# Radiomics based on habitat analysis in predicting parametrial invasion of early stage cervical cancer

## TL;DR

This study shows that analyzing tumor sub-regions using radiomics improves the prediction of parametrial invasion in early-stage cervical cancer compared to whole-tumor analysis.

## Contribution

The novel approach uses habitat-based radiomics to better predict parametrial invasion in cervical cancer than traditional methods.

## Key findings

- Habitat 3 model showed higher diagnostic performance than whole-tumor and other habitat models.
- The habitat 3 model achieved 100% specificity and 97.9% sensitivity in the training cohort.
- Radiomics based on habitat analysis outperformed traditional whole-tumor radiomics in predicting parametrial invasion.

## Abstract

To evaluate radiomics based on habitat analysis for preoperatively predicting parametrial invasion (PMI) in clinically early-stage cervical cancer (CC).

This retrospective study included 110 consecutive patients clinically staged as IB-IIA before treatment. Patients were randomly divided into the training and testing cohorts in an 8:2 ratio. Regions of interest were manually delineated on T2-weighted images, which were then segmented into sub-regions using a k-means clustering algorithm based on voxel intensity and entropy values. Radiomic features were then extracted from both the whole tumor and each sub-region. Feature selection was performed using correlation analysis, recursive feature elimination, and the least absolute shrinkage and selection operator method. Subsequently, models were constructed based on valid radiomics features extracted from the whole tumor and from each sub-region. The diagnostic accuracy of the models was evaluated using receiver operating characteristic analysis. The area under the curve (AUC) was compared descriptively, and the analysis was supplemented by net reclassification improvement and comprehensive discrimination improvement measures.

Tumors were divided into three sub-regions (habitat 1-3). A total of 2260 and 1890 radiomics features were extracted from whole tumor and each habitat, respectively. After selection, 10, 10, 7 and 9 valid features were selected from whole tumor and habitats 1-3, respectively. All models had good classification performance for positive and negative PMI in the training and testing cohorts, with an AUC ranging from 0.777 to 1.00 in the training cohort and from 0.750 to 0.850 in the testing cohort. In addition, the diagnostic performance of habitat 3 was higher than that of the whole tumor, habitat 1, habitat 2 models in the training and testing cohorts, and the difference was statistically significant (p<0.05). The sensitivity, specificity, and AUC (95% confidence interval) of habitat 3 model in the training and testing cohorts were 97.9%, 100%, 1.00 (0.999–1.00) and 75.0%, 100%, 0.850 (0.649–1.00), respectively.

Radiomics based on habitat analysis effectively predicts PMI in early-stage CC, with diagnostic performance superior to that of traditional whole-tumor radiomics. This approach provides a promising method for preoperative prediction of PMI in CC and aids clinicians and patients in treatment decisions.