# Multiparametric magnetic resonance imaging-based comprehensive model on prediction of lymphovascular space invasion in cervical cancer

## TL;DR

This study creates a model using MRI scans to predict lymphovascular space invasion in cervical cancer before surgery, which could improve treatment planning.

## Contribution

The novel RDL model combines radiomics and deep learning features from MRI scans to predict lymphovascular space invasion more accurately than individual models.

## Key findings

- The RDL model achieved an AUC of 0.968 in the training cohort, outperforming both the radiomics and deep learning models.
- In the validation cohort, the RDL model showed significantly better predictive performance than the single models with an AUC of 0.859.
- The RDL model demonstrated superior calibration and clinical benefit compared to other models.

## Abstract

To develop and validate a comprehensive model integrating multiparametric magnetic resonance imaging (MRI) radiomics and deep learning features for preoperative prediction of LVSI in early-stage cervical cancer.

155 patients from January 2019 to December 2023 were enrolled in this study and divided into the training and validation cohorts randomly at a ratio of 7:3. Radiomics and deep learning features were extracted from T2-weighted images (T2WI), apparent diffusion coefficient (ADC) maps, and late contrast-enhanced T1-weighted images (CE-T1WI). Mann–Whitney U test, the least absolute shrinkage and selection operator regression (LASSO) were used to select radiomics and deep learning features. Radiomics model (Rad model), deep learning model (DL model), and radiomics-deep learning model (RDL model) were derived from the training cohort using support vector machines (SVM) classifier. The prediction performances of the three models were evaluated with the area under the curve (AUC), calibration curve, decision curve analysis (DCA) and tested in the validation cohort.

The RDL model achieved predictive performance for LVSI in cervical cancer with an AUC of 0.968 (95% confidence interval (CI): 0.938-0.999) in the training cohort, higher than 0.801(95% CI: 0.712-0.891) of Rad model and 0.902(95 CI: 0.845-0.959) of DL model with statistical significance after Bonferroni correction. In the validation cohort, the predictive performance of the fusion model (RDL)(AUC = 0.859, 95% CI 0.751-0.967) was significantly superior to that of the single model (AUC of DL Model = 0.745 95% CI 0.595-0.894; AUC of Rad Model = 0.686 95% CI 0.525-0.847, P < 0.001), however, the DL and radiomics models did not demonstrate statistically significant differences in performance within the validation cohort (Delong test, P>0.05). Analysis of the calibration and decision curves indicated superior predictive precision and net clinical benefit for the RDL model relative to the others.

The advanced RDL model demonstrated strong predictive accuracy for LVSI in cervical cancer, suggesting its promising role as a noninvasive imaging biomarker. This tool could significantly enhance preoperative treatment planning by providing reliable insights without invasive procedures.