Performance of a semi-automatic machine leaning method for discriminating HER2 2+ status of breast cancers based on DCE-MRI (Preprint)

Abstract

BACKGROUND Amplification status of human epidermal growth factor receptor2 (HER2) 2+ is currently tested by fluorescence in situ hybridization (FISH). However, the FISH technique is expensive, time consuming, and requires off-site testing. The requirement for alternative low-cost and accurate surrogate measures to formal genetic analysis is urgent. In addition, machine learning is broadly accepted for its ability to decipher complicated connections between medical image features and gene expression status. OBJECTIVE To investigate the potential association between texture features extracted from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and HER2 2+ amplification status of breast cancer. METHODS 92 patients with HER2 2+ breast cancer who underwent 3T MRI and FISH detection in 2018 were retrospectively selected, including 52 HER2 2+ positive and 40 negative cases. The lesion area was delineated semi-automatically with MATLAB, and a total of 307 texture features were extracted from precontrast, postcontrast, and subtraction images, independently. The Student’s t-test or Mann-Whitney U test was performed to identify significant features between different HER2 2+ amplification status. Principal component analysis was used to eliminate the feature correlations. Three machine learning classifiers, logistic regression analysis, quadratic discriminant analysis, and support vector machine (SVM), were with a leave-one-outcross validation method used to establish the classification models of HER2 2+ amplification status. Classification performance was evaluated by receiver operating characteristic (ROC) analysis. RESULTS Texture features calculated from subtraction images showed more promising results than those obtained from pre- and postcontrast images. The model with the SVM based on features from subtraction image achieved the best performance, with an area under the ROC curve of 0.890, sensitivity of 80.77%, specificity of 85.00%, and accuracy of 82.61%. CONCLUSIONS To a certain extent, texture features of breast cancer extracted from DCE-MRI are associated with HER2 2+ amplification status. Additional studies are necessary to confirm the present preliminary findings.