which was still relatively high among several indicators; the DLR's AUC was 0.704; the DTL's AUC was the lowest, 0.654. (Fig 3-g)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Calibration curves of different LVSI models for endometrial cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Hosmer-Lemeshow (HL) test is used to quantify the difference between the predicted probability and the observed results, generating calibration graphs (Fig 3-h and 3-i).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3.1 Training dataset calibration curve (Cohort train Calibration):\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe training dataset calibration curve shows that most models match the predicted probability with the actual positive frequency well. Among them, the Combined and DLR models perform best. In the prediction probability interval of 0.2-0.8, the actual positive frequency is highly consistent with the predicted probability. For example, when the prediction probability is 0.4, the actual frequency is close to 0.4; when the prediction probability is 0.6, the actual frequency is stable at 0.58-0.62. Although Clinic, Rad and other models have slight deviations, the overall trend is close to the ideal line. (Fig 3-h).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3.2 Test set calibration curve (Cohort test Calibration):\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the test set calibration curve, the actual positive frequency of the Combined model is closer to the ideal line in the prediction probability range of 0.4-0.6, which shows the rationality of the evaluation of this risk segment; the DLR model also shows the actual positive frequency consistent with the prediction probability at some probability points (such as 0.3-0.4), reflecting a certain calibration potential.(Fig 3-i). \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. DeLong test\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis \"Cohort test Delong\" graph presents the comparisons among different models on the test set. The Combined model stands out with extremely small P - values when compared with most models, indicating significantly superior performance, and it can be used as the main applied model. The Clinic model has differences in performance from some models and can leverage its own advantages in combination with clinical practice.(Fig 3-j)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6. Clinical application of LVSI in endometrial cancer: decision curve analysis (DCA)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTwo DCA graphs show different models' performance in training and test sets(Fig 4-a, Fig 4-b).\u003c/p\u003e\n\u003cp\u003eIn the training set, for 0 - 0.2, models are close to the perfect calibration line with similar performance; for 0.2 - 0.8, Combined and DLR excel, offering high net benefits, and Clinic and DTL also work; for 0.8 - 1.0, models near the line again.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn the test set, 0 - 0.2 shows deviation. For 0.2 - 0.6, Combined still has good net benefits in parts, though reduced compared to the training set. For 0.6 - 1.0, models fluctuate greatly, yet this indicates improvement directions, and generalization ability can be enhanced.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7. Endometrial cancer LVSI nomogram\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Nomogram uses Myometrial_invasion (myometrial infiltration), Histopathologic_subtype (histopathological subtype), Para_aortic_lymph_node_metastasis (paraaortic lymph node metastasis) and DLR as variables to assign corresponding Points scores to each characteristic and different states; by accumulating Total Points, it is finally mapped to the Risk axis to realize the visual prediction of risk probability (Fig 5). \u0026nbsp; \u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study constructed multimodal MRI imaging, deep learning, and feature extraction fusion models for predicting lymphovascular space invasion (LVSI) in endometrial cancer. The LR model showed the best performance in LVSI prediction on the training dataset and test set. Combining clinical and radiomics features using the LR algorithm in column plots had excellent performance. The