SlideShare ist ein Scribd-Unternehmen logo
1 von 49
Automated left ventricle segmentation
in SAX CMR images using cost-volume filtering (CVF)
and novel myocardial contour processing framework
Graduate Institute of Communication Engineering
College of EECS, NTU
Oct 22, 2014
Speaker: An-Cheng Chang, MSc student
What is CMR segmentation and why?
• Tracing the chamber volume gives insight into
how well the heart functions.
• CMR segmentation is involved as a major part of
the analysis.
2
Left ventricle (LV)
Heart
Cardiac Magnetic Resonance (CMR) images
Epicardium
border
Endocardium
border
Long axis
Apex
Base
Long axis
Objective
Automatically trace the LV endo∙card∙ium border
3
Left ventricle (LV)
Endocardium
border
heartinner tissue
(it is not a trivial task)
Difficulties in CMR segmentation
• Endocardium border is often obscured by papillary muscles and trabeculae
carneae. (Fig A)
• Variation between individuals. Wide pathological variations. (Fig B)
• Low image quality: noises and distortions, e.g., field inhomogeneity, partial
volume effect. (Fig C and Fig D)
4
C. Artifacts
(field inhomogeneity)
B. Wide subject
variations
D. Artifacts
(partial volume)
A. Endocardium border
obscured by PMTC
Ground truth
Auto
Related works on CMR segmentation
Ngo et al., IEEE ICIP 2013
• Deep learning (pre-trained) + level set
• Algorithm initialize by cropping ROI
5
Cropped by operator
Pre-trained
deep learning
network
Cropped by operator
Hu et al., Magn Reson Imaging (Elsevier, 2013)
• GMM + dynamic programming
• Algorithm initializes by cropping ROI
Method
SR (%)
Mean(StD)
APD (mm)
Mean(StD)
DM
Mean(StD)
Hu 2013 91.1(9.4) 2.24(0.40) 0.89(0.03)
Ngo 2013 97.9(6.18) 2.08(0.40) 0.90(0.03)
Ours 94.1(6.1) 1.75(0.42) 0.91(0.03)
Ours (2014)
• CVF + contour processing
• Algorithm initializes by one click on the LV
Result highlight
6
Auto contours rejected: 0 out of 18
Mean error: 2.59mm (+0%)
EF underestimated by 4%
Subject SC-HYP-40; hypertrophic heart.
Auto contours rejected: 7 out of 18
Mean error: 3.23mm (+25%)
EF underestimated by 11%
With proposed CVF Without CVF
Endocardium delineation using
CVF and proposed myocardial contour processing
7
LV localization &
ROI refinement
Blood pool
classification
by CVF
Myocardial
contour
processing
3D+T
volume
Result
(BASE)
(APEX)
slice 1
slice 2
slice M
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
t=1 t=2 t=N
Polar transformation
Inverse polar transformation
Proposed method
Endocardium delineation using
CVF and proposed myocardial contour processing
8
LV localization &
ROI refinement
Blood pool
classification
by CVF
Myocardial
contour
processing
3D+T
volume
Result
(BASE)
(APEX)
slice 1
slice 2
slice M
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
t=1 t=2 t=N
Polar transformation
Proposed method
Inverse polar transformation
Blood pool classification by CVF
9
Signal intensity
Occurrence
LV
blood poolOthers
T
LV blood pool
Others
Binary label image:
Local variance map:
Cost slice for selecting ‘LV blood pool’
Cost slice for selecting ‘Others’
Proposed
cost volume
Refined
blood pool segment:
Cost aggregation &
Label selection
Polar transformation
Proposed
cost initialization
Cost-volume filtering
(CVF)
Cost-volume-filtering (CVF) based
image segmentation
• CVF is originally used for refining stereo matching results. Recently been
generalized for discrete labeling problems (Rhemann et al.; appears in PAMI 2013,
CVPR 2011).
• Method: Initializing cost(x, y, label)  cost aggregation  label selection.
• Cost initialization scheme depends on applications.
10
*Rhemann et al.; appears in PAMI 2013, CVPR 2011
(a) Stereo matching (depth from stereo) (b) Interactive image segmentation
Applications of CVF
Why CVF?
• Histogram-based labeling (Otsu’s method, GMM-based thresholding) ignores
spatial relationship.
11
LV blood pool
Others
LV
blood pool
Others
T
2D image 1D histogram 2D image
Spatial
information
lost
• CVF considers both spatial relationship and intensity similarity when outputting
labeled results.
-> Good for handling bias field caused by MR field inhomogeneity.
• We proposed a new cost initialization scheme for CVF to address the partial
volume effect in MR images.
• CVF is fast. Can be O(N) time and non-approximate.
Effectiveness of CVF-based segmentation
12
• The proposed cost initialization scheme addresses the partial volume issue.
• Compare the result between Fig B and Fig C.
13
• Robust against MR field inhomogeneity
Robustness of CVF-based segmentation
.
PA
.
PB
PA: Brighter PB: Dimmer
Polar
transformation
Inverse polar
transformation
Principle of CVF:
using proposed CMR segmentation as an example
14
Cost(x, y, label)
Image
data
Cost
aggregation
Guide
image Labels(x,y)

Cost initialization
Image data
Original image
Cost volume
Cost(x, y, ’blood pool’)
Cost(x, y, ’others’)
High cost
Low cost
Binary label
image
Local
variance map
Label selection
(reduce dimension)
Filtered_cost(x, y, label)
LV blood pool
Others
Principle of CVF:
using proposed CMR segmentation as an example
15
Cost(x, y, label)
Image
data
Cost
aggregation
Guide
image
Cost initialization
Labels(x,y)

Label selection
(reduce dimension)
Filtered_cost(x, y, label)
Kernel
(Box filter)
× =
Cost slice Ci Guide image I Weighting W
Principle of CVF:
using proposed CMR segmentation as an example
16
Cost(x, y, label)
Image
data
Cost
aggregation
Guide
image
Cost initialization
Labels(x,y)

Label selection
(reduce dimension)
Filtered_cost(x, y, label)
Kernel
(Box filter)
× =
Cost slice Ci
The principle: cost is aggregated from similar* neighbors.
*similar in guide image I
Guide image I Weighting W
(shift-variant)
Principle of CVF:
using proposed CMR segmentation as an example
17
Cost(x, y, label)
Image
data
Cost
aggregation
Guide
image
Cost initialization
Labels(x,y)

Label selection
(reduce dimension)
Filtered_cost(x, y, label)
Guide image I Kernel
(Gaussian)
× =
Weighting W
The principle: cost is aggregated from similar* neighbors.
*similar in guide image I
Cost slice C1 for
‘blood pool’
Principle of CVF:
using proposed CMR segmentation as an example
18
Cost(x, y, label)
Image
data
Cost
aggregation
Guide
image
Cost initialization
Filtered_cost(x, y, label)
Labels(x,y)

