Title: Automated neuroanatomical microvessel segmentation, nuclei labeling, and geometric computations with
1Automated neuroanatomical microvessel
segmentation, nuclei labeling, and geometric
computations with 3D two photon laser scanning
microscopyJoint Work J. P. Kaufhold, P. S.
Tsai, P. Binder, B. Friedman, P. D. Lyden, D.
Kleinfeld and A.I. Ifarraguerri
- John Kaufhold, Ph.D.
- SAIC Intelligent Systems Division
- February 1st, 2008
Grant Manipulating Neural Tissue With Ultrashort
Laser Pulses NIH NIBIB BRP R01 EB003832
22000s
See Paul Frenchs and Jelena Kovacevics talks,
e.g.
3The Microscope
4Microvessel segmentation, nuclei labeling, and
geometric computations
RESULT
Tsai, Blinder, Kaufhold, Friedman, Ifarraguerri,
Lyden, Kleinfeld (in progress)
5Related work Francis Cassot, Frederic Lauwers,
Celine Fouard, Prohaska, and Lauwers-cances
Valérie. Microcirculation, 13(1)1-18, 2006.
6Enabling technology 1 All Optical Histology
High Fluence Ultrashort Laser Pulses for
Ablation Threshold Fluence 1J/cm2
1mJ/(3mm)2 Combined with Low Fluence Pulses for
TPLSM Imaging
7Enabling technology 2 Multiply Labeled
Tissue (vasculature nuclei)
8Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
9Sample control (age, weight, sex)
10Sample control (age, weight, sex)
legend
11Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
12Gain Correction
Before Gain Correction
Mean projection across 400 z-frames
13Registration And Tiling
Block 1
Block 2
Maximize correlation coefficient of
intensities in overlap
Block n
14Channel Subtraction
D DAPI channel only somata
F FITC channel mostly vessels some cell nuclei
A Alexa 594 channel labeled neuronal nuclei
some vessels some cell
nuclei
D F A
15(No Transcript)
16Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
17Grid volume into overlapping blocks
Registered and Tiled Volume
Block size 100x100x100 voxels
Strip data size 300x700x1000 voxels AOH data
size 1000x1000x1000 voxels
18Restitching Binary Volumes
This process is repeated for all block overlaps,
reducing segmentation edge effects.
19Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
20Produce Binary Somata Masks
INPUT
Raw Block
Filtering 1) regularizes somata shape and 2)
separates SM nuclei intensities from background
21Extract Nuclear Centroids From Mask
INPUT
OUTPUT
2D illustration of 3D steps to isolate centroids
from masks
22Validate Nuclei Centroid Estimates
Nuclei Mask Overlays
Nuclei Centroid Overlay (single slice)
Nuclei Centroid Overlay (10 m projection)
23Nuclei Centroid Estimation Performance
automated algorithm
hand-labeled centroids
11 correspondence, Automated method to hand label
mT
False Alarms
Misses
Scoring ground truth centroids vs. automatic
centroid location estimates
24Association distance lt 10 microns
25Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
26Classification of Nuclei Types
Aim Separate neuronal nuclei from non-neuronal
nuclei Method Iteratively reestimate classes
average intensity in jth background mask
Nucleus to background difference is computed per
jth centroid
27Classification of Nuclei Types
, compute decision boundary
At ith iteration,
28Classification of Nuclei Types
Counts
Ratio v1 / mean (v2, v3, vn)
29Classification of Nuclei Types
Counts
Ratio v1 / mean (v2, v3, vn)
30Classification of Nuclei Types
Counts
Ratio v1 / mean (v2, v3, vn)
31Classification of Nuclei Types
OUTPUT
Counts
Ratio v1 / mean (v2, v3, vn)
32Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
33Vessel Segmentation Cascade
INPUT
See also Frangi, Niessen, Vincken, Viergever,
1998 Fridman, Pizer, Aylward, Bullitt, 2003
34Vessel Segmentation Cascade
OUTPUT
Kirbas Quek Review, 2000 Viola Jones Detector
Cascade, CVPR 2001
35Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
36Morphometry Restoring Anisotropy Corrections
Cortical Surface
R into Cortex
INPUT
Vascular Mask, VM Orthographic projection (prior
to aniostropy correction)
Observation Vessel footprints at left thinner
than footprints at right. This is due to PSF
elongation in imaging depth direction. Issue Bias
es microvascular density estimates vs. depth into
cortex and complicates capillary isolation.
