Detecting Cancer Metastases on Gigapixel Pathology Images

TL;DR

The paper tackles automatic detection and localization of breast cancer lymph node metastases in gigapixel pathology images using a patch-based CNN framework. By evaluating on Camelyon16 with sliding-window heatmaps and non-maximum suppression, it achieves state-of-the-art tumor detection (92.4% at 8 FP per image) and strong slide-level AUC (>97%), while showing that larger-scale inputs, pretraining on natural images, or color normalization offer limited benefits. The approach demonstrates robustness across independent data and highlights practical improvements for clinical workflow, including reduced false negatives and near pathologist-level performance. The work also analyzes data quality issues and emphasizes the importance of data augmentation and model efficiency, suggesting directions for future expansion with larger datasets.

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.

Figures (8)

• Figure 1: Left: three tumor patches and right: three challenging normal patches.
• Figure 2: Difficulty of pixel-accurate annotations for scattered tumor cells. Ground truth annotation is overlaid with a lighter shade. Note that the tumor annotations include both tumor cells and normal cells e.g., white space representing adipose tissue (fat).
• Figure 3: The three colorful blocks represent Inception (V3) towers up to the second-last layer (PreLogit). Single scale utilizes one tower with input images at 40X magnification; multi-scale utilizes multiple ( e.g., 2) input magnifications that are input to separate towers and merged.
• Figure 4: Left to right: sample image, ground truth (tumor in white), and heatmap outputs (40X-ensemble-of-3, 40X+20X, and 40X+10X). Heatmaps of 40X and 40X-ensemble-of-3 look identical. The red circular regions at the bottom left quadrant of the heatmaps are unannotated tumor. Some of the speckles are either out of focus patches on the image or non-tumor patches within a large tumor.
• Figure 5: Left: a patch from a H&E-stained slide. The darker regions are tumor, but not the lighter pink regions. Right: the corresponding predicted heatmap that accurately identifies the tumor cells while assigning lower probabilities to the non-tumor regions.