Digital Pathology

Spatio-Net Research Project

We have developed AI models intended to enhance pathologists’ accuracy and efficiency in staging a variety of cancers.

Detecting metastasis of lymph nodes is critical to judging cancer prognosis, particularly for breast cancer. However, detection can be challenging for the pathologist because slide images are high resolution (e.g., 100k X 200k pixels), and tissue is highly variable. We developed a spatially structured deep network (Spatio-Net) that improves the detection of metastases by more than 5% over current state-of-the-art for breast cancer.

The Spatio-Net research project achieved this advantage by taking spatial variation into account. In contrast, other approaches divide slide images into small areas and analyze each individually, discarding spatial structure information that is vital to detection inference. Spatio-Net has a further advantage from its rich training dataset, which includes over 1 million image patches from 300 patients.

Challenge 1: Large, High-Resolution Images

Challenge 2: High Variance within Each Slide