Hippo volume parsing PPT
Download
Report
Transcript Hippo volume parsing PPT
Machine Learning Based Approaches to Subcortical
Segmentation
Jonathan Morra Ph.D.
Outline
Problem
• Bayesian Formulation
Methods
• AdaBoost
• Auto Context Model
Clinical Motivation
Pipeline Demos
Problem
Define
• X = (x1 … xN) and Y = (y1 … yN) be examples and
labels with yi = (1 … K) being one of K labels (2
in our case)
However, this is difficult because of the
complexities of p(Y|X), p(X|Y), and p(Y)
AdaBoost
Alpha update rule
1 t
t log
2 1 t
Weight update rule
Dt 1 (i )
1
Dt (i )e t yi ht ( xi )
Zt
Auto Context Model
Auto Context Model
The auto context model is provably minimizing
the error during each iteration
• AdaBoost is guaranteed to choose features which
minimize a convex error function
• At each iteration of the auto context model we
augment the image based feature pool
• If these features improve the classification
ability, AdaBoost selects them; otherwise, it
ignores them
Features
Signed
Distance
Prior (P(0))
P(0) formulation
Calculate signed distance maps
for each training mask
Transform to (0 1) by Logit
Average over all masks
Features
Intensity
Position
Neighborhood Filters
Various
Haar
Filters
Mean
Standard Deviation
Curvature
Haar (various shapes)
Filters range from 1x1x1 to
7x7x7
Each of the above features was
calculated on each channel
Clinical Motivation
Subcortical segmentation is
an important step in many
areas
• Drug trials
• Studies of disease
• Population statistics
Time consuming by hand
• 2 hours per hippocampus
• In ADNI we have 800 subjects
x 4 time points x 2
hippocampi per individual =
6400 hippocampi
• This results in 12,800 man
hours for this one study
Figure from Fischl et al. “Whole Brain Segmentation: Automated Labeling of Neuroanatomical
Structures in the Human Brain.” Neurotechnique 31, pp 341-355, 2002.
Pipeline
References
Morra JH, Tu Z, Apostolova LG, Green AE, Avedissian C,
Madsen SK, Parikshak N, Hua X, Toga AW, Jack CR, Jr.
and others. (2008): Validation of a fully automated 3D
hippocampal segmentation method using subjects with
Alzheimer's disease mild cognitive impairment, and
elderly controls. Neuroimage 43(1):59-68.
Morra JH, Tu Z, Apostolova LG, Green AE, Toga AW,
Thompson PM (2008): Automatic Subcortical
Segmentation Using a Contextual Model. Medical Image
Computing and Computer Assisited Intervention
(MICCAI). 5241:191-201.
Grant support for this work was provided by the National Institute for Biomedical Imaging and Bioengineering,
the National Center for Research Resources, National Institute on Aging, the National Library of Medicine, and
the National Institute for Child Health and Development (EB01651, RR019771, HD050735, AG016570, LM05639 to
Paul M. Thompson) and by the National Institute of Health Grant U54 RR021813 (UCLA Center for Computational
Biology).