In this study, we developed and validated various radiomics-based machine learning diagnostic models for predicting lymph node metastasis in cervical cancer using multiparametric magnetic resonance imaging (MRI). Leveraging preoperative MRI with multiple parameters, we extracted 4065 imaging features, ultimately selecting 18 features and constructing multiple machine learning diagnostic models. The XGBoost model demonstrated superior performance in predicting lymph node metastasis in the validation set, with an accuracy of 0.9268 and an AUC of 0.8969, outperforming other models.
Previous studies have indicated that radiomics analysis can extract quantitative image features from medical images, thereby improving tumor diagnosis, staging, and prognosis(Lambin et al. 2017; Fang et al. 2020; Yuan et al. 2021). Fang et al.'s research(Fang et al. 2020) demonstrated that radiomics scores derived from MRI as prognostic biomarkers for early-stage cervical cancer patients. Kan et al.(Kan et al. 2019) utilized T2WI and ceT1WI to extract radiomic features and employed an SVM model(AUC(95% CI), 0.754(0584–0.924)) to predict lymph node metastasis in early-stage cervical cancer. However, some studies only utilized features extracted from a single MRI image(Xu et al. 2019; Yamada et al. 2019), or performed radiomics analysis solely on 2D slices(Ytre-Hauge et al. 2018). In contrast, our approach involved using three common 3D MRI sequences in cervical cancer patients for feature extraction, allowing for the extraction of rich radiomic features and enhancing the model's reliability. Moreover, we systematically compared multiple machine learning models and identified the best-performing model, showing consistent and clinically excellent performance between the training and validation sets (AUC = 0.897 in the validation set), which suggests the excellence and reliability of our research methodology.
Our research utilized a diverse set of machine learning algorithms, such as LR, Naïve Bayes, SVM, AdaBoost and XGBoost. XGBoost is an efficient implementation of the widely-used gradient boosting decision tree algorithm in the field of science. XGBoost is an efficient gradient boosting algorithm that iteratively trains decision trees, incorporates regularization and parallel processing, and constructs a powerful ensemble model suitable for regression and classification tasks. In various machine learning problems, especially in scenarios with a large number of features, high dimensions, and complex structures, XGBoost often exhibits excellent performance. In another study, based on another way of model construction, Sheng et al.(Sheng et al. 2022) found the XGBoost model demonstrated a better performance in Invasive ductal breast cancer molecular subtype prediction, especially in the triple-negative and non-triple-negative groups with an AUC of 0.903. Song(Song 2021) used the XGBoost algorithm to build a radiomics model for predicting axillary LNM in invasive ductal breast cancer with an AUC of 0.890. In our study, XGBoost model demonstrated superior prediction performance, achieving an AUC of 0.897, surpassing the SVM model which obtained an AUC of 0.791.
Preceding investigations have elucidated the prospective efficacy of radiology in prognosticating the survival outcomes of individuals with cervical cancer, boasting superior accuracy than traditional clinical parameters(Liu et al. 2022; Zheng et al. 2022). Our study aimed to evaluate the prognostic significance of MRI-derived radiological characteristics in cervical cancer patients, with whom higher rad-score exhibit lower overall survival rates (HR 1.9, 95% CI 1.2-3.0, P = 0.006). The precise anticipation of disease progression via radiology holds pivotal importance in clinical settings, guiding the selection of optimal therapeutic strategies and thereby enhancing patient prognoses.
Despite the positive results of this research, it is also fundamental to consider the limitations of this study. This study is retrospective and conducted in a single center, lacking external validation, which may introduce selection bias. To enhance the robustness of the models, a larger sample size from multiple centers is necessary. Furthermore, a manual segmentation approach is time-consuming and may impact the precision of feature extraction in certain cases. Previous research has indicated that utilizing an automatic segmentation method could streamline radiomics applications in clinical settings, promoting high levels of intra- and inter-observer reproducibility(Liu et al. 2019; Chen et al. 2020).