Label selection
(reduce dimension)
Filtered cost slice C1’
Filtered cost slice C2’
Labels fSelect the label with the least cost
CVF-refined
Binary label image
Shift-
variant
filter W(I)
19
CVF: one iteration
CVF: 100 iterationsOtsu‘s bi-level thresholding
Original image
Endocardium delineation using
CVF and proposed myocardial contour processing
20
LV localization &
ROI refinement
Blood pool
classification by
CVF
Myocardial
contour
processing
3D+T
volume
Result
(BASE)
(APEX)
slice 1
slice 2
slice M
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
t=1 t=2 t=N
Polar transformation
Inverse polar transformation
Proposed method
Endocardial Contour Processing Framework
21
■ Corrected contour (assigned to set B)
■ Good contour (assigned to set G)
Contour based on Canny’s edge detector
Contour based on CVF
papillary muscletrabaculae
Inverse polar
transform
C(p)
I(p)
E(p)
Combine two raw contours I and C and return a regularized final contour
• I: based on CVF-labeled binary image; intensity similarity
• C: based on Canny’s edge detector; gradient information
Goodcontour
Correctcontour
Contour function generation and correction
22
1D contour function f
First-order derivative f’
Fluctuations
2D labeled image
Fluctuations are detected and
corrected by linear interpolation
Corrected contour
Fluctuations are a result of the
presence of papillary muscle and
trabeculae
Detecting the fluctuations
23
Box filter w2
Box filter w3
Impulse wδ
∗
∗
∗
𝑓′
𝑓′
𝑓′
Type 1
Type 2
Type 3
Contour function
1st order
derivative
ResponseFilter bank
Endocardial Contour Processing Framework
24
■ Corrected contour (assigned to set B)
■ Good contour (assigned to set G)
Contour based on Canny’s edge detector
Contour based on CVF
papillary muscletrabaculae
Inverse polar
transform
C(p)
I(p)
E(p)
Combine two raw contours I and C and return a regularized final contour
• I: based on CVF-labeled binary image; intensity similarity
• C: based on Canny’s edge detector; gradient information
Complementary contour generation
25
Original image
Non-maximum suppressed
edge response
Edge pruned
Complementary contour function C
Contour form CVF acts as a guide to pick up
Canny’s edge response
Endocardial Contour Processing Framework
26
■ Corrected contour (assigned to set B)
■ Good contour (assigned to set G)
Contour based on Canny’s edge detector
Contour based on CVF
papillary muscletrabaculae
Inverse polar
transform
C(p)
I(p) E(p)
min 𝒪 𝐸 =
𝑝 ∈ 𝐵
𝐸 𝑝 − 𝐶 𝑝 2
+
𝑝 ∈𝐺
𝐸 𝑝 − 𝐼 𝑝 2
+ 𝜆
𝑝
𝐸′′
𝑥 2
Combine C(p) and I(p) by minimizing the following objective function:
data term smoothness term
Least squares problem  re-formulate to Ax=b and solve for x
Endocardial Contour Processing Framework
• The additional constraint will ensure the final contour to enclose the
blood pool
27
𝒪 𝐸 =
𝑝 ∈ 𝐵
𝐸 𝑝 − 𝐶 𝑝 2
+
𝑝 ∈𝐺
𝐸 𝑝 − 𝐼 𝑝 2
+ 𝜆
𝑝
𝐸′′
𝑥 2
subject to 𝐸 𝑥 > 𝐼(𝑥)
Auto Auto /w constraint By expert
28
Inverse polar
transform
Review of the system processing flow
Lock down ROI
Polar mapping
Otsu’s
thresholding
After CVF
Myocardial contour processing
Test dataset:
Sunnybrook cardiac MR database
• The first (in 2009) publically accessible cardiac MR database
• Provides 45 MR datasets including one healthy plus three pathological cases
• Includes an evaluation tool & ground truth at end-diastole (ED) and end-
systole (ES)
29
Technical details:
• Acquisition protocol: SSFP MR SAX images are obtained during 10-15 second breath-holds with a
temporal resolution of 20 cardiac phases over the heart cycle, and scanned from the ED phase. Six to 12
SAX images were obtained from the atrioventricular ring to the apex
• MRI scanner: 1.5T GE Signa MRI.
(thickness=8mm, gap=8mm, FOV=320mm*320mm, matrix= 256*256)
Segmentation example:
Case of hypertrophy
30
Left: computed. Right: expert-drawn
Segmentation example:
Case of normal heart
31
Left: computed. Right: expert-drawn
Segmentation example:
Case of heart failure /w infarction
32
Left: computed. Right: expert-drawn
CMR volume segmentation at ED and ES phase
Patient: SC-HF-I-01 (Heart failure with infarct)
33
-Red: Auto. - - - Purple: expert
CMR volume segmentation at ED and ES phase
Patient: SC-HYP-09 (Hypertrophy)
34
-Red: Auto. - - - Purple: expert
CMR volume segmentation at ED and ES phase
Patient: SC-HF-NI-04 (Non-ischemic heart failure)
35
-Red: Auto. - - - Purple: expert
Evaluation metrics
Average perpendicular distance (APD)
• The computed contour for a given slice is qualified if
APD < 5mm
Overlapping dice metric (DM)
DM =
2𝐴 𝑎𝑚
𝐴 𝑚 + 𝐴 𝑎
Success rate (SR)
• Number of qualified contours (APD < 5mm) in one
CMR scan.
36
SR=75%
Performance evaluation
with respect to pathological groups
SR (%) APD (mm) DM
Group Mean StD Mean StD Mean StD
SC-HF-I 94.2 7.7 1.542 0.296 0.93 0.02
SC-HF-NI 95.1 4.1 1.736 0.459 0.92 0.02
SC-HYP 93.6 6.5 1.900 0.420 0.88 0.03
SC-N 93.4 5.3 1.834 0.377 0.89 0.02
Overall 94.1 6.1 1.748 0.417 0.91 0.03
37
A total of 800 MR images from 45 patients are evaluated
Result comparison
Method
SR (%)
Mean(StD)
APD (mm)
Mean(StD)
DM
Mean(StD)
Huang 2011 (auto) 81.5(18.0) 2.19(0.44) 0.91(0.03)
Hu 2013 (auto) 91.1(9.4) 2.24(0.40) 0.89(0.03)
Constantinides
2012 (semi-auto)
91.0(8.0) 1.94(0.42) 0.89(0.04)
Constantinides
2012 (auto)
80.0(16.0) 2.44(0.56) 0.86(0.05)
Ngo 2013
(semi-auto)
97.9(6.18) 2.08(0.40) 0.90(0.03)
Ours (auto) 94.1(6.1) 1.75(0.42) 0.91(0.03)
38
Ours against others. All use the same 45-patient dataset.
Additional result comparison
Method
SR (%)
Mean(StD)
APD (mm)
Mean(StD)
DM
Mean(StD)
Jolly 20091 95.62(8.83) 2.26(0.59) 0.88(0.04)
Lu 20091 72.45(18.86) 2.07(0.61) 0.89(0.03)
Huang 20091 -- 2.10(044) 0.89(0.04)
Wijnhout 20091 86.47(11) 2.29(0.57) 0.89(0.03)
Constantinides
20091 92.28(--) 2.04(0.47) 0.89(0.04)
Marák 20091 -- 3.00(0.59) 0.86(0.04)
Feng 2013 92.8(9.2) 1.93(0.37) 0.86(0.04)
Ngo 2013 96.58(3.66) 2.22(0.46) 0.89(0.03)
Ours 96.31(4.85) 1.67(0.40) 0.91(0.03)
39
Ours against others. All use the same 15-patient ‘validation’ dataset.
1 Results reported in MICCAI LVSC 2009.
Evaluating ejection fraction
40
EF error
Group Mean StD
SC-HF-I -0.15 2.89
SC-HF-NI 2.27 4.64
SC-HYP 2.13 5.01
SC-N 2.39 5.70
Overall 1.61 4.72
y = 0.9547x + 0.3663
R² = 0.9421
0
10
20
30
40
50
60
70
80
90
0 10 20 30 40 50 60 70 80 90
AUTOEF
MANUAL EF
AUTO VS. MANUAL EF
SC-HF-NI
SC-HF-I
SC-HYP
SC-N
Method R2 for EF
Cocosco 2008 0.90
Lu 2013 0.92
Lorenzo-Valdés 2004 0.92
Cordero-Grande 2011 0.92
Constantinides 2012 0.83
Ours 0.94
-10
-5
0
5
10
15
20
0 20 40 60 80 100
AUTO-MANUAL
(AUTO + MANUAL) /2
BLAND-ALTMAN PLOT FOR EF
Mean: 1.61
SD: 4.72
Measures the proportion of blood ejected with each cardiac cycle
• EF = (ED volume – ES volume)/ED volume
Conclusion
• Developed an algorithm that detects the left ventricular endocardial contour in
CMR images, with top-tier accuracy.
• Use cost-volume filtering (CVF) to combat MR inhomogeneity.
• Proposed a novel cost initialization scheme that handles partial volume effect.
• Proposed a contour processing framework, in which information from gradient and
intensity similarity are encoded along with a smoothness constraint
• Clinical aspect: highly correlated (R2 = 0.94) between auto and manual EF. No
systematic bias is observed.
• Future work includes incorporating inter-slice and inter-frame relationships to
increase detection rate.
41
LV localization
& ROI
refinement
Blood pool
classification
by CVF
Myocardial
contour
processing
3D+T
volume
Result
(BASE)
(APEX)
slice 1
slice 2
slice M
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
t=1 t=2 t=N
Polar transformation
Inverse polar transformation
Proposed method
Supplement A:
Automated localization of the left ventricle (LV)
43
• Find the area that covers LV blood pool and LV muscle
• Lock down the region of interest (ROI) and hands off the ROI to the rest
of the algorithm
• This will exclude the influence of nearby tissues when making initial
estimate of LV blood pool
Iteratively refining the region of interest (ROI)
44
Blood
pool
Myocardium
Others
Signal intensity
Occurrence
Myocardium LV
blood pool
Others
T
Class=‘others’if signal < T
Class=‘LV blood pool’if signal > T
The rationale
T?
(Fig A) Initialize a ROI inside the LV, then:
• (Fig B) Classify the pixels in the ROI (using Otsu’s method)
• (Fig C) Retain the primary connected component
• (Fig E) Compute LV blood pool’s convex hull
• (Fig D) Dilate the convex hull. This has become the new ROI
Repeat B->C->E->D until convergence
45
Step-by-step breakdown
46
47
• Low contrast area can be recovered regardless of detected LV centroid
PB
.
PC
.
PC
.
PB
.
PB
.
PC
.
PC
PC
.
PB
.
PB
PB
PC
Polar transformation
vs.
vs.
Ours
Robustness of CVF-based segmentation
What brings down the success rate?
48
Supplement B:
Extreme case of hypertrophy (SC-HYP-08)
49
-Red: automated. - - - Purple: expert-drawn
What brings down the success rate?
• Mostly in apical slices & extremely hypertrophic hearts.
• Severe partial volume effect in apical slices violates our assumption.
• In hypertrophic cases, endocardium border can be completely obscured
by papillary muscle and trabeculae carneae.
50
(A) (B)

