Reproduce “Detecting Cancer Metastases on Gigapixel Pathology Images” on OpenResearch

About this paper

Yun Liu, Krishna Gadepalli, Mohammad Norouzi, George E. Dahl, Timo Kohlberger, Aleksey Boyko, Subhashini Venugopalan, Aleksei Timofeev and 5 more

Researchers from Google Brain developed a deep learning framework to detect cancer metastases on gigapixel pathology images, achieving 92.4% sensitivity at 8 false positives per image on the Camelyon16 dataset. The system surpassed previous automated methods and human pathologist performance, demonstrating robust generalization across independent datasets.

Key insights

• Simplified slide-level classification using only the maximum value from the tumor probability heatmap achieved high AUCs (>97%), comparable to more complex classifiers, indicating strong inherent patch-level prediction capabilities.
• Pre-training on ImageNet, multi-scale inputs, and explicit color normalization did not provide performance benefits, suggesting that comprehensive data augmentation in the target domain can suffice for learning robust features.
• Smaller Inception architectures (with approximately 3% of the parameters) achieved comparable performance to full models, indicating potential for more efficient and less resource-intensive deployment.

Summary by alphaXiv.