HER2 Molecular Marker Scoring Using Transfer Learning and Decision Level Fusion.

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning in oncology

Background:

• Accurate human epidermal growth factor receptor 2 (HER2) scoring is crucial for breast cancer prognosis and treatment decisions.
• Manual HER2 scoring by pathologists is time-consuming, subjective, and prone to inter-observer variability.

Purpose of the Study:

• To develop and evaluate deep learning transfer learning models for automated HER2 scoring in breast cancer tissue samples.
• To compare the performance of various pre-trained deep learning architectures for HER2 scoring.

Main Methods:

• Utilized five pre-trained deep learning architectures (VGG16, VGG19, ResNet50, MobileNetV2, NASNetMobile) for 3-class HER2 classification.
• Applied transfer learning and statistical voting (mode operator) for enhanced scoring accuracy.
• Trained and tested models on the HER2 Challenge dataset, comprising 2130 training and 800 testing image patches.

Main Results:

• Transfer learning models demonstrated significant accuracy in HER2 scoring.
• VGG19 achieved the highest accuracy of 93% on test images.
• Statistical voting further improved accuracy to 98%, indicating a robust automated quantification pipeline.

Conclusions:

• Deep learning, particularly VGG19 with statistical voting, offers a highly accurate and efficient automated solution for HER2 scoring.
• This automated approach can assist in clinical decision-making, potentially reducing diagnostic variability and improving patient care.