This document discusses modeling anatomical variability in populations using biomedical images like MRI and CT scans. It covers:
1) Common solutions involve geometric transformations to capture variability and probabilistic generative models of image and signal variability. Examples include active shape/appearance models and mixture models.
2) Outstanding challenges include making patient-specific predictions and modeling anatomical variability.
3) Examples of applications discussed are image segmentation, registration, fMRI analysis, and modeling hippocampal and atrial shape differences between populations.
2. Biomedical Image Analysis
(MRI or CT)
• Compared to Computer Vision:
– More constrained problem, defined metrics of success
– Understanding and interpretation
» Population studies vs. patient-specific predictions
• Examples:
– Image segmentation
» Well defined organ/tissue, goal: match or exceed expert
– Image registration/matching
» Metric: how well it supports localization
– fMRI analysis
» How well the model predicts behavior
3. Common solutions
• Geometry and Statistics
– Geometric transformations
» real deformation and capture variability
– Probabilistic generative models
» variability of images (MRI, CT) and signals (fMRI)
• Examples:
– Active shape/appearance models (Cootes and Taylor)
– Mixture models with spatial priors (Wells, Van Leemput)
– Mutual Information registration (Viola and Wells)
– Diffeomorphic transformations (Miller, Thirion, Pennec)
– Hierarchical models (Friston, Penny, Worsley, Genovese, Wells)
• Outstanding challenges
– Patient-specific predictions
– Anatomical variability
4. Anatomical Shape Analysis
• Image-based descriptors of shape
• Model of differences between two populations
– Discriminative modeling
• Explicit representation of differences
– Perturbation analysis
Hippocampal shape in schizophrenia z x
NIPS’01, MedIA’05
5. Anatomical Heterogeneity
• Population as a collection of homogeneous sub-groups
– Correlate clinical and demographic information with the partition
• Mixture model
– Clustering of subjects into sub-populations
– Average template for each sub-population
– Simultaneously estimate the partition and the templates
» Generalized EM algorithm with image registration
Anatomical templates in aging
Young Middle Old
IEEE TMI ’09,’10
6. Nonparametric Atlases for
Segmentation: Left Atrium
• Left atrium segmentation:
• Spatial prior for ablation scars:
MICCAI Workshop on Statistical Atlases and
Computational Models of the Heart ‘10
7. Functional Organization of the Brain
• Cluster functional response profiles Body
– Hierarchical model for population
Face
– Enables discovery of
» Novel functional systems ?
» Novel stimulus categories
– Removes the need for spatial alignment Scene
– Enables novel fMRI experiments
Supported by NSF
CRCNS
NIPS’10,
NeuroImage’10
8. Lessons Learned
• Not a computer vision problem, but an interesting problem
that involves images
– Need to understand the problem and the goals
– Work closely with the scientist
• Well defined goals and (at times painful) ways to validate
– Might change overtime
• Closely related to computer vision, machine learning and
probabilistic inference