sitem-insel AG

Participants will learn about the relationship between a clinical problem and a machine learning problem as well as the influence of training data quality and annotation errors on the ability of algorithms to learn. Participants will be introduced to state-of-the-art AI models and learn how to apply and validate them in practice.

Main learning objectives

• Understand the formulation of a clinical problem as a machine-learning problem

• Understand the challenges and practicalities of assembling a training set of medical images

• Understand the tradeoffs between manual or automatic labelling of training data

• Understand how to select a state-of the art model, and have hands-on experience training such a model

• Understand and have hands on experience with selecting and applying appropriate metrics for the performance of a medical AI system

• Understand how the results of validating a machine-learning model may affect design choices in a subsequent experiment

Learning content

• Introduction to image classifiers, ChexNet dataset, image segmentation and Decathlon datasets
• Labeling of image data, automatic vs. manual methods
• Adapting existing code to a new problem: classification
• Modern CNN architectures explained
• Adapting code to a new problem: segmentation
• Deep Learning Classifiers versus old fashioned modelling
• Interpreting model performance
• Issues in adapting code, 3D data, multimodal data, modern deep learning