7/8/2015 Machine Intelligence Unit, ISI 1
Soft Computing: Excursion
through Fuzzy Sets, Neural
Networks, Genetic Algorithms,
and Rough Sets
SUSHMITA MITRA
Machine Intelligence Unit
Indian Statistical Institute
Email: [email protected]
7/8/2015 Machine Intelligence Unit, ISI 2
Contents
Pattern Recognition
Soft Computing
Fuzzy Sets
Artificial Neural Networks
Genetic Algorithms
Rough Sets
Role in Biomedical Image Analysis
Radiomics & Radiogenomics
7/8/2015 Machine Intelligence Unit, ISI 3
Basics of Pattern Recognition
Classification: Supervised learning
Clustering: Unsupervised learning
Feature selection/ Feature extraction
7/8/2015 Machine Intelligence Unit, ISI 4
Classification: predicts categorical class labels
classifies data (constructs a model) based on the training set and the values (class labels) in a classifying attribute and uses it in classifying new data
Prediction: models continuous-valued functions, i.e., predicts unknown
or missing values
Typical Applications credit approval
target marketing
medical diagnosis
treatment effectiveness analysis
Classification vs. Prediction
7/8/2015 Machine Intelligence Unit, ISI 5
Supervised vs. Unsupervised Learning
Supervised learning (classification)
Supervision: The training data (observations,
measurements, etc.) are accompanied by labels indicating
the class of the observations
New data is classified based on the training set
Unsupervised learning (clustering)
The class labels of training data is unknown
Given a set of measurements, observations, etc. with the
aim of establishing the existence of classes or clusters in
the data
7/8/2015 Machine Intelligence Unit, ISI 6
M : Height, Weight, Complexion, Diet….
…..
….……
xxxxxxxxxxx
xxxx B
P
F:
Weight
Height
D : Straight Line
D Classifier Design
7/8/2015 Machine Intelligence Unit, ISI 7
Classification methods
Goal: Predict class Ci = f(x1, x2, .. Xn)
Regression: (linear or any other polynomial)
a*x1 + b*x2 + c = Ci.
Nearest neighour
Decision tree classifier: divide decision space
into piecewise constant regions.
Probabilistic/generative models
Neural networks: partition by non-linear
boundaries
7/8/2015 Machine Intelligence Unit, ISI 9
Pictorially…
Inter-cluster distances are maximized
Intra-cluster distances are
minimized
7/8/2015 Machine Intelligence Unit, ISI 10
What Is Good Clustering?
A good clustering method will produce high quality
clusters with
high intra-class similarity
low inter-class similarity
The quality of a clustering result depends on both
the similarity measure used by the method and its
implementation.
The quality of a clustering method is also measured
by its ability to discover some or all of the hidden
patterns.
7/8/2015 Machine Intelligence Unit, ISI 11
Two different K-means Clusterings
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
0
0.5
1
1.5
2
2.5
3
x
y
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
0
0.5
1
1.5
2
2.5
3
x
y
Sub-optimal Clustering
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
0
0.5
1
1.5
2
2.5
3
x
y
Optimal Clustering
Original Points
Soft Computing
7/8/2015 Machine Intelligence Unit, ISI 13
Soft Computing
Consortium of methodologies that works
synergistically and provides flexible information
processing
Exploits tolerance for imprecision, uncertainty,
approximate reasoning and partial truth to
achieve tractability, low cost solutions and close
resemblance with human-like decision making
Provides acceptable solution at low cost(car parking example)
7/8/2015 Machine Intelligence Unit, ISI 14
Main Components
Fuzzy Sets
Artificial Neural Networks
Genetic Algorithms
Rough Sets
Fuzzy Sets
Role of Fuzzy Sets (FS)
Modeling of imprecise/qualitative knowledge
Transmission and handling of uncertainties at
various stages
Supporting, to an extent, human type
reasoning in natural form
Understandability of patterns
Providing approximate solution faster
Precision vs. Imprecision
Lines 2-3 m long
Convex areas
People 7 ft. tall
Full membership
Crisp sets: {0,1}
Crisp models
Probability
Fairly short lines
Almost convex areas
People about 7 ft. tall
Partial membership
Fuzzy sets: [0,1]
Fuzzy models
Membership
Probabilities vs. Possibilities
7/8/2015 Machine Intelligence Unit, ISI 18
When a tiger is about to pounce, only one thing is certain:
If you take time to compute probabilities, you'll be eaten.
but if you decide in a split second whether to climb
(i) the tree to the right or
(ii) the one ahead,
Then you have used possibilities
(as in measures - one uses the min/max operations).
