Development and Validation of MRI Radiomics Models to Differentiate HER2-Zero, -Low, and -Positive Breast Cancer
Yuqin Peng,Xiang Zhang,Ya Qiu,Baoxun Li,Zehong Yang,Jiayi Huang,Jinru Lin,Chushan Zheng,Lanxin Hu,Jun Shen
DOI: https://doi.org/10.2214/AJR.23.30603
Abstract:BACKGROUND. Breast cancer HER2 expression has been redefined using a three-tiered system, with HER2-zero cancers considered ineligible for HER2-targeted therapy, HER2-low cancers considered candidates for novel HER2-targeted drugs, and HER2-positive cancers treated with traditional HER2-targeted medications. OBJECTIVE. The purpose of this study was to assess MRI radiomics models for a three-tiered classification of HER2 expression of breast cancer. METHODS. This retrospective study included 592 patients with pathologically confirmed breast cancer (mean age, 47.0 ± 18.0 [SD] years) who underwent breast MRI at either of a health system's two hospitals from April 2016 through June 2022. Three-tiered HER2 status was pathologically determined. Radiologists assessed the conventional MRI features of tumors and manually segmented the tumors on multiparametric sequences (T2-weighted images, DWI, ADC maps, and T1-weighted delayed contrast-enhanced images) to extract radiomics features. Least absolute shrinkage and selection operator analysis was used to develop two radiomics signatures, to differentiate HER2-zero cancers from HER2-low or HER2-positive cancers (task 1) as well as to differentiate HER2-low cancers from HER2-positive cancers (task 2). Patients from hospital 1 were randomly assigned to a discovery set (task 1: n = 376; task 2: n = 335) or an internal validation set (task 1: n = 161; task 2: n = 143); patients from hospital 2 formed an external validation set (task 1: n = 55; task 2: n = 50). Multivariable logistic regression analysis was used to create nomograms combining radiomics signatures with clinicopathologic and conventional MRI features. RESULTS. AUC, sensitivity, and specificity in the discovery, internal validation, and external validation sets were as follows: for task 1, 0.89, 99.4%, and 69.0%; 0.86, 98.6%, and 76.5%; and 0.78, 100.0%, and 0.0%, respectively; for task 2, 0.77, 93.8%, and 32.3%; 0.75, 92.9%, and 6.8%; and 0.77, 97.0%, and 29.4%, respectively. For task 1, no nomogram was created because no clinicopathologic or conventional MRI feature was associated with HER2 status independent of the MRI radiomics signature. For task 2, a nomogram including an MRI radiomics signature and three pathologic features (histologic grade of III, high Ki-67 index, and positive progesterone receptor status) that were independently associated with HER2-low expression had an AUC of 0.87, 0.83, and 0.80 in the three sets. CONCLUSION. MRI radiomics features were used to differentiate HER2-zero from HER2-low cancers or HER2-positives cancers as well as to differentiate HER2-low cancers from HER2-positive cancers. CLINICAL IMPACT. MRI radiomics may help select patients for novel or traditional HER2-targeted therapies, particularly those patients with ambiguous results of immunohistochemical staining results or limited access to fluorescence in situ hybridization.