Prediction of myelosuppression in cervical cancer after concurrent chemoradiotherapy by CT radiomics-based model.

Introduction

Cervical cancer (CC) ranks as the fourth most common malignancy in women globally, with over 500,000 annual diagnoses and 300,000 deaths, disproportionately affecting developing countries [ 1 , 2 ]. Standard treatment combines surgery, radiotherapy, and chemotherapy, with concurrent chemoradiotherapy (CCRT) emerging as a cornerstone for locally advanced disease due to improved local control and survival [ 3 ]. However, CCRT carries significant risks, particularly myelosuppression, which occurs in 44–68% of patients and can lead to treatment delays, infections, and mortality [ 4 ].

In recent years, radiomics, a field within medical imaging, has gained considerable attention. Radiomics involves the extraction and analysis of quantitative features from medical images to characterize tumor phenotypes, assess prognosis, and guide personalized treatment decisions. Various imaging modalities, including computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET), are employed in radiomics studies [ 5 , 6 ]. CT imaging has shown particular relevance in the evaluation and management of cervical cancer. It provides detailed anatomical information and can assist in tumor staging, identifying lymph node involvement, and assessing treatment response [ 7 ]. The application of radiomics using CT images in cervical cancer holds promise for enhancing diagnostic accuracy, prognostic assessment, and treatment planning [ 8 ].

Although concurrent chemoradiotherapy has become the standard treatment for locally advanced cervical cancer, accurately predicting and managing treatment-related toxicities (especially myelosuppression) remains a significant clinical challenge. Myelosuppression, characterized by reduced blood cell production, not only impacts treatment efficacy through dose reductions or delays but also increases risks of infection, hemorrhage, and anemia, significantly compromising patient quality of life and survival outcomes [ 9 ]. Currently, dosimetric parameters based on the Dose-Volume Histogram (DVH), such as the volume of pelvic bone marrow receiving 10 Gy (V10) and 20 Gy (V20), have been confirmed as important predictors of hematologic toxicity [ 10 – 12 ]. In recent years, with advancements in imaging and dose analysis technologies, novel predictive methods have continuously emerged. On one hand, dosiomics, which extracts high-dimensional spatial features from three-dimensional dose distributions, enables a more detailed characterization of dose distribution patterns and can be used to predict tumor response under different treatment modalities [ 13 ]. On the other hand, imaging biomarkers derived from pretreatment CT scans have emerged as a novel approach, with multiple studies reporting associations between radiomic features and hematological toxicities [ 14 , 15 ]. Specifically in cervical cancer, Ren et al. utilized radiomic analysis of treatment planning CT scans to successfully predict complete response rates (balanced accuracy of 80.8% in the validation set) and hematological toxicities (leukopenia prediction with balanced accuracy of 69.9%) in locally advanced cervical cancer patients undergoing intensity-modulated radiotherapy [ 16 ]. This study not only confirmed the feasibility of radiomics in predicting treatment response in cervical cancer but also provided new insights and methods for precise prediction of myelosuppression.

Existing research is mostly based on clinical parameters (such as age, chemotherapy dose) or single imaging features (such as CT dose volume parameters), lacking multidimensional data integration. This study combines radiomics features of CT images with clinical features to construct a comprehensive prediction model. This study selected regions of interest (ROIs) from CT images and extracted radiological features, and then used the Least Absolute Shrinkage and Selection Operator (LASSO) method to identify basic imaging features. Construct a comprehensive prediction model based on 19 clinical parameters closely related to cervical cancer radiotherapy and chemotherapy. In addition, eight machine learning algorithms are utilized to construct radiomics and clinical models. Finally, univariate and multivariate Cox logistic regression models are applied to identify independent prognostic factors, and a nomogram is constructed based on these factors. The findings of this study have the potential to provide valuable insights for the treatment of patients with cervical cancer, contributing to improved therapeutic decision-making and patient outcomes.

Materials and methods

Patients

The study protocol was reviewed and approved by the Ethics Committee in Clinical Research (ECCR) of the Second Affiliated Hospital of Zhengzhou University. It was conducted following the Declaration of Helsinki. As this study was a retrospective cohort study, informed consent was waived, and all patient data were anonymized and desensitized.

Patients with cervical cancer from 1st January 2017 to 31st December 2023 in the Second Affiliated Hospital of Zhengzhou University were collected for this study. These 300 patients received a cervical cancer diagnosis, oncological surgery, and postoperative radiotherapy at The Second Affiliated Hospital of Zhengzhou University. The flowchart of patient exclusion and inclusion was shown in Fig. 1 .

The inclusion criteria included [ 17 ] (a) performance status (PS) ＜ 2 (b) patients with relatively intact organ functions, basically standard hematological system, and no contraindications to treatment; (c) no previous history of other malignant tumor diseases and radiotherapy; (d) postoperative pathological examination results confirming the diagnosis of cervical cancer; (e) postoperative radiotherapy; (f) the patient had complete pathology, imaging, and radiation therapy dose information; (g) Received concurrent chemoradiotherapy only, with no prior or subsequent immunotherapy of any form.

The exclusion criteria included (a) no definite postoperative pathological findings; (b) no complete clinicopathological data; (c) no CT scan was performed before postoperative radiotherapy; (d) patient’s pathology, imaging, and radiation treatment dose data are missing; (e) intolerance to radiotherapy or chemotherapy, and treatment plan was not completed due to severe acute toxic reactions during treatment.

