For more information on this session, please click here.
Can Deep Learning Models be Trained with Annotations Collected via Crowdsourcing?
Shadi Albarqouni, PhD | Technische Universität München, Germany
One of the major challenges facing researchers nowadays in applying deep learning (DL) models to Medical Image Analysis is the limited amount of annotated data. Collecting such ground-truth annotations requires domain knowledge (expertise), cost, and time, making it infeasible for large-scale databases. We presented a novel concept for training DL models from noisy annotations collected through crowdsourcing platforms, i.e., Amazon Mechanical Turk, Crowdflower, by introducing a robust aggregation layer to the convolutional neural networks. Our proposed method was validated on a publicly available database on Breast Cancer Histology Images showing interesting results of our robust aggregation method compared to baseline methods, i.e., Majority Voting. In follow-up work, we introduced a novel concept of an image to game-object translation in biomedical Imaging allowing medical images to be represented as star-shaped objects that can be easily embedded to readily available game canvas. The proposed method reduces the necessity of domain knowledge for annotations. Exciting and promising results were reported compared to the conventional crowdsourcing platforms.
The Challenge of Making Encoded Data Clinically Actionable
Ulysses Balis, MD | University of Michigan
While application of machine learning techniques to histological images and, more recently, libraries of histological images, holds the promise to augment the contemporary diagnostic paradigm in digital WSI use for conventional surgical pathology, there yet exist substantial barriers between simply having available a computational pipeline capable of generating medically actionable knowledge and implementing a robust clinical solution intended for direct use by busy surgical pathologists. To address this challenge, the Pathology Informatics Division at the University of Michigan has been involved in the long-term development of a generalizable model by which both anatomic and clinical pathology computational pipelines can be deployed at scale, and in a fashion that is immediately accessible to pathologists, who might not otherwise possess specific machine vision or programming skills. This presentation will focus specifically on a panel of mage-based analytical tools and computational pipelines that have been developed by the Michigan team, in specific support of actual surgical pathology workflow challenges in the setting of a busy, high volume practice. Example applications will include: generalizable image-based search, automated annotation, and automated segmentation tools.
How Artificial Intelligence May Change the Way Pathologists Work
Jeroen van der Laak, PhD | Radboud University Medical Center, Nijmegen, The Netherlands
Recent research has shown the large potential of artificial intelligence for analysis of digitized histopathological slides. This field of research, often referred to as 'computational pathology', may offer solutions for the challenges that pathologists face, now and in the near future. Computational pathology may relieve the pathologists' workload by automating routine tasks (e.g. tumor detection in biopsies, assessment of tumor size and distance to surgical margins). Expectedly, the first algorithms of this kind will be commercially available within the next few years. In addition, the increased complexity of histopathological diagnostics required for personalized healthcare may be facilitated by automated assessment of tumor grade and other morphological biomarkers (e.g. tumor-stroma-ratio, tumor budding) as well as by identification of morphological biomarkers not accessible by the human eye.
Shadi Albarqouni is Senior Research Scientist at Chair for Computer Aided Medical Procedures (CAMP) at Technical University of Munich (TUM), Germany. He received his Ph.D. in Computer Science from Technical University of Munich, Germany. He has been working on machine learning with an emphasis on deep learning for medical applications. Albarqouni has published more than 30 papers in both Medical Imaging Computing and Computer Assisted Interventions and presented in IEEE TMI, MICCAI, IPCAI, IJCARS, BMVC, and ICRA. He serves as a reviewer for dozens of top-tier conferences and journals including IEEE TMI, IEEE JBHI, IJCARS and Pattern Recognition. Further, he has been serving as a PC member for a couple of MICCAI workshops during the period 2015-2018. His current research interests include Semi-/Weakly Supervised Deep Learning, Domain Adaptation, and Uncertainty and Explainability of Deep Learning Models.
Ul Balis is professor of Pathology at the University of Michigan and currently serves as the director of the Division of Pathology Informatics, in the Department of Pathology. He is a board-certified Pathology Informaticist, with longstanding interest in the intersection of computational approaches and the practice of medicine. This division he directs is noteworthy for being one of the few such academic information technology divisions operating in support of pathology while being housed wholly within the pathology department itself. He has active, NIH R01-supported research initiatives in several areas of pathology and medical informatics, including machine learning and use of encoded data, image-based analytics, machine vision tools for histopathology, image-based search algorithms and federated enterprise data architectures, with all of these areas serving as rich training substrate for a growing and thriving pathology informatics fellowship â€“ one of only five such programs in the U.S. Dr. Balis has had a longstanding interest in pathology informatics education, and currently serves as a standing member on the Clinical Informatics Subspecialty Boards Exam Committee. Dr. Balis is the author of over 100 publications, multiple patents, numerous book chapters and is co-editor of a contemporary text on the topic of Pathology Informatics (along with Drs. Mark Tuthill and Liron Pantanowitz). He has delivered over 180 invited presentations, nationally and internationally, on various topics related to: pathology informatics, image analysis, data analytics and automation.
Jeroen van der Laak is associate professor in computational Pathology at the Department of Pathology of the Radboud University Medical Center in Nijmegen, The Netherlands and guest professor at the Center for Medical Image Science and Visualization (CMIV) in Linkoping, Sweden. His research focuses on the use of machine learning for the analysis of whole slide images. Application areas include: improvement of routine pathology diagnostics, objective quantification of immunohistochemical markers, and study of novel imaging biomarkers for prognostics. Dr van der Laak has an MSc in computer science and acquired his PhD from the Radboud University in Nijmegen. He co-authored over 95 peer-reviewed publications and is member of the editorial boards of Laboratory Investigation and the Journal of Pathology Informatics. He is member of the board of directors of the Digital Pathology Association and organizer of sessions at the European Congress of Pathology and the Pathology Visions conference. He coordinated the highly successful CAMELYON grand challenges in 2016 and 2017. Dr van der Laak acquired research grants from the European Union and the Dutch Cancer Society, among others.