Parcellation of Human Inferior Parietal Lobule Based on Diffusion MRI Bilge Soran 1 Zhiyong Xie 2...

Parcellation of Human Inferior Parietal Lobule

Based on Diffusion MRI

Bilge Soran1 Zhiyong Xie2 Rosalia Tungaraza3 Su-In Lee1 Linda Shapiro1,2 Thomas Grabowski3

University of Washington1Dept. of Computer Science and Engineering

2Dept. of Electrical Engineering3Integrated Brain Imaging Center

Aug 2012

This work was supported by NIH-NINDS Grant No. RC4-NS073008 (PI: T. Grabowski).

2

Motivation

• The function of a brain can be studied by analyzing its anatomical connections.

• The importance of this work is in its investigation of methods for parcellation of the brain of a living subject, rather than manual parcellation of a post-mortem subject.

3

Outline• Background – Human inferior parietal lobule – Diffusion MRI

• Current Approaches• Methods– Unsupervised Clustering Approaches

• Evaluation Metric• Results

4

Human Inferior Parietal Lobule

• The IPL is the cortical region with marked functional heterogeneity involved in visuospatial attention, memory, and mathematical cognition functions.

5

AimTo parcellate the IPL data of living

subjects and evaluate the parcellation quality by means of overlap with clusters

of a standard atlas.

Juelich atlas A parcellation result in sagittal view

6

Diffusion MRI• Measures the diffusion of water molecules in

biological tissues.

Tractography• Method for identifying anatomical connections

in the living human brain.

• Offers an overall view of brain anatomy, including the degree of connectivity between different regions of the brain.

7

Current Parcellation Approaches

1. Methods relying on the known connection patterns of the functional fields.

2. Methods using a statistical model, based on the assumption of the statistical distribution of the data.

3. Methods using unsupervised machine learning methods.

In this project we used unsupervised machine learning techniques to parcellate the IPL region.

Our Method

8

PreprocessingGenerate the connectivity map for each seed

point using Probabilistic Tractography

Cluster the voxels of the IPL into functional fields based on the

connectivity pattern

Using K-Means, EM, Spectral Clustering

Verification with the Juelich atlas

Using proposed metrics

9

Juelich Atlas(Juelich histological (cyto- and myelo-architectonic) atlas)

• The Juelich atlas is a probabilistic map of the areas of the brain.

• It was created based on the microscopic and quantitative histological examination of ten human post-mortem brains.

lh-IPC, Sagittal View

1

23

4 5

67

10

Preprocessing

• Regions of interests (ROIs) were extracted.

• For each voxel in the IPL, a connectivity profile is calculated using probabilistic tractography.

5000 particles initiated from each seed voxel

# of particles that reached each target region counted.

Destrieux atlas1IPL region in Destrieux atlas1

1Destrieux C., et al. Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature, NeuroImage, 2010

11

Data for Clustering

• 3D coordinates of the IPL voxels.

• Target connectivity probabilities of each seed voxel.

Target regions showing no connectivity patterns with any of the seed voxels are discarded.

12

Parcellation Methods

• K-means Clustering• Mean Shift Clustering• Expectation Maximization (EM)

Clustering• Spectral Clustering– Standard Normalized Graph Cuts– Normalized Graph Cuts with Feature Selection– Normalized Graph Cuts with K-means

13

Similarity Matrix Computation

Seed Voxels

Target Regions

Connectivity Matrix

Seed Voxels

Seed Voxels

Distance Matrix

Seed Voxels

Seed Voxels

Connectivity Similarity

Matrix

Seed Voxels

Seed Voxels

Spatial Affinity Matrix

Seed Voxels

Seed Voxels

Composite Similarity

Matrix

+

14

Similarity Computation

16

Evaluation Metric• Based on the number of intersecting voxels between a cluster of

the atlas and a normalized cut cluster.

• Defined as the average of the sum of the largest overlap with the computed clusters of each atlas cluster and the sum of the largest overlap with the atlas clusters of each computed cluster.

AC 1 AC 2 AC 3 AC 4 AC 5 AC 6 AC 7CC 1 0 0 0 91 15 3 0CC 2 0 0 0 24 0 79 8CC 3 0 53 67 9 0 0 0CC 4 6 78 14 4 0 0 2CC 5 91 31 0 0 0 0 0

An example table used in evaluation.

17

Evaluation Metric

Else the resulting parcellation is unacceptable.

18

Experiments and Results

• The data set consisted of the left and right IPL regions, which have similar connectivity patterns, of 19 subjects.

• The right IPL data set was divided into two with 10 subjects for training and 9 for testing; the left IPL data set was used only for testing.

• This resulted in 10 subjects for training and 28 subjects for testing.

19

Experiments and Results

• The three different normalized cut algorithms were run on the training set, and the best parameter set for each algorithm was determined according to the proposed metric.

• Because the Juelich atlas contains 7 manually determined clusters, experiments were run with +/- 2 margin of 7, namely 5 to 9 clusters.

• Experimented with α varying from 0.5 to 0.9 with 0.1 intervals.

20

Normalized Graph Cuts

• After the best parameters were determined on the training set, both NGC and NGC with feature selection algorithms were run on the whole training and testing sets.

• The normalized graph cuts with the K -means approach did not produce any acceptable parcellation results according to the proposed metric.

21

Subject Right Hemisphere Left HemisphereNGC-std NGC-fea NGC-std NGC-fea

1 0.625392 0.626959 0.615437 0.620902

2 0.753577 0.751987 0.609091 0.610909

3 0.718487 0.710084 0.604126 0.608055

4 0.569952 0.570747 0.672694 0.674503

5 0.652021 0.655536 0.675652 0.672174

6 0.637883 0.646240 0.706388 0.707617

7 0.640599 0.640599 0.553792 0.554674

8 0.593415 0.601072 0.592657 0.599650

9 0.627622 0.628497 0.636494 0.635057

10 0.683212 0.682482 0.566836 0.564298

11 0.590669 0.590669 0.627559 0.633071

12 0.611452 0.615542 0.640625 0.637336

13 0.747917 0.746875 0.620948 0.622195

14 0.657725 0.653433 0.691321 0.683432

15 0.666667 0.666667 0.592619 0.592619

16 0.657588 0.701362 0.660417 0.658333

17 0.600768 0.599808 0.642747 0.642747

18 0.611494 0.611494 0.716960 0.712555

19 0.670802 0.669847 0.644431 0.643819AVERAGE 0.648276 0.651047 0.635305 0.635471

Parcellation performances

22

Juelich atlas and parcellation results in sagittal view

23

Conclusions• This work investigates methods for parcellating

the IPL region of a living subject.

• We used a new atlas-based metric to evaluate the quality of the parcellation.

• Normalized graph cuts showed the best performance among the clustering methods applied.

• Our work shows the feasibility of this approach for parcellation of brain regions of living subjects.