Predicting Pathological Characteristics of HER2-Positive Breast Cancer from Ultrasound Images: a Deep Ensemble...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• HER2-positive breast cancer (HER2+ BC) requires accurate prognostic biomarker identification for effective treatment.
• Current methods for assessing prognostic biomarkers can be invasive or time-consuming.
• Novel approaches are needed to efficiently predict these critical characteristics.

Purpose of the Study:

• To evaluate the feasibility of using ultrasound (US) images to identify prognostic biomarkers in HER2+ BC.
• To develop and assess deep learning models for classifying axillary lymph node involvement (ALNM), lymphovascular invasion (LVI), and histological grade (HG).

Main Methods:

• Utilized a dataset of 512 female patients with pathologically validated HER2+ BC.
• Trained five deep convolutional neural networks (DCNNs) and a deep ensemble (DE) approach.
• Evaluated model performance using accuracy, sensitivity, specificity, PPV, NPV, ROC curves, AUCs, and heat maps. DeLong test was used for AUC comparisons.

Main Results:

• The deep ensemble (DE) model demonstrated superior performance, achieving high AUCs and accuracy for LVI (AUC=0.869, accuracy=69.7%) and HG (AUC=0.973, accuracy=73.8%), especially with imbalanced data.
• Performance on ALNM (AUC=0.780, accuracy=77.5%) was comparable to single models.
• The DE model showed significant classification performance (p < 0.05).

Conclusions:

• Pretreatment US-based DE models show potential as a clinical tool for predicting pathological characteristics in HER2+ BC patients.
• This approach may facilitate timely treatment strategy adjustments.
• Ultrasound imaging combined with deep learning offers a promising non-invasive method for prognostic biomarker identification.