According to the Common Terminology Criteria for Adverse Events (CTCAE version 5.0), the grading definitions for bone marrow suppression are as follows: Grade I:Leukocyte count less than the lower limit of normal (LLN) but ≥3.0 × 10 9 /L; hemoglobin less than LLN but ≥100 g/L; platelet count less than LLN but ≥75 × 10 9 /L. Grade II: Leukocyte count < 3.0 to 2.0 × 10 9 /L; hemoglobin < 100 to 80 g/L; platelet count < 75 to 50 × 10 9 /L. Grade III: Leukocyte count < 2.0 to 1.0 × 10 9 /L; hemoglobin < 80 g/L (transfusion indicated); platelet count < 50 to 25 × 10 9 /L. Grade IV: Leukocyte count < 1.0 × 10 9 /L; hemoglobin reduced to life-threatening levels requiring urgent intervention; platelet count < 25 × 10 9 /LAll patients underwent weekly hematologic monitoring from the initiation of concurrent chemoradiotherapy until 4 weeks post-treatment completion, with the lowest recorded blood cell counts (leukocytes, hemoglobin, and platelets) used to determine the grade of myelosuppression. Based on these criteria, patients were categorized into mild myelosuppression (Grade 0–I: leukocyte count > 3.0 × 10 9 /L, hemoglobin > 100 g/L, and platelet count > 75 × 10 9 /L; n = 216) and severe myelosuppression (Grade II–IV: at least one cell line meeting Grade II or higher criteria; n = 84).

All patients underwent postoperative pelvic intensity-modulated radiotherapy (IMRT) using conventional IMRT planning techniques and standard dose constraints for organs at risk, without specific bone marrow sparing techniques. The clinical target volume (CTV) encompassed the vaginal cuff, parametrial tissue, presacral region, and the bilateral common, internal, and external iliac as well as obturator lymph node drainage areas, and included the upper half of the vagina. The planning target volume (PTV) was generated by adding a 5–7 mm margin to the CTV. The prescribed external beam radiation dose ranged from 45 to 50.4 Gy, delivered in 25–28 fractions of 1.8–2.0 Gy each. The dose constraints for organs at risk were as follows: spinal cord Dmax < 40 Gy, bilateral femoral heads V50 < 5%, bladder V50 < 50%, rectum V50 < 50%, and small bowel V50 < 10% and Dmax < 52 Gy. Although dosimetric parameters of the pelvic bone marrow closely associated with myelosuppression (e.g., V10, V20, V40, and Dmean) were not mandated as hard constraints during plan optimization, they were recorded and evaluated within the treatment planning system. Detailed dosimetric data are presented in Table S1 . For patients receiving brachytherapy, high-dose-rate intracavitary brachytherapy was performed using an iridium-192 source in accordance with the GEC-ESTRO guidelines, delivering 2–4 fractions of 5–6 Gy each to the high-risk clinical target volume (HR-CTV). All radiotherapy plans were reviewed and approved by senior radiation oncologists prior to treatment delivery.

For all study participants, the baseline data were collected. The baseline data included Age, clinical staging according to the 2018 International Federation of Gynecology and Obstetrics (FIGO) criteria [ 18 ], Performance Status, status of HPV, pathological type, Surgical approach, maximal tumor diameter, status of LN metastasis, Deep stromal invasion, lymph-vascular space invasion (LVSI), nerves invasions, strategy of radiotherapy and the application of adjuvant chemotherapy.

Radiomic feature extraction

All patients underwent Siemens SOMATOM Drive quantum dual source 64 row CT enhanced localization scan with a scanning layer thickness of 2 mm and reconstruction of 1 mm before treatment. The tube voltage was 100–120 kV and the tube current was 200–380 mA; Extract DICOM format CT data from patients and import enhanced CT images of 300 patients into 3D Slicer software (version 5.2.2). Two attending doctors with more than 5 years of clinical experience independently delineate the ROIs of primary cervical cancers manually with 3D Slicer software. The decision was made after consultation due to differing opinions. The ROI in this study is defined as the clinical target volume (CTV) for radiotherapy following cervical cancer surgery, with all patients having no primary tumour (already surgically removed). Delineation adhered to standard postoperative radiotherapy guidelines, encompassing the common iliac, internal iliac, external iliac, presacral, obturator, and upper half of the vaginal lymphatic drainage regions. This region anatomically encompasses substantial pelvic hematopoietic bone marrow. The delineation boundaries directly determine the bone marrow’s radiation dose and volume, thereby closely correlating with the risk of developing myelosuppression following concurrent chemoradiotherapy [ 19 , 20 ]. Additionally, CT images from 50 randomly selected patients were used to assess the consistency of the two physicians’ ROI delineation using the Dice coefficient.

The features of 300 cervical cancer samples in the ROI of Computed Tomography (CT) were extracted using the PyRadiomics database (v 3.0.1, http://pyradiomics.readthedocs.io ) in Python (v 3.7) [ 21 ], containing 216 mild and 84 severe samples. First, the original CT images undergo preprocessing, with all voxels resampled to an isotropic resolution of 1 × 1 × 1 mm 3 . Subsequently, the images are Z-score normalised to eliminate grey-scale variations between scans. Finally, the normalised images are grey-level discretised at 25 HU intervals to obtain integer grey-level images. Based on the preprocessed images, nine transformations are applied to generate derived images, including: - Original image - Discrete wavelet transform (level 1 decomposition, 8 subbands) Laplacian of Gaussian (LoG) filtering (sigma = 1.0, 2.0, 3.0 mm), square transformation, square root transformation (with prior non-negative range shift), exponential transformation, logarithmic transformation (also with prior shift), local binary pattern (LBP3D), and gradient transformation. Subsequently, radiomics features are extracted from both the original image and each derived image. These encompass shape features, first-order statistical features, and texture features (such as the grey-level co-occurrence matrix (GLCM)). For texture feature extraction, parameters are set as follows: distance of 1 voxel, 13 three-dimensional directions, symmetric matrix, and normalisation applied. Then, Z-score standardization was utilized to normalize the data with a mean of 0 and variance of 1. Subsequently, Welch T-test and Two Sample t-test were conducted on the imaging histological features of patients with mild and severe cervical cancer, with the selection criteria being features with a significance level of p value < 0.05 for subsequently analysis.