Weitere ähnliche Inhalte

Andere mochten auch

Slides on Photosynth.net, from my MSc at Imperial
Slides on Photosynth.net, from my MSc at ImperialSlides on Photosynth.net, from my MSc at Imperial
Slides on Photosynth.net, from my MSc at ImperialKevin Keraudren
 
Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...
Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...
Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...Kevin Keraudren
 
Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...
Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...
Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...Kevin Keraudren
 
Region-based Semi-supervised Clustering Image Segmentation
Region-based Semi-supervised Clustering Image SegmentationRegion-based Semi-supervised Clustering Image Segmentation
Region-based Semi-supervised Clustering Image SegmentationOnur Yılmaz
 
Cardiac Image Analysis based on K Means Clustering
Cardiac Image Analysis based on K Means ClusteringCardiac Image Analysis based on K Means Clustering
Cardiac Image Analysis based on K Means ClusteringNAVEEN TOKAS
 
Automatic left ventricle segmentation
Automatic left ventricle segmentationAutomatic left ventricle segmentation
Automatic left ventricle segmentationahmad abdelhafeez
 
Marker Controlled Segmentation Technique for Medical application
Marker Controlled Segmentation Technique for Medical applicationMarker Controlled Segmentation Technique for Medical application
Marker Controlled Segmentation Technique for Medical applicationRushin Shah
 
Breast Lesion Segmentation in Ultrasound Images
Breast Lesion Segmentation in Ultrasound ImagesBreast Lesion Segmentation in Ultrasound Images
Breast Lesion Segmentation in Ultrasound ImagesMohamed Elawady
 
Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...
Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...
Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...AIST
 
Exposé segmentation
Exposé segmentationExposé segmentation
Exposé segmentationDonia Hammami
 
Echo assessment of lv systolic function and swma
Echo assessment of lv systolic function and swmaEcho assessment of lv systolic function and swma
Echo assessment of lv systolic function and swmaFuad Farooq
 
Image segmentation
Image segmentationImage segmentation
Image segmentationDeepak Kumar
 
Applications of Digital image processing in Medical Field
Applications of Digital image processing in Medical FieldApplications of Digital image processing in Medical Field
Applications of Digital image processing in Medical FieldAshwani Srivastava
 
Biomedical image processing ppt
Biomedical image processing pptBiomedical image processing ppt
Biomedical image processing pptPriyanka Goswami
 
PPT on BRAIN TUMOR detection in MRI images based on IMAGE SEGMENTATION
PPT on BRAIN TUMOR detection in MRI images based on  IMAGE SEGMENTATION PPT on BRAIN TUMOR detection in MRI images based on  IMAGE SEGMENTATION
PPT on BRAIN TUMOR detection in MRI images based on IMAGE SEGMENTATION khanam22
 
Image segmentation ppt
Image segmentation pptImage segmentation ppt
Image segmentation pptGichelle Amon
 

Andere mochten auch (20)

Master Presentation
Master PresentationMaster Presentation
Master Presentation
 
Slides on Photosynth.net, from my MSc at Imperial
Slides on Photosynth.net, from my MSc at ImperialSlides on Photosynth.net, from my MSc at Imperial
Slides on Photosynth.net, from my MSc at Imperial
 
Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...
Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...
Endocardial 3D Ultrasound Segmentation using Autocontext Random ForestsPresen...
 
Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...
Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...
Sparsity Based Spectral Embedding: Application to Multi-Atlas Echocardiograph...
 
Region-based Semi-supervised Clustering Image Segmentation
Region-based Semi-supervised Clustering Image SegmentationRegion-based Semi-supervised Clustering Image Segmentation
Region-based Semi-supervised Clustering Image Segmentation
 
Cardiac Image Analysis based on K Means Clustering
Cardiac Image Analysis based on K Means ClusteringCardiac Image Analysis based on K Means Clustering
Cardiac Image Analysis based on K Means Clustering
 
Automatic left ventricle segmentation
Automatic left ventricle segmentationAutomatic left ventricle segmentation
Automatic left ventricle segmentation
 
Marker Controlled Segmentation Technique for Medical application
Marker Controlled Segmentation Technique for Medical applicationMarker Controlled Segmentation Technique for Medical application
Marker Controlled Segmentation Technique for Medical application
 
Breast Lesion Segmentation in Ultrasound Images
Breast Lesion Segmentation in Ultrasound ImagesBreast Lesion Segmentation in Ultrasound Images
Breast Lesion Segmentation in Ultrasound Images
 
Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...
Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...
Andrey Mukhtarov - The Study of Applicability of the Decision Tree Method for...
 
Exposé segmentation
Exposé segmentationExposé segmentation
Exposé segmentation
 
Echo assessment of lv systolic function and swma
Echo assessment of lv systolic function and swmaEcho assessment of lv systolic function and swma
Echo assessment of lv systolic function and swma
 
Segmentation
SegmentationSegmentation
Segmentation
 
Image segmentation
Image segmentationImage segmentation
Image segmentation
 
Applications of Digital image processing in Medical Field
Applications of Digital image processing in Medical FieldApplications of Digital image processing in Medical Field
Applications of Digital image processing in Medical Field
 
Biomedical image processing ppt
Biomedical image processing pptBiomedical image processing ppt
Biomedical image processing ppt
 
IMAGE SEGMENTATION.
IMAGE SEGMENTATION.IMAGE SEGMENTATION.
IMAGE SEGMENTATION.
 
PPT on BRAIN TUMOR detection in MRI images based on IMAGE SEGMENTATION
PPT on BRAIN TUMOR detection in MRI images based on  IMAGE SEGMENTATION PPT on BRAIN TUMOR detection in MRI images based on  IMAGE SEGMENTATION
PPT on BRAIN TUMOR detection in MRI images based on IMAGE SEGMENTATION
 
Image segmentation ppt
Image segmentation pptImage segmentation ppt
Image segmentation ppt
 
Medical image analysis
Medical image analysisMedical image analysis
Medical image analysis
 

Ähnlich wie Automated quantitative assessment of left ventricular functions by MR image segmentation

Basics of ct lecture 2
Basics of ct  lecture 2Basics of ct  lecture 2
Basics of ct lecture 2Gamal Mahdaly
 
ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...
ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...
ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...CSCJournals
 
Haemodynamic monitoring
Haemodynamic monitoringHaemodynamic monitoring
Haemodynamic monitoringguest5c708a
 
Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...Usatyuk Vasiliy
 
Investigation of repeated blasts at Aitik mine using waveform cross correlation
Investigation of repeated blasts at Aitik mine using waveform cross correlationInvestigation of repeated blasts at Aitik mine using waveform cross correlation
Investigation of repeated blasts at Aitik mine using waveform cross correlationIvan Kitov
 
Wiener Filter Hardware Realization
Wiener Filter Hardware RealizationWiener Filter Hardware Realization
Wiener Filter Hardware RealizationSayan Chaudhuri
 
Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...Usatyuk Vasiliy
 
CT based Image Guided Radiotherapy - Physics & QA
CT based Image Guided Radiotherapy - Physics & QACT based Image Guided Radiotherapy - Physics & QA
CT based Image Guided Radiotherapy - Physics & QASambasivaselli R
 
Iaetsd a review on ecg arrhythmia detection
Iaetsd a review on ecg arrhythmia detectionIaetsd a review on ecg arrhythmia detection
Iaetsd a review on ecg arrhythmia detectionIaetsd Iaetsd
 
CyberSec_JPEGcompressionForensics.pdf
CyberSec_JPEGcompressionForensics.pdfCyberSec_JPEGcompressionForensics.pdf
CyberSec_JPEGcompressionForensics.pdfMohammadAzreeYahaya
 

Ähnlich wie Automated quantitative assessment of left ventricular functions by MR image segmentation (20)

SPIE_2015_Fahmi
SPIE_2015_FahmiSPIE_2015_Fahmi
SPIE_2015_Fahmi
 
03raster 1
03raster 103raster 1
03raster 1
 
Cardiac CT
Cardiac CTCardiac CT
Cardiac CT
 
Basics of ct lecture 2
Basics of ct  lecture 2Basics of ct  lecture 2
Basics of ct lecture 2
 
ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...
ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...
ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-...
 
Haemodynamic monitoring
Haemodynamic monitoringHaemodynamic monitoring
Haemodynamic monitoring
 
Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...
 
Investigation of repeated blasts at Aitik mine using waveform cross correlation
Investigation of repeated blasts at Aitik mine using waveform cross correlationInvestigation of repeated blasts at Aitik mine using waveform cross correlation
Investigation of repeated blasts at Aitik mine using waveform cross correlation
 
introduction
introductionintroduction
introduction
 
Wiener Filter Hardware Realization
Wiener Filter Hardware RealizationWiener Filter Hardware Realization
Wiener Filter Hardware Realization
 
Implication of 3D Mapping in EP
Implication of 3D Mapping in EP Implication of 3D Mapping in EP
Implication of 3D Mapping in EP
 
Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...Cycle’s topological optimizations and the iterative decoding problem on gener...
Cycle’s topological optimizations and the iterative decoding problem on gener...
 
xldb-2015
xldb-2015xldb-2015
xldb-2015
 
CT based Image Guided Radiotherapy - Physics & QA
CT based Image Guided Radiotherapy - Physics & QACT based Image Guided Radiotherapy - Physics & QA
CT based Image Guided Radiotherapy - Physics & QA
 
Iaetsd a review on ecg arrhythmia detection
Iaetsd a review on ecg arrhythmia detectionIaetsd a review on ecg arrhythmia detection
Iaetsd a review on ecg arrhythmia detection
 
JFEF encoding
JFEF encodingJFEF encoding
JFEF encoding
 
Cardiac MRI
Cardiac MRICardiac MRI
Cardiac MRI
 
CyberSec_JPEGcompressionForensics.pdf
CyberSec_JPEGcompressionForensics.pdfCyberSec_JPEGcompressionForensics.pdf
CyberSec_JPEGcompressionForensics.pdf
 
