Transcript Introduction to Segmentation, Registration and

eEdE-73 Introduction to Segmentation, Registration and Volume
Analysis for Imaging Genomics
Ginu A. Thomas, MD; Ahmed Elakkad, MD; Mohamed Elbanan, MD;
Pascal O. Zinn, MD; Rivka R. Colen, MD.
Section of Neuroradiology
Department of Diagnostic Radiology
MD Anderson Cancer Center
Houston, Texas.
Disclosures
•
No Disclosures
Questions
1. Zinn et al. in 2011 uncovered
which Gene/miRNA that is
associated with high FLAIR signal
in GBM?
2. Diehn et al. discovered the
association between high contrast
enhancement to necrosis ratio and
which of the following?
1.
2.
3.
4.
1.
2.
3.
4.
FHL2
PDGF
POSTN
TNFSF4
CAV1
EGFR
PGK
VEGF
Purpose
• Image processing is essential for extracting data for any
type of research study dealing with quantitative imaging
variables.
• Before any type of volumetric analysis, the volumes in
question should be well defined and presented as data
types that are easy to deal with.
• In this exhibit we look at the methodology for image
registration, segmentation and model building that we
used for our study imaging genomic mapping in
Glioblastoma (GBM).
Background
RSNA’s Quantitative Imaging Biomarkers Alliance (QIBA)
definition of Quantitative imaging:
• It is the extraction of quantifiable features from
medical images for the assessment of normal or the
severity, degree of change, or status of a disease,
injury, or chronic condition relative to normal.”
Background
• Across different modalities, MRI gives the largest amount
of quantifiable data.
• Radiologists usually perform some sort of quantitative
imaging in their daily clinical practice, from measuring
and annotating abnormalities to determining tissue
perfusion or diffusion and tracer uptake.
Approach/Methods
•
We used 3D Slicer software for
image visualization, analysis,
manipulation, segmentation,
obtaining tumor volumetry and
3D visualization.
•
3D Slicer is a product of
continuous international
collaboration led by the
Surgical Planning Lab at
Brigham & Women’s Hospital.
Approach/Methods
• The platform has about 125
different modules and provides
functionality for segmentation,
registration and three
dimensional visualization of
multimodal image data, as well
as advanced image analysis
algorithms for diffusion tensor
imaging, functional magnetic
resonance imaging and image
guided therapy.
Approach/Methods Cont’d
We used two volumes for our project:
1. FLAIR series for segmentation of the
edema/invasion (Fast Spin Echo (FSE) T2 or PD if FLAIR is
not available).
2. Post-contrast T1 weighted imaging (T1WI) series for
segmentation of the tumor enhancing portion and the
necrosis. We used Spoiled Gradient Echo Recaller (SPGR) or
PWI series if T1 Post contrast wasn’t available).
Approach/Methods Cont’d
T1 POST with enhancing portion depicted using yellow label and
necrosis using red label in this Left Temporal GBM
Approach/Methods Cont’d
FLAIR Image registered over the T1POST to get accurate
quantification of edema/Invasion (Blue label)
Approach/Methods Cont’d
• Each sequence was loaded in Slicer as a volume.
• The 'Parse Directory' function was used to parse the
series and show all the series available in the data.
Approach/Methods Cont’d
Image visualization:
•
We selected all the desired series separately to be
opened in the Volume module.
•
The Volume module loads saves and adjusts the display
parameters of volume data.
•
A volume can be set for raw images of a certain
sequence or label map and various options are available
to adjust the way a volume is displayed.
Approach/Methods Cont’d
• It provides information about the volumes displayed such
as image dimensions, image spacing, image origin, scan
order, number of scalars, type of scalar and window/level
presets.
• Various options are available to adjust the way a volume
is displayed, and to change the levels of opacity &
positioning of each.
Approach/Methods Cont’d
• The Volumes module also enables users to choose the
appropriate color map, manually adjust window/level sets
and display a histogram which plots the number of pixels
against image intensity over a background of the ongoing
window/level.
Findings/Discussion Cont’d
I. Registration
• Different scans may be recorded at different
angles, with different voxel sizes and use different
number of slices.
• We co-register these scans, brought them into
spatial alignment using mutual information
optimization.
Findings/Discussion Cont’d
Modules used in Registration
Transformation module:
• We used Transformation module to manually create and
edit Slicer transformation nodes.
• Transformation nodes are matrices of data that map the
voxels of different volume.
• They can be both linear or nonlinear warp transformations.
• We usually specified the T1 volume as the fixed base
volume, and the FLAIR as the moving 'transformed'
volume. We used the 'Translation' and 'Rotation' tools to
manually align every 2 volumes in space coordinates.
Findings/Discussion Cont’d
Manual registration
module in Slicer
Findings/Discussion Cont’d
Resample Scalar/Vector/DWI Volume :
•
This module implements image and vector-image
resampling through the use of ITK Transforms.
•
'Resampling' is performed in space coordinates, not
pixel/grid coordinates.
•
It is quite important to ensure that image spacing is
properly set on the images involved.