Screening of important features

The key steps of the imaging feature analysis process are illustrated in Fig. 2 . For the radiomics features preliminarily screened via Z-score standardisation, Welch’s t-test, and two-sample t-test, we further conducted least absolute shrinkage and selection operator (LASSO) logistic regression analysis using the glmnet software package (version 4.1–7) [ 22 ]. We set the “family” parameter to “binomial” for feature selection and weight adjustment. During the model construction process, we adopted the method of 10-times repeated cross-validation to determine the optimal model. Finally, the Lasso regression selected the Lambda value that minimized the error of the model on the test set. The model corresponding to this value performed the best in terms of balancing model complexity and prediction accuracy. It effectively avoided overfitting and ensured that the model had good predictive ability. In order to remove the reproducible features selected by LASSO, the Spearman correlation coefficient was used to calculate the correlation between the features. Subsequently, to maximize the retention of feature characterization, a greedy recursive deletion strategy for feature filtering was employed, which removed highly relevant independent variables according to a set threshold (cor > 0.9), and important features were ultimately retained.

Acquisition of radiomics feature scores (rad_score)

The 300 cervical cancer samples were divided into training set (mild = 152, severe = 59) and test set (mild = 64, severe = 25) in a ratio of 7:3. Subsequently, a cervical cancer bone marrow suppression response classification imaging model was constructed using eight machine learning algorithms available caret package (1.0.2) [ 23 ], including regularized logistic regression (regLogistic), support vector machine (SVM), random forest (RF), k-nearest neighbors (KNN), Extreme gradient boosting (xgboost), naive Bayes, neural network (nnet), and LASSO. The final parameter settings for each algorithm used in prediction are as follows: RF: n_estimators = 100, max_depth = 10, min_samples_split = 2, min_samples_leaf = 1, random_state = 42; SVM: kernel=“rbf”, C = 1, gamma=“scale”, probability = TRUE, random_state = 42; KNN: n_neighbors = 5, weights=“distance”, metric=“minkowski”; xgboost: n_estimators = 100, max_depth = 6, learning_rate = 0.1, subsample = 0.8, colsample_bytree = 0.8, random_state = 42; Naive Bayes: var_smoothing = 1e-09; regLogistic: C = 1, penalty=“l2”, solver=“lbfgs”, max_iter = 1000, random_state = 42; nnet and LASSO utilise the default parameters of the R package. In addition, 10-fold cross-validation was conducted on the training set to assess the performance of each model, which was evaluated based on the Area Under the receiver operating characteristic (ROC) curves (AUC) value, sensitivity, and specificity.

Following this, based on the models obtained from the training set, ROC curves using pROC package (1.18.0) [ 24 ] and confusion matrices were plotted for each model using the test set to validate the diagnostic efficacy of each model. Additionally, line plots were used to compare the AUC of different models in the training set and test set. Accuracy, sensitivity, specificity, recall, precision, recall and F1 score were calculated for each model to identify the best-performing model. Finally, after comprehensive analysis of all these parameters, the model with the best performance in the test set was selected. The predicted probabilities from the best machine learning model were used as radiomics feature scores (rad_score), and the features of the best model were ranked based on their importance.

Acquisition of clinical feature scores (clinical_score)

Subsequently, Welch Two Sample t-test (heteroscedastic) or chi-square test (categorical data) was employed to compare the clinical features of the patients. Following that, in training set, a cervical cancer post-radiochemotherapy bone marrow suppression response classification clinical model was constructed in a similar manner as in section “ Screening of important features ”, and clinical feature scores (clinical_score) were calculated and the importance of clinical features was ranked.

Construction of a nomogram model

The univariate and multivariate cox logistic regression models for rad_score and clinical_score in the training set were built to screen for independent prognostic factors using the rms package (6.7–0) ( p value < 0.05) [ 25 ]. Subsequently, a nomogram model was constructed using the rms package based on the aforementioned multivariate cox logistic regression analysis. The calibration curve, ROC curve, and decision curve were plotted in the training set based on the nomogram predictive model to evaluate the predictive performance of the model. To further validate the model’s generalisation ability, we repeated the above evaluation process on an independent test set.

Construction and comparison of dosimetric models and combined models

To evaluate the incremental predictive value of radiomics scores relative to conventional dosimetric parameters, we constructed a dosimetric model and a combined model based on random forests within the training set, comparing them with the radiomics model. All models employed the identical random forest algorithm (with parameter settings as per section “ Acquisition of radiomics feature scores (rad_score) ”). The input features for the dosimetric models comprised four conventional dosimetric parameters (Dmean, V10, V20, V40); the input features for the combined model comprised the Rad_Score alongside the aforementioned four dosimetric parameters. Model performance was assessed using the AUC and its 95% confidence interval.

Results

Feature selection and correlation analysis in radiomics data

To assess the consistency of ROI delineation, we randomly selected 50 patients and had two attending physicians independently delineate the target regions. The Dice coefficient was calculated, revealing an average coefficient exceeding 0.91 ± 0.05 (range: 0.78–0.94), indicating excellent delineation consistency. Subsequently, radiological features were extracted from all patients’ CT images, yielding a total of 1,834 imaging histologic features were obtained, encompassing 360 first-order features, 14 shape features, and 1,460 textural features, which were classified into 7 types (glcm, firstorder, glrlm, glszm, glcm, firstorder, glrlm, glszm, gldm, ngtdm, shape) (Figs. 3 A–B, Table 1 ). Then, when lambda. min is equal to 0.0333, 15 strongly correlated features were screened through LASSO (Figs. 3 C–D). Following the LASSO selection of 15 features that may exhibit redundancy, Spearman correlation coefficients were employed to assess inter-feature relationships. Among them, the logarithm_glcm_Correlation and squareroot_glcm_Correlation exhibited a strong positive correlation between them (Fig. 3 E). Then, employing a greedy recursive elimination strategy with cor > 0.9, highly correlated variables (lbp_3D_m2_firstorder_Maximum) were removed, resulting in the retention of 14 important features (Fig. 3 F).

