journalclub - robust statistical fusion of image labels

Upload: telnet2

Post on 05-Apr-2018

219 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    1/28

    ROBUST STATISTICAL

    FUSION OF IMAGE LABELSLandman et al

    IEEE Transactions on Medical Imaging

    VOL. 31, NO. 2, FEBRUARY 2012

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    2/28

    Labeling in MR image analysis

    Labeling Problem

    Identifying class membership of voxels

    Currently no true answer

    Manual Voxel-by-voxel Labeling

    Considered as a gold standard

    Exceptionally time consuming and resource intensive

    Difference in interpretation between raters

    Validating automatic or semi-automatic methods

    The study of structures for which no automated method exists

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    3/28

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    4/28

    Limitation of STAPLE algorithm

    STAPLE requires that all raters delineate all voxelswithin a given region

    Raters are often requested to label datasets more thanoncein order to establish a measure of intra-raterreliability.

    Raters are often divided into a class of experts and

    novices

    Label inversion problem due to highly inaccurate raters

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    5/28

    Label inversion problem

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    6/28

    STAPLER algorithm

    Algorithm

    Estimate a maximum a posterioriof both rater reliability and truelabels in the Expectation Maximization framework

    Evaluation

    Random rater simulation

    Boundary random rater simulation

    Simulations to characterize the occurrence of label inversionproblem

    Note Minimally trained raters and large number of participants

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    7/28

    STAPLER algorithm

    How to estimate true labels?

    Majority voting

    :

    How to estimate rater performance?

    Rater performance is not considered to be perfect but to be fuzzy.

    | : [ , , ,,()]

    Confusion matrix (hidden variable in EM framework)

    1 2

    1 0.9 0.2

    2 0.1 0.8

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    8/28

    EM algorithm

    Expectation Maximization

    Parameter estimation in probabilistic models with incomplete data

    Computes iteratively the Maximum Likelihood estimationwith theassumption of hidden variable

    Toy Example: Coin flipping with two different coins, A and B

    {, } : 10 trials of flipping.

    ,,,,,,,,,,, , , ,

    {, }

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    9/28

    EM algorithm

    Assume that we know which coin is flipped, for example:

    How to estimate , maximizing the likelihood of(| , )?

    A: 24 Head 6 Tail

    B: 9 Head 11 Tail

    Coin Results

    B H T T T H H T H T H (5H 5T)

    AH H H H T H H H H H (9H 1T)

    A H H H H T H T H H H (8H 2T)

    B H T T T T H H T H T (4H 6T)

    A H T H H T T H H H H (7H 3T)

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    10/28

    EM algorithm

    + 0.8

    + 0.45

    Maximum Likelihood Estimation

    arg max

    (| , )

    30

    24

    1

    24

    1 6

    1

    1

    24 24 6 2 43 0 0

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    11/28

    EM algorithm

    What if we dont know which coin is used?

    {, , , } is called hidden variable or latent factor

    Maximizing log with respect to

    Efficient iterative process and guarantees to converge

    Repeat E-step and M-step until the algorithm converges

    E-step: , [log (, |)]

    M-step: argmax , [log (, |)]

    EM algorithm becomes very slow as the number of variables increases.

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    12/28

    EM algorithm

    Begin with initial parameters

    , 0.6, 0.5

    Compute the probability of the event for A and B

    Event (, , ) (, , ) ( |, )

    #1 (5H 5T) 0.2007

    (105 0.60.4)

    0.2495

    (105 0.50.5)0.2/(0.2+0.25)=0.45

    #2 (9H 1T) 0.04 0.0098#3 (8H 2T) 0.1209 0.0439

    #4 (4H 6T) 0.1115 0.2051

    #5 (7H 3T) 0.2150 0.1172

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    13/28

    EM algorithm

    Compute expectation (E-step)

    Estimate new parameter (M-step)

    .

    .+. 0.71,

    .

