Modelling and Propagating Uncertainties in Machine Learning for Medical Images of Patients with Neurological Diseases
Tal Arbel
April 28, 2021, Wednesday, 3:00 PM - 4:00 PM EDT
Abstract
Although deep learning (DL) models have been shown to outperform other frameworks for a variety of medical contexts, inference in the presence of pathology in medical images presents challenges to popular networks. Errors in deterministic outputs lead to distrust by clinicians and hinders the adoption of DL methods in the clinic. Moreover, given that medical image analysis typically requires a sequence of inference tasks to be performed, this results in an accumulation of errors over the sequence of outputs. This talk will describe recent work exploring (MC-dropout) measures of uncertainty in DL lesion and tumour detection and segmentation models in patient images and illustrate how propagating uncertainties across cascaded medical imaging tasks can improve DL inference. The models have been successfully applied to large-scale Multiple Sclerosis clinical trial datasets and to the MICCAI BRaTs brain tumour segmentation challenge datasets. Finally, we describe a new hierarchical adversarial knowledge distillation network (HAD-Net) that improves enhanced tumour segmentation in the absence of post-contrast enhanced images (e.g. post Gadolinium injection). We show that the estimated uncertainties associated with the HAD-Net outputs do correlate with segmentation errors, paving the way for clinical review and potentially for future integration into clinical workflow.
Bio
Tal Arbel is Professor of Department of Electrical and Computer Engineering at McGill University. Her research goals are to develop new probabilistic machine learning frameworks in computer vision and in medical imaging, particularly in the context of neurology and neurosurgery. This includes the development of probabilistic graphical models for pathology (lesion, tumour) detection and segmentation in large, multi-center patient images dataset, on automatically identifying imaging biomarkers that predict disease progression in patients as well as potential responders to treatment. She has worked extensively on developing fast and efficient multi-modal image registration techniques for clinical interventions, such as image-guided neurosurgery.

Key topics of interest: Bayesian inference, statistical models, statistical pattern recognition, information theory, face detection and trait classification, medical image analysis, neurology and neurosurgery, including multi-modal image registration and lesion and tumour, detection, segmentation, classification and prediction.