RF algorithm was machine learning model in imaging model

First, the samples were divided into a training set (mild = 152, severe = 59) and a test set (mild = 64, severe = 25) at a ratio of 7:3. In the training set, based on the 14 important features mentioned above, a classification imaging model for bone marrow suppression response after cervical cancer radiotherapy and chemotherapy was further constructed using 8 machine learning algorithms. It was found that naive bayes had the highest AUC value and sensitivity, while RF had the highest specificity (Figs. 4 A–B). Subsequently, the performance of each model was evaluated in the test set, revealing that the RF model had the highest AUC value (AUC = 0.735 (95% CI = 0.624–0.846)), and could effectively distinguish between mild and severe samples (Figs. 4 C–D). Additionally, the RF model showed high AUC values in both datasets (Fig. 4 E). Furthermore, the RF model exhibited the highest scores for various parameters in the test set, except for specificity and precision, with AUC = 0.735 (95% CI = 0.624–0.846), accuracy = 0.719 (95% CI = 0.618–0.802), sensitivity = 0.703 (95% CI = 0.507–0.845), recall = 0.703 (95% CI = 0.507–0.845), and F1 score = 0.782 (95% CI = 0.513–1.000) (Table 2 ). After considering these parameters comprehensively, the RF algorithm was determined to be the best machine learning algorithm to conduct model. Finally, based on the RF algorithm, the feature importance was ranked, revealing that wavelet_LHL_glcm_Idmn was the most important (Fig. 4 F).

Furthermore, we compared the predictive performance of radiomics scores against conventional dosimetric parameters (Dmean, V10, V20, V40). The dosimetric model based on RF achieved an AUC of 0.478 (95% CI: 0.344–0.612), indicating limited predictive efficacy of conventional dosimetric indicators alone. The radiomics score model significantly outperformed the dosimetric model. The combined model yielded an AUC of 0.624 (95% CI: 0.501–0.746), which is presumed to be associated with the introduction of noise from incorporating weakly predictive features (Table S1 – S2 ).

RF algorithm was machine learning model in clinical model

Subsequently, it was further analyzed the differences in clinical characteristics of patients between the training and test sets using two different statistical methods, and a significant difference was found in adjuvant chemotherapy (adjuvant CHT) between the two groups ( p = 0.009) (Table S2 ). After that, an oncotherapy-induced bone marrow suppression classification clinical model for cervical cancer was further constructed using eight machine learning algorithms. The results revealed that in the training set, RF exhibited the highest AUC value and specificity, whereas nnet demonstrated the highest sensitivity (Figs. 5 A–B). Lastly, it was observed that the important feature predicted based on the RF algorithm was PLT (Fig. 5 C).

Nomogram model for bone marrow suppression prediction evaluation

In training set, it was found that both rad_score and Clinical_Score, predicted by the best model obtained above, were significantly associated ( p value < 0.05) in both the univariate and multivariate logistic regression models (Table S3 – S4 , Figs. 6 A–B). Based on Clinical_Score and rad_score, a nomogram model was constructed to predict the likelihood of bone marrow suppression reaction in patient samples within training set (Fig. 6 C). Subsequently, the predictive performance of the nomogram was evaluated. The slopes of the calibration curves were close to 1 in both datasets, indicating acceptable predictive performance of the model (Fig. 6 D). Additionally, the AUC values of the nomogram model was 0.724 (95% CI = 0.650–0.798), suggesting a strong fit of the nomogram model (Fig. 6 E). According to the Decision Curve Analysis (DCA）for the model, the higher net benefit rate was observed for the nomogram model (Fig. 6 F). The model also performed well on the test set, with a calibration curve slope close to 1 and an AUC of 0.712 (95% CI = 0.631–0.793), demonstrating discriminatory power comparable to that observed in the training set. DCA analysis indicated that, within a certain range of threshold probabilities, the nomogram continues to provide a positive net benefit (Supplementary Figure 1 ).

Discussion

Cervical cancer ranks as the fourth most common cancer in women. Concurrent chemoradiotherapy (CCRT) is the preferred treatment for locally advanced cases, reducing recurrence/metastasis risks but significantly increasing bone marrow suppression (BM) risk [ 26 , 27 ]. BM often causes treatment interruption, anemia, infections, and life-threatening neutropenic fever [ 19 ]. Existing prediction models for BM induced by CCRT are limited in accuracy, and traditional prognostic factors have declining clinical utility. Meng et al. constructed a nomogram incorporating histological grade, pre-radiotherapy chemotherapy, and radiotherapy modality (CCRT vs. IMRT) to predict acute hematologic toxicity. Traditional risk factors like age and BMI showed limited accuracy (AUC = 0.67) [ 28 ]. Therefore, it is very important to construct a comprehensive prediction model that combines radiomics features with clinical features. This study developed a radiomics-based predictive model using CT images to assess BM risk in cervical cancer patients undergoing chemoradiotherapy. By extracting texture, morphological, and spatial features from tumor and bone marrow regions, radiomics technology uncovers hidden imaging biomarkers to complement traditional models.

After screening in the training set and verifying in the validation set, we selected the Random Forest algorithm, which has the best performance among the machine learning methods, and used it to predict probability as the radiomics feature score (rad_score). The random forest algorithm has been recognized by many scholars for its powerful data processing capabilities. Many studies have reported that its predictive accuracy for disease prognosis is higher than traditional logistic regression [ 29 – 32 ].

