MIT AI Knowledge Based 3D Medical Image Segmentation Tina Kapur MIT Artificial Intelligence Laboratory
MIT AI Outline Goal of Segmentation Applications Why is segmentation difficult? My method for segmentation of MRI Future Work
MIT AI The Goal of Segmentation
MIT AI The Goal of Segmentation
MIT AI Applications of Segmentation Image Guided Surgery
MIT AI Applications of Segmentation Image Guided Surgery
MIT AI Applications of Segmentation Image Guided Surgery Surgical Simulation
MIT AI Applications of Segmentation Image Guided Surgery Surgical Simulation
MIT AI Applications of Segmentation Image Guided Surgery Surgical Simulation Neuroscience Studies Therapy Evaluation
MIT AI Limitations of Manual Segmentation slow (up to 60 hours per scan) variable (up to 15% between experts) [Warfield 95, Kaus98]
MIT AI The Automatic Segmentation Challenge An automated segmentation method needs to reconcile –Gray-level appearance of tissue –Characteristics of imaging modality –Geometry of anatomy
MIT AI How to Segment? i.e. Issues in Segmentation of Anatomy
MIT AI How to Segment? i.e. Issues in Segmentation of Anatomy Tissue Intensity Models
MIT AI How to Segment? i.e. Issues in Segmentation of Anatomy Tissue Intensity Models –Parametric [Vannier] –Non-Parametric [Gerig] –Point distribution Models [Cootes] –Texture [Mumford]
MIT AI How to Segment? i.e. Issues in Segmentation of Anatomy Tissue Intensity Models Imaging Modality Models
MIT AI How to Segment? i.e. Issues in Segmentation of Anatomy Tissue Intensity Models Imaging Modality Models –MRI inhomogeneity [Wells]
MIT AI How to Segment? i.e. Issues in Segmentation of Anatomy Tissue Intensity Models Imaging Modality Models Anatomy Models: Shape, Geometric/Spatial
MIT AI How to Segment? i.e. Issues in Segmentation of Anatomy Tissue Intensity Models Imaging Modality Models Anatomy Models: Shape, Geometric/Spatial –PCA [Cootes and Taylor, Gerig, Duncan, Martin] –Landmark Based [Evans] –Atlas [Warfield]
MIT AI Typical Pipeline for Segmentation of Brain MRI Pre-processing for noise reduction EM Segmentation Morphological or other post-processing pre-processing (noise removal)
MIT AI Typical Pipeline for Segmentation of Brain MRI Pre-processing for noise reduction EM Segmentation Morphological or other post-processing pre-processing (noise removal) intensity-based classification
MIT AI Typical Pipeline for Segmentation of Brain MRI Pre-processing for noise reduction EM Segmentation Morphological or other post-processing pre-processing (noise removal) post-processing (morphology/other) intensity-based classification
MIT AI Contributions of Thesis Developed an integrated Bayesian Segmentation Method for MRI that incorporates de-noising and global geometric knowledge using priors into EM- Segmentation Applied integrated Bayesian method to segmentation of Brain and Knee MRI.
MIT AI Contributions of Thesis The Priors –de-noising: novel use of a Mean-Field Approximation to a Gibbs random field in conjunction with EM-Segmentation (EM-MF) –geometric: novel statistical description of global spatial relationships between structures, used as a spatially varying prior in EM- Segmentation
MIT AI Background to My Work Expectation-Maximization Algorithm EM-Segmentation
MIT AI Expectation-Maximization Relevant Literature: –[Dempster, Laird, Rubin 1977] –[Neal 1998]
MIT AI Expectation-Maximization (what?) Search Algorithm for Parameters of a Model to Maximize Likelihood of Data Data: some observed, some unobserved
MIT AI Expectation-Maximization (how?) Initial Guess of Model Parameters Re-estimate Model Parameters: –E Step: compute PDF for hidden variables, given observations and current model parameters –M Step: compute ML model parameters assuming pdf for hidden variables is correct
MIT AI Notation –Observed Variables: –Hidden Variables : –Model Parameters: Expectation-Maximization (how exactly?)
MIT AI Initial Guess: Successive Estimation of –E Step: –M Step: Expectation-Maximization (how exactly?)
