# Prediction Model of Lymph Node Metastasis in Cervical Cancer Based on MRI Habitat Radiomics

## TL;DR

A new MRI-based model using tumor heterogeneity improves the prediction of lymph node metastasis in cervical cancer, helping guide treatment decisions.

## Contribution

A novel habitat-based radiomics model outperforms traditional methods in predicting pelvic lymph node metastasis in cervical cancer.

## Key findings

- The combined habitat radiomics and clinical model achieved an AUC of 0.895 in predicting lymph node metastasis.
- Habitat-based radiomics outperformed conventional whole-tumor radiomics and clinical models alone.
- The model shows strong calibration and clinical utility for preoperative decision-making.

## Abstract

Lymph node metastasis is an important factor affecting treatment decisions and prognosis in patients with cervical cancer, but it is difficult to accurately assess before surgery using conventional imaging methods. In this study, we developed a new prediction model based on magnetic resonance imaging (MRI) radiomics that takes tumor heterogeneity into account. By dividing tumors into different intratumoral subregions (habitats) and combining imaging features with clinical information, we were able to more accurately predict pelvic lymph node metastasis in patients with early-stage cervical cancer. Our results show that this habitat-based radiomics model performs better than traditional clinical or whole-tumor radiomics models and may help clinicians better plan individualized treatment strategies before surgery.

Background: Radiomics provides a non-invasive approach for predicting lymph node metastasis (LNM) in cervical cancer, but conventional whole-tumor analysis often overlooks intratumoral heterogeneity. Methods: This study aimed to develop and validate an MRI-based habitat radiomics model for preoperative prediction of pelvic LNM in early-stage cervical cancer. Tumor regions were delineated on diffusion-weighted imaging, and intratumoral habitats were generated using unsupervised K-means clustering. Radiomic features were extracted from whole tumors and habitat subregions, combined with clinical variables, and selected using correlation analysis and LASSO regression. Four models—clinical, conventional radiomics, habitat radiomics, and combined—were constructed and evaluated. Results: In internal validation, the combined model achieved the best performance (AUC = 0.895), outperforming the clinical (AUC = 0.799), conventional radiomics (AUC = 0.611), and habitat models (AUC = 0.872). Calibration and decision curve analyses demonstrated good agreement and clinical utility. Conclusions: Integrating habitat-based radiomics with clinical factors significantly improves the preoperative prediction of LNM, providing a robust and clinically applicable tool for individualized management of cervical cancer patients.