Supervised Nonparametric Image Parcellation

DOI: 10.1007/978-3-642-04271-3_130 In book: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2009, pp.1075-1083


Segmentation of medical images is commonly formulated as a supervised learning problem, where manually labeled training data
are summarized using a parametric atlas. Summarizing the data alleviates the computational burden at the expense of possibly
losing valuable information on inter-subject variability. This paper presents a novel framework for Supervised Nonparametric
Image Parcellation (SNIP). SNIP models the intensity and label images as samples of a joint distribution estimated from the
training data in a non-parametric fashion. By capitalizing on recently developed fast and robust pairwise image alignment
tools, SNIP employs the entire training data to segment a new image via Expectation Maximization. The use of multiple registrations increases robustness
to occasional registration failures. We report experiments on 39 volumetric brain MRI scans with manual labels for the white
matter, cortex and subcortical structures. SNIP yields better segmentation than state-of-the-art algorithms in multiple regions
of interest.

Download full-text


Available from: Mert R Sabuncu, Sep 24, 2015
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: 51 52 Bayesian learning of latent variable models 2.1 Bayesian modeling and variational learning Unsupervised learning methods are often based on a generative approach where the goal is to find a latent variable model which explains how the observations were generated. It is assumed that there exist certain latent variables (also called in different contexts source signals, factors, or hidden variables) which have generated the observed data through an unknown mapping. The goal of generative learning is to identify both the latent variables and the unknown generative mapping. The success of a specific model depends on how well it captures the structure of the phenomena underlying the observations. Various linear models have been popular, because their mathematical treatment is fairly easy. However, in many realistic cases the observations have been generated by a nonlinear process. Unsupervised learning of a nonlinear model is a challenging task, because it is typically computationally much more demanding than for linear models, and flexible models require strong regularization for avoiding overfitting.
    Full-text · Article · May 2011
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Image registration is typically formulated as an optimization problem with multiple tunable, manually set parameters. We present a principled framework for learning thousands of parameters of registration cost functions, such as a spatially-varying tradeoff between the image dissimilarity and regularization terms. Our approach belongs to the classic machine learning framework of model selection by optimization of cross-validation error. This second layer of optimization of cross-validation error over and above registration selects parameters in the registration cost function that result in good registration as measured by the performance of the specific application in a training data set. Much research effort has been devoted to developing generic registration algorithms, which are then specialized to particular imaging modalities, particular imaging targets and particular postregistration analyses. Our framework allows for a systematic adaptation of generic registration cost functions to specific applications by learning the "free" parameters in the cost functions. Here, we consider the application of localizing underlying cytoarchitecture and functional regions in the cerebral cortex by alignment of cortical folding. Most previous work assumes that perfectly registering the macro-anatomy also perfectly aligns the underlying cortical function even though macro-anatomy does not completely predict brain function. In contrast, we learn 1) optimal weights on different cortical folds or 2) optimal cortical folding template in the generic weighted sum of squared differences dissimilarity measure for the localization task. We demonstrate state-of-the-art localization results in both histological and functional magnetic resonance imaging data sets.
    Full-text · Article · Jul 2010
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: We present a new non-parametric model constraint graph min-cut algorithm for automatic kidney segmentation in CT images. The segmentation is formulated as a maximum a-posteriori estimation of a model-driven Markov random field. A non-parametric hybrid shape and intensity model is treated as a latent variable in the energy functional. The latent model and labeling map that minimize the energy functional are then simultaneously computed with an expectation maximization approach. The main advantages of our method are that it does not assume a fixed parametric prior model, which is subjective to inter-patient variability and registration errors, and that it combines both the model and the image information into a unified graph min-cut based segmentation framework. We evaluated our method on 20 kidneys from 10 CT datasets with and without contrast agent for which ground-truth segmentations were generated by averaging three manual segmentations. Our method yields an average volumetric overlap error of 10.95%, and average symmetric surface distance of 0.79mm. These results indicate that our method is accurate and robust for kidney segmentation. Additional material can be found at .
    Full-text · Article · Sep 2010
Show more