DCE-MRI for Oncology in R
DCE-MRI for Oncology in RDCE-MRI for Oncology in R
DCE-MRI for Oncology in R
 
4.pptx
4.pptx4.pptx
4.pptx
 

Kürzlich hochgeladen

Data Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt RobisonData Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt RobisonAnna Loughnan Colquhoun
 
AWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAndrey Devyatkin
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerThousandEyes
 
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...Miguel Araújo
 
GenAI Risks & Security Meetup 01052024.pdf
GenAI Risks & Security Meetup 01052024.pdfGenAI Risks & Security Meetup 01052024.pdf
GenAI Risks & Security Meetup 01052024.pdflior mazor
 
presentation ICT roal in 21st century education
presentation ICT roal in 21st century educationpresentation ICT roal in 21st century education
presentation ICT roal in 21st century educationjfdjdjcjdnsjd
 
A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?Igalia
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerThousandEyes
 
Developing An App To Navigate The Roads of Brazil
Developing An App To Navigate The Roads of BrazilDeveloping An App To Navigate The Roads of Brazil
Developing An App To Navigate The Roads of BrazilV3cube
 
Strategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a FresherStrategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a FresherRemote DBA Services
 
A Domino Admins Adventures (Engage 2024)
A Domino Admins Adventures (Engage 2024)A Domino Admins Adventures (Engage 2024)
A Domino Admins Adventures (Engage 2024)Gabriella Davis
 
Tata AIG General Insurance Company - Insurer Innovation Award 2024
Tata AIG General Insurance Company - Insurer Innovation Award 2024Tata AIG General Insurance Company - Insurer Innovation Award 2024
Tata AIG General Insurance Company - Insurer Innovation Award 2024The Digital Insurer
 
TrustArc Webinar - Stay Ahead of US State Data Privacy Law Developments
TrustArc Webinar - Stay Ahead of US State Data Privacy Law DevelopmentsTrustArc Webinar - Stay Ahead of US State Data Privacy Law Developments
TrustArc Webinar - Stay Ahead of US State Data Privacy Law DevelopmentsTrustArc
 
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...apidays
 
Powerful Google developer tools for immediate impact! (2023-24 C)
Powerful Google developer tools for immediate impact! (2023-24 C)Powerful Google developer tools for immediate impact! (2023-24 C)
Powerful Google developer tools for immediate impact! (2023-24 C)wesley chun
 
The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024Rafal Los
 
Axa Assurance Maroc - Insurer Innovation Award 2024
Axa Assurance Maroc - Insurer Innovation Award 2024Axa Assurance Maroc - Insurer Innovation Award 2024
Axa Assurance Maroc - Insurer Innovation Award 2024The Digital Insurer
 
Scaling API-first – The story of a global engineering organization
Scaling API-first – The story of a global engineering organizationScaling API-first – The story of a global engineering organization
Scaling API-first – The story of a global engineering organizationRadu Cotescu
 
🐬 The future of MySQL is Postgres 🐘
🐬  The future of MySQL is Postgres   🐘🐬  The future of MySQL is Postgres   🐘
🐬 The future of MySQL is Postgres 🐘RTylerCroy
 

Kürzlich hochgeladen (20)

Data Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt RobisonData Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt Robison
 
AWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of Terraform
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected Worker
 
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
 
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
 
GenAI Risks & Security Meetup 01052024.pdf
GenAI Risks & Security Meetup 01052024.pdfGenAI Risks & Security Meetup 01052024.pdf
GenAI Risks & Security Meetup 01052024.pdf
 
presentation ICT roal in 21st century education
presentation ICT roal in 21st century educationpresentation ICT roal in 21st century education
presentation ICT roal in 21st century education
 
A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected Worker
 
Developing An App To Navigate The Roads of Brazil
Developing An App To Navigate The Roads of BrazilDeveloping An App To Navigate The Roads of Brazil
Developing An App To Navigate The Roads of Brazil
 
Strategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a FresherStrategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a Fresher
 
A Domino Admins Adventures (Engage 2024)
A Domino Admins Adventures (Engage 2024)A Domino Admins Adventures (Engage 2024)
A Domino Admins Adventures (Engage 2024)
 
Tata AIG General Insurance Company - Insurer Innovation Award 2024
Tata AIG General Insurance Company - Insurer Innovation Award 2024Tata AIG General Insurance Company - Insurer Innovation Award 2024
Tata AIG General Insurance Company - Insurer Innovation Award 2024
 
TrustArc Webinar - Stay Ahead of US State Data Privacy Law Developments
TrustArc Webinar - Stay Ahead of US State Data Privacy Law DevelopmentsTrustArc Webinar - Stay Ahead of US State Data Privacy Law Developments
TrustArc Webinar - Stay Ahead of US State Data Privacy Law Developments
 
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
 
Powerful Google developer tools for immediate impact! (2023-24 C)
Powerful Google developer tools for immediate impact! (2023-24 C)Powerful Google developer tools for immediate impact! (2023-24 C)
Powerful Google developer tools for immediate impact! (2023-24 C)
 
The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024
 
Axa Assurance Maroc - Insurer Innovation Award 2024
Axa Assurance Maroc - Insurer Innovation Award 2024Axa Assurance Maroc - Insurer Innovation Award 2024
Axa Assurance Maroc - Insurer Innovation Award 2024
 
Scaling API-first – The story of a global engineering organization
Scaling API-first – The story of a global engineering organizationScaling API-first – The story of a global engineering organization
Scaling API-first – The story of a global engineering organization
 
🐬 The future of MySQL is Postgres 🐘
🐬  The future of MySQL is Postgres   🐘🐬  The future of MySQL is Postgres   🐘
🐬 The future of MySQL is Postgres 🐘
 

Automated quantitative assessment of left ventricular functions by MR image segmentation

