an automatic method for change detection in serial dti-derived...

An automatic method for change detection in

serial DTI-derived scalar images

H. Boisgontier1,2, V. Noblet1, F. Heitz1,L. Rumbach2,3, J.-P. Armspach2

1LSIIT, UMR CNRS-ULP 7005, Strasbourg, France, 2LINC UMR CNRS ULP 7191 St b F

Work supported by2LINC, UMR CNRS-ULP 7191, Strasbourg, France,

3Service de Neurologie, CHU Minjoz, Besançon, France

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

ObjectiveObjective

1st scan 2nd scan

=> Detect significant changes between diffusion-weighted scalar images

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

General schemeGeneral scheme

Examination 1Detection map

... CorrectEddy- Compute

...

Compute

Examination 1

... Eddycurrent

distortion

ComputeTensor

...Compare :StatisticalChange

Computeindex

Change Detection

E i ti 2

... ...

Examination 2

Correct

... Computeindex

ComputeTensor

... RegisterEddy-current

distortion

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Statistical change detectionStatistical change detection

We consider the Generalized Likelihood Ratio Test (GLRT):We consider the Generalized Likelihood Ratio Test (GLRT):Likelihood ratio under 2 hypothesis:

H0: there is no significant change between images I1 and I2.0 g g g 1 2

H1: there is a significant change between images I1 and I2.

I1 and I2 are considered as realizations of random variables drawn according to parametric probability density functions:parametric probability density functions:

• with the same parameters θ0 under assumption H0

• with two sets of different parameters θ1≠ θ2 under assumption H1

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

with two sets of different parameters θ1≠ θ2 under assumption H1

Bosc’s methodBosc s method

Bosc et al (Neuroimage vol 20 (2) pp 643 656 2003) use a similarBosc et al (Neuroimage, vol 20 (2), pp 643-656, 2003) use a similarframework for detecting MS lesion evolution in conventional multimodal MRIsequencesq

Hypothesis: • Intensities are modeled as a constant value µ in a window Ws of 3×3×3 voxelsµ s

• additive stationary Gaussian noise

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Bosc’s methodBosc s method

However Bosc’s approach fails when coping with DTI derived scalar imagesHowever, Bosc s approach fails when coping with DTI-derived scalar imagesbecause of the non-stationarity of noise.

MDMD FAMD FA

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Proposed approachProposed approach

We propose to account for the non stationarity of the noise by estimating aWe propose to account for the non stationarity of the noise by estimating avariance at each voxel:

Estimation of the variance is done using the method described inChang et al, Magnetic Resonance in Medicine, vol 57(1), pp 141 - 149, 2007

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Variance estimation (1/2)Variance estimation (1/2)

Variance on the tensor elements is obtained as a by product of least squaresVariance on the tensor elements is obtained as a by-product of least squaresestimation:

with

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Variance estimation (2/2)Variance estimation (2/2)

The covariance matrix is propagated to the eigenvalues of XThe covariance matrix is propagated to the eigenvalues of X

and then to FA or MDand then to FA or MD

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

GLRT: H1 hypothesis1 yp

1st1st scan

Tensor, index and varianceand variance estimation

2nd scan

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

GLRT: H0 hypothesis0 yp

ConcatenationConcatenation of the two scans

Tensor, index and variance estimation

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Spatial informationSpatial information

ThresholdingMedian filtering

GLRT RGLRT Detection

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

ValidationValidation

Validation of such method is difficult due to the lack of ground truthtruth

Even for experts, it is difficult to detect changes in DT-images

We resort to simulation of MS lesion appearance

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

MS lesion simulationMS lesion simulation

Experimental studies have shown that demyelination leads to an increase of theExperimental studies have shown that demyelination leads to an increase of the radial diffusivity (D ) in DW-images.

[Harsan et al, Journal of neuroscience research, vol 83(3), pp 392-402, 2006]⊥

• Two scans of a healthy subject are considered• The two last eigenvalues (corresponding to the radial diffusion) of the second examination are increased in several regions of Interest to simulate lesion gapparitions

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

MS lesion simulationMS lesion simulation

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Follow-up of MS patientsFollow up of MS patients

1st scan 2nd scan Proposedmethod

Standard method

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”

Conclusion and perspectivesConclusion and perspectives

Further validation should be done on real cases

Automatic threshold selection: p-value estimation, multiple comparison problem

Extension for multimodal detection

Extension to tensor images and to other model of diffusiong

MIAMS’08: MICCAI workshop on “Medical Image Analysis on Multiple Sclerosis”