A%20Classification-based%20Glioma%20Diffusion%20Model%20Using%20MRI%20Data - PowerPoint PPT Presentation

About This Presentation
Title:

A%20Classification-based%20Glioma%20Diffusion%20Model%20Using%20MRI%20Data

Description:

– PowerPoint PPT presentation

Number of Views:103
Avg rating:3.0/5.0
Slides: 38
Provided by: csUal
Category:

less

Transcript and Presenter's Notes

Title: A%20Classification-based%20Glioma%20Diffusion%20Model%20Using%20MRI%20Data


1
A Classification-based Glioma Diffusion Model
Using MRI Data
  • Marianne Morris1,2
  • Russ Greiner1,2, Jörg Sander2,
  • Albert Murtha3, Mark Schmidt1,2
  • 1 Alberta Ingenuity Centre for Machine Learning
  • 2 University of Alberta
  • 3 Cross Cancer Institute, Alberta Cancer Board

2
Predict Tumour Growth
initial tumour
tumour 6 months later
  • Why?
  • Study tumour growth patterns
  • Improve treatment planning

3
Outline
  • Introduction
  • Incremental Growth Modeling
  • Features
  • Models (UG, GW, CDM)
  • Experiments

4
Incremental Growth Model
  • Iteratively assign each voxel around the active
    tumour border to tumour vsnon-tumour
  • Stops at termination condition
  • Reaching a specified size of tumour
  • theres no more voxels to add
  • Several Approaches

5
Incremental Growth Model









Tumor
6
Incremental Growth Model
Neighbours









Tumor
7
Incremental Growth Model


-
-
-
-
-


Tumor
8
Incremental Growth Model


-
-
-
-
- - -


Neighbours
Tumor
9
Incremental Growth Model

-
-
-
- -
- -
- - -


Tumor
10
Incremental Growth Model

-
-
-
- -
- -
- - -


Tumor
11
Which New Voxels to Add
  • UG Uniform Growth
  • GW Growth based on tissue types
  • CDM Classification-based diffusion

12
Tumour growth modeling uniform diffusion (UG)
  • Radial uniform growth
  • (in all directions alike)

Original tumour
Final tumour volume
13
Tumour growth modeling White vs. Grey matter
(GW)
  • A 51 ratio for diffusion in white matter vs.
    grey matter (Sawnson et al., 2000)

White matter
Grey matter
Original tumour
Final tumour volume
14
Tumour growth modeling
  • Uniform growth
  • Yes!
  • GW model
  • If White matter Yes!
  • If Grey matter 20
  • CDM model
  • Learn tumour growth pattern

voxel
Active tumour border
15
Classification-Based Diffusion Model (CDM)
  • Preprocessing
  • Noise reduction
  • Spatial registration
  • Intensity Standardization
  • Tissue segmentation
  • Tumour segmentation
  • Feature extraction
  • Classification
  • Tumour growth modeling

16
Features
  • Patient features
  • Tumour properties
  • Voxel features
  • Features of neighbouring voxels
  • A total of 75 features

17
Features Patient
  • Age
  • Correlation between age and glioma grade(more
    aggressive tumours occur in older patients
    benign tumours in children)

patient
18
Features Tumour
  • Area-volume ratio
  • Volume increase between 2 scans
  • Percentage of edema

19
Features Voxel
tumour
voxel
  • Min Distance from tumour border
  • Tissue type derived from template
  • Tissue type derived from patients image
  • Image intensities (T1, T1-contrast, T2)
  • Template intensity
  • Edema region
  • Coordinates Tissue Map
  • Distance-Area ratio

edema
tumour
voxel
tumour
20
Features Neighborhood
y
z
2
1
0
3
6
x
4
5
  • For each of 6 neighbors
  • Edema
  • Image intensities
  • Tissue type derived from template
  • Tissue type derived from patients image
  • A neighbourhood in 3D is the 6 voxels immediately
    adjacent to some voxel v
  • (not including diagonal ones)