  • 1. Automated left ventricle segmentation in SAX CMR images using cost-volume filtering (CVF) and novel myocardial contour processing framework Graduate Institute of Communication Engineering College of EECS, NTU Oct 22, 2014 Speaker: An-Cheng Chang, MSc student
  • 2. What is CMR segmentation and why? • Tracing the chamber volume gives insight into how well the heart functions. • CMR segmentation is involved as a major part of the analysis. 2 Left ventricle (LV) Heart Cardiac Magnetic Resonance (CMR) images Epicardium border Endocardium border Long axis Apex Base Long axis
  • 3. Objective Automatically trace the LV endo∙card∙ium border 3 Left ventricle (LV) Endocardium border heartinner tissue (it is not a trivial task)
  • 4. Difficulties in CMR segmentation • Endocardium border is often obscured by papillary muscles and trabeculae carneae. (Fig A) • Variation between individuals. Wide pathological variations. (Fig B) • Low image quality: noises and distortions, e.g., field inhomogeneity, partial volume effect. (Fig C and Fig D) 4 C. Artifacts (field inhomogeneity) B. Wide subject variations D. Artifacts (partial volume) A. Endocardium border obscured by PMTC Ground truth Auto
  • 5. Related works on CMR segmentation Ngo et al., IEEE ICIP 2013 • Deep learning (pre-trained) + level set • Algorithm initialize by cropping ROI 5 Cropped by operator Pre-trained deep learning network Cropped by operator Hu et al., Magn Reson Imaging (Elsevier, 2013) • GMM + dynamic programming • Algorithm initializes by cropping ROI Method SR (%) Mean(StD) APD (mm) Mean(StD) DM Mean(StD) Hu 2013 91.1(9.4) 2.24(0.40) 0.89(0.03) Ngo 2013 97.9(6.18) 2.08(0.40) 0.90(0.03) Ours 94.1(6.1) 1.75(0.42) 0.91(0.03) Ours (2014) • CVF + contour processing • Algorithm initializes by one click on the LV
  • 6. Result highlight 6 Auto contours rejected: 0 out of 18 Mean error: 2.59mm (+0%) EF underestimated by 4% Subject SC-HYP-40; hypertrophic heart. Auto contours rejected: 7 out of 18 Mean error: 3.23mm (+25%) EF underestimated by 11% With proposed CVF Without CVF
  • 7. Endocardium delineation using CVF and proposed myocardial contour processing 7 LV localization & ROI refinement Blood pool classification by CVF Myocardial contour processing 3D+T volume Result (BASE) (APEX) slice 1 slice 2 slice M . . . . . . . . . . . . . . . t=1 t=2 t=N Polar transformation Inverse polar transformation Proposed method
  • 8. Endocardium delineation using CVF and proposed myocardial contour processing 8 LV localization & ROI refinement Blood pool classification by CVF Myocardial contour processing 3D+T volume Result (BASE) (APEX) slice 1 slice 2 slice M . . . . . . . . . . . . . . . t=1 t=2 t=N Polar transformation Proposed method Inverse polar transformation
  • 9. Blood pool classification by CVF 9 Signal intensity Occurrence LV blood poolOthers T LV blood pool Others Binary label image: Local variance map: Cost slice for selecting ‘LV blood pool’ Cost slice for selecting ‘Others’ Proposed cost volume Refined blood pool segment: Cost aggregation & Label selection Polar transformation Proposed cost initialization Cost-volume filtering (CVF)
  • 10. Cost-volume-filtering (CVF) based image segmentation • CVF is originally used for refining stereo matching results. Recently been generalized for discrete labeling problems (Rhemann et al.; appears in PAMI 2013, CVPR 2011). • Method: Initializing cost(x, y, label)  cost aggregation  label selection. • Cost initialization scheme depends on applications. 10 *Rhemann et al.; appears in PAMI 2013, CVPR 2011 (a) Stereo matching (depth from stereo) (b) Interactive image segmentation Applications of CVF
  • 11. Why CVF? • Histogram-based labeling (Otsu’s method, GMM-based thresholding) ignores spatial relationship. 11 LV blood pool Others LV blood pool Others T 2D image 1D histogram 2D image Spatial information lost • CVF considers both spatial relationship and intensity similarity when outputting labeled results. -> Good for handling bias field caused by MR field inhomogeneity. • We proposed a new cost initialization scheme for CVF to address the partial volume effect in MR images. • CVF is fast. Can be O(N) time and non-approximate.
  • 12. Effectiveness of CVF-based segmentation 12 • The proposed cost initialization scheme addresses the partial volume issue. • Compare the result between Fig B and Fig C.
  • 13. 13 • Robust against MR field inhomogeneity Robustness of CVF-based segmentation . PA . PB PA: Brighter PB: Dimmer Polar transformation Inverse polar transformation
  • 14. Principle of CVF: using proposed CMR segmentation as an example 14 Cost(x, y, label) Image data Cost aggregation Guide image Labels(x,y)  Cost initialization Image data Original image Cost volume Cost(x, y, ’blood pool’) Cost(x, y, ’others’) High cost Low cost Binary label image Local variance map Label selection (reduce dimension) Filtered_cost(x, y, label) LV blood pool Others
  • 15. Principle of CVF: using proposed CMR segmentation as an example 15 Cost(x, y, label) Image data Cost aggregation Guide image Cost initialization Labels(x,y)  Label selection (reduce dimension) Filtered_cost(x, y, label) Kernel (Box filter) × = Cost slice Ci Guide image I Weighting W
  • 16. Principle of CVF: using proposed CMR segmentation as an example 16 Cost(x, y, label) Image data Cost aggregation Guide image Cost initialization Labels(x,y)  Label selection (reduce dimension) Filtered_cost(x, y, label) Kernel (Box filter) × = Cost slice Ci The principle: cost is aggregated from similar* neighbors. *similar in guide image I Guide image I Weighting W (shift-variant)
  • 17. Principle of CVF: using proposed CMR segmentation as an example 17 Cost(x, y, label) Image data Cost aggregation Guide image Cost initialization Labels(x,y)  Label selection (reduce dimension) Filtered_cost(x, y, label) Guide image I Kernel (Gaussian) × = Weighting W The principle: cost is aggregated from similar* neighbors. *similar in guide image I Cost slice C1 for ‘blood pool’
  • 18. Principle of CVF: using proposed CMR segmentation as an example 18 Cost(x, y, label) Image data Cost aggregation Guide image Cost initialization Filtered_cost(x, y, label) Labels(x,y)  Label selection (reduce dimension) Filtered cost slice C1’ Filtered cost slice C2’ Labels fSelect the label with the least cost CVF-refined Binary label image Shift- variant filter W(I)
  • 19. 19 CVF: one iteration CVF: 100 iterationsOtsu‘s bi-level thresholding Original image
  • 20. Endocardium delineation using CVF and proposed myocardial contour processing 20 LV localization & ROI refinement Blood pool classification by CVF Myocardial contour processing 3D+T volume Result (BASE) (APEX) slice 1 slice 2 slice M . . . . . . . . . . . . . . . t=1 t=2 t=N Polar transformation Inverse polar transformation Proposed method
  • 21. Endocardial Contour Processing Framework 21 ■ Corrected contour (assigned to set B) ■ Good contour (assigned to set G) Contour based on Canny’s edge detector Contour based on CVF papillary muscletrabaculae Inverse polar transform C(p) I(p) E(p) Combine two raw contours I and C and return a regularized final contour • I: based on CVF-labeled binary image; intensity similarity • C: based on Canny’s edge detector; gradient information
  • 22. Goodcontour Correctcontour Contour function generation and correction 22 1D contour function f First-order derivative f’ Fluctuations 2D labeled image Fluctuations are detected and corrected by linear interpolation Corrected contour Fluctuations are a result of the presence of papillary muscle and trabeculae
  • 23. Detecting the fluctuations 23 Box filter w2 Box filter w3 Impulse wδ ∗ ∗ ∗ 𝑓′ 𝑓′ 𝑓′ Type 1 Type 2 Type 3 Contour function 1st order derivative ResponseFilter bank
