# Augmented Prediction of N Parameter in Breast Cancer: Is It Possible with Shear-Wave Elastography Ultrasound Radiomics?

## TL;DR

This study explores using ultrasound and shear-wave elastography with machine learning to predict lymph node status in breast cancer patients.

## Contribution

The novel contribution is integrating SWE-derived radiomics features into a machine learning pipeline for predicting axillary lymph node status in breast cancer.

## Key findings

- The ML classifier achieved AUCs of 0.685 (training) and 0.677 (test) for predicting ALN status.
- Expert radiologists outperformed the ML model (AUC = 0.817), but the difference was not statistically significant.
- SWE-derived radiomics features showed potential for inclusion in predictive models for ALN status.

## Abstract

Ultrasound (US) is still the most sensitive modality to predict axillary lymph node (ALN) status in patients with breast cancer (BC) but suffers from a low and variable specificity. A Simple Logistic Machine Learning algorithm was used with US B-mode and SWE-derived radiomics features of 133 primary BC lesions to identify cases with positive ALN status. The classifier showed AUC of 0.685 and 0.677, MCC of 0.387 and 0.375 in the training and test set, respectively. The performance of ML was lower, even if not significantly (p = 0.481) from that of an expert radiologist (AUC = 0.817) who evaluated US images of ALN in the test set. Although the accuracy of ML was relatively low compared to the values reported in the literature, our findings support the inclusion of SWE-derived radiomics features of the primary BC lesion in a radiomics pipeline for the prediction of ALN status.

Background/Objectives: The aim was to assess whether a machine learning (ML) algorithm could empower the ability of ultrasound (US) integrated with shear-wave elastography (SWE) to preoperatively define the ALN status in breast cancer (BC). Methods: Patients with at least one histologically proven BC lesion, who underwent preoperative breast US and SWE were retrospectively enrolled. BC lesions were segmented on US and SWE images by three different operators and radiomics features were extracted. A multi-step US and SWE feature selection was performed. A Simple Logistic ML classifier was applied to the dataset to predict the ALN status, its performance assessed through the AUC and Matthews Correlation Coefficient (MCC). The performance of the ML classifier was compared to that of an expert radiologist, who evaluated the US B-mode lymph-node features included in the test set. Results: A total of 133 BC lesions were included and divided into a training set, composed of 89 BC lesions (ALN−: 52; ALN+: 37), and a test set, including 44 BC lesions (ALN−: 24; ALN+: 20). Eight features out of the 1098 radiomics features extracted from US and SWE images were selected to build the predictive model. Simple Logistic classifier showed AUC of 0.685 and 0.677, MCC of 0.387 and 0.375 in the training and test set, respectively. The performance of the expert radiologist was higher than that of the ML classifier (AUC = 0.817), but not significantly different (p = 0.481). Conclusions: The inclusion of SWE-derived radiomics features could aid in the preoperative assessment of ALN status in BC using an ML approach.