Findings/Discussion Cont’d
II. Segmentation
•
Image segmentation is a means to simplify
image
data by dividing it into useful and
relevant segments.
Findings/Discussion Cont’d
Hui Li, Jianfei Cai, Thi Nhat Anh nguyen, Jianmin Zheng. A benchmark for semantic Image
Segmentation. IEEE ICME 2013
Findings/Discussion Cont’d
• Segmentation was performed after perfectly coregistering the different sequences loaded on Slicer.
• Segmentation was carried out in a simple hierarchical
model of anatomy, proceeding from the peripheral to the
central.
• Three different structures were segmented and later
modeled, namely, edema/invasion, enhancing tumor and
necrosis, respectively.
Segmentation FLAIR
signal hyperintensity
Segmentation of
enhancing tumor
portion on Post-Gd
T1WI
One label map
showing three tumor
compartments
Findings/Discussion Cont’d
• After segmentation was
performed, 3D models
were created to give
better visualization and
understanding of spatial
anatomic relationships
between different tumor
compartments.
Findings/Discussion Cont’d
3D Slicer Editor module:
• This is a core module for manual and semi-automatic
segmentation of images.
• Some of the tools mimic digital painting applications like
Photoshop or GIMP, but Slicer tools work on 3D arrays of
voxels rather than on 2D pixels, considering slice
thickness and slice spacing.
• The overall goal is to efficiently and precisely define
structures within their volumes as “label map volumes”.
Findings/Discussion Cont’d
Label map showing segmented tumor
model in 3D slicer.
Findings/Discussion Cont’d
• These label maps can be used for building 3D models,
further processing and also for quantification of the
segmented portions.
• Label maps can be exported and re-used in different
other softwares, thus, expanding the usage of such
image analysis.
• We used the single label map approach to create models
of different tumor volumes on the same label map. Of
note, Slicer gives only one label to each pixel.
Findings/Discussion Cont’d
• Manual segmentation is very precise and accurate in
comparison to automated and semi-automated methods
but it is more laborous and time consuming.
• Semi-automated methods such as:
1. Thresholding†
2. Region-growing and competitive region-growing†‡
3. Atlas-based techniques
4. Edge-detection techniques†
5. Clustering†‡ *
†Srinivasan G, Shobha G. Segmentation techniques for target recognition. International Journal of Computers and Communications 2007; 1:313333.
‡Tirpude N, Welekar R. A study of brain magnetic resonance image segmentation techniques. IJARCCE 2012; 2:958
* Bauer S, May C, Dionysiou D, et al. Multiscale modeling for image analysis of brain tumor studies. IEEE Trans Biomed Eng 2012; 59:25-29
Findings/Discussion Cont’d
III. Model Making
 Using the 'Make Model' tool of the 'Editor' module, the
models of edema/invasion, tumor enhancement and
necrosis can be generated from the previously performed
segmentation.
 This module is used for loading, saving, layering, changing
the appearance of, and organizing 3D surface models. The
'Info' section provides the general information about the
models like Surface Area, Volume, etc.
3D Model making in Slicer
Findings/Discussion Cont’d
IV. Data Display
The Data Module:
• It gives a complete view of the hierarchical (tree)
representation of the MRML Scene and allows you to
modify the MRML Scene and the individual "MRML
nodes" that represent its constituent data.
•
It allows you to create and edit transformation
hierarchy, and to fix and modify models and volumes at
a desired registration under transform nodes.
Findings/Discussion Cont’d
Obtaining the volumes:
• After the models of edema, tumor
and necrosis have been generated
from the segmentation, the
corresponding volumes can be
calculated and noted from the
‘Label statistics' module.
• Label Statistics module is used for
obtaining image and volume
properties from the label maps
Summary/Conclusion
• This quantitative approach minimizes intra-reader and
inter-reader subjectivity and allows more objectivity.
Quantitative imaging provides a noninvasive method for
initial characterization and assessment of abnormalities
and evaluating progression/regression and therapy
response.
Summary/Conclusion
• The complementary information provided by Quantitative
Imaging will help improve outcome prediction and
evaluation of tumor response to therapy, especially, with
certain pathologies where progression is very hard to be
detected on longitudinal MRI scans.
Summary/Conclusion
• Future directives in quantitative image analysis will focus
on :
1. reducing manual user interaction, improving accuracy
and reproducibility of the automated and semiautomated techniques
2. Optimization of Segmentation algorithms to suit a
wider scope of lesions such as Ground Glass Opacity,
Mixed Nodules where contour detection is very
challenging.
Summary/Conclusion
• Large scale comparative studies are needed to
compare manual segmentations versus semiautomatic and automatic algorithms, and to
compare different segmentation algorithms and
softwares.
Answers
1. Zinn et al. in 2011 uncovered
which Gene/miRNA that is
associated with high FLAIR signal
in GBM?
2. Diehn et al. discovered the
association between high contrast
enhancement to necrosis ratio and
which of the following?
1.
2.
3.
4.
1.
2.
3.
4.
FHL2
PDGF
POSTN
TNFSF4
CAV1
EGFR
PGK
VEGF
Thank you