Previous studies have established prediction models for lung cancer and esophageal cancer based on random forest algorithms, with prediction accuracy, sensitivity, and specificity exceeding 70%. These results demonstrated the random forest algorithms was the good predictive ability for bone marrow suppression in cancer chemotherapy patients [ 33 , 34 ]. In our research queue, the results are basically consistent with previous studies. The ROC AUC of the Random Forest model is 0.74, The accuracy is 0.72, the Sensitivity is 0.70, The specificity is 0.76, The precision is 0.88, the recal is 0.70, and F1 is 0.78. Thus, the Random Forest algorithm is the best machine learning method for predicting bone marrow suppression after radiotherapy and chemotherapy in cervical cancer surgery, and used to calculate the radiomics score of the model. Furthermore, feature importance ranking based on the Random Forest (RF) algorithm revealed that wavelet_LHL_glcm_Idmn was the most significant predictor. This feature represents the Inverse Difference Moment Normalized (Idmn) derived from the Gray-Level Co-occurrence Matrix (GLCM) and is primarily used to quantify local textural uniformity. Previous studies have demonstrated that CT texture analysis can effectively assess the microstructural status of bone marrow. For instance, in patients with multiple myeloma, CT texture parameters have been shown to correlate closely with the degree of bone marrow infiltration [ 35 ]. Additionally, in the context of cervical cancer, pretreatment CT-based radiomics has successfully predicted hematologic toxicity following chemoradiotherapy, suggesting that CT imaging can capture information regarding bone marrow functional reserve [ 36 ]. Therefore, wavelet_LHL_glcm_Idmn may reflect the heterogeneity of the pelvic bone marrow microenvironment, such as fatty infiltration or alterations in cellular density, and may thereby be associated with the risk of myelosuppression after chemoradiotherapy.

Clinical-radiomics model is an emerging interdisciplinary field in the medical field that combines knowledge from medical imaging, computer science, and statistics [ 37 ]. Its aim is to extract valuable information through the analysis and mining of a large amount of medical imaging data to assist clinical decision-making and personalized treatment [ 38 ]. Some published studies have explored the relationship between the clinical-radiomics model and cervical cancer lymph node metastasis, staging prediction or deep stromal invasion [ 39 – 42 ].

François Lucia et al. applied 18F-FDG PET/CT and MRI radiomics, combined with clinical parameters, to establish a machine learning model to predict the involvement of para aortic lymph nodes (PALN) in locally advanced cancer (LACC) patients before radiotherapy and chemotherapy (CRT). The final research results indicate that compared to clinical models, The clinical radio model can better predict the involvement of para aortic lymph nodes in cervical cancer before radiotherapy and chemotherapy [ 43 ]. In previous studies on CC, many researchers have also used machine learning, combining clinical and radiomics, to construct predictive models of CC under different conditions. For example, Huang et al. constructed a clinical radiomics model that outperformed the clinical model for predicting cervical cancer staging, thereby facilitating clinical decision-making [ 44 ]. To predict deep mesenchymal infiltration in patients with early-stage CC, Yan et al. also constructed a predictive model based on clinical and radiomics data, which had good predictive performance [ 45 ]. Therefore, this study included clinical features of cervical cancer, screened for differences in bone marrow suppression levels, and once again found that the clinical model constructed by random forest had the highest AUC value and specificity, while Nnet had the highest sensitivity. The random forest algorithms was used to predict probability as a clinical feature score.

The most important clinical features selected in this study were frequency of radiotherapy, parametrial invasion, type of pathology, adjuvant chemotherapy, and age. The frequency of radiotherapy was closely related to bone marrow suppression. Chemoradiotherapy not only directly damaged granulocytes, but also caused chromosomal mutations in hematopoietic cells, accelerated the apoptosis rate of blood cells, and reduced the content of blood cells [ 46 ]. And previous study reported that radiation time, dose, range, frequency etc. were closely related to the degree of bone marrow damage [ 47 ]. Adenocarcinoma, parametrial invasion, and nerve infiltration are risk factors for postoperative radiotherapy of cervical cancer, which are directly related to the prognosis of cervical cancer [ 48 ]. Age was also a factor affecting bone marrow suppression in cervical cancer patients. Studies have shown that the probability of bone marrow suppression in cervical cancer patients over 65 years old after radiotherapy and chemotherapy was significantly higher than in patients under 65 years old [ 49 ].

Subsequently, a nomogram was constructed by integrating radiomics signature scores with clinical feature scores to predict the risk of myelosuppression after radiotherapy for cervical cancer. Calibration curves and decision curve analysis demonstrated that the model possessed a certain predictive capability. Based on the model’s predictions, corresponding intervention strategies could be implemented in clinical practice. For example, for low‑risk patients, standard radiotherapy plans are sufficient; for high‑risk patients, stricter dose constraints on the pelvic bone marrow (e.g., exploring lower V10/V20 constraints) could be applied during plan optimization while maintaining target coverage as much as possible. Alternative strategies include adjusting the concurrent chemotherapy regimen (e.g., reducing cisplatin dose) or prophylactic use of hematopoietic growth factors. However, this study has not yet prospectively validated these strategies in conjunction with the model’s predictions, nor has it clearly defined optimal dose‑constraint thresholds based on risk stratification. Therefore, the clinical translation pathway of this model remains incomplete, and future studies should establish a research design that links prediction with intervention to verify its true value in guiding individualized treatment.

This study used multiple machine learning algorithms to classify and model the omics features of images and combined with logistic regression models to construct a comprehensive prediction model, fully utilizing the advantages of different algorithms to improve the accuracy and stability of predictions. In terms of model evaluation, various indicators such as ROC, accuracy, sensitivity, specificity, recall, and F1 value were used to evaluate model performance. However, this study has certain limitations. First, the radiomic feature selection process, which included LASSO regression, correlation analysis, and greedy elimination, was performed on the complete dataset prior to splitting the data into training and testing sets. This means that information from the testing set was indirectly incorporated during the feature selection stage, which may lead to overestimation of the test set performance metrics and fail to fully reflect the model’s generalizability to independent real-world data. It should be noted that the primary purpose of the testing set in this study was to compare the performance of different machine learning algorithms using the same set of 14 features, rather than to validate the generalizability of a fixed model independently. Consequently, while the comparison among algorithms remains fair, the predictive performance of the models may still be overestimated. Second, constrained by the sample size, particularly the limited number of cases in the severe myelosuppression group ( n = 84), the generalizability and stability of the models were restricted, posing risks of either overfitting or underfitting. Third, this study employed a single-center retrospective design and lacked an independent external validation cohort. Given that radiomic features are highly sensitive to variations in CT scanners and acquisition protocols, this further limits the generalizability of the models. Finally, decision curve analysis of the nomogram model showed that the net clinical benefit of all models approached zero at medium-to-high risk thresholds (>0.4), suggesting that the models have limited clinical value in the high-risk threshold range. Therefore, future multi-center collaborations are needed to expand the sample size and validate the external applicability of the model, along with the design of prospective studies to avoid information leakage and more accurately assess predictive performance, with a particular focus on evaluating the stability of the nomogram model across different risk thresholds. Concurrently, we aim to integrate deep learning and multimodal fusion strategies, combining clinical and genomic information, to ultimately develop robust and interpretable clinical decision support tools.