0.0
0.5
1.0
x
x
0.0
0.5
1.0
x
x
S – membership function
1.0
0.5
0.0
x
1.0
0.5
0.0
x
1.0
0.5
0.0
x
- membership function
SINGULAR VS. GRANULAR VALUES
7.3% high
102.5 very high
160/80 high
Singular Granular
unemployment
temperature
blood pressure
GRANULATION OF A VARIABLE
continuous quantized granulated
Example: Age
quantized Age0
1
µ1
0
young
middle
-aged old
Age
µ
granulated
7/8/2015 Machine Intelligence Unit, ISI 23
Artificial Neural Networks
7/8/2015 Machine Intelligence Unit, ISI 24
Background
Analogous to Biological Nervous System
7/8/2015 Machine Intelligence Unit, ISI 25
Salient Features
Biological neuron: Impulse triggered by cell, travels along axon – releases chemical transmitter raising/lowering electric potential – if greater than threshold, causes a pulse
Neural network: Not programmed, learn by examples; non-parametric
More efficient than powerful digital computers (number-crunchers) in cognitive tasks
7/8/2015 Machine Intelligence Unit, ISI 26
Primary task of all biological neural systems is to control various
functions (mainly behavioral).
Human being can do it almost instantaneously and without much
effort. e.g., recognizing a scene or music immediately.
Artificial Neural Network (ANN) or Neural Network (NN)
models try to simulate the biological neural network with electronic circuitry.
Also known as Connectionists Model/ Parallel Distributed
Processing (PDP).
Purpose : To achieve human-like performance particularly in
cognitive tasks like pattern recognition & image processing.
7/8/2015 Machine Intelligence Unit, ISI 27
Definition : Massively parallel interconnected network of
simple (usually adaptive) processing elements that interact with
objects of the real world in a manner similar to biological systems.
NN models are extreme simplifications of human neural
systems.
Computational elements (neurons/nodes/ processors) are
analogous to that of the fundamental constituents (neurons) of
the biological nervous system.
Definition
7/8/2015 Machine Intelligence Unit, ISI 28
Comparison
Von Neumann machines: single, fast, powerful CPU; error in 1 bit affects result; memory
Parallel processing architecture: multiple, complex CPUs; shared memory
ANN: very simple, slow processing elements; massively parallel interconnections; graceful degradation (fault tolerance); no memory –information encoded redundantly, distributed among connection weights
7/8/2015 Machine Intelligence Unit, ISI 29
Brief History
Rosenblatt (’58-’61): Perceptron --
Linear separability
Minsky & Papert (’68): XOR problem
Rumelhart, Hinton, Williams (’86):
Multilayer Perceptron (MLP) with
backpropagation of error
7/8/2015 Machine Intelligence Unit, ISI 30
Processing units
Receives input from connected neurons, compute an output
value and sends it to other connected neurons.
Three types of units - input, output, hidden.
Output value - oi(t)=f(Ii(t))
Total input for ith neuron is Ii.
f is a threshold or squashing function.
Unidirectional connections (wij)
wij < 0 unit uj inhibits unit ui.
wij = 0 unit uj has no direct effect on unit ui.
wij > 0 unit uj excites unit ui.
General framework of neural networks
7/8/2015 Machine Intelligence Unit, ISI 31
Learn from example - shown a set of inputs, they self-
adjust to produce consistent response.
Generalize from previous examples to new ones - once
trained, a network's response is mostly insensitive to
variations in input.
Cross validation and Over learning;
Abstract essential characteristics from inputs - find the
ideals (prototype) from imperfect inputs.
Feedforward (no loops) /Feedback (recurrent)
Characteristics of neural networks
7/8/2015 Machine Intelligence Unit, ISI 32
Types of Learning
Three learning paradigms :
Supervised (with desired output during training)
Unsupervised (without desired output during training)
Reinforcement ( “reward” -- reinforcement signal for correct output)
Various learning rules :
Error correction
Boltzmann learning
Hebbian learning
Competitive learning
7/8/2015 Machine Intelligence Unit, ISI 33
Benefits of Neural Networks
Nonlinearity : To model a physical phenomenon which is in general nonlinear
Adaptivity : To cope with change in environment
Massive parallelism : To enable faster computation
Robustness : Ability to handle missing, confusing and/or noisy data
Fault tolerance : Ability to work, at least to some extent, even if in component failure
Input-output mapping : Ability to model an arbitrary input-output mapping
7/8/2015 Machine Intelligence Unit, ISI 34
Some common feature are there; but differ in finer details.
Multi-layer perceptron (hetero associator/supervised classifier)
Hopfield's model of associative memory
Kohonen's model of self-organizing neural network
(regularity detector/ unsupervised classifier)
Radial basis function network (supervised)
Adaptive resonance theory (regularity detector) -- ART
Cellular neural network
Neocognitron
Popularly used NN models
7/8/2015 Machine Intelligence Unit, ISI 35
Fields of applications PRclassification, clustering, feature
evaluation,
image preprocessing,
character recognition,
speech recognition,
retrieval by content,
function approximation, optimization, prediction
data mining,
expert system design/rule generation,
financial forecasting,
bioinformatics.
