Segmentation Workshop - Télécom ParisTech › angelini › shared_files › ... · 2008-11-17 ·...
Transcript of Segmentation Workshop - Télécom ParisTech › angelini › shared_files › ... · 2008-11-17 ·...
• 3 Challenges: MS lesions, Liver tumors, Coronary arteries
• Training and evaluation data, quantitative evaluation
• 36 teams submitted results (120 registered)
Segmentation Workshop
• Organized by Wiro Niessen, Theo van Walsum, Coert Metz, Michiel Schaap (Erasmus, Rotterdam NL)
• 32 datasets with ground truth by experts (with variability)
• 8 training
• 16 pre-workshop testing
• 8 on-site testing
• Quantitative criteria (overlap and accuracy)
• 3 categories: fully auto (5), minimal inter. (3), interactive (5)
Coronary Challenge
• Winners:
• fully auto: Zambal et. al. (VrVis, Austria)
• min. interactive: Dikici and O’Donnell (SCR)
• interactive: Friman et. al. (Mevis Lab, Germany)
• Siemens’ team 2 (Tek et. al.): 2nd of fully auto, 3rd overall
Coronary Challenge
• Zambal et. al.
• heart model optimized by diffeomorphisms
• sampling feature with histogram separability measure
Coronary Challenge
• Friman et. al. (cf. ISBI’08)
• template model optimization
• multi-hypothesis deterministic tracking
Coronary Challenge
• Dikici and O’Donnell
• axis symmetry voting + graph-cuts
Coronary Challenge
• Bauer and Bischof (Graz, Austria)
• GVF-based medialness
Coronary Challenge
• Keynote: John Condeelis (A. Einstein College, Yeshiva U., NY)
“High resolution optical in vivo imaging of tumor cell mobility, chemotaxis, invasion and metastasis in breast tumors”
•Multi-photon imag.
•Cell tagging•Interactions
• Effect of graph topology (neighborhoods)
• 6-c, 26-c, 10-c
• Effect of weighting function
• Graph algorithms: graph-cut and random walker
• Tests on 62 CT datasets (lymph nodes and tumor mainly)
• Results: topology
• increased connectivity stabilizes parameter sensitivity
• isotropic neighborhoods (6-, 26-c) more stable than anisotropic ones
• Results: weighting
• histogram-based best, but large deviation (app.-dependent)
• reciprocal weighting consistently more stable than Gaussian
• better absolute performance
• lower std. deviation
• significantly lower parameter sensitivity
• T1 MR (tested on IBSR Harvard)
• Tree dependencies between structuresVentricles
Caudate Nuclei
Putamina Thalami
• Markovian dependencies from manual binary segmentations
• Model: 1 binary segmentation
• Auto-context model
• iterative training of classifiers (Adaboost and PBTs)
• start with image patches -> probability maps
• training set is augmented with these maps
• training is iterated
• Proof of convergence (error decreasing)
• Rely on Adaboost for discrimination
• Simultaneous co-registration and construction of image templates from a set of images
• Gaussian mixture model
• Optimization by Generalized E.M. (B-Spline registration)
• Similar (identical?) to S. Allassonniere’s model (Poly.)
• Various optimizations: stochastic sampling, registration “anchoring”, etc...
• Notion of “stability” of the templates
• Experiment on the OASIS database (416 brain MR)(age and Alzheimer)
K=2
K=3
“Some brains age faster than others”
• Dementia experiment (60 datasets from OASIS)
“Dementia not a binary state”
• Texture models for Active Shape Models
• Comparison with Active Appearance Models (AAM, Cootes)
• Application to heart, brain ventricles, vertebraes
• Idea: model uncertainty through entropy
• Optimize intensity mappings to s levels (normalization)
• Competition between model specificity and image information
• Optimization of the model by simulated annealing
• Analogy with MDL
• Optimization of unseen pictures
• iterative max a posteriori
• normalization (mapping) by voting by the models (max likelihood)
• prior from the ASM
• exhaustive search on geometric parameters