[Paper Review] Detecting Cancer Metastases on Gigapixel Pathology Images

Yun Liu, Krishna Gadepalli+11 more|arXiv (Cornell University)|Mar 3, 2017

AI in cancer detection3 references521 citations

TL;DR

The paper develops a CNN-based framework to detect and localize small metastases in gigapixel pathology slides, achieving state-of-the-art lesion-level detection and high slide-level AUC on Camelyon16. It argues against pre-training, color normalization, and multi-scale patching, showing slim models can match performance.

ABSTRACT

Each year, the treatment decisions for more than 230,000 breast cancer patients in the U.S. hinge on whether the cancer has metastasized away from the breast. Metastasis detection is currently performed by pathologists reviewing large expanses of biological tissues. This process is labor intensive and error-prone. We present a framework to automatically detect and localize tumors as small as 100 x 100 pixels in gigapixel microscopy images sized 100,000 x 100,000 pixels. Our method leverages a convolutional neural network (CNN) architecture and obtains state-of-the-art results on the Camelyon16 dataset in the challenging lesion-level tumor detection task. At 8 false positives per image, we detect 92.4% of the tumors, relative to 82.7% by the previous best automated approach. For comparison, a human pathologist attempting exhaustive search achieved 73.2% sensitivity. We achieve image-level AUC scores above 97% on both the Camelyon16 test set and an independent set of 110 slides. In addition, we discover that two slides in the Camelyon16 training set were erroneously labeled normal. Our approach could considerably reduce false negative rates in metastasis detection.

Motivation & Objective

• Address the challenge of detecting small metastases in extremely large whole-slide images (gigapixel scale).
• Develop a patch-based CNN framework with efficient inference over slides to localize tumors for pathologists.
• Evaluate on Camelyon16 and an independent dataset to assess slide-level accuracy and tumor-level localization.

Proposed method

• Use Inception (V3) architecture to classify 299x299 patches predicting center 128x128 tumor regions.
• Train with balanced patch sampling and extensive data augmentation (rotations, flips, color jitter, jitter in patch extraction).
• Perform slide-level prediction by sliding a 128-pixel-stride heatmap over the slide and taking the maximum heatmap value.
• Compare with prior methods using FROC and AUC metrics, and explore model size (small vs full), pre-training, color normalization, and multi-scale inputs.
• Apply non-maximum suppression over heatmaps to extract tumor coordinates for evaluation.

Experimental results

Research questions

• RQ1 Can a CNN-based patch classifier detect and localize small metastases in gigapixel pathology slides with high sensitivity and low false positive rate?
• RQ2 Is pre-training on natural images, color normalization, or multi-scale inputs beneficial for this task?
• RQ3 What is the impact of model size on performance, and how effective is a heatmap-based sliding window approach for slide-level and tumor-level evaluation?

Key findings

• Achieved high slide-level AUCs (>97%) on Camelyon16 test set and an independent dataset (NHO-1).
• For 40X scale, FROC of 98.1% with 100 max FPs per slide during validation; test FROC 87.3 (with CI) and AUC 96.7.
• Small Inception-based models (3% of full parameter count) achieved nearly equal performance to full models.
• Color normalization and multi-scale inputs did not improve performance under extensive data augmentation and training.
• Ensembling across rotations/flips provided modest improvements; ensemble of 3 models yielded best results (AUC 97.7 on test).
• The approach detected tumors in two slides labeled normal due to data processing errors, indicating robustness to label noise.