ConvXGB: A novel deep learning model to predict recurrence risk of early-stage cervical cancer following surgery using multiparametric MRI images

Abstract

Background: Accurate estimation of recurrence risk for cervical cancer plays a pivot role in making individualized treatment plans. We aimed to develop and externally validate an end-to-end deep learning model for predicting recurrence risk in cervical cancer patients following surgery by using multiparametric MRI images.

Methods: The clinicopathologic data and multiparametric MRI images of 406 cervical cancer patients from three institutions were collected. We designed a novel deep learning model called "ConvXGB" for predicting recurrence risk by combining the convolutional neural network (CNN) and eXtreme Gradient Boost (XGBoost). The predictive performance of the ConvXGB model was evaluated using time-dependent area under curve (AUC), compared with the deep learning radio-clinical model, clinical model, conventional radiomics nomogram and an existing histology-specific tool. The potential of the ConvXGB model in predicting the recurrence-free survival (RFS) and overall survival (OS) was assessed.

Results: The ConvXGB model outperformed other models in predicting recurrence risk, with AUCs for 1 and 3 year-RFS of 0.872(95% CI, 0.857-0.906) and 0.882(95% CI, 0.860-0.904) respectively in the test cohort. This model showed better discrimination, calibration and clinical utility. Grad-CAM analysis was adopted to help clinicians better understand the predictive results. Moreover, Kaplan-Meier survival analysis revealed that patients who were stratified into high-risk group by the ConvXGB model were significantly susceptible to higher cumulative recurrence risk rates and worse outcome.

Conclusion: The ConvXGB model allowed for predicting postoperative recurrence risk in cervical cancer patients and for stratifying the risk of RFS and OS.

Keywords: Cervical cancer; Deep learning; MRI scan; Recurrence-free survival;Overall survival.

Fig. 1

Flow diagram of study design. Abbreviations: MRI, magnetic resonance imaging; CNN, convolutional neural network; XGBoost, eXtreme Gradient Boost; ConvXGB, a novel architecture by combing CNN and XGBoost; ConvXGB&clin, a combined model in which independent clinical features were incorporated into a prediction layer of the ConvXGB; RFS, recurrence-free survival; OS, overall survival; ROC, receiver operator characteristic curve.

Fig. 2

The workflow of model development in the current study. Abbreviations: MRI, magnetic resonance imaging; CNN, convolutional neural network; XGBoost, eXtreme Gradient Boost; ConvXGB, a novel architecture by combing CNN and XGBoost; ConvXGB&clin, a combined model in which independent clinical features were incorporated into a prediction layer of the ConvXGB; T2WI/FS, T2-weighted imaging with fat suppression; ADC, apparent diffusion coefficient; ROI, region of interests; FIGO, Federation of Gynecology and Obstetrics; ICC, intra- and inter-class correlation coefficients; LASSO, least absolute shrinkage and selection operator.

Fig. 3

Model evaluation and interpretation. Time-dependent ROC for the ConvXGB, ConvXGB&clin, clinical model and radiomics nomogram in the derivation (A) and test cohorts (B) at different time points. (C-D) Representative examples of MRI images and visualization of ConvXGB prediction using Grad-CAM. Calibration plots of the ConvXGB, ConvXGB&clin, clinical model, radiomics nomogram and histology-specific tool for 3-year recurrence-free survival in the derivation (E) and test cohorts (F). (G) decision curve analysis revealed that the ConvXGB model showed better clinical utility. Abbreviations: AUC, area under curve; ROI, region of interests; Grad-CAM, gradient-weighted class activation mapping.

Fig. 4

Cumulative recurrence probability (A, B) and Kaplan–Meier survival curves for OS (C, D) by the ConvXGB model between high- (red line) and low-risk (blue line) groups from the derivation and test cohort.