Multimodal PA/US imaging and radiomics for the prediction of HER2-zero, -low, and -positive breast cancers: A novel approach for targeted therapy selection

Abstract

Purpose

This study evaluates the efficacy of photoacoustic/ultrasound (PA/US) imaging-based radiomics for distinguishing HER2-zero, HER2-low, and HER2-positive breast cancer (BC), aiming to enhance targeted therapy selection.

Methods

We analyzed 346 pathologically confirmed BC patients who underwent multimodal PA/US imaging at Shenzhen People’s Hospital from January 2022 to January 2025. HER2 status was determined pathologically and classified into three levels. Radiologists assessed conventional US features and manually segmented tumors on PA-images for radiomics feature extraction. Using the Least Absolute Shrinkage and Selection Operator analysis, we developed radiomics models for differentiating between HER2-zero versus HER2-low/positive cancers (Task 1), and HER2-low versus positive cancers (Task 2), and HER2-zero versus low cancers (Task 3). Patients were randomly divided into training sets and testing sets. Multivariate logistic regression was used to integrate radiomics, clinical-pathological, and US features into nomograms.

Results

In testing set, radiomics features demonstrated an AUC of 0.846 with sensitivity of 79.3 % and specificity of 72.7 % for Task 1, and an AUC of 0.801 with sensitivity of 64.0 % and specificity of 82.8 % for Task 2, and an AUC of 0.767 with sensitivity of 80.7 % and specificity of 72.7 % for Task 3. For Task 1, 2 and 3, nomograms including PA imaging radiomics features combined with clinical-pathological features achieved AUCs of 0.848, 0.881 and 0.780, respectively.

Conclusion

PA radiomics features effectively differentiate between HER2-zero and HER2 low/positive, and between HER2-low and HER2-positive BC, offering potential utility in guiding targeted therapy decisions.

Summary

This study demonstrates the potential of PA imaging-based radiomics for accurately classifying HER2 expression statuses in BC, enhancing the selection process for targeted therapies. By integrating multi-modal imaging and pathology data, the developed radiomics models show robust performance, promising a non-invasive diagnostic supplementary for clinical application where traditional methods are limited.