AUC of the deep learning hybrid feature model on the training dataset and test set and the AUC of the hybrid model on the training dataset and test set are consistent with previous studies showing that multimodal data fusion can capture information on tumor heterogeneity\u003csup\u003e\u0026nbsp;[15]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eCompared with Liu et al.'s study \u003csup\u003e[17]\u003c/sup\u003e, although the training performance of our combined model was slightly highe, the test set AUC in our study was lower. The possible reason is that deep learning in our study can mine deep associations in multimodal data, but it has high requirements for data homogeneity, and data collection bias and sample heterogeneity in the test set challenge its generalization ability.\u003c/p\u003e\n\u003cp\u003eWe explored the ResNet101 network to enhance the performance of traditional CNN models and compared different models. Feature extraction and selection in this study ensured feature effectiveness. The good performance of the independent LR model may be related to its complex data processing ability. radiomics lineup diagrams are useful for clinics, yet the study has limitations such as a small sample size affecting generalization and subjectivity in ROI sketching causing errors.\u003c/p\u003e\n\u003cp\u003eRadiomicsfeature selection is crucial. Texture features (e.g., texture complexity, contrast, uniformity) can reflect tumor tissue heterogeneity and cell structure complexity related to LVSI \u003csup\u003e[18, 19]\u003c/sup\u003e. These features are closely related to tumor aggressiveness, such as high contrast texture suggesting internal structural disorder of tumor, and long-run emphasis reflecting abnormal cell arrangement\u003csup\u003e\u0026nbsp;[20]\u003c/sup\u003e. Shape features (e.g., tumor area, perimeter) can reflect tumor growth and invasiveness \u003csup\u003e[21]\u003c/sup\u003e. We selected various Radiomicsfeatures including first - order statistical, texture, and shape features, which can reflect tumor microstructures and biological characteristics from different angles \u003csup\u003e[22]\u003c/sup\u003e. First - order statistical features mainly describe the distribution of image grey release values, such as mean, standard deviation, skewness, and kurtosis. These features can reflect the density and uniformity of tumor tissue, and may have a certain relationship with the occurrence of LVSI \u003csup\u003e[23]\u003c/sup\u003e. Some features like angular second - order moment in texture features and maximum diameter in shape features are significantly correlated with LVSI\u003csup\u003e\u0026nbsp;[24, 25]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eDeep learning models, through attention mechanisms like GAM modules, focus on tumor margins and vascular structures, identifying microscopic infiltrations \u003csup\u003e[26]\u003c/sup\u003e. Our innovation is the \"complementary feature fusion\" strategy, combining radiomics interpretability features with deep learning abstract features. The experimental results show that the fusion model has a slightly higher test - set AUC (0.720) than the individual radiomics model (0.687) and the deep learning feature fusion model (0.704), possibly due to the complementarity of the two types of features \u003csup\u003e[27, 28]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResearch limitations and prospects\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eThis study has the following limitations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study, though insightful, has areas for improvement.Study data—clinical, pathological, and imaging—show heterogeneity. Tumor - related imaging data and subjectivity in LVSI evaluation of pathological sections may affect finding precision and reliability \u003csup\u003e[27 - 29]\u003c/sup\u003e.The model has good predictive performance, yet its biological mechanism is unclear. Further study with gene expression or immunohistochemical data is needed \u003csup\u003e[28 - 30]\u003c/sup\u003e.The multimodal MRI radiomics - deep - learning model is complex, requiring professional skills for operation. This complexity could impede its clinical use.