  • 24. Endocardial Contour Processing Framework 24 ■ Corrected contour (assigned to set B) ■ Good contour (assigned to set G) Contour based on Canny’s edge detector Contour based on CVF papillary muscletrabaculae Inverse polar transform C(p) I(p) E(p) Combine two raw contours I and C and return a regularized final contour • I: based on CVF-labeled binary image; intensity similarity • C: based on Canny’s edge detector; gradient information
  • 25. Complementary contour generation 25 Original image Non-maximum suppressed edge response Edge pruned Complementary contour function C Contour form CVF acts as a guide to pick up Canny’s edge response
  • 26. Endocardial Contour Processing Framework 26 ■ Corrected contour (assigned to set B) ■ Good contour (assigned to set G) Contour based on Canny’s edge detector Contour based on CVF papillary muscletrabaculae Inverse polar transform C(p) I(p) E(p) min 𝒪 𝐸 = 𝑝 ∈ 𝐵 𝐸 𝑝 − 𝐶 𝑝 2 + 𝑝 ∈𝐺 𝐸 𝑝 − 𝐼 𝑝 2 + 𝜆 𝑝 𝐸′′ 𝑥 2 Combine C(p) and I(p) by minimizing the following objective function: data term smoothness term Least squares problem  re-formulate to Ax=b and solve for x
  • 27. Endocardial Contour Processing Framework • The additional constraint will ensure the final contour to enclose the blood pool 27 𝒪 𝐸 = 𝑝 ∈ 𝐵 𝐸 𝑝 − 𝐶 𝑝 2 + 𝑝 ∈𝐺 𝐸 𝑝 − 𝐼 𝑝 2 + 𝜆 𝑝 𝐸′′ 𝑥 2 subject to 𝐸 𝑥 > 𝐼(𝑥) Auto Auto /w constraint By expert
  • 28. 28 Inverse polar transform Review of the system processing flow Lock down ROI Polar mapping Otsu’s thresholding After CVF Myocardial contour processing
  • 29. Test dataset: Sunnybrook cardiac MR database • The first (in 2009) publically accessible cardiac MR database • Provides 45 MR datasets including one healthy plus three pathological cases • Includes an evaluation tool & ground truth at end-diastole (ED) and end- systole (ES) 29 Technical details: • Acquisition protocol: SSFP MR SAX images are obtained during 10-15 second breath-holds with a temporal resolution of 20 cardiac phases over the heart cycle, and scanned from the ED phase. Six to 12 SAX images were obtained from the atrioventricular ring to the apex • MRI scanner: 1.5T GE Signa MRI. (thickness=8mm, gap=8mm, FOV=320mm*320mm, matrix= 256*256)
  • 30. Segmentation example: Case of hypertrophy 30 Left: computed. Right: expert-drawn
  • 31. Segmentation example: Case of normal heart 31 Left: computed. Right: expert-drawn
  • 32. Segmentation example: Case of heart failure /w infarction 32 Left: computed. Right: expert-drawn
  • 33. CMR volume segmentation at ED and ES phase Patient: SC-HF-I-01 (Heart failure with infarct) 33 -Red: Auto. - - - Purple: expert
  • 34. CMR volume segmentation at ED and ES phase Patient: SC-HYP-09 (Hypertrophy) 34 -Red: Auto. - - - Purple: expert
  • 35. CMR volume segmentation at ED and ES phase Patient: SC-HF-NI-04 (Non-ischemic heart failure) 35 -Red: Auto. - - - Purple: expert
  • 36. Evaluation metrics Average perpendicular distance (APD) • The computed contour for a given slice is qualified if APD < 5mm Overlapping dice metric (DM) DM = 2𝐴 𝑎𝑚 𝐴 𝑚 + 𝐴 𝑎 Success rate (SR) • Number of qualified contours (APD < 5mm) in one CMR scan. 36 SR=75%
  • 37. Performance evaluation with respect to pathological groups SR (%) APD (mm) DM Group Mean StD Mean StD Mean StD SC-HF-I 94.2 7.7 1.542 0.296 0.93 0.02 SC-HF-NI 95.1 4.1 1.736 0.459 0.92 0.02 SC-HYP 93.6 6.5 1.900 0.420 0.88 0.03 SC-N 93.4 5.3 1.834 0.377 0.89 0.02 Overall 94.1 6.1 1.748 0.417 0.91 0.03 37 A total of 800 MR images from 45 patients are evaluated
  • 38. Result comparison Method SR (%) Mean(StD) APD (mm) Mean(StD) DM Mean(StD) Huang 2011 (auto) 81.5(18.0) 2.19(0.44) 0.91(0.03) Hu 2013 (auto) 91.1(9.4) 2.24(0.40) 0.89(0.03) Constantinides 2012 (semi-auto) 91.0(8.0) 1.94(0.42) 0.89(0.04) Constantinides 2012 (auto) 80.0(16.0) 2.44(0.56) 0.86(0.05) Ngo 2013 (semi-auto) 97.9(6.18) 2.08(0.40) 0.90(0.03) Ours (auto) 94.1(6.1) 1.75(0.42) 0.91(0.03) 38 Ours against others. All use the same 45-patient dataset.
  • 39. Additional result comparison Method SR (%) Mean(StD) APD (mm) Mean(StD) DM Mean(StD) Jolly 20091 95.62(8.83) 2.26(0.59) 0.88(0.04) Lu 20091 72.45(18.86) 2.07(0.61) 0.89(0.03) Huang 20091 -- 2.10(044) 0.89(0.04) Wijnhout 20091 86.47(11) 2.29(0.57) 0.89(0.03) Constantinides 20091 92.28(--) 2.04(0.47) 0.89(0.04) Marák 20091 -- 3.00(0.59) 0.86(0.04) Feng 2013 92.8(9.2) 1.93(0.37) 0.86(0.04) Ngo 2013 96.58(3.66) 2.22(0.46) 0.89(0.03) Ours 96.31(4.85) 1.67(0.40) 0.91(0.03) 39 Ours against others. All use the same 15-patient ‘validation’ dataset. 1 Results reported in MICCAI LVSC 2009.
  • 40. Evaluating ejection fraction 40 EF error Group Mean StD SC-HF-I -0.15 2.89 SC-HF-NI 2.27 4.64 SC-HYP 2.13 5.01 SC-N 2.39 5.70 Overall 1.61 4.72 y = 0.9547x + 0.3663 R² = 0.9421 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 AUTOEF MANUAL EF AUTO VS. MANUAL EF SC-HF-NI SC-HF-I SC-HYP SC-N Method R2 for EF Cocosco 2008 0.90 Lu 2013 0.92 Lorenzo-Valdés 2004 0.92 Cordero-Grande 2011 0.92 Constantinides 2012 0.83 Ours 0.94 -10 -5 0 5 10 15 20 0 20 40 60 80 100 AUTO-MANUAL (AUTO + MANUAL) /2 BLAND-ALTMAN PLOT FOR EF Mean: 1.61 SD: 4.72 Measures the proportion of blood ejected with each cardiac cycle • EF = (ED volume – ES volume)/ED volume
  • 41. Conclusion • Developed an algorithm that detects the left ventricular endocardial contour in CMR images, with top-tier accuracy. • Use cost-volume filtering (CVF) to combat MR inhomogeneity. • Proposed a novel cost initialization scheme that handles partial volume effect. • Proposed a contour processing framework, in which information from gradient and intensity similarity are encoded along with a smoothness constraint • Clinical aspect: highly correlated (R2 = 0.94) between auto and manual EF. No systematic bias is observed. • Future work includes incorporating inter-slice and inter-frame relationships to increase detection rate. 41 LV localization & ROI refinement Blood pool classification by CVF Myocardial contour processing 3D+T volume Result (BASE) (APEX) slice 1 slice 2 slice M . . . . . . . . . . . . . . . t=1 t=2 t=N Polar transformation Inverse polar transformation Proposed method
  • 42. Supplement A: Automated localization of the left ventricle (LV) 43 • Find the area that covers LV blood pool and LV muscle • Lock down the region of interest (ROI) and hands off the ROI to the rest of the algorithm • This will exclude the influence of nearby tissues when making initial estimate of LV blood pool Iteratively refining the region of interest (ROI)
  • 43. 44 Blood pool Myocardium Others Signal intensity Occurrence Myocardium LV blood pool Others T Class=‘others’if signal < T Class=‘LV blood pool’if signal > T The rationale T?
  • 44. (Fig A) Initialize a ROI inside the LV, then: • (Fig B) Classify the pixels in the ROI (using Otsu’s method) • (Fig C) Retain the primary connected component • (Fig E) Compute LV blood pool’s convex hull • (Fig D) Dilate the convex hull. This has become the new ROI Repeat B->C->E->D until convergence 45 Step-by-step breakdown
  • 45. 46
  • 46. 47 • Low contrast area can be recovered regardless of detected LV centroid PB . PC . PC . PB . PB . PC . PC PC . PB . PB PB PC Polar transformation vs. vs. Ours Robustness of CVF-based segmentation
  • 47. What brings down the success rate? 48 Supplement B:
  • 48. Extreme case of hypertrophy (SC-HYP-08) 49 -Red: automated. - - - Purple: expert-drawn
  • 49. What brings down the success rate? • Mostly in apical slices & extremely hypertrophic hearts. • Severe partial volume effect in apical slices violates our assumption. • In hypertrophic cases, endocardium border can be completely obscured by papillary muscle and trabeculae carneae. 50 (A) (B)

