Management and post-processing of prostate perfusion MRI data for tumor/cancer localization and diagnosis (using MATLAB GUI)

1. Management and Post-Processing of
Prostate Perfusion MRI
By:
Vanya Vabrina Valindria (V3)
Vega Valentine (V2)
VIBOT
2010

2. Introduction
Prostate cancer  leading cause of cancer death in males
Diagnosis:
Conventional MR 
DCE (Dynamic Contrast Enhance) – MRI : Perfusion imaging 
Link between contrast material up-take in tumors and micro-vascular 
observation from signal-intensity time curve
The anatomy

3. Aim
 Generate Parametric Images on a pixel- by –
pixel basis
 Follow the kinetics of the contrast agent within
the prostate to localize cancer/tumor area
 Build user interface for post-processing of
Perfusion MR Imaging

4. Med_Toolbox_V3V2
 Aim to manage and post-process Prostate Perfusion MRI
 Available in MATLAB GUI

5. How do we come
up with this
Toolbox?

6. Toolbox ‘Browse patient’:
Select Patient  Slice  ROI of prostate
ROI

7. Normalization
Parametric calculation:
Rectangular ROI of the prostate –
selected by the user
We use Standardized signal-intensity
curves  divide the signals by the
muscle
Muscle:
Same coordinates for each patient
Average of intensity inside ROI muscle
before the contrast injection
(in time series ~ 1 -4)

8. Signal Intensity-Time Curve
2.4  Comparison between tissues curves
2.2 Cancer
2
1.8
Normal PZ
1.6
Intensity
1.4
1.2
Muscle
1
0.8
0.6
Unenhanced
0.4
0 50 100 150 200 250 300
Series Time(s)

9. Parametric Calculation
2.6
S_max
2.4
2.2
2
Intensity
1.8
1.6
1.4 S_end
Max.Contrast
Enhancement
1.2
S_base 1
0.8
0 t_onset 50 t_max 100 150 200 250 300
Series Time (s)

10. Toolbox ‘Image Post-Processing’:
Parametric Images
Select Analyses  Wash IN

11. Wash In
 Wash-in : the mean rate of increase in intensity between the
onset time and the maximum-signal time.
 Estimated the degree of early strong enhancement of
cancerous tissue.
 Formula:
MATLAB command:
x_in = double(delta_t*t_onset: delta_t : delta_t*tmax);
coeff_wi = polyfit(x_in, double(TIC(t_onset:tmax)),1);
wash_in(i,j) = double(coeff_wi(1))*100;

12. Wash-in Results
Patient 313; Slice 11 Patient 405, Slice 14
Patient 409, Slice 9

13. Toolbox ‘Image Post-Processing’:
Parametric Images
Select Analyses  Wash OUT

14. Wash Out
 Wash Out: the decreasing slope after the
maximum intensity signal
 Formula:
 MATLAB code:
x_out = double(delta_t*tmax: delta_t :
delta_t*series);
coeff_wo = polyfit(x_out,
double(TIC(tmax:series)),1);
wash_out(i,j) = double(-1*coeff_wo(1))*100;

15. Wash-Out Results
Patient 313, Slice 12 Patient 405, Slice 15
Patient 409, Slice 7

16. Toolbox ‘Image Post-Processing’:
Parametric Images
Select Analyses  Maximum Contrast Enhancement

17. Maximal Contrast Enhancement
(MCE)
MCE: Relative difference between the maximum
signal and the baseline signal
Formula:
MATLAB Code:
max_enh(i,j) = double(S_max - S_base)/double(S_base)*100;

18. MCE Results
Patient 313, Slice 12 Patient 405, Slice 15
Patient 409, Slice 10

19. Toolbox ‘Image Post-Processing’:
Select Analyses  Combine Parametric Images

20. Combined Parametric Image
 Combination of all normalized parametric
images
MATLAB Code:
washin_norm = (wash_in - min(min(wash_in)))/(max(max(wash_in)) -
min(min(wash_in)));
washout_norm = (wash_out - min(min(wash_out)))/(max(max(wash_out))
- min(min(wash_out)));
max_enh_norm = (max_enh - min(min(max_enh)))/(max(max(max_enh)) -
min(min(max_enh)));
params_norm = (washin_norm + washout_norm + max_enh_norm)./3;