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe fusion model of multimodal MRI radiomics and deep learning features has shown good advantages in EC-LVSI prediction, providing a new technical path for precise risk stratification before surgery.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eY. T. participated in the study design, data collection and manuscript writing. F.F. S. conducted a critical review of the manuscript. Y. T.,F.F. S. authors have contributed equally to this work .W. Z. participated in the data collection and data analytics. K. W. assisted in the collection and collation of references. Y. C. participated in data collection and typesetting. J.Q. S. (corresponding author) supervised the entire research project, participated in the study design, case collection and was responsible for the final approval of the manuscript. All authors read and approved the final version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eBray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. \u003cem\u003eCA Cancer J Clin\u003c/em\u003e. 2024;74(3):229-263. doi:10.3322/caac.21834\u003c/li\u003e\n \u003cli\u003eBosse T, Peters EE, Creutzberg CL, et al. Substantial lymph-vascular space invasion (LVSI) is a significant risk factor for recurrence in endometrial cancer--A pooled analysis of PORTEC 1 and 2 trials. \u003cem\u003eEur J Cancer\u003c/em\u003e. 2015;51(13):1742-1750. doi:10.1016/j.ejca.2015.05.015\u003c/li\u003e\n \u003cli\u003eGuntupalli SR, Zighelboim I, Kizer NT, et al. Lymphovascular space invasion is an independent risk factor for nodal disease and poor outcomes in endometrioid endometrial cancer. \u003cem\u003eGynecol Oncol\u003c/em\u003e. 2012;124(1):31-35. doi:10.1016/j.ygyno.2011.09.017\u003c/li\u003e\n \u003cli\u003eLiu D, Huang J, Zhang Y, et al. Multimodal MRI-based radiomics models for the preoperative prediction of lymphovascular space invasion of endometrial carcinoma. \u003cem\u003eBMC Med Imaging\u003c/em\u003e. 2024;24(1):252. Published 2024 Sep 20. doi:10.1186/s12880-024-01430-1\u003c/li\u003e\n \u003cli\u003eIroni G, Mapelli P, Bergamini A, et al. Hybrid PET/MRI in Staging Endometrial Cancer: Diagnostic and Predictive Value in a Prospective Cohort. \u003cem\u003eClin Nucl Med\u003c/em\u003e. 2022;47(3):e221-e229. doi:10.1097/RLU.0000000000004064\u003c/li\u003e\n \u003cli\u003eSaleh GA, Abdelrazek R, Hassan A, Hamdy O, Tantawy MSI. Diagnostic utility of apparent diffusion coefficient in preoperative assessment of endometrial cancer: are we ready for the 2023 FIGO staging?. \u003cem\u003eBMC Med Imaging\u003c/em\u003e. 2024;24(1):226. Published 2024 Aug 28. doi:10.1186/s12880-024-01391-5\u003c/li\u003e\n \u003cli\u003eZhang Q, Ouyang H, Ye F, et al. Multiple mathematical models of diffusion-weighted imaging for endometrial cancer characterization: Correlation with prognosis-related risk factors. \u003cem\u003eEur J Radiol\u003c/em\u003e. 2020;130:109102.doi:10.1016/j.ejrad.2020.109102.\u003c/li\u003e\n \u003cli\u003eLi S, Wang Y, Sun Y, et al. Both intra- and peri-tumoral radiomics signatures can be used to predict lymphatic vascular space invasion and lymphatic metastasis positive status from endometrial cancer MR imaging. \u003cem\u003eAbdom Radiol (NY)\u003c/em\u003e. 2024;49(11):4140-4150. doi:10.1007/s00261-024-04432-3\u003c/li\u003e\n \u003cli\u003eYan B, Jia Y, Li Z, et al. Preoperative prediction of lymphovascular space invasion in endometrioid adenocarcinoma: an MRI-based radiomics nomogram with consideration of the peritumoral region. \u003cem\u003eActa Radiol\u003c/em\u003e. 2023;64(9):2636-2645. doi:10.1177/02841851231181681\u003c/li\u003e\n \u003cli\u003eLambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. \u003cem\u003eEur J Cancer\u003c/em\u003e. 2012;48(4):441-446. doi:10.1016/j.ejca.2011.11.036\u003c/li\u003e\n \u003cli\u003eMa C, Zhao Y, Song Q, et al. Multi-parametric MRI-based radiomics for preoperative prediction of multiple biological characteristics in endometrial cancer. \u003cem\u003eFront Oncol\u003c/em\u003e. 2023;13:1280022. Published 2023 Dec 15. doi:10.3389/fonc.2023.1280022\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eJin X, Yan R, Li Z, et al. Evaluation of Amide Proton Transfer-Weighted Imaging for Risk Factors in Stage I Endometrial Cancer: A Comparison With