Conclusions

This study developed a computed tomography (CT)-based radiomics model, integrated with clinical features, to predict myelosuppression in patients with cervical cancer undergoing concurrent chemoradiotherapy. Through radiomics analysis and machine learning methods, we identified imaging features associated with myelosuppression and constructed a predictive model. A nomogram integrating clinical parameters and the radiomics signature demonstrated favorable predictive performance in the internal validation cohort. However, given the single-center retrospective design and the lack of external validation, the generalizability of this model remains to be verified. Future research should focus on multi-center prospective validation, standardization of imaging protocols, and further optimization of feature selection methods to enhance the robustness and clinical applicability of the model.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Supplementary Material 2

Supplementary Material 3

Abbreviations

Concurrent chemoradiotherapy

Cervical cancer

Computed tomography

Region of interest

Random forest

Magnetic resonance imaging

Positron emission tomography

Least absolute shrinkage and selection operator

Ethics committee in clinical research

Performance status

Federation of gynecology and obstetrics

Lymph-vascular space invasion

Laplacian of Gaussian

Support vector machine

K-nearest neighbors

Receiver operating characteristic

Decision curve analysis

Bone marrow

Para aortic lymph nodes

Author contributions

Conceptualization, Chen Sun and Alan Chu; Data curation, Jinghui Yang; Formal analysis, Shijia Liu; Funding acquisition, Mengxi Li,Shijia Liu,Xin Wang and Alan Chu; Investigation, Rui Song; Methodology, Xin Wang,Zongwen Liu; Project administration, Chen Sun; Resources, Xiao Liu; Software, Mengxi Li, Xin Wang; Supervision, Zongwen Liu; Validation, Chen Sun, Xin Wang and Alan Chu; Visualization, Chen Sun and Alan Chu,Mengxi Li; Writing – original draft, Chen Sun; Writing – review & editing, Xin Wang,Mengxi Li, Shijia Liu, Rui Song, Jinghui Yang, Xiao Liu, Lanlan Gan, Yongtai Wang, Zongwen Liu and Alan Chu.

Funding

This research was funded by Key Research and Promotion Special Project of Henan Province, grant number (262102311184) and (262102520006). Joint Construction Project of Medical Science and Technology in Henan Province, grant number, (LHGJ20230325) and (LHGJ20250360). Zhengzhou City Healthcare Sector Guided Science and Technology Innovation Program, grant number, (2025YLZDJH356).

Data availability

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

Declarations

Ethics approval and consent to participate

The study protocol was reviewed and approved by the Ethics Committee in Clinical Research (ECCR) of the Second Affiliated Hospital of Zhengzhou University. The IRB number was 4,101,035,716,743，the protocol code was KY2024162.It was conducted following the Declaration of Helsinki.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s Note

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

Chen Sun, Alan Chu, Xin Wang and Mengxi Li contributed equally to this work and share first authorship.

Contributor Information

Xin Wang, Email: wangxin1117ico@gmail.com.

Mengxi Li, Email: lmx60333@126.com.

References

1. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clinicians. 2021;71(3):209–49. 10.3322/caac.21660.

2. Moore DH. Cervical cancer. Obstet Gynecol. 2006;107(5):1152–61. 10.1097/01.AOG.0000215986.48590.79.

3. Yang J, Yin J, Yan G, Huang D, Wang J. Postoperative chemoradiotherapy versus radiotherapy alone for cervical cancer: a systematic review and meta-analysis. J Obstet Gynaecol. 2016;36(5):641–48. 10.3109/01443615.2015.1134458.

4. Corbeau A, Kuipers SC, de Boer SM, Horeweg N, Hoogeman MS, Godart J, et al. Correlations between bone marrow radiation dose and hematologic toxicity in locally advanced cervical cancer patients receiving chemoradiation with cisplatin: a systematic review. Radiother Oncol. 2021;164:128–37. 10.1016/j.radonc.2021.09.009.

5. Rowe SP, Nair S, Bomanji JB, et al. CT radiomics predicts radium-223-induced myelosuppression in metastatic castration-resistant prostate cancer. Eur J Nucl Med Mol Imag. 2022;49(10):3351–61.

6. Xu C, Liu W, Zhao Q, Zhang L, Yin M, Zhou J, et al. CT-based radiomics nomogram for overall survival prediction in patients with cervical cancer treated with concurrent chemoradiotherapy. Front Oncol. 2023;13:1287121. 10.3389/fonc.2023.1287121.

7. Zhou X, Li Y, Lu F, Yin Y. Applying logistic LASSO regression for the diagnosis of atypical Crohn’s disease. Fudan Univ J Med Sci. 2017;44(3):385–90.

8. Guo Y, Wu J, Wang Y, Jin Y. Development and validation of an ultrasound-based radiomics nomogram for identifying HER2 status in patients with breast carcinoma. Diagn (Basel). 2022;12(12):3075. 10.3390/diagnostics12123130.

9. Deng B, Kong W, Han C, Zhou C, Li J, Song D, et al. Study of long-term effects of pelvic radiotherapy on the function of bone marrow in recurrent cervical cancer patients. Int J Med Sci. 2024;21(10):2000–10. 10.7150/ijms.95900.

10. Mell LK, Kochanski JD, Roeske JC, Haslam JJ, Mehta N, Yamada SD, et al. Dosimetric predictors of acute hematologic toxicity in cervical cancer patients treated with concurrent cisplatin and intensity-modulated pelvic radiotherapy. Int J Multiling Radiat Oncol*Biol*Phys. 2006, Dec, 1;66(5):1356–65. 10.1016/j.ijrobp.2006.03.018.