21. Combined Image Results
 Patient 313, Slice 12 Patient 405, Slice 15
Patient 409, Slice 10

22. Toolbox ‘Extra Features’:
Check signal intensity-time curve for a selected point

23. Toolbox ‘Extra Features’:
Check signal intensity-time curve for a selected point

24. Toolbox ‘Extra Features’:
Display parametric images value in Zoom ROI

25. Toolbox ‘Extra Features’:
Display parametric images value in Zoom ROI

26. Toolbox ‘Extra Features’:
SAVE in DICOM format

27. Save all as DICOM
Save all images in DICOM
Same DICOM information
MATLAB Code:
metadata = dicominfo([direct 'Image00' id]);
dicomwrite(I_wi, ['D:MEDICAL
IMAGINGProjectDatasetsParametricImg' dir(37:47)
'_Slice' num2str(slice) '_wash_in.dcm'], metadata);

28. Toolbox ‘Extra Features’:
SAVE all in JPEG format

29. Toolbox ‘Image Post-Processing’:
See the Segmented Cancer Result

30. Cancer Segmentation
 Method: Region Growing
 Manually selected seeds (for each patient)
 Homogeneity criteria:
a,b : is the position of the evaluated pixel
f : the intensity value of the image in Wash In, Wash Out and MCE
Images (3 features)
Mean :the value of n-region mean

31. Cancer Segmentation
 Criteria for Cancer Region:
Wash in value > 4
Wash out value > 0.2
MCE value > 200
Max.Area < 500 pixels
 Global threshold for all of the images:
T(cancer) = 50
The region is consider as cancer region and keep growing when:

32. Segmentation Results
 Patient 313

33. Segmentation results
 Patient 405

34. Segmentation Results
 Patient 405

35. Segmentation Results
 Patient 409

36. Segmentation Results
 Patient 409

37. Registration
 Problem of Patient 407 in time series: 25 – 29
 Deformable registration
 Method:
 Control point selection  Dense SIFT
 Wrapping image  Thin Plate Spline

38. Registration Flow Chart
Reference
Extract ROI SIFT Dense
Image
Find
TPS
Matching
Warping
Points
Unregistered
Extract ROI SIFT Dense
Image
Registered
Image

39. Reference Image
Registration Result 50
 For patient 407, Slice: 8 100
150
Original Images
Time series: 25 Time series: 26 Time series: 28
50 100 150 200 250
50 50
100 100
150 150
Registered Images 50 100 150 200 250
50 100 150 200 250 50 100 150 200 250
Time series: 25 Time series: 26 Time series: 28
50 50
100 100
150 150

40. Conclusion
 User interface for prostate cancer in MR-perfusion
images had been done using MATLAB GUI
 Parametric images obtained are useful to characterize
the prostate tissues.
 Robust for cancer segmentation and deformable
registration for prostate region

41. Time for
Toolbox
DEMO!!

42. References
 Shen, Kaikai. Parametric Image Formation of Human Prostate Cancer
Vascularisation from MRI Perfusion Data. 2008. Laboratoire
Electronique, Informatique et Image University of Burgundy.
 Jeong Kon Kim, Seong Sook Hong. Rate on the Basis of Dynamic
ContrastEnhanced MRI: Usefulness for Prostate Cancer Detection and
Localization. Journal of magnetic resonance imaging 22:639–646 (2005)
 Olivier Rouvière, et al. Characterization of time-enhancement curves of
benign and malignant prostate tissue at dynamic MR imaging. Eur
Radiol (2003) 13:931–942 DOI 10.1007/s00330-002-1617-6.
 Deanna Lyn Langer . Multi-parametric Magnetic Resonance Imaging
(MRI) in Prostate Cancer. 2010. University of Toronto.
 Baowei Fei. Image registration for the prostate . 2009. Case Western
Reserve University .
 Satish Viswanath, et al. A Comprehensive Segmentation, Registration,
and Cancer Detection Scheme on 3 Tesla In Vivo Prostate DCE-MRI.
2008. American Cancer Society

43. Thank You...

44. Signal-intensity Time Curve
Signal enhancement seen on a DCE-MR images can be
assessed in two ways:
 Analysis of signal intensity changes (semi quantitative)
 Quantifying contrast agent concentration change
(pharmacokinetics modeling)
Enhancement  parameterized by examining changes in
signal intensity over time

