Detecting Cancer Metastases on Gigapixel Pathology Images

Each year, the treatment decisions for more than 230,000 breast cancerpatients in the U.S. hinge on whether the cancer has metastasized away from thebreast. Metastasis detection is currently performed by pathologists reviewinglarge expanses of biological tissues. This process is labor intensive anderror-prone. We present a framework to automatically detect and localize tumorsas small as 100 x 100 pixels in gigapixel microscopy images sized 100,000 x100,000 pixels. Our method leverages a convolutional neural network (CNN)architecture and obtains state-of-the-art results on the Camelyon16 dataset inthe challenging lesion-level tumor detection task. At 8 false positives perimage, we detect 92.4% of the tumors, relative to 82.7% by the previous bestautomated approach. For comparison, a human pathologist attempting exhaustivesearch 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. Inaddition, we discover that two slides in the Camelyon16 training set wereerroneously labeled normal. Our approach could considerably reduce falsenegative rates in metastasis detection.