Multi-Modal Quantitative Analysis of Pediatric Focal Epilepsy Andy Eow Medical Vision Group CSAIL, MIT.

Multi-Modal Quantitative Analysis of

Pediatric Focal EpilepsyAndy Eow

Medical Vision GroupCSAIL, MIT

Project Background

Focal Epilepsy

• 2.5 million Americans suffer from some form of epilepsy.A large percentage being symptomatic partial epilepsy / focal epilepsy.

• 25% of these people are unresponsive to anti-epileptic medication and surgery is the last alternative.

• Success of surgery is highly dependent on the surgeon’s ability to locate the epileptic foci.

EEG Techniques

• Sub-dural EEG (Gold Standard)

• 32-channel Surface EEG

• 128-channel Surface EEG

Sub-dural EEG

32-channel EEG128-channel EEG

Project Background

EEG Source Localization

• Repeatedly solving the forward problem to obtain a set of dipole parameters that produces electrical potential that matches observed EEG data.

• Ill-posed inverse problem where multiple solutions can generate similar scalp potentials.

• Source localization with scalp EEG and visual correlation with imaging modalities guide placement of subdural grid.

Goal

• Improve EEG source localization with:

Patient-specific Head Model (Accuracy of forward solution)

Hotspots Prior Probability Map (Constraint on inverse solution)

• Focal cortical dysplasia (FCD) is one of the most epileptogenic lesions associated with early onset medically refractive focal epilepsy. These are the cases which I focused on.

Pipeline Framework

Sus

pici

ous

Reg

ion

Det

ectio

nP

re-p

roce

ssin

g

AlignedPET / SPECT

AlignedT2w MRI

Fun

ctio

nal

Met

abol

ism

Asy

mm

etry

Hotspots PriorProbability Map

T2w MRI PET / SPECT

Skull-StrippedT1w MRI

Patient-specificHead Model

T1w MRI

Str

uctu

ral

Asy

mm

etry

Inverse EEG SourceLocalization

Multi-modal Image Alignment(Rigid Volume Correspondence)

Skull-StrippedPET / SPECT

AlignedT1w MRI

Brain / GM / WM / CSFSegmentation

Fun

ctio

nal

Met

abol

ism

Asy

mm

etry

PET / SPECT

AlignedPET / SPECT

T1w MRI

AlignedT2w MRI

Multi-modal Image Alignment(Rigid Volume Correspondence)

Skull-StrippedT1w MRI

Brain / GM / WM / CSFSegmentation

AlignedT1w MRI

T2w MRI

Patient-specificHead Model

Skull-StrippedPET / SPECT

Str

uctu

ral

Asy

mm

etry

Asymmetry Analysis

Plane of SymmetryDetection

Asymmetry Analysis

Plane of SymmetryDetection

Hotspots PriorProbability Map

Inverse EEG SourceLocalization

Patient Example

Details

• 16 years when image acquisition was done.

• Focal cortical dysplasia in the right inferior frontal lobe

• T1-weighted & T2-weighted structural MRI volumes

T1-weighted MRI T2-weighted MRI

Pre-processing

Multi-modal Image Registration

• ITK’s rigid registration framework with a MI metric

Tissue Segmentation

• Intensity Correction (Neil Weisenfeld’s entropy min. implementation)

• Noise Smoothing (ITK’s Curvature Flow anisotropic diffusion filter)

• Modified Watershed Segmentation

Intensity Correction

Noise SmoothingSegmentation Label Map

Symmetry Plane Detector

*RT 2

1*RT

cII scI

H

• Choose arbitrary target symmetry plane. Obtain chiral, Ic, from reflection about this plane.

• Find optimal rigid transformation, TR*, that maps Ic to I.

• “Half” the transform TR* and apply to Ic to obtain

symmetrically aligned chiral.

• Aligns plane of maximal inter-hemispheric similarity with target symmetry plane.

Symmetry Plane Robustness

Assumption

• In the absence of large pathologies, plane of maximal inter-hemispheric similarity matches the anatomical mid-sagittal plane.

MCA Infarction (Worst Case)

• Left hemisphere largely absent.

• Left shifting of mid-line structures.

• Plane of maximal inter-hemispheric similarity no longer coincides well with the mid-sagittal plane.

• Matches remaining cortical tissues and other major structural features such as the skull, ocular cavities etc.

• Holistically, symmetry plane detected still reasonable.

• Unlikely for FCD cases to present with such large pathologies.

Asymmetry Analysis

Deformation Field, F

• Find optimal deformation that best matches symmetrically aligned volume and its chiral (about the target symmetry plane)

• ITK’s non-rigid registration framework with a SSD metric

Asymmetry

• |F|(1 + .F)

• Emphasis is placed on |F|.

Deformation Magnitude, |F|

Modified Deformation Divergence, 1+.F

Asymmetry, |F|(1+.F)

Asymmetry Analysis

Asymmetry

• Thresholding: + S

• Only voxels of significant asymmetry remaining

Hotspots Label Map

• Clustering of significant asymmetry voxels

• Delineation of separate hotspot regions

• Further removal of small noise clusters

Thresholded Asymmetry Hotspots Label Map

Hotspots Prior Probability Map

Building the Probability Map …

• Assigning probability values

• Photspot = 0.9

• PGM = 0.2

• PWM = 0.1

• PCSF = 0.01

• Poutside = 0

• Gaussian smoothing

Prior Probability Map

Hotspots Label Map

EEG Source Localization

Experiments

• Simulate dipole within detected hotspot to obtain EEG measurements.

• Perturbations:

EEG voltage noise, v

Electrode location noise, e

Tissue conductivity, c

Results

Simulated Noise

v / V e / mm c / %Location Error / mm

NeuroFEM Prior

Zero Noise NA NA NA 0.07 Not necessary

Low Noise 5 x 10-4 3 5 2 Not necessary

Medium Noise 1 x 10-3 4 7.5 4.9 0.5

High Noise 5 x 10-3 5 10 14.6 0.5

Future Work

Functional Neuroimaging

• PET / SPECT

• Detect functional asymmetries such as regions of decreased glucose metabolism for PET and cerebral blood flow for SPECT etc.

• Complements structural asymmetries provided by MRI.

Diffusion-Tensor Imaging

• Anisotropic patient-specific head model

• Reduced FA in dysplastic neurons within white-matter tissue structures. Complements structural and functional asymmetries to generate a more comprehensive prior probability map.

• Detects cortical abnormalities at an earlier stage.

• Analyze neural connectivity between potential surgical sites and eloquent

Thank you!