Neuroimage Analysis Center An NCRR National Resource Center Time Series MRI Core Analysis, Modeling - toward Dynamic Surrogates.

Neuroimage Analysis Centerhttp://www.spl.harvard.edu/nac

An NCRR National Resource Center

Time Series MRI Core Analysis, Modeling -toward Dynamic Surrogates of Disease

Dominik S. Meier, Ph.D.

Center for Neurological ImagingBWH Radiology & Neurology

Neuroimage Analysis Centerhttp://www.spl.harvard.edu/nac-2-


TSA Paradigm: Capture Processes

Reversible focal tissue

changes

Irreversible focal tissue

changes

Reversible diffuse tissue changes

Irreversible diffuse tissue

changes

Irreversible tissue loss

Total global irreversible tissue

loss

ATROPHY

Reversible mass effect and therapy-related shrinkage

Reversible focal tissue

changes

Irreversible focal tissue

changes

Reversible diffuse tissue changes

Irreversible diffuse tissue

changes

Irreversible tissue loss

Total global irreversible tissue

loss

ATROPHY

Reversible mass effect and therapy-related shrinkage

• Research and technology development for longitudinal studies of neurodegenerative disease involving MRI morphometry as outcome measure.

• Core work will explore the ability of serial in vivo MRI to illuminate the timing and sequence of the individual pathological processes underlying neurodegenerative disease.

• Segmentation of Change vs. Change of Segmentation

•Current/Common paradigm: Segmentation -> Trend Analysis

•TSA paradigm: Trend Analysis -> Segmentation



Aims

•Aim 1: Time Series Fusion•Develop integrated methods for serial image data fusion•concatenates multiple 3-D MRI datasets into a single coherent 4-D space. •spatial and intensity normalization •voxel-based "chronobiopsy"

•Aim 2: Time Series Change Detection•Develop a new hierarchical framework for change detection and delineation. •3-level hierarchy of (1) detection, (2) delineation, and (3) segmentation.•specificity in detection and precision in segmentation•Detection requires high levels of expert knowledge•enhanced precision for delineation requires automation

•Aim 3: Time Series Modeling•Develop a framework for change characterization and visualization•parametric models of MRI intensity change•on each voxel time-series profile within the areas of change•investigate the serial MRI data from the viewpoint of a specific biological or clinical hypothesis•"temporal differentiation before spatial integration"

•Aim 4: Time Series Validation•Investigate ways to obtain error estimates and sensitivity to change. •scan-rescan data, automated calculation of residual from the fused 4D set•confidence intervals on the model parameters•areas of reference with no pathologic change •sensitivity analyses:sensitivity to change in both the spatial and the intensity domain



Prelim. Results

Application: In preparation

Neuroimage Analysis Centerhttp://www.spl.harvard.edu/nac


Spectrum of Serial Morphometry

Differentiation -> Classification:“new/enlarging” (red), “stable” (green)“resolving” (blue)

V(t1) V(t2) V(t3)

€

∫

€

∫

€

∫

€

∫

€

C (Ix )∫

Serial Volumetry Differential Morphometry Segmentation of Change Time Series Modeling

- Spatially nonspecific - Sensitive to Registration Error

-Greater Expert Input-+segmentation of change+ Controlled Sensitivity

-Model Required-+ Controlled Sensitivity+Segmentation implied

€

∂∂t €

∂∂t

€

∂∂t

€

t → Ix (t, x)

€

C (Ix )

€

∫

€

∫

€

∫

€

∫

€

∫

€

C (Ix )

€

C (Ix ) 1. Classifier/Segmentation 2. Differentiation 2. TS Modeling

2. Integration

4. Integration 4. Integration 3.Classifier +Integration

3. Differentiation3.Classifier / Segmentation

€

Ix a Ix' 1. Registration

€

Ix a I 'x' 1. Normalization 1. Normalization

€

Ix a I 'x'

0 R2=0.670R2=0.943R2=0.943

weeks

T2 intensity

5 10

15 20 30 40 50

2. Classifier/Segmentation

3. Differentiation



Technological Biological / Clinical

•can dynamic metrics derived from serial MRI provide

surrogates with stronger pathological specificity

(inflammatory, degenerative, reparatory processes ) ?

•Different pathol. processes have different time

signatures, even if their morphological footprints

remain the same..

•E.g. Inflammation creates mass effect and occurs

rapidly.

Inflammation•Blood Brain Barrier breakdown•Edema•Cellular Infiltration

Degeneration•Demyelination•Axonal Damage

Repair•Macrophage activity•Astrocytosis•Remyelination•Axonal Repair?

~ weeks

~ months - years

~ months

The cross-sectionalconcept revisited

Avoid data reductioncompare first – reduce later

The longitudinalconcept revisited

Avoid data reductiondifferentiate first – integrate later

•Segmentation of Change vs. Change of Segmentation



Data Fusion Pipeline

Effective spatial resolution loss in serial imaging

for tissue-specific normalization

Registration

Segmentation

Bias Field Correction

Partial Volume Filter

Intensity Normalization

Baseline Normalization

t1baseline

t2follow-up

t3follow-up

t4follow-up

coil sensitivity bias

variable head positioning

variable gain, scanner drift, upgrades etc.

Differential: detection of change



Two-Process Time Series Model

0 10 20 30 40 50 weeks

Y1: Inflammation / Degeneration

Y2: Resorbtion / Repair

Y1 + Y2

MRI intensity

Example: New MS lesion formationWe model MRI intensity change as the superposition of two opposing processes, one causing T2 prolongation, another T2 shortening.



Time Series Modeling Example: MS Lesion Formation

0

R2=0.670

R2=0.943

R2=0.943

weeks

T2 intensity

5 10 15 20 30 40 50

F1 = Level of hyperintensityF2 = Level or recoveryF3 = Duration

weeks

complete recovery

partial recovery

no recovery

F1

F1

F1

F2

F2

F2=0

F3

F3

F3

MRI intensity



Example: Feature Maps of Change

F1: Hyperintensity , F2: residual damage , F3: duration [weeks]



• sensitivity to change• precision of trend assessment

• estimated error in measuringnew lesion change

Differentiation before Segmentation



Error Accumulation / Sensitivity Analysis / Pipeline Design

1 dimension of variation: add and show all results

1009080706050403020100

N=191 N=59 N=43 N=39

p=0.25 p=0.003* p=0.33

DataPreprocessing VolumetryModeling

How one parameter at last step of pipeline affects results is easily tested.

The effect of a parameter early in the pipeline is much more difficult to assess.



Conclusions:

•Repair does occur in MS, varying in extent by location & subject

•MRI intensity dynamics provide reliable metrics of activity

•Short-term T2 lesion recovery shows links to progression in both

atrophy and disability

•SPMSS shows trends to different lesion patterns than RRMS

•Dissociation between new lesion size and residual damage“big lesion small damage”, NO equivalence in total lesion burden

•Spatial patterns that match histopathological observations

Neuroimage Analysis Center An NCRR National Resource Center Time Series MRI Core Analysis, Modeling - toward Dynamic Surrogates.

Documents