Segmentation: Basic Concepts and Modern Methods

This lecture provides a general overview of segmentation and a hands-on practical session. This lecture starts by defining segmentation as the partitioning of images into meaningful regions, such as organs or lesions, to improve clinical workflows and measurement accuracy. The lecture covers essential quality criteria, including accuracy (measured using  Dice coefficients and Hausdorff distances) and trustworthiness. The lecture also mentions current state-of-the-art solutions, such as nnU-Net, which leverage convolutional neural networks to achieve superior robustness with less manual algorithm tuning.

Learning objectives:
By the end of this lecture, students will be able to:

• Grasp the technical definition of partitioning images into semantically meaningful sub-regions (anatomical structures, organs, lesions) 
• Evaluate segmentation quality
• Measure accuracy using metrics like the Dice coefficient, Jaccard coefficient, and Hausdorff distance
• Use building blocks like thresholding, morphological operators (erosion/dilation), and convolution kernels for smoothing and edge detection 
• Understand the transition to CNNs (Convolutional Neural Networks) and how to utilize tools like nnU-Net and TotalSegmentator for automated tasks
• Recognize the limitations of modern methods, such as "shortcut learning" and the importance of independent test sets