Transcript Document

COMPUER-AIDED CLASSIFICATION OF CELLS IN COMPLEX BRAIN TISSUE FROM 5-CHANNEL 3-D CONFOCAL DATASETS
Bio-Med
Yousef Al-Kofahi1, Gang Lin1, Christopher Bjornsson2, Karen Smith3, Arunachalam Narayanaswamy1, Badrinath Roysam1, William Shain3
1 Electrical,
Computer, and Systems Engineering Department, Rensselaer Polytechnic Institute, Troy, NY 12180-3590
L2
2 Center for Biotechnology and Interdisciplinary Studies, Rensselaer Polytechnic Institute, Troy, NY 12180-3590
3 Laboratory of Nervous System Disorders, Wadsworth Center, Albany, NY 12201-0509.
Abstract
Spectral Unmixing and Image Segmentation
Segmentation
Output
Class 4:
Endothelials
• Initial Clustering
 Fuzzy c-means clustering is used to divide the cells into 3 clusters
Trace
Processes
 Each cell ( j ) has a degree of membership to each
cluster:
3
2
77
3
3
8
9.4%
Neurons
(341 found)
28
0
310
10
38
11.1%
Endothelials (245 found)
49
2
9
195
60
24.5%
SUM
139
13.6%
Summary cell classification results after 5 manual corrections to the
training set (using all the cells)
0
6
22
28
8.1%
Astrocytes (324 found)
318
0
0
6
6
1.9%
Machine
Classification
(85 found)
2
77
3
3
8
9.4%
Microglia (73 found)
0
73
0
0
0
0.0%
Neurons
(341 found)
23
0
310
8
31
9.1%
Neurons
(294 found)
16
0
278
0
16
5.4%
Endothelials (245 found)
44
1
5
195
50
20.0%
Endothelials (197 found)
19
0
2
176
21
10.7%
SUM
117
11.5%
SUM
43
4.8%
j
The training set contained 289 cells
326
Human
Microglia
d  (max(d )  median(d )) / 2
j
% of Errors
Astrocytes (348 found)
Total Errors
Machine
Classification
Summary cell classification results after 5 manual corrections to the
training set (correctly segmented cells)
Endothelials
Human
Number of Manual corrections to the training set vs. the
number of classification errors. The number of errors dropped
from 139 to 117 after 5 corrections.
Neurons
 Cells close to the centroid of the cluster tend to have higher degrees of membership
 The training set for class j contains all the cells with:
(85 found)
Microglia
Processes Tracing
Microglia
Astrocytes
Surface Segmentation
• Training Set Extraction
 Most of the clustering errors are expecteds to be at the boundaries of the clusters
9.5%
% of Errors
 Each cell ( j ) is initially assigned to the cluster with the highest degree of
j
membership dmax
33
Total Errors
i 1
25
Endothelials
Convergence
Analysis
j
d
 i 1
8
Neurons
Trace
Processes
di j  [0,1] ,
Some Examples:
Nuclear Segmentation
Class 3:
Astrocytes
0
Microglia
GFAP
Channel
Class 2:
Microglia
326
Astrocytes
5-channel 3-D image of 100 um thick rat
hippocampal slice: blue:nuclei, purple: Nissl
green: vasculature, red: astrocytes processes
yellow: microglia.
Class 1:
Neurons
% of Errors
Inspect &
Edit
Total Errors
Nuclear
Segmentation
Astrocytes (348 found)
Endothelials
Final k-means clustering
of Astrocytes and Endothelials
Nissl
Channel
Iba-1
Channel
•
Surface
Segmentation
Machine
Classification
Neurons
Spectral
Unmixing
SVM Classification
Microglia
CyQuant
Channel
Summary cell classification results using all the cells
Human
/
R3
• Classification Results
Astrocytes
EBA
Channel
Class 3:
Astrocytes + Endothelials
Training Set Extraction
SVM Training
R2
 The classification results were validated against a human expert validation
 The validation process starts with validating the nuclear segmentation
 Some of the segmentations errors include over-/under-segmented nuclei and nuclei near
the image edge
 We used a dataset of 1019 cells. Among them, 888 are correctly segmented.
 Our classifier is applied on all the cells.
 Validation considered all the cells at first, and then just the correctly segmented ones
(Optional Step): Manual Correction of Training Set Errors
 The image is decomposed into 5 channels
 Each channel is segmented independently
Validating
TestBEDs
•The Validation Process
Initial Fuzzy c-means clustering
Class 2:
Microglia
S5
Experimental Results and Validation
Nuclei
Feature Vectors
Class 1:
Neurons
S4
Fundamental
L1 Science
R1
Cells Classification
The cellular organization of brain tissue is truly complex. This work
presents a computational method to identify the principal cell types in threedimensional (3-D) confocal image stacks with multiple fluorescent
channels. The cells are classified into four major classes (Neurons,
Microglia, Astrocytes and Endothelials) by using a two-step classifier that
applies fuzzy c-means clustering followed by Support Vector Machines
(SVM). The resulting classification results were validated against a human
expert, and the accuracy of the classifier was %95.5 in the correctly
segmented nuclei.
S2 S3
S1
L3
EnviroCivil
A
C
B
• SVM Training and classification
A multi-category SVM classifier is first trained using the training set extracted in
the previous step
 The support vectors are then used to classify the unlabeled cells
 The SVM’s solution approximates the Bayes decision rules without estimating
conditional probabilities.
Features Extraction
•
Intrinsic Features:
 Intrinsic features include a set of morphometric, topological, and intensitybased features for each nucleus.
 We have a total of 23 intrinsic features
• Optional Manual corrections of Training Set Errors
 We implemented an interactive scheme by incorporating the user’s feedback into
our classifier.
 An expert user can correct the some of all of the errors in the training set to
provide better classification results.
• Associative Features:
 Associative features are computed
for each nucleus to quantify
relationships with other channels.
 We have a total of 8 associative
features
 Examples:
A) Neuronal nuclei are surrounded by
Nissl signal
B) Microglial nuclei are surrounded by
IBA-1 signal
C) Astrocyte nuclei are proximal to a
point of convergence of processes
D) Endothelial cells are adjacent vessels
D
Color Code:
Astrocytes
• Final Clustering of Astrocytes and Endothelials
AA
 Each cell is represented by 31 features, but only 28 features were used in the
classification done so far.
 In the final clustering step, we used 8 features only, including the 3 unused features
 In this step, k-means clustering is used.
B
Microglia
Neurons
Endothelials
Blood Vessels
References
C
C
• The Feature Vector:
 Each Nucleus is represented by a feature vector
 We have a total of 31 intrinsic and Associative features
D
D
[1] Bezdek J., Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum, NY,
1981.
[2] R.-E. Fan, P.-H. Chen, and C.-J. Lin. Working set selection using the second order
information for training SVM. Journal of Machine Learning Research 6, 1889-1918, 2005
ACKNOWLEDGEMENTS
Supported in part by NIBIB, EB-000359, NINDS, NS-044287, & NSF, EEC-9986821 is
gratefully acknowledged.
Classification Output:
(A) Automatic classification
(B) Manual Classification
(C) Classification Errors
(D) Composite 3D rendering