# Value of intra- and peritumoral ultrasound radiomics for predicting axillary lymph node burden in breast cancer

## TL;DR

This study shows that combining ultrasound-based radiomic features from tumor and surrounding areas can accurately predict lymph node involvement in breast cancer before surgery.

## Contribution

The novel integration of intratumoral and peritumoral ultrasound radiomics with clinical variables improves preoperative prediction of axillary lymph node burden.

## Key findings

- A combined model using intratumoral and 3mm peritumoral features achieved a testing AUC of 0.818.
- The integrated nomogram outperformed individual radiomics or clinical models in predicting lymph node burden.
- Decision curve analysis confirmed the clinical utility of the combined model across threshold probabilities.

## Abstract

To evaluate ultrasound-based radiomic features, derived from both intratumoral and peritumoral regions, for noninvasive preoperative prediction of axillary lymph node(ALN) burden in breast cancer.

This retrospective study analyzed data from 300 pathologically confirmed breast cancer patients undergoing preoperative ultrasound. The cohort was randomly divided into a training set (n = 210) and a testing set (n = 90) at a 7∶3 ratio. Primary tumor regions of interest (ROIs) were manually delineated on preoperative ultrasound images using ITK-SNAP. Peritumoral ROIs were generated by radially expanding the intratumoral ROI by 2mm, 3mm, and 4mm. A comprehensive set of radiomic features was extracted from each ROI, with feature selection via LASSO based methods. Six machine-learning classifiers were trained on intratumoral features to identify the optimal algorithm. Using this algorithm, we built: (1) A radiomics model based solely on intratumoral or peritumoral features. (2) Combined models incorporating intratumoral and peritumoral features at each expansion margin (2mm, 3mm, and 4mm). The best-performing radiomics model was then integrated with significant clinical and conventional imaging variables to construct a composite nomogram. Model discrimination was evaluated by area under the receiver operating characteristic curve (AUC), calibration was assessed via calibration curves, and clinical utility was appraised using decision curve analysis (DCA). Model interpretability was facilitated through Shapley additive explanation (SHAP) values and visualized in a nomogram.

A Random Forest classifier applied to combined intratumoral and 3mm peritumoral features yielded the highest AUCs (training set: 0.825; testing set: 0.746). Multivariable logistic regression identified lesion location and ultrasonographic axillary lymph node status as independent clinical predictors (p<0.05). The integrated nomogram—combining these clinical factors with the optimal radiomics signature—demonstrated superior performance (training AUC: 0.906; testing AUC: 0.818). DCA confirmed that the combined model conferred the greatest net clinical benefit across a range of threshold probabilities, and calibration curves indicated excellent agreement between predicted and observed probabilities.

A composite model integrating intratumoral and 3mm peritumoral ultrasound radiomic features with key clinical and imaging variables enables accurate, noninvasive preoperative prediction of ALN burden in breast cancer. This approach may serve as a valuable decision support tool to guide individualized surgical planning.