Imaging Depth into plane
Imaging Depth
500u
37Morphometry Restoring Anisotropy Corrections
Cortical Surface
R into Cortex
OUTPUT
Vascular Mask Corrected, VC Orthographic
projection (post aniostropy correction)
Result Vessel footprints at right same width on
average as those at left. Vessel locations are
mostly preserved. Assertion Vessel diameter more
useful post-correction.
Imaging Depth into plane
Imaging Depth
500u
38Morphometry Restoring Anisotropy Corrections
METHOD impose isotropy via binary deconvolution
Imaging Depth
39Morphometry Restoring Anisotropy Corrections
40Morphometry Restoring Anisotropy Corrections
Amount to symetrically shave from local vessel
segment x-secs as a function of imaging depth
9
8
7
Major to Minor Axis Diff
6
5
4
3
0
100
200
300
400
500
600
700
Imaging Depth Bin
41Morphometry Restoring Anisotropy Corrections
INPUT
Imaging Depth
42Morphometry Restoring Anisotropy Corrections
OUTPUT Local shaving
Imaging Depth
43- Binary Deconvolution self-calibration issues
- Vessels parallel to the sampled plane should be
ignored for estimating major-minor axes.
Alternatively, a calibration phantom of small
FITC-filled tubes (2-15 m radii) would be ideal. - Does not take into account potential increases
in estimated vessel diameter with increased FITC
intensity. This factor could be absorbed into the
local correction in all directions. - Constrained to thin to no less than 2 voxels
horizontally. - Vessels x-sections with 2 components on a
horizontal strip may fragment.
44Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
45Capillary Network Isolation
Cortical Surface
R into Cortex
INPUT Anisotropy-corrected microvascular
mask Observation Larger vessels contribute
significantly to volume fraction, but participate
significantly less in gas exchange. Aim Remove
vessels larger than capillaries.
46Capillary Network Isolation
Cortical Surface
R into Cortex
OUTPUT An approximation to capillary-only
microvasculature isolation. METHOD Only vessel
segments whose centerline distances to the
closest non-vessel point lt 3u are kept (i.e. lt6u
lumenal diameters).
47Capillary Network Isolation
INPUT
All vessels projected, VC
48Capillary Network Isolation
49Capillary Network Isolation
50Capillary Network Isolation
51Capillary Network Isolation
52Capillary Network Isolation
53Capillary Network Isolation
54Capillary Network Isolation
55Capillary Network Isolation
56Capillary Network Isolation
57Capillary Network Isolation
58Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
59Aim Extract vectorized model of centerline points
60Vectorization of Microvascular Networks
Surface Rendering (2.7 of previous volume)
61Vectorization of Microvascular Networks
Surface Rendering (2.7 of previous volume)
Centerlines
Method Francis Cassot, Frederic Lauwers, Celine
Fouard, Prohaska, and Lauwers-cances Valérie.
Microcirculation, 13(1)1-18, 2006.