11. Albuquerque K, Giangreco D, Morrison C, Siddiqui M, Sinacore J, Potkul R, et al. Radiation-related predictors of hematologic toxicity after concurrent chemoradiation for cervical cancer and implications for bone marrow–sparing pelvic IMRT. Int J Multiling Radiat Oncol*Biol*Phys. 2011, Mar, 15;79(4):1043–47. 10.1016/j.ijrobp.2009.12.025.

12. Xiang X, Ding Z, Zeng Q, Feng L, Qiu C, Chen D, et al. Dosimetric parameters and absolute monocyte count can predict the prognosis of acute hematologic toxicity in cervical cancer patients undergoing concurrent chemotherapy and volumetric-modulated arc therapy. Radiat Oncol. 2022, Mar, 5;17(1):48. 10.1186/s13014-022-02018-1.

13. Lei Y, Bai H, Yu J, Zhang Z, Wang L, Li B, et al. Research progress of dosiomics in precision radiotherapy. J Cancer Res Ther. 2025, Sep, 1;21(4):787–95. 10.4103/jcrt.jcrt_132_25.

14. Yue H, Li X, You J, Feng P, Du Y, Wang R, et al. Acute hematologic toxicity prediction using dosimetric and radiomics features in patients with cervical cancer: does the treatment regimen matter? Front Oncol. 2024;14:1365897. 10.3389/fonc.2024.1365897.

15. Liu Y, Guo L, Wang Y, Xu Q, Zhang J, Meng X. Radiation hematologic toxicity prediction in rectal cancer: a comparative radiomics-based study on CT image and dose map. Front Oncol. 2025;15:1516855. 10.3389/fonc.2025.1516855.

16. Ren K, Shen L, Qiu J, Sun K, Chen T, Xuan L, et al. Treatment planning computed tomography radiomics for predicting treatment outcomes and haematological toxicities in locally advanced cervical cancer treated with radiotherapy: a retrospective cohort study. BJOG: Int J Obstet Gynaecol. 2023;130(2):222–30. 10.1111/1471-0528.17285.

17. Abu-Rustum NR, Yashar CM, Arend R, Barber E, Bradley K, Brooks R, et al. NCCN Guidelines® insights: cervical cancer, version 1. 2024. J Natl Compr Cancer Network: JNCCN. 2023;21(12):1224–33. 10.6004/jnccn.2023.0062.

18. Wright JD, Matsuo K, Huang Y, Tergas AI, Hou JY, Khoury-Collado F, et al. Prognostic performance of the 2018 international Federation of gynecology and obstetrics cervical cancer staging Guidelines. Prognostic Perform The 2018 Int Fed Gynecology Obstet Cerv Cancer Staging Guidelines. Obstet Gynecology. 2019;134(1):49–57. 10.1097/AOG.0000000000003311.

19. Vitzthum LK, Heide ES, Park H, Williamson CW, Sheridan P, Huynh-Le MP, et al. Comparison of hematologic toxicity and bone marrow Compensatory response in Head and Neck vs. Cervical cancer patients undergoing chemoradiotherapy. Front Oncol. 2020, Jul, 21;10:1179. 10.3389/fonc.2020.01179.

20. Shao L, Luo Y, Zhou D. Hematopoietic stem cell injury induced by ionizing radiation. Antioxid Redox Signal. 2014, Mar, 20;20(9):1447–62. 10.1089/ars.2013.5635. 2014 Feb 10.

21. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational radiomics System to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104–07. 10.1158/0008-5472.CAN-17-0339.

22. Li Y, Lu F, Yin Y. Applying logistic LASSO regression for the diagnosis of atypical Crohn’s disease. Sci Rep. 2022;12(1):11340. 10.1038/s41598-022-15609-5.

23. Zhang Z, Zhao Y, Canes A, Steinberg D, Lyashevska O. Predictive analytics with gradient boosting in clinical medicine. Annals of translational medicine 2019. Ann. Transl. Med. 2019;7(7):152. 10.21037/atm.2019.03.29.

24. Yan P, Ke B, Song J, Fang X. Identification of immune-related molecular clusters and diagnostic markers in chronic kidney disease based on cluster analysis. Front Genet. 2023;14:1111976. 10.3389/fgene.2023.1111976.

25. Sachs MC. plotROC: a tool for plotting ROC curves. J Stat Soft. 2017;79(Code Snippet 2):79. 10.18637/jss.v079.c02.

26. Burmeister CA, Khan SF, Schäfer G, Mbatani N, Adams T, Moodley J, et al. Cervical cancer therapies: current challenges and future perspectives. Tumour Virus Res. 2022;13:200238. 10.1016/j.tvr.2022.200238.

27. Mayadev JS, Ke G, Mahantshetty U, Pereira MD, Tarnawski R, Toita T. Global challenges of radiotherapy for the treatment of locally advanced cervical cancer. Int J Gynecological Cancer: Off J The In-Ternational Gynecological Cancer Soc. 2022;32(3):436–45.22. 10.1136/ijgc-2021-003001.

28. Meng X, Wang N, Yu M, Kong D, Zhang Z, Chang X, et al. Development of a nomogram for predicting grade 2 or higher acute hematologic toxicity of cervical cancer after the pelvic bone marrow sparing radiotherapy. Front Public Health. 2022;10:993443. 10.3389/fpubh.2022.993443.

29. Liu D, Zhang X, Zheng T, Shi Q, Cui Y, Wang Y, et al. Optimisation and evaluation of the random forest model in the efficacy prediction of chemoradiotherapy for advanced cervical cancer based on radiomics signature from high-resolution T2 weighted images. Arch Gynecol Obstet. 2021;303(3):811–20. 10.1007/s00404-020-05908-5.

30. Guo Y, Wu J, Wang Y, Jin Y. Development and validation of an ultrasound-based radiomics nomogram for identifying HER2 status in patients with breast carcinoma. Diagn (Basel, Switz). 2022;12(12):3130. 10.3390/diagnostics12123130.