Diffusion-Weighted Imaging and Diffusion Kurtosis Imaging. \u003cem\u003eFront Oncol\u003c/em\u003e. 2022;12:876120. Published 2022 Apr 14. doi:10.3389/fonc.2022.876120\u003c/li\u003e\n \u003cli\u003eWang Y, Liu W, Lu Y, et al. Fully Automated Identification of Lymph Node Metastases and Lymphovascular Invasion in Endometrial Cancer From Multi-Parametric MRI by Deep Learning. \u003cem\u003eJ Magn Reson Imaging\u003c/em\u003e. 2024;60(6):2730-2742. doi:10.1002/jmri.29344\u003c/li\u003e\n \u003cli\u003eJiang X, Zhai W, Song J, et al. Associations between MRI radiomic phenotypes and clinical outcomes in endometrial cancer: Implications for preoperative risk stratification. \u003cem\u003eMagn Reson Imaging\u003c/em\u003e. 2025;117:110298. doi:10.1016/j.mri.2024.110298\u003c/li\u003e\n \u003cli\u003eZhang Y, Chen S, Wang Y, et al. Deep learning-based methods for classification of microsatellite instability in endometrial cancer from HE-stained pathological images. \u003cem\u003eJ Cancer Res Clin Oncol\u003c/em\u003e. 2023;149(11):8877-8888. doi:10.1007/s00432-023-04838-4\u003c/li\u003e\n \u003cli\u003eZhang Y, Chen S, Wang Y, et al. Deep learning-based methods for classification of microsatellite instability in endometrial cancer from HE-stained pathological images. \u003cem\u003eJ Cancer Res Clin Oncol\u003c/em\u003e. 2023;149(11):8877-8888. doi:10.1007/s00432-023-04838-4\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eLiu D, Huang J, Zhang Y, et al. Multimodal MRI-based radiomics models for the preoperative prediction of lymphovascular space invasion of endometrial carcinoma. \u003cem\u003eBMC Med Imaging\u003c/em\u003e. 2024;24(1):252. Published 2024 Sep 20. doi:10.1186/s12880-024-01430-1\u003c/li\u003e\n \u003cli\u003eYan BC, Li Y, Ma FH, et al. Radiologists with MRI-based radiomics aids to predict the pelvic lymph node metastasis in endometrial cancer: a multicenter study. \u003cem\u003eEur Radiol\u003c/em\u003e. 2021;31(1):411-422. doi:10.1007/s00330-020-07099-8\u003c/li\u003e\n \u003cli\u003eMao W, Chen C, Gao H, Xiong L, Lin Y. A deep learning-based automatic staging method for early endometrial cancer on MRI images. \u003cem\u003eFront Physiol\u003c/em\u003e. 2022;13:974245. Published 2022 Aug 30. doi:10.3389/fphys.2022.974245\u003c/li\u003e\n \u003cli\u003eLiu XF, Yan BC, Li Y, Ma FH, Qiang JW. Radiomics feature as a preoperative predictive of lymphovascular invasion in early-stage endometrial cancer: A multicenter study. \u003cem\u003eFront Oncol\u003c/em\u003e. 2022;12:966529. Published 2022 Aug 18. doi:10.3389/fonc.2022.966529\u003c/li\u003e\n \u003cli\u003eBonatti M, Pedrinolla B, Cybulski AJ, et al. Prediction of histological grade of endometrial cancer by means of MRI. \u003cem\u003eEur J Radiol\u003c/em\u003e. 2018;103:44-50. doi:10.1016/j.ejrad.2018.04.008\u003c/li\u003e\n \u003cli\u003eLiu XF, Yan BC, Li Y, Ma FH, Qiang JW. Radiomics Nomogram in Assisting Lymphadenectomy Decisions by Predicting Lymph Node Metastasis in Early-Stage Endometrial Cancer. \u003cem\u003eFront Oncol\u003c/em\u003e. 2022;12:894918. Published 2022 May 31. doi:10.3389/fonc.2022.894918\u003c/li\u003e\n \u003cli\u003eFeng M, Zhao Y, Chen J, et al. A deep learning model for lymph node metastasis prediction based on digital histopathological images of primary endometrial cancer. \u003cem\u003eQuant Imaging Med Surg\u003c/em\u003e. 2023;13(3):1899-1913. doi:10.21037/qims-22-220\u003c/li\u003e\n \u003cli\u003eCoada CA, Santoro M, Zybin V, et al. A Radiomic-Based Machine Learning Model Predicts Endometrial Cancer Recurrence Using Preoperative CT Radiomic Features: A Pilot Study. \u003cem\u003eCancers (Basel)\u003c/em\u003e. 2023;15(18):4534. Published 2023 Sep 13. doi:10.3390/cancers15184534\u003c/li\u003e\n \u003cli\u003eMeng X, Zhang X, Tian S, et al. Evaluation of lymphovascular space invasion in endometrial carcinoma by APTw and mDixon-Quant. \u003cem\u003eActa Radiol\u003c/em\u003e. 2024;65(11):1440-1446. doi:10.1177/02841851241277339\u003c/li\u003e\n \u003cli\u003eHodneland E, Dybvik JA, Wagner-Larsen KS, et al. Automated segmentation of endometrial cancer on MR images using deep learning. Sci Rep. 2021;11(1):179. Published 2021 Jan 8. doi:10.1038/s41598-020-80068-9\u003c/li\u003e\n \u003cli\u003eTakahashi Y, Sone