Localized Contrastive and Attention Based Multiple Instance Learning for Automatic Staging of Histopathology Images.

Document Type


Publication Date



oregon; chiles


Cancer is the second deadliest disease in the US. Each year, lung cancer alone causes the deaths of over 125,000 people . A key tool in the diagnosis and treatment of cancer lies in the analysis of digital Histopathology images. Currently this is done using a highly trained specialist, known as a pathologist, who examines these high resolution images to determine the type and severity of a cancerous tumor. Of particular interest is the stage of the cancer, which correlates with cancer severity and overall life expectation of cancer patients. To categorize the tumor, the pathologist evaluates a number of morphological characteristics, such as the size and shape of cancer cells, their level of differentiation, and whether or not they have invaded nearby tissues. Due to the costly nature of such evaluations, the field of digital pathology has emerged to develop automated platforms for doing this same analysis using automated processes. In this context, we have developed a novel deep learning pipeline to classify digitized Breast Cancer and Lung Adenocarcinoma histopathological images based on their pathological stage. Features are extracted from image tiles using a multi-scale contrastive learning algorithm with a novel method of including incorporating spatial concordance within the target instance. Afterward, multi-Instance learning is applied for slide level classification. The pipeline has been trained and validated using data downloaded from The Cancer Imaging Archive where an AUC of over 0.7 was achieved.

Clinical Institute





Pulmonary Medicine


Naik N, Crabtree A, Bifulco C, Piening B, Srinivasa G, Matlock K. Women in Computer Vision Annual Meeting; June 18-22; Vancouver, BC. 2023: 30.a

This document is currently not available here.