62Vectorization of Microvascular Networks
63Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
64Cortical Flattening
INPUT
METHOD Cortical surface map estimation and
voxel-column shifting
OUTPUT
Before Cortical Surface Correction
After Cortical Surface Correction
65(No Transcript)
66Cell Population Count andVessel Density
Statistics
Neuronal Density
All Microvessels
N10 mice
All Cell Density
Capillaries Only
Non-Neuronal Density
67Cell Population Count andVessel Density
Statistics
Neuronal Density
All Microvessels
N10 mice
All Cell Density
Capillaries Only
Non-Neuronal Density
68Cell Population Count andVessel Density
Statistics
Neuronal Density
All Microvessels
N10 mice
All Cell Density
Capillaries Only
Non-Neuronal Density
69Cell Population Count andVessel Density
Statistics
70Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
71Neuroanatomy Result(not my expertise! See
coauthor list!)
Median Cell Density (count/mm3 tissue)
72Neuroanatomy Result(not my expertise! See
coauthor list!)
Number of cells related to capillary volume
Median Cell Density (count/mm3 tissue)
73Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Vessel segmentation short cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
74ACTIVE LEARNING
75Traditional treadmill CAD development process (12
step program)
- Take medical image data (xi are medical image
volumes, e.g.) - Label image (cancer/no cancer) via radiologist,
for better diagnosis/treatment/therapy - Realize binary cancer/no cancer labels (yi) could
be generated automatically via segmentation
algorithms - Realize segmentation algorithms are untweakable
to desired accuracy (simultaneous
under/oversegmentation) - Realize with enough training data (x labeled with
y), CAD could predict next label (requires
machine learning/segmentation algorithms) - Label a database of medical images (x,y)
- Train a classifier mapping x?y based on
training data, (x,y) - Estimate detection performance and confidence via
validation (Cross Fold Validation, e.g.) - Realize errors in classifier can be traced to
specific clusters/issues - Derive new features or tweak CAD training to fix
error clusters/issues - Add new class to accumulate confuser classes
(vessel/airway/chest wall, e.g., for lung CAD) - Incremental progress until all classes are
included - New modality developed to deal with difficult
issues in legacy system (design to CAD still less
profitable than design to M.D.)
The last decade
76Main idea in Active Learning Focus labeling
effort on most informative (most difficult)
exemplars
77- Classifier Accuracy Depends on
- Amount of Training Data
- Selection of Informative Examples
100 Accuracy (separable classes)
Classifier Accuracy (area under ROC curve, e.g.)
100
Percent of database instances labeled (distance
to right is proportional to labeling cost, )
Y. Abramson and Y. Freund, Active learning for
visual object detection, UCSD Technical Report.
78Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Short vessel segmentation cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
79Recent Work and Community Needs/Issues
- Recent System Architecture Work NIH NIBIB R01,
Reducing gigabytes of voxels to metrics on tens
of thousands of object relationships - Image formation, dark/gain correction,
registration, tiling, channel subtraction - Blocking large datasets
- Segmentation of cell nuclei
- Classification of nuclei types
- Short vessel segmentation cascade
- Morphometry-restoring anisotropy corrections
- Capillary network isolation
- Vectorization of microvasculature networks
- Cell population count and vessel density
statistics (cortical flattening) - Neuroanatomy resultlarger animals have more
nuclei and blood vessel fractions - Need for machine learning to add value Labeled
datasets - Issues
- Expensive
- Use active learning for truly expensive labeling
efforts - No glory. Academically invisible. Effort/Effect
ratio high. - Embarrassment of riches (i.e. data)
- Great microscopy labs produce microscopes which
produce enormous quantities of data (giant cubes
of numbers), but require concomitant processing - Not immediately publishable biological
theories/experimental results
Message 1 Theres not one algorithm to segment
everything. These algorithms are domain-specific.
Message 2 1) Create labeled databases 2)
you dont have to label all your data to get most
of the learning benefit (as long as you label
wisely).
Message 3 Data explosions are going
around. Budget more image analysis into your
proposal.
80(No Transcript)
81Technical Focus Advance the utility of
high-fluence ultrashort pulses, plus enabling
hardware and software tools, for
discovery-based studies in cortical
neurovasodynamics
Scientific Focus Connection between vascular
topology, flow dynamics, and neuronal control