MIT AI Expectation-Maximization Summary/Intuition: –If we had complete data, maximize likelihood –Since some data is missing, approximate likelihood with its expectation –Converges to local maximum of likelihood
MIT AI EM-Segmentation [Wells 1994] Observed Signal is modeled as a product of the true signal and a corrupting gain field due to the imaging equipment Expectation-Maximization is used on log- transformed observations for iterative estimation of –tissue classification –corrupting bias field (inhomogeneity correction)
MIT AI M-Step E-Step EM-Segmentation [Wells 1994]
MIT AI Estimate intensity correction using residuals based on current posteriors. Compute tissue posteriors using current intensity correction. M-Step E-Step EM-Segmentation [Wells 1994]
MIT AI Observed Variables –log transformed intensities in image Hidden Variables –indicator variables for classification Model Parameters –the slowly varying corrupting bias field ( refer to variables at voxel s in image) EM-Segmentation [Wells 1994]
MIT AI Initial Guess: Successive Estimation of –E Step: –M Step: EM-Segmentation [Wells 1994]
MIT AI Initial Guess: Successive Estimation of –E Step: –M Step: EM-Segmentation [Wells 1994]
MIT AI Situating My Work Prior in EM-Segmentation: –Independent and Spatially Stationary My contribution is addition of two priors: –a spatially stationary Gibbs prior to model local interactions between neighbors (thermal noise) –spatially varying prior to model global relationships between geometry of structures
MIT AI The Gibbs Prior Gibbs Random Field (GRF) –natural way to model piecewise homogeneous phenomena –used in image restoration [Geman&Geman 84] –Probability Model on a lattice –Partially Relaxes independence assumption to allow interactions between neighbors
MIT AI EM-MF Segmentation: EM + Gibbs Prior We model tissue classification W as a Gibbs random field:
MIT AI We model tissue classification W as a Gibbs random field: EM-MF Segmentation: Gibbs Prior on Classification
MIT AI To fully specify the Gibbs model: –define neighborhood system as a first order neighborhood system i.e. 6 closest voxels –use to define EM-MF Segmentation: Gibbs Prior on Classification
MIT AI EM-MF Segmentation: Gibbs form of Posterior Gibbs prior and Gaussian Measurement Models lead to Gibbs form for Posterior:
MIT AI Gibbs prior and Gaussian Measurement Models lead to Gibbs form for Posterior: EM-MF Segmentation: Gibbs form of Posterior
MIT AI EM-MF Segmentation For E-Step: Need values for
MIT AI EM-MF Segmentation For E-Step: Need values for Cannot compute directly from Gibbs form
MIT AI EM-MF Segmentation For E-Step: Need values for Cannot compute directly from Gibbs form Note
MIT AI EM-MF Segmentation For E-Step: Need values for Cannot compute directly from Gibbs form Note Can approximate – Mean-Field Approximation to GRF
MIT AI Mean-Field Approximation Deterministic Approximation to GRF [Parisi84] –the mean/expected value of a GRF is obtained as a solution to a set of consistency equations Update Equation is obtained using derivative of partition function with respect to the external field g. [Elfadel 93] Used in image reconstruction [Geiger, Yuille, Girosi 91]
MIT AI Mean-Field Approximation to Posterior GRF Intuition: denominator is normalizer numerator captures: effect of labels at neighbors measurement at voxel itself
MIT AI Summary of EM-MF Segmentation Modeled piecewise homogeneity of tissue using a Gibbs prior on classification Lead to Gibbs form for Posteriors Posterior Probabilities in E-Step are approximated as a Mean-Field solution
MIT AI EM-MF Results Application: Brain MRI –white matter, gray matter, fluid/air, skin/scalp Results Comparison with Manual Segmentation
MIT AI Some Results EM EM-MF
MIT AI Some Results EM EM-MF
MIT AI More Results Noisy MRIEM Segmentation EM-MF Segmentation
MIT AI Posterior Probabilities (EM) White matter Gray matter
MIT AI Posterior Probabilities (EM-MF) White matter Gray matter
MIT AI Results
MIT AI Modeling Global Geometric Relationships between Structures
MIT AI Relative Geometry Models Motivate Using Knee MRI Brain MRI Example Modeling Global Geometric Relationships between Structures
MIT AI Segmented Knee MRI Femur Tibia Femoral Cartilage Tibial Cartilage MERL, SPL, MIT, CMU Surgical Simulation (Sarah Gibson, PI)
MIT AI Motivation Primary Structures –image well –easy to segment Secondary Structures –image poorly –relative to primary Tibial Cartilage Femoral Cartilage Tibia Femur
MIT AI Relative Geometric Prior Approach Select primary/secondary structures Measure geometric relation between primary and secondary structures from training data Given novel image –segment primary structures –use geometric relation as prior on secondary structure in EM-MF Segmentation
MIT AI Segment Primary Structures: Femur, Tibia SeedRegion GrowingBoundary Localization
MIT AI Status Have Bone Want Cartilage
MIT AI Measure Geometric Relationship between Primary and Secondary Structures Femur Tibia Femoral Cartilage Tibial Cartilage Using primitives such as –distances between surfaces –local normals of primary structures –local curvature of primary structures –etc.
