Cancer Metastasis Detection Through Multiple Spatial Context Network

Wutong Zhang, Chuang Zhu, Jun Liu, Ying Wang, Mulan Jin

Abstract:

Breast cancer is one of the leading causes of death by cancer in women, and it often requires accurate detection of metastasis in lymph nodes through Whole-slide Images (WSIs). At present, there are many algorithms of cancer metastasis detection based on CNN, which are generally patch-level models, aiming for increasing the sensitivity, speed, and consistency of metastasis detection. However, most of these algorithms use patch as an independent individual to train, which leads to the neglect of much important spatial context information in WSI.
In this paper, we propose a multiple spatial context network (MSC-Net) which considers the spatial correlations between neighboring patches through fusing the spatial information probability maps obtained from the two novel networks we propose, the self-surround spatial context stacked network (SSC-Net) and the center-surround spatial context shared network (CSC-Net). The SSC-Net is a deep mining of continuous information between patches, while CSC-Net strengthens the influence of the neighborhood information to the central patch.
Furthermore, for saving memory overhead and reducing computational complexity, we propose a framework which can quickly scan the WSI through the mechanism of the patch feature sharing.
We demonstrate evaluations on the camelyon16 dataset and compare with the state-of-the-art trackers. Our method provides a superior result.

TL;DR: A novel and efficient method for enhancing the spatial relationship of pathological images to detect cancer metastasis

Keywords: Deep Learning, Spatial Context Relation, Cancer Metastasis Detection