Preoperative Differentiation of HER2-Zero and HER2-Low from HER2-Positive Invasive Ductal Breast Cancers Using...

Area of Science:

• Oncology
• Radiology
• Biomedical Imaging
• Machine Learning in Medicine

Background:

• Accurate human epidermal growth factor receptor 2 (HER2) determination is crucial for guiding HER2-targeted therapies.
• HER2-low status, currently classified as HER2-negative, may qualify patients for novel anti-HER2 drug conjugates.

Purpose of the Study:

• To classify three HER2 expression levels (HER2-zero, HER2-low, HER2-positive) using breast MRI BI-RADS features.
• To develop models distinguishing HER2-zero from HER2-low/positive (Task-1) and HER2-low from HER2-positive (Task-2).

Main Methods:

• Retrospective analysis of 621 invasive ductal cancer cases.
• Utilized Random Forest for MRI feature selection and four machine learning algorithms (DT, SVM, k-NN, ANN) to build classification models.
• Statistical analysis included Chi-square, ANOVA, and Kruskal-Wallis tests; ROC curves evaluated model performance.

Main Results:

• Peritumoral edema, multiple lesions, and non-mass enhancement (NME) were significant differentiating features.
• For HER2-zero vs. non-zero, k-NN achieved the highest AUC (0.86 training, 0.79 testing).
• For HER2-low vs. HER2-positive, DT achieved the highest AUC (0.79 training, 0.69 testing).

Conclusions:

• Preoperative MRI BI-RADS features, analyzed with advanced feature selection and ML, can effectively classify HER2 status.
• This approach aids in identifying HER2-low breast cancer patients eligible for targeted therapies.