6 neighbors
21
Classification-Based Diffusion Model (CDM)
  • Preprocessing
  • Noise reduction
  • Spatial registration
  • Intensity Standardization
  • Tissue segmentation
  • Tumour segmentation
  • Feature extraction
  • Classification
  • Tumour growth modeling

22
CDM Classifier
  • Voxel v becomes tumour given
  • qv PT (class (v) tumour epatient,etumour,ev
    )
  • Features of the patient epatient
  • the tumour etumour
  • the voxel and its neighbours ev

23
Learning Parameters (Classifier)
  • How to learn T ?
  • Naïve Bayes
  • Logistic Regression
  • Linear-kernel SVM
  • Trained on other brain images

24
Outline
  • Introduction
  • Incremental Growth Modeling
  • Experiments
  • Evaluation Measure
  • Model Comparison
  • Best Case
  • Average Case
  • Special Cases
  • Average P/R

25
Experimental Procedure
  • Training data
  • Sample of voxels in volume-difference between two
    scans including 2-voxel border around the volume
    at the 2nd time scan
  • Volume-pairs for 17 patients
  • Total of ½ million voxels
  • We evaluate voxels encountered in diffusion
    process
  • Cross-validation (17 patients)

Original tumour
Additional tumour growth
26
Tumour growth modeling CDM (wrt Neighbours)
  • Voxel v becomes tumour based on
  • Features epatient,etumour,ev
  • Compute
  • qv PT (class (v) tumour epatient,etumour,ev
    )
  • Neighbours of voxel v
  • If k tumour-voxel neighbours,probability that
    voxel v becomes tumour
  • pv 1 (1 qv)k
  • Decision
  • Declare voxel v is tumour if pv ? 0.65

v1 v2 v6 v7 v5
v3 v4

v6,v7 k 0 v1,v2,v5 k 1 v3,v4 k 2
27
System Performance
True positives False positives False negatives
Left to right Slices from lower to upper brain
Time 1 scan
Time 2 scan
CDM prediction
28
Evaluation
  • Precision, Recall
  • for tumour, non-tumour voxels
  • nt truth pt prediction Precision
    Recall

29
Diffusion Modeling Process
  • We grow tumour from initial volume at 1st time
    scan to size of tumour volume at 2nd time scan
  • Precision Recall
  • because predicted volume ? truth volume

?
Tumour volume at 2nd time scan
Tumour at 1st time scan
30
Results Model Comparison
31
Results (Best case)
  • GBM_7 CDM beats UG by 20 and GW by 12

Grew tumour along edema regions but didnt
predict other wing of butterfly
True positives False positives False negatives
32
Results (Average case)
  • GBM_1 CDM beats UG by 6 and GW by 8

Need a more accurate brain atlas
True positives False positives False negatives
33
Results (Special case)
  • GBM_10 CDM beats UG by 8 and GW by 2

Resection Recurrence
True positives False positives False negatives
34
Results
CDM performs significantly better than UG and GW!
  • T-test the probability that the means are not
    significantly different
  • Paired data (same data sample different models)
  • CDM vs UG p 0.001
  • CDM vs GW p 0.001
  • (UG vs GW p 0.034)

X is the mean Var the variance n the number of
samples
35
Future work
  • More expressive features
  • Spectroscopy, DTI, genetic data
  • Larger dataset (treatment effect)
  • Brain atlas (highways vs. barriers)

36
Conclusion
  • Challenge
  • Predicting how brain tumours will grow
  • Answer
  • Learned model CDM performs significantly better
    than other existing models!
  • can improve with additional data

37
Acknowledgements
  • The University of AlbertaDept of Computing
    Science
  • The Alberta Ingenuity Centre for Machine Learning
  • Cross Cancer InstituteAlberta Cancer Board
  • Brain Tumor Growth Prediction team
Write a Comment
User Comments (0)
About PowerShow.com