45. Advantages using Muscle
Norm??
 The enhancement is normalized (same range) for all patients.
 MR signal units are arbitrary and not reproducible from one
patient to another
 Ease to used for segmentation on post-processing because all
patients have the same threshold values in signal-time
intensity curve

46. Signal-intensity Time Curve
 Comparison between un-normalized and standardized curve
 Interval time between series: delta_t = 6.828 s  from
str2num(info.AcquisitionTime)
 There are 40 series in each slice  Series time = 40*delta_t
Un-normalized Curve Standardized Curve
2.6
650
2.4
600
2.2
550
2
500
Intensity
1.8
Intensity
450
400 1.6
350 1.4
300 1.2
250 1
200 0.8
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Series Time (s) Series Time (s)

47. Comparison between other
Equation??
Wash In Linear least square fitting of signal data from
10% to 90% maximum intensity
enhancement
100
0.06 50
0.04 0
0.02
-50
0
Wash Out -50 0 50 100
0
Our calculation
0.02 0.04 0.06 0.08 0.1
Our calculation
Use in Patient 313, Slice 11 and Patient 409,
Slice 15

48. Save all as JPEG
Save all images in
JPEG format
To visualize directly
MATLAB Code:
imwrite(I_wi ,
['D:MEDICAL
IMAGINGProjectDatase
tsColorImg'
dir(37:47) '_Slice'
num2str(slice)
'_WashIn.jpg'],
'jpg');

49. Flowchart Segmentation by Region Growing
Image (u*v)
Mark Visited and calculate
homogeneity criteria
Label Visited
(u*v) (u*v)
NO
Does it fit the
criteria?
Starts for each Regions
(1:r)
YES
Queue Add pixel to the region and
(1:uv) mark Label
Initialize the seed (in the top Update region size and
Queue) region’s statistics
Retrive the pixel from the Add the pixel to the
Queue Queue list
Check its neighbours/
adjacency pixels (1:n)
NO
YES
Have not been
Visited?
NO YES
All pixels in Queue
END
are visited?
NO

50. REGISTRATION
 Problem  Deformable image registration,:
particularly because of bladder and rectum
filling
 Hard to use only rigid-body registration
 Hard to select control points and find the
matching points
 Register the prostate back inside the ROI the
reference position

51. Control point selection  SIFT
 Dense Scale Invariant Feature Transform (vl_feat
toolbox) to automatically select control points
 Descriptors are obtained for densely sampled key points
with identical size and orientation.
 Optimal parameter 
Extract a descriptor each STEP = 5 pixels
A spatial bin covers SIZE = 5 pixels
 Matching points  the minimum squared Euclidean
distance between the matches.

52. Automatic Control Points Selection
Matching Control Points from Dense SIFT Descriptors.
Some examples in Patient 407:
Input Prostate Reference Prostate Input Prostate Reference Prostate
Input Prostate Reference Prostate Input Prostate Reference Prostate

53. Wrapping: TPS (Thin Plate Shape)
 The transformation is modeled using TPS
 Used as the non-rigid transformation model in image alignment
 Fits a mapping function between corresponding control points by
minimizing the energy function
Control Points in Input Image Control Points in Ref.Image Wrapped Image by TPS
0 50
50 50
50
100 100
0 1 1
100 100
3 3
150 150
0 150 150
200 200
0 200 200
4
4
2 2
250 250
0 250 250
300 300
0 300 300
50 100 150 200 250 300 50 100 150 200 250 300
50 100 150 200 250 300 50 100 150 200 250 300 50 100 150 200 250 300

54. Reference Image
Registration Result 50
 For Patient 407, Slice 11
100
150
Original Images
Time series: 25 Time series: 27 Time series: 28 100
50 150 200 250
50 50
100 100
150 150
Registered Images
50 100 150 200 250 50 100 150 200 250 50 100 150 200 250
Time series: 25 Time series: 27 Time series: 28
50 50
100 100
150 150

55. Parametric Images after registration
 Patient 407, Slice 8
Wash In Wash Out MCE
Combined Segmented

56. Non cancer detection…?
 Patient 313, Slice 3
Wash In Wash Out MCE
Combined Segmented