MIT AI Femur Tibia Femoral Cartilage Tibial Cartilage Measure Geometric Relationship between Primary and Secondary Structures
MIT AI Estimate of
MIT AI Status Have Bone Have spatial relation between Bone and Cartilage Need Cartilage
MIT AI Use Relative Geometric Prior in EM Segmentation Replace stationary prior with relative geometric prior:
MIT AI Results: Segmentation of Femoral & Tibial Cartilage MRI Image Model-Based Segmentation Manual Segmentation
MIT AI Relative Geometric Priors for Brain Tissue Prior Estimation –Select primary structures (boundary of skin, ventricles) –Estimate Using Prior in Segmentation –Segment primary structures: skin, ventricles –Use as geometric prior
MIT AI White MatterGray Matter Estimate distance to Ventricles distance to Skin
MIT AI Resultant Segmentation MRIEM Segmentation EM-MF with Geometric Prior
MIT AI Posterior Probabilities Gray Matter White Matter EM-MF+Geometric PriorEM
MIT AI In Summary Incorporated robustness to thermal noise by using Mean-Field Approximation to Gibbs model in conjunction with EM Segmentation. Applied to Brain MRI. Introduced Relative-Geometry Models and applied to Brain and Knee MRI.
MIT AI Future Work Further development of Relative-Geometry Models: –Automatic selection of primary/secondary structures –Additional primitives for Spatial Relationships
MIT AI STOP
MIT AI d1 (distance to Ventricles) d2 (distance to Skin) White Matter Grey Matter Fluid Air Left Caudate Right Caudate Class Conditional Density
MIT AI Unified Bayesian Segmentation Method Simultaneous noise reduction intensity-based classification use of geometric information for segmentation.
MIT AI Rest of Talk: 1.The Unified Segmentation Method 2.Two Priors 3.Results on Brain, Knee Segmentation 4.Conclusions
MIT AI The Gibbs Prior To fully specify the Gibbs model: –define neighborhood system as a first order neighborhood system i.e. 6 closest voxels –use to define
MIT AI Components of EM Framework Measurement models for tissue Prior models for tissue Model for bias field –piecewise smooth
MIT AI Addition of Two Priors Gibbs prior on tissue appearance –models tissue as piecewise constant Geometric Prior to encode spatial relations –gray matter is outside ventricles and inside skull
MIT AI Gibbs Model
MIT AI Gibbs Model probability model on a lattice independence assumption is partially relaxed spatial range of interaction is local neighborhood
MIT AI
MIT AI Gibbs Model probability model on a lattice independence assumption is partially relaxed spatial range of interaction is local neighborhood Mean-Field Approximation Approximates neighboring random variables with their mean values: - algebraic and computational simplicity
MIT AI Contributions of Thesis MRI+NoiseEM Segmentation EM-MF with Geometric Prior
MIT AI Proposed MRI Segmentation Method Bayesian Statistical Classification Scheme that uses Expectation-Maximization Replaces pipeline with Priors on intensity and geometry
MIT AI Proposed MRI Segmentation Method Previous Work [Wells 1994, 1996] –derived as a special case –spatially stationary, independent priors –piecewise smooth inhomogeneity model This Work: –locally interacting prior for intensity –spatially varying prior for geometry
MIT AI Next Background on Expectation-Maximization
MIT AI 1. Use Bayes’ rule, and independence between and to write: 2. Specify Measurement Models as Gaussian Computation of
MIT AI 3. Specify Prior on Classification as a Gibbs Random Field. 4. Gibbs Prior + Gaussian Measurement Model imply is also a GRF. 5. Approximate using Mean- Field Solution for the GRF. Computation of
MIT AI Recap: 1. Bayes’ Rule to rewrite … 2. Gaussian Measurement Models 3. Gibbs form of Prior 4. Gibbs form of Posterior 5. Mean-Field Approximation to Gibbs form Computation of
MIT AI The Gibbs Prior Tissue-class interaction Matrix Stationary Prior
MIT AI The Mean-Field Solution Gibbs models can be solved using –[Metropolis 1953], [Geman and Geman 1984], [Besag 1986] etc. We use Mean-Field Approximation to estimate the expected value of the posterior GRF as a solution to a set of consistency equations.
MIT AI The Mean-Field Approximation Deterministic Approximation Update Equation is obtained using derivative of partition function with respect to the external field g.
MIT AI The Mean-Field Solution Intuition: denominator is normalizer numerator captures: effect of labels at neighbors measurement at voxel itself
MIT AI Initial Guess: Successive Estimation of –E Step: Estimate as Mean-Field Solution to a Gibbs Random Field –M Step: Compute same as [Wells 1996] EM-MF Summary