K, Noda K, et al. Automated system for diagnosing endometrial cancer by adopting deep-learning technology in hysteroscopy. \u003cem\u003ePLoS One\u003c/em\u003e. 2021;16(3):e0248526. Published 2021 Mar 31. doi:10.1371/journal.pone.0248526\u003c/li\u003e\n \u003cli\u003eDobrzycka B, Terlikowska KM, Kowalczuk O, Niklinski J, Kinalski M, Terlikowski SJ. Prognosis of Stage I Endometrial Cancer According to the FIGO 2023 Classification Taking into Account Molecular Changes. \u003cem\u003eCancers (Basel)\u003c/em\u003e. 2024;16(2):390. Published 2024 Jan 17. doi:10.3390/cancers16020390\u003c/li\u003e\n \u003cli\u003eBosse T, Peters EE, Creutzberg CL, et al. Substantial lymph-vascular space invasion (LVSI) is a significant risk factor for recurrence in endometrial cancer--A pooled analysis of PORTEC 1 and 2 trials. \u003cem\u003eEur J Cancer\u003c/em\u003e. 2015;51(13):1742-1750. doi:10.1016/j.ejca.2015.05.015\u003c/li\u003e\n \u003cli\u003eFasmer KE, Hodneland E, Dybvik JA, et al. Whole-Volume Tumor MRI Radiomics for Prognostic Modeling in Endometrial Cancer. \u003cem\u003eJ Magn Reson Imaging\u003c/em\u003e. 2021;53(3):928-937. doi:10.1002/jmri.27444\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 5 are available in the Supplementary Files section.\u003c/p\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":"Endometrial carcinoma, lymphatic vascular space invasion, MRI, Radiomics, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-6281099/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6281099/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eLymphovascular space invasion (LVSI) is a key prognostic indicator in endometrial cancer, impacting disease progression, treatment strategies and overall survival of patients. The accurate identifying of LVSI of endometrial cancer before surgery is challenging due to certain limitations in radiological methods.\u003c/p\u003e\u003ch2\u003eObjective\u003c/h2\u003e \u003cp\u003eTo develop an efficient approach for predicting LVSI in endometrial cancer before surgery using multimodal MRI Radiomics and deep learning, offering a crucial foundation for clinical treatment.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eTwo radiologists, unaware of pathology, manually outlined the region of interest (ROI) on preoperative DWI, T1WI\u0026thinsp;+\u0026thinsp;C, and T2WI images. The intersection of their ROIs was the final one. radiomics features were extracted from the ROI, and depth features from the largest - area ROI slice in MRI images. In the training set, 8 models were built via logistic regression after feature reduction. These models were tested against pathology, and the area under the receiver operating characteristic curve (AUC) was calculated. The best - performing imaging model was combined with a deep - learning model to form a hybrid, and its performance was evaluated.​\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003e308 patients were split 7:3 into a training set (n\u0026thinsp;=\u0026thinsp;215) and a test set (n\u0026thinsp;=\u0026thinsp;93). The LR model performed best in LVSI prediction. Combining clinical and radiomics with the LR algorithm enhanced performance. Deep - learning mixed - feature models had AUCs of 0.948 (training) and 0.704 (test), while mixed models had 0.953 (training) and 0.720 (test).​\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eDeep - learning mixed - feature and mixed models are most effective for preoperative LVSI prediction in endometrial cancer, guiding clinical decision - making.\u003c/p\u003e","manuscriptTitle":"Multimodal MRI Radiomics Features and Deep Learning Features in Predicting Lymphovascular Space Invasion in Endometrial Carcinoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-21 09:54:56","doi":"10.21203/rs.3.rs-6281099/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":"932476cc-ad58-4b3a-b034-168ab3bac3cf","owner":[],"postedDate":"April 21st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-05-01T13:08:23+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-21 09:54:56","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6281099","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6281099","identity":"rs-6281099","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}