31. S SN. Prediction of breast cancer through random forest. CMIR. 2023;19(10):1144–55. 10.2174/1573405618666220930150625.

32. Jin Y, Lan A, Dai Y, Jiang L, Liu S. Development and testing of a random forest-based machine learning model for predicting events among breast cancer patients with a poor response to neoadjuvant chemotherapy. Eur J Med Res. 2023;28(1):394. 10.1186/s40001-023-01361-7.

33. Dong Y, Hu C, Liu J, Lv H. Construction of an auxiliary scoring model for myelosuppression in patients with lung cancer chemotherapy based on random forest algorithm. Am J Transl Res. 2023;15(6):4155–63.

34. Zheng Z, Zhang Q, Han Y, Wu T, Zhang Y. Predictive model of chemotherapy-induced myelosuppression for patients with esophageal cancer. Cancer Control: J Moffitt Cancer Cent: J The Moffitt Cancer Cent. 2022;29:10732748221126929. 10.1177/10732748221126929.

35. Reinert CP, Krieg E, Esser M, Nikolaou K, Bösmüller H, Horger M. Role of computed tomography texture analysis using dual-energy-based bone marrow imaging for multiple myeloma characterization: comparison with histology and established serologic parameters. Eur Radiol. 2021, Apr;31(4):2357–67. 10.1007/s00330-020-07320-8.

36. Ren K, Shen L, Qiu J, Sun K, Chen T, Xuan L, et al. Treatment planning computed tomography radiomics for predicting treatment outcomes and haematological toxicities in locally advanced cervical cancer treated with radiotherapy: a retrospective cohort study. BJOG. 2023, Jan;130(2):222–30. 10.1111/1471-0528.17285.

37. Avanzo M, Wei L, Stancanello J, Vallières M, Rao A, Morin O, et al. Machine and deep learning methods for radiomics. Med Phys. 2020;47(5):e185–202. 10.1002/mp.13678.

38. Li X, Zhang L, Yang J, Teng F. Role of Artificial Intelligence in medical image analysis: a review of Current trends and future directions. J Med Biol Eng. 2024;44(2):231–43. 10.1007/s40846-024-00863-x.

39. Yu Y, He Z, Ouyang J, Tan Y, Chen Y, Gu Y, et al. Magnetic resonance imaging radiomics predicts preoperative axillary lymph node metastasis to support surgical decisions and is associated with tumor microenvironment in invasive breast cancer: a machine learning, multicenter study. EBioMedicine. 2021;69:103460. 10.1016/j.ebiom.2021.103460.

40. Huang YQ, Liang CH, He L, Tian J, Liang CS, Chen X, et al. Development and validation of a radiomics nomogram for Preoperative prediction of lymph node metastasis in colorectal cancer. JCO. 2016;34(18):2157–64. 10.1200/JCO.2015.65.9128.

41. Dong D, Fang MJ, Tang L, Shan XH, Gao JB, Giganti F, et al. Deep learning radiomic nomogram can predict the number of lymph node metastasis in locally advanced gastric cancer: an international multicenter study. Ann Oncol. 2020;31(7):912–20. 10.1016/j.annonc.2020.04.003.

42. Bian Y, Zheng Z, Fang X, Jiang H, Zhu M, Yu J, et al. Artificial Intelligence to predict lymph node metastasis at CT in pancreatic Ductal Adenocarcinoma. Artif. Intell To Predict Lymph Node Metastasis CT Pancreat Ductal Adenocarcinoma. Radiol. 2023;306(1):160–69. 10.1148/radiol.220329.

43. Lucia F, Bourbonne V, Pleyers C, Dupré PF, Miranda O, Visvikis D, et al. Multicentric development and evaluation of 18F-FDG PET/CT and MRI radiomics models to predict para-aortic lymph node involvement in locally advanced cervical cancer. Eur J Nucl Med Mol Imag. 2023;50(8):2514–28. 10.1007/s00259-023-06180-w.

44. Huang Q, Deng B, Wang Y, Shen Y, Hu X, Feng C, et al. Reduced field-of-view DWI‑derived clinical–radiomics model for the prediction of stage in cervical cancer. Insights Imag. 2023;14(1):18. 10.1186/s13244-022-01346-w.

45. Yan H, Huang G, Yang Z, Chen Y, Xiang Z. Machine learning-based multiparametric Magnetic resonance imaging radiomics model for Preoperative predicting the deep stromal invasion in patients with early cervical cancer. J Digit Imag Inf Med. 2024;37(1):230–46. 10.1007/s10278-023-00906-w.

46. Tangjitgamol S, Katanyoo K, Laopaiboon M, Lumbiganon P, Manusirivithaya S, Supawattanabodee B. Adju-vant chemotherapy after concurrent chemoradiation for locally advanced cervical cancer. The Cochrane Database Sys-Tematic Rev. 2014;2014(12):Cd010401.

47. Zhong L, Li K, Song L, Yin R. The effect of consolidation chemotherapy after concurrent chemoradiation on the prognosis of locally advanced cervical cancer: a systematic review and meta-analysis. J Obstet Gynaecol. 2022;42(5):830–37. 10.1080/01443615.2021.2012437.

48. Luo HC, Lin GS, Liao SG, Wang FM, Cheng HH, Feng J, et al. Cervical cancer treated with reduced-volume intensity-modulated radiation therapy base on Sedlis criteria (NCCN VS RTOG). The Br J Radiol. 2018;91(1081):20170398. 10.1259/bjr.20170398.

49. Huang J, Gu F, Ji T, Zhao J, Li G. Pelvic bone marrow sparing intensity modulated radiotherapy reduces the incidence of the hematologic toxicity of patients with cervical cancer receiving concurrent chemoradiotherapy: a single-center prospective randomized controlled trial. Radiat Oncol. 2020;15(1):180. 10.1186/s13014-020-01606-3.

Associated Data

Supplementary Materials

Supplementary Material 1

Supplementary Material 2

Supplementary Material 3

Data Availability Statement

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