    .+. 0.58

    Event ( |, ) ( |, ) ( , ) ( , )

    #1 (5H 5T) 0.45 0.55 2.2H 2.2T 2.8H 2.8T

    #2 (9H 1T) 0.80 0.2 7.2H 0.8T 1.8H 0.2T

    #3 (8H 2T) 0.73 0.27 5.9H 1.5T 2.1H 0.5T

    #4 (4H 6T) 0.35 0.65 1.4H 2.1T 2.6H 3.9T

    #5 (7H 3T) 0.65 0.35 4.5H 1.9T 2.5H 1.1T

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    14/28

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    15/28

    Label Fusion

    An image of N voxels

    : the number of training voxels

    : the number of undetermined voxels

    , , arethe set of all voxels

    {0,2,3, , L 1}is all possible labels

    {, , }is a collection of raters

    represents the r-th observation of voxel iby raterj

    represents the hidden true segmentation

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    16/28

    STAPLE algorithm

    Estimate true segmentation in a probabilistic framework

    : [ , , ,,()]

    :

    Probability distribution function of the true label

    () ( matrix) represents the probability that voxel ihas true

    label son the k-th iteration

    Hidden variable that control the p.d.f

    (|) ( matrix) represents the probability that raterj

    observes label swhen the true label is son the k-th iteration (raterperformance)

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    17/28

    STAPLE algorithm

    EM algorithmto estimate the hidden true segmentation

    E-step: the calculation of the conditional probability ofthe true segmentation

    () , (

    =)

    (|)

    (=) (|)

    M-step: the calculation of the rater performanceparameters

    +(|)

    ():

    ()

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    18/28

    STAPLE algorithm

    Example

    1 ...

    ...+... 0.9545

    2 ...

    ...+... 0.6667

    Rater A B C

    Label 1 1 1 2

    Label 2 2 1 1

    A 1 2

    1 0.9 0.2

    2 0.1 0.8

    B 1 2

    1 0.7 0.4

    2 0.3 0.6

    C 1 2

    1 0.5 0.5

    2 0.5 0.5

    1 1 .9545

    0.9545 0.3333

    2 1 .3333

    0.9545 0.3333

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    19/28

    STAPLER algorithm

    New E-step

    ,

    (=)

    (|):

    (=) ):

    where + is a global prior

    +

    New M-step

    +

    =

    ()

    ()

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    20/28

    Results

    Data

    A high-resolution MPRAGE

    149 x 81 x 39 voxels

    0.82 x 0.82 x 1.5 mm resolution

    Expert labeled the cerebellum from each dataset with 12 divisionsof the cerebellar hemispheres

    Simulation

    Voxel-wise random raters

    Boundary random raters Evaluation

    Jaccard Index

    Dice coefficient

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    21/28

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    22/28

    Result

    Voxel-wise random simulated labels

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    23/28

    Result

    Boundary-random simulated labels

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    24/28

    Result

    Empirical simulation results

    38 undergraduate raters

    Raters labeled between 10 and 100 slices for the axial set

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    25/28

    Discussion

    STAPLER extends the applicability of the STAPLEtechnique to common research situations with missing,partial, and repeated data.

    STAPLER facilitates use of training data and reliability

    priors to improve accuracy.

    STAPLER enables parallel manual labeling and reducesdetrimental impacts

    STAPLER can readily be augmented by introducingspatially adaptive, unconditional label probabilities.

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    26/28

    EM algorithm

    Log likelihood function defined as

    log (|)

    log is strictly concave and also strictly increasing because

    < 0

    Definition of convex function

    (1 ) (1 )()

    Compute an updated estimate such that,

    ln ln (|)

    Introduce a hidden variable , (|)

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    27/28

    EM algorithm

    Rewrite the equation

    ln , ln

    ln , ,

    , , ln

    ,

    ln ,

    , (|

    )

    Thus,

    + arg max arg max , ln ,

    arg max , ln ,

  • 8/2/2019 JournalClub - Robust Statistical Fusion of Image Labels

    28/28