Hinweis der Redaktion

  1. “C” “MR”
  2. 照心臟MRI除了可以觀察病理特徵之外,還可以獲得其他額外資訊
  3. A. 因為左心室內部不是平滑的 心肉柱會附著在心肌內壁 還有一個獨立的肌肉組織叫做乳狀肌非常靠近心肌內壁 B. 無法用複雜模型完美描述所有病例和個體的差異。愈少假設愈好
  4. Deep learning and level set 先不管技術本身
  5. CVF 用來解決partial volume所造成的像素分類錯誤
  6. *Briefly describe the workflow* The input is the CMR 4D volume, and we will process each image individually. Each image will go through three main stages of the algorithm, and the endocardial contour that goes along the inner muscle wall of the left ventricle will be produced. (First thing we need to do is to locate the left ventricle Then we classify the picture elements (pixels) and identify which are inside the LV and which are not In the third stage, we will process the contour of the left ventricle )
  7. *Briefly describe the workflow* The input is the CMR 4D volume, and we will process each image individually. Each image will go through three main stages of the algorithm, and the endocardial contour that goes along the inner muscle wall of the left ventricle will be produced. (First thing we need to do is to locate the left ventricle Then we classify the picture elements (pixels) and identify which are inside the LV and which are not In the third stage, we will process the contour of the left ventricle )
  8. Right after we locate the left ventricle, we do polar coordinates mapping (why?) to transform the original image to what we call here a polar image. Next, we detect the blood pool using the decision boundary we just determined. However, this simple decision process does not give us the best segmentation result yet in many cases. So we will further refine it using a technique called cost-volume filtering process. And the cost-volume filtering *If asked what CVF is: it is a
  9. Discrete labeling problems: spatially
  10. Field inhomogeneity=gradual intensity roll-off. Blood in LV should have the same level of gray,
  11. 讓受到partial volume影響的地方 選擇兩個label的cost都相等 舉例:
  12. Each candidate label
  13. 舉出CVF特性: not space-invariant
  14. Each candidate label
  15. Each candidate label
  16. *Briefly describe the workflow* The input is the CMR 4D volume, and we will process each image individually. Each image will go through three main stages of the algorithm, and the endocardial contour that goes along the inner muscle wall of the left ventricle will be produced. (First thing we need to do is to locate the left ventricle Then we classify the picture elements (pixels) and identify which are inside the LV and which are not In the third stage, we will process the contour of the left ventricle )
  17. Detecting the blood pool in the left ventricle is only part of the segmentation. Because inside the muscle wall, there are some irregularities that can't be detected so far. These irregularities, by clinical practice, need to be included in the blood pool as well.  So we propose a contour processing framework to deal with this. Here is the
  18. Detecting the blood pool in the left ventricle is only part of the segmentation. Because inside the muscle wall, there are some irregularities that can't be detected so far. These irregularities, by clinical practice, need to be included in the blood pool as well.  So we propose a contour processing framework to deal with this. Here is the
  19. Detecting the blood pool in the left ventricle is only part of the segmentation. Because inside the muscle wall, there are some irregularities that can't be detected so far. These irregularities need to be included in the LV blood pool as well.  So we propose a contour processing framework, 
  20. Section closing remarks: That was our proposed method for segmenting the LV blood pool. Next we will evaluate the performance of our method.
  21. We use the Sunnybrook cardiac MR database to evaluate our algorithm. A lot researchers also evaluate their algorithms with this database, so we are able to directly compare against each other.
  22. Visual examinations of the result Using the algorithm we developed, we are able to closely match the expert’s delineation of endocardial contour.
  23. Here are a few more examples
  24. APD measures the error
  25. That means it can be used in practical situations. It saves human time, it involves minimal human control, and the result is highly accurate.
  26. Our method starts with locating the target, the left ventricle, in the CMR images. The localization process starts with a small region at the center of the image, and eventually the region will grow to fully cover the left ventricle This step is very crucial to the rest of the algorithm. The first reason is that we want to apply the algorithm on the right target. And the second reason is like this:
  27. If we do a statistical analysis of the signal intensity, we find that each tissue type has its own intensity distribution. For example, The left ventricle blood pool is generally brighter than the left ventricle muscle wall; so you will see two distinctive hills here on the histogram. And its clear to us that we can minimize the decision error by putting a decision boundary here to classify these two tissues. However, the outside tissues may interfere with the decision process. Now it won't be easy to find a good decision boundary that divide the blood pool and the muscle wall because now the distribution function of so many tissue types are mixed up together.  And that is the second reason why locating the left ventricle is important; by doing so we remove external tissues' influence on the decision process so we are able to get a better classification result that correctly labels the blood pool
  28. (Skip)