Related Experiment Video
Updated: Sep 22, 2025
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
6.9K
Determining breast cancer biomarker status and associated morphological features using deep learning
Paul Gamble1, Ronnachai Jaroensri1, Hongwu Wang1
1Google Health, Palo Alto, CA USA.
Communications Medicine
|May 23, 2022
Summary
Deep learning models can now predict breast cancer biomarkers (estrogen receptor, progesterone receptor, human epidermal growth factor receptor 2) directly from H&E stained slides, improving efficiency and interpretability in cancer management.
Area of Science:
- Computational pathology
- Digital pathology
- Artificial intelligence in oncology
Background:
- Breast cancer management relies on biomarkers like estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2).
- Current scoring systems, while validated, can be costly and subject to interpretation variability.
- Discrepancies in biomarker results necessitate repeat testing, impacting workflow and patient care.
Purpose of the Study:
- To develop and validate deep learning systems (DLS) for predicting ER/PR/HER2 status directly from routine hematoxylin-and-eosin (H&E) stained slides.
- To assess the performance of DLS at both patch-level and slide-level analysis.
- To ensure interpretability of the DLS by correlating computational findings with known histopathological features.
Main Methods:
- Three independent deep learning systems were developed to predict ER/PR/HER2 status using H&E images.
- Models were trained and evaluated on pathologist-annotated slides from diverse data sources.
- Performance was measured using area under the receiver operator characteristic curve (AUC) at patch and slide levels, with interpretability analyses including TCAV and saliency mapping.
Main Results:
- Patch-level AUCs for ER, PR, and HER2 prediction were 0.939, 0.938, and 0.808, respectively.
- Slide-level AUCs reached 0.86 for ER, 0.75 for PR, and 0.60 for HER2.
- Interpretability analyses confirmed known histomorphological associations, such as ER/PR positivity with low-grade/lobular histology and triple-negative status with increased inflammatory infiltrates.
Conclusions:
- Deep learning systems can rapidly estimate breast cancer biomarker status from standard H&E slides.
- This approach offers a computationally efficient alternative to traditional biomarker testing.
- Prioritizing interpretability ensures that AI-driven insights align with established pathological knowledge, enhancing clinical utility.
