Explore the critical role of causality in medical imaging through this insightful lecture by Ben Glocker from Imperial College London. Delve into the challenges of data scarcity and data mismatch in medical image analysis, and discover how causal reasoning can provide new perspectives on these issues. Learn about the causal relationships between images, annotations, and data-collection processes, and their impact on predictive model performance and learning strategies. Examine surprising insights, such as the potential unsuitability of semi-supervision for image segmentation. Investigate real-world examples in skin lesion classification, brain tumor segmentation, and radiology reports. Gain knowledge on topics including semi-supervised learning, data augmentation, dataset shift, acquisition shift, domain adaptation, and domain generalization. Understand the importance of considering causal relationships in machine learning-based image analysis for improved success in clinical practice. Conclude with guidelines, regulations, and recommendations for implementing causal reasoning in medical imaging research and applications.
Overview
Syllabus
Intro
Predictive Modelling
Challenges: Data Scarcity & Mismatch
A Causal Perspective
Example: Skin Lesion Classification
Example: Brain Tumour Segmentation
Example: Radiology Reports
Semi-supervised learning
Data Augmentation
Dataset Shift
Acquisition Shift A Little Experiment
Acquisition Shift: A Little Experiment
Domain Adaptation
Domain Generalisation
PACS Benchmark
Episodic Training
Global Class Alignment
Local Sample Clustering
Back to Causality
Guidelines and Regulation
Recommendations
Taught by
Institute for Pure & Applied Mathematics (IPAM)
