# Habitat radiomics based on CT for assessing BRCA mutation status in patients with high-grade serous ovarian cancer: a multicenter study

## TL;DR

This study shows that CT-based habitat radiomics can non-invasively predict BRCA mutation status in ovarian cancer patients, with SVM being the most reliable model.

## Contribution

The study introduces habitat radiomics as a novel method to predict BRCA mutations in ovarian cancer by analyzing tumor heterogeneity.

## Key findings

- The SVM model achieved an AUC of 0.952 in training and 0.841 in testing, showing stable performance.
- The combined habitat model outperformed traditional clinical and whole-tumor radiomic models.
- Habitat radiomics effectively captures tumor heterogeneity for BRCA mutation prediction.

## Abstract

This study aims to evaluate the potential of CT-based habitat radiomics in predicting BRCA mutation status in patients with high-grade serous ovarian cancer (HGSOC). The goal is to identify radiomic features from distinct tumor habitats that correlate with BRCA mutations and assess the predictive accuracy of various machine learning models.

A total of 228 patients with histologically confirmed HGSOC were included in this multicenter, retrospective study, with 168 patients in the training cohort and 60 patients in the test cohort. Radiomic features were extracted from the entire tumor and subdivided into five distinct “habitats” based on local tumor features. Predictive models were developed for each of the following: clinical model, radiomics model (based on whole tumor characteristics), five habitat models (habitat1, habitat2, habitat3, habitat4, habitat5), and a combined habitat model (integrating habitat1–5). Five machine learning algorithms (logistic regression (LR), support vector machine (SVM), light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost)) were applied to each model. The model with the highest average area under the curve (AUC) across the algorithms in the training cohort was selected as the optimal model. Further comparison and evaluation of the optimal models from different algorithms were performed to determine the most reliable one.

Among the five machine learning algorithms, XGBoost showed the highest AUC in the training cohort but exhibited a significant drop in the test cohort, indicating overfitting. In contrast, the SVM model demonstrated more consistent performance across both cohorts, with an AUC of 0.952 in the training cohort and 0.841 in the test cohort, making it the most stable performer among the tested algorithms for predicting BRCA mutation status. Calibration and net benefit analyses further confirmed the potential of the SVM-based habitat model as a non-invasive exploratory tool.

CT-based habitat radiomics offers a promising, non-invasive method for predicting BRCA mutation status in HGSOC. The combined habitat model outperformed traditional clinical and whole-tumor radiomic models by more effectively capturing tumor heterogeneity. SVM, demonstrating stable and reliable performance across datasets, emerged as the most robust model for clinical use. These findings support the integration of habitat radiomics, particularly SVM, for personalized, non-invasive molecular assessment in clinical practice.