A Study of Deep Learning Colon Cancer Detection in Limited Data Access Scenarios

• DOI: 10.48550/arxiv.2005.10326
• Publication Date: 2020-01-01
• Authors: Apostolia Tsiriko...; Karin Stacke; Gabriel Eilertsen; Martin Lindvall; Jonas Unger
• Keywords: Digital Pathology; Digitization

Digitization of histopathology slides has led to several advances, from easy data sharing and collaborations the development digital diagnostic tools. Deep learning (DL) methods for classification detection have shown great potential, but often require large amounts training that are hard collect, annotate. For many cancer types, scarceness creates barriers DL models. One such scenario relates detecting tumor metastasis in lymph node tissue, where low ratio non-tumor cells makes task time-consuming. DL-based tools can allow faster diagnosis, with potentially increased quality. Unfortunately, due sparsity cells, annotating this type demands a high level effort pathologists. Using weak annotations slide-level images demand access substantial amount as well. In study, we investigate mitigation strategies limited scenarios. Particularly, address whether it is possible exploit mutual structure between tissues develop general techniques, wherein one particular tissue could value other cancers tissues. Our case exemplified by model metastatic colon nodes. Could be trained little or even no data? As alternative sources, 1) taken primary 2) different organ (breast), either transformed target domain (colon) using Cycle-GANs. We show suggested approaches make detect very data, opening up possibility existing, annotated generalize domains.