MICCAI 2014 Tutorial

Medical Image Segmentation: Beyond Level-Sets

Date/time: Sunday, September 14, morning (8:30-12:30)

Place: MIT, Stratton Building (Student center), room 491 (W20-491), access 4th floor by elevator

Organizers:

• Yuri Boykov, University of Western Ontario, Canada
• Ismail Ben Ayed, GE Healthcare, Canada
• Daniel Cremers, Technical University of Munchen, Germany
• Milan Sonka, University of Iowa, USA

Outline: This tutorial overviews recent developments in optimization-based geometric methods for image segmentation that should be accessible to the general MICCAI audience. First, we describe common mathematical models behind standard segmentation techniques and, then, illustrate how different surface representation techniques (discrete or continuous) lead to different algorithms with different properties. We emphasize graph-theoretic and continuous relaxation techniques that had significant impact in computer vision community in the last 10 years due to ability to compute globally optimal solutions for many problems. In particular, our goal is to juxtapose them with standard level-set techniques, which have a long history in the MICCAI community. While presenting the basic numerical ideas behind each approach, our emphasis is on conceptual understanding of presented methods so that tutorial attendees gain sufficient knowledge to use publicly available code. We will demonstrate a number of practical applications and compare different methods.

Academic objectives: overview of recent advances in optimization-based geometric approaches to image segmentation with emphasis on general concepts and public code libraries. The emphasis is on global optimization techniques and medical applications. Target audience: general MICCAI attendees.

Slides:

• part 1: (yuri) basic binary segmentation
• part 2: (yuri + ismail) advanced binary segmentation
• part 3: (milan) multi-region segmentation LOGISMOS Master, single/multiple surfaces, costs, complex/multi objects,
• extra materials on level-sets: ismail
• extra materials on convex relaxations: daniel (see Daniel's video)

Tentative Schedule:

• Part 1: Overview of objective functions (70 min):
  • 1.1 Basic binary segmentation
    --- continuous and discrete formulations (meshes, level-sets, graphs, etc.)
    --- regional and boundary terms
    --- linear, quadratic, higher-order terms
    --- shape priors (connectivity, convexity, shape templates, shape moments, etc)
  • 1.2 Multi-region segmentation
    --- ray-based
    --- grid-based

  ------- 15 min break -------

• Part 2: Overview of standard optimization techniques (70 min)
  • 2.1 Basics of continuous methods (active contours, level-sets, convex relaxations)
  • 2.2 Graph cuts: exact and approximate techniques
    --- submodularity and graph cuts
    --- approximating methods: trust region, bound and pseudo-bound optimization

  ------- 15 min break --------

• Part 3: Comparisons and Applications (70 min)
  • 3.1 - Graph Cuts vs Level Sets vs Convex Relaxation
  • 3.2 - Cardiac, spine, brain tumors
  • 3.3 - Pulmonary, cardiac, ophthalmic, etc.