HER2 Domain Adaptation

Overview

This research project at Mayo Clinic extends HER2 biomarker detection from breast cancer-only applications to 7 different cancer types using advanced domain adaptation techniques, achieving >90% accuracy across all domains and reducing manual pathologist assessment time by >75%.

Background: HER2 Biomarker

Human Epidermal Growth Factor Receptor 2 (HER2) is a critical biomarker in oncology:

• Overexpression correlates with aggressive tumor behavior
• Guides targeted therapy decisions (e.g., Herceptin/trastuzumab)
• Traditionally assessed through immunohistochemistry (IHC) and FISH testing
• Manual pathologist review is time-consuming and subject to inter-observer variability

Problem Statement

Challenge

Existing HER2 detection models are trained exclusively on breast cancer histopathology images and fail to generalize to other cancer types due to:

• Domain shift: Different tissue morphology and staining patterns
• Limited labeled data: Expensive pathologist annotations for each cancer type
• Computational resources: Training separate models for each domain is inefficient

Goal

Develop a domain adaptation framework that transfers knowledge from breast cancer HER2 detection to multiple cancer types without requiring extensive re-labeling or retraining.

Solution: Multi-Domain Transfer Learning

Architecture

1. Source Domain: Breast Cancer HER2 Detection
   • Pre-trained ResNet-50 on 10,000+ annotated breast cancer slides
   • Achieved 96% accuracy on breast cancer validation set
   • Feature extractor frozen for transfer learning
2. Target Domains (7 cancer types):
   • Gastric cancer
   • Colorectal cancer
   • Ovarian cancer
   • Lung cancer
   • Bladder cancer
   • Esophageal cancer
   • Uterine cancer
3. Domain Adaptation Techniques:
   • Adversarial Domain Adaptation: Domain-invariant feature learning
   • Maximum Mean Discrepancy (MMD): Minimize distribution shift
   • Self-supervised pre-training: Contrastive learning on unlabeled slides
   • Fine-tuning with limited labels: 100-500 labeled samples per domain

Training Pipeline

Breast Cancer Dataset (10K+ slides)
         ↓
   ResNet-50 Backbone
         ↓
Domain-Adversarial Training
         ↓
Feature Alignment (MMD Loss)
         ↓
Fine-tuning on Target Domains
         ↓
Evaluation on 7 Cancer Types

Technical Implementation

Data Preprocessing

• Whole Slide Imaging (WSI) at 40x magnification
• Patch extraction (512×512 pixels) with overlapping windows
• Color normalization using Reinhard method
• Data augmentation: rotation, flipping, color jittering

Model Architecture

• Backbone: ResNet-50 pre-trained on ImageNet + breast cancer
• Domain Discriminator: 3-layer MLP for adversarial training
• Feature Alignment: MMD kernel with Gaussian RBF
• Classification Head: 2-layer FC network (HER2+ vs HER2-)

Training Strategy

• Phase 1: Supervised training on breast cancer (source domain)
• Phase 2: Adversarial adaptation with unlabeled target slides
• Phase 3: Fine-tuning with limited labeled samples (100-500 per domain)
• Optimizer: AdamW with learning rate scheduling
• Loss Function: Cross-entropy + adversarial + MMD

Results & Performance

Accuracy Across Cancer Types

Average across all domains: >90% accuracy

Clinical Impact

✅ Time Savings: Reduced manual pathologist review time by >75%
✅ Consistency: Eliminated inter-observer variability
✅ Scalability: Single model supports 7+ cancer types
✅ Cost Efficiency: No need for extensive labeling for each new domain
✅ Clinical Validation: Concordance with pathologist ground truth >92%

Comparison with Baselines

Technical Stack

Deep Learning: PyTorch, torchvision, timm
Medical Imaging: OpenSlide, Pillow, scikit-image
Domain Adaptation: PyTorch Domain Library, MMD implementations
Data Processing: NumPy, Pandas, OpenCV
Cloud Infrastructure: Google Cloud Vertex AI, Compute Engine
Visualization: Matplotlib, Seaborn, Plotly
Experiment Tracking: Weights & Biases, TensorBoard

Key Innovations

1. Multi-Domain Transfer: First study to extend HER2 detection across 7 cancer types
2. Label Efficiency: Achieved >90% accuracy with only 100-500 samples per domain
3. Adversarial + MMD: Hybrid domain adaptation outperforms single-method approaches
4. Clinical Integration: Seamless integration into pathology workflows

Collaboration

Worked closely with:

• Pathologists for ground truth annotations and clinical validation
• Oncologists for treatment decision integration
• Bioinformatics Teams for genomic data correlation
• IT Infrastructure for secure deployment in clinical environments

Future Directions

• Additional Cancer Types: Expanding to pancreatic, prostate, and head/neck cancers
• Multi-Task Learning: Simultaneous prediction of HER2, ER, PR, and Ki-67 biomarkers
• Weakly Supervised Learning: Reduce annotation requirements using slide-level labels
• Real-Time Inference: Deploy on edge devices for intraoperative decision support
• Publication: Manuscript in preparation for submission to top medical imaging journal

Publication Plans

Target Journal: Medical Image Analysis or Nature Medicine
Expected Submission: Q4 2025
Co-authors: Mayo Clinic pathologists, bioinformatics team, ML researchers

Status: Research Complete - Publication in Preparation
Organization: Mayo Clinic, Rochester, MN
Role: Data Science Intern - Research Engineer
Duration: May 2025 - Present