MICCAI'09 Tutorial

Discrete Optimization in Biomedical Image Analysis: Methodologies and Applications

Nikos Paragios, Pushmeet Kohli &  Nikos Komodakis

 

  • Slides/Presentation of N. Paragios & N. Komodakis (here)

  • Slides/Presentation of P. Kohli (here)

Ecole Centrale de Paris/INRIA Saclay, Ile-de-France/Microsoft Research/University of Crete

Context: Optimization is a critical component of biomedical image analysis (segmentation, registration, classification, computer-aided diagnosis) both in terms of accuracy with respect to the obtained results as well as in terms of computational constraints. Biomedical image analysis tasks have been predominately addressed through variational formulations and gradient-like optimization methods that can be sensitive to the initial conditions, computationally expensive while at the same time suffer from portability from one application to another. In the recent years we have witnessed a revolution in the field of discrete optimization and computer vision. On one hand we have observed the introduction of novel optimization techniques to solve well known formulations (like MRFs) which can offer guarantees both in terms of the optimality properties as well as in terms of computational efficiency.

Scope: The aim of this proposal is to introduce the notion of discrete modeling of biomedical image analysis tasks. The main focus will be on explaining how well known tasks can be expressed as discrete optimization problems. Discrete Markov Random Fields (MRFs) can model a wide variety of problems in medical imaging and related fields. It is exactly for this reason that MRF optimization is considered to be a task of fundamental importance, which has attracted a significant amount of research over the last years. The goal of this course will be to provide an overview for some of the recent developments in the field of MRF optimization. To this end, we will review a wide range of state-of-the-art discrete optimization algorithms such as graph-cut based techniques, efficient linear programming, tight relaxations, belief propagation networks and higher order interactions. Furthermore, we will show how such formulations can be used to address well studied problems in medical image analysis, like feature selection, model-free and model-based segmentation, deformable pair-wise & population registration, classification in higher dimensional spaces, etc.

Attendance: 75 people

Program [3:30]:

PART 1 (Introduction/Optimization)

  • Mathematical Modeling of Biomedical Image Analysis

  • Markov Random Fields, Theory, Practice and Applications

  • MRFs optimization techniques

    • Graph-Cuts, Max Flow/Min Cuts

    • Efficient Linear Programming, Primal Dual Strategies

    • Belief Propagation Networks, Message Passing Algorithms

    • Tighter Relaxations

PART 2 (Biomedical Image Analysis and MRFs)

  • Computed Tomography Reconstruction

  • Feature Extraction, Feature Structuring, Dimensionality Reduction

  • Model-free and Model-based segmentation with appearance and geometric priors

  • Pair-wise and group-wise rigid and deformable registration using efficient linear programming, local and global metrics, with or without prior knowledge

  • Unsupervised Clustering with automatic cluster selection in high-dimensional spaces

  • Clustering of Skeletal Muscle Fibers

PART 3 [15min]

  • Discussion/Questions