SDiPath Trailblazers in AI Series – Christine Decaestecker

Learning from imperfect annotations in digital pathology

Abstract : Many machine learning applications in the field of digital pathology rely on images annotated by human experts. However, as in other areas of medical imaging, these experts are not infallible on these often laborious tasks, and may even disagree with each other on complex tasks. These annotations are therefore subject to various types of imperfection (imprecision, lacuna, error and discordance) and their use undermines confidence in the quality of the model that is trained on them. This talk will present research carried out within our department (LISA, ULB) aimed at analyzing the impact of imperfect annotations on tasks involving the detection, segmentation and classification of specific objects or structures in whole-slide imaging, as well as strategies for improving the robustness of deep networks trained on such data.

Bio :Christine Decaestecker is senior research associate at the FNRS, director of the Laboratory of Image Synthesis and Analysis (lisa.polytech.ulb.be) at the École Polytechnique de Bruxelles (ULB), and co-head, with Pr. Isabelle Salmon, of the multidisciplinary department of digital pathology (DIAPath) at the Centre of Microscopy and Molecular Imaging (www.cmmi.be). She is the author of over 250 scientific publications. Her current research interests are multidisciplinary and involve image analysis, machine learning and data analysis, particularly in the field of digital pathology.
She is currently involved in the Belgian TRAIL consortium (TRusted AI Labs) and its ARIAC project (Applications and Research for Trusted Artificial Intelligence), where she coordinates the “Human-AI interaction” work package. She is also involved in the Walloon Region’s MedResyst (Network and Systems Medicine) strategic innovation initiative.

Zoom link: https://unibe-ch.zoom.us/j/64244437918?pwd=ME0zc1JlTFZZMTQ5UERpZGpXY0xRdz09

Register here: https://docs.google.com/forms/d/e/1FAIpQLSeElhR830Xyf2qrZYgHhdGtO7Yd1bRCZmmBgg2fpZdJ5xUXKA/viewform