Kurzfassung

Artificial intelligence and machine learning are advancing rapidly and transforming many aspects of modern life. One requirement for machine learning approaches is large amounts of data in a digital format. In contrast to the field of radiology, pathologists have thus far almost exclusively relied on analogous technology such as bench top microscopes and glass slides. With the recent introduction of whole slide scanners an ever increasing amount of image data arises and digital pathology is...Artificial intelligence and machine learning are advancing rapidly and transforming many aspects of modern life. One requirement for machine learning approaches is large amounts of data in a digital format. In contrast to the field of radiology, pathologists have thus far almost exclusively relied on analogous technology such as bench top microscopes and glass slides. With the recent introduction of whole slide scanners an ever increasing amount of image data arises and digital pathology is set to become a reality.Two important applications of machine learning in pathology are (I) automatic classification of tissue type (i.e. cancerous vs. non-cancerous) and (II) prediction of future clinical events (i.e. relapse). We have recently characterized a large clinical cohort of 600 patients with colorectal cancer. From all of these patients we have produced high-quality whole slides images (WSI) that depict the histomorphology of the primary tumor as well as an advanced tissue microarray (TMA). Furthermore, we obtained pathological as well as epidemiological data on this cohort. In a preliminary experiment, we have used image feature extraction and a machine learning approach on a subset of samples and were able to correctly perform grading of colorectal carcinomas in approximately 80 percent of prior unseen cases. In the proposed project, we want to build on these experiences and expand our analysis to the complete cohort as well as a large collection of H&E slides that are part of a TCGA dataset. More importantly, we want to investigate the potential use of more sophisticated machine learning algorithms (i. e. deep neural networks) for classification and prediction. In a last step, we want to “reverse engineer” the respective algorithms and identify image features and regions, which are most relevant to the model.» weiterlesen» einklappen