7/8/2015 Machine Intelligence Unit, ISI 36
Perceptron
Single layer perceptron: A single layer of neurons
connected by weights to a set of inputs.
THRESHOLD
OUTx2
x1
xn
w2
w1
wn
Let x1,x2,…,xn be the set of inputs
and w1,w2,…,wn be the weights.
If wixi > then the output is 1
else 0, where = threshold
wixi = separating linex1
x2
7/8/2015 Machine Intelligence Unit, ISI 37
x2
x1
xn
OUT1
OUT2
OUTn
Multiple decision linesx2
x1
7/8/2015 Machine Intelligence Unit, ISI 38
Learning rule
Learning: Present a set of input patterns, adjust the weights
until the desired output occurs for each of them.
wi(t+1) = wi(t) + i;
i = xi;
= T – A (i.e., target – actual).
If the sets of patterns are linearly separable, the single
layer perceptron algorithm is guaranteed to find a separating
hyperplane in a finite number of steps.
Reward if correct, punish if incorrect by updating
7/8/2015 Machine Intelligence Unit, ISI 39
Overcoming constraint of linear separability
Cascading layersTwo layers
Generates convex decision regions
S1
S2
X
Y LAYER 2 NEURON
IS 1 ONLY IN
THIS REGION
w11
w12w21
w22
S1
S2
LAYER 1
LAYER 2X
Y
7/8/2015 Machine Intelligence Unit, ISI 40
Three layers
Decision regions of any shape
TRIANGLE
ATRIANGLE
B
NON –CONVEX REGION A AND NOT B
LAYER 1 LAYER 2
LAYER 3
Y
X1
X2
X3
Multi-layer networks
7/8/2015 Machine Intelligence Unit, ISI 41
Multilayer perceptron
OUTPUT LAYER
HIDDEN LAYER
INPUT LAYER
INPUT PATTERN
OUTPUT PATTERN
Wkj
Wji
k
j
i
7/8/2015 Machine Intelligence Unit, ISI 42
Nodes of two different consecutive layers are
connected by links or weights.
There is no connection among the elements of the same
layer.
The total input (Ii) to the ith unit
Ii =
oj is the output of the jth neuron.
j
jijow
7/8/2015 Machine Intelligence Unit, ISI 43
The output of a node i is obtained as
oi = f(Ii), f is the activation function.
Mostly the activation function is sigmoidal/squashing,
with the form,
f(x) = 1/(1+e-(x-)/0).
f(x)
0 2-2-4 4
0.5
1.0
x
Initially very small random values assigned to weights.
7/8/2015 Machine Intelligence Unit, ISI 44
Parameter updating
For learning (training) we present the input pattern X={xi}, and ask
the net to adjust its set of weights/biases in the connecting links such
that the desired output T={ti} is obtained at the output layer.
Then another pair of X and T is presented for learning.
Learning tries to find a simple set of weights and biases that will be
able to discriminate among all the input/output pairs presented to it.
The output {oi} will not be the same as the target {ti}.
Error is, E = 2
For learning the correct set of weights error is E is reduced as
rapidly as possible.
Use gradient descent technique.
i
ii ot )(2
1
7/8/2015 Machine Intelligence Unit, ISI 45
Neuro-Fuzzy Hybridization
7/8/2015 Machine Intelligence Unit, ISI 46
Neuro-Fuzzy Hybridization
ANN emulates
architecture and
information
representation of
human brain
FS mimics human
reasoning and
uncertainty handling
FNN: ANN equipped
with capability of
handling fuzzy
information
NFS: Fuzzy system
augmemented by
ANN to enhance its
flexibility, speed and
adaptibility
7/8/2015 Machine Intelligence Unit, ISI 47
Generic & Application-specific merits
ANN: massive
parallelism, robustness,
learning
FL: modeling imprecise
qualitative knowledge,
and transmission of
uncertainty
ANN: generating
highly nonlinear
decision regions
FL: handling
uncertainty in input
description and
output decision
7/8/2015 Machine Intelligence Unit, ISI 48
(for More Intelligent System)
Neuro-Fuzzy Computing
7/8/2015 Machine Intelligence Unit, ISI 49
Integration Schemes
Fuzziness at input and/or output levels
ANN implements fuzzy decision system
Fuzziness at neuronal level
Measures of fuzziness used as error or
energy function
Genetic Algorithms
Role of GAs
Adaptive, robust, efficient,
global search methods
suitable in large search
space.
Select a model based on
optimizing some preference
criterion/objective function.
Useful in regression.
Role of GAs
Adaptive, robust, efficient,
global search methods
suitable in large search
space.
Select a model based on
optimizing some preference
criterion/objective function.
Useful in regression.
GAs are adaptive and robust computational procedures modeled on the
mechanics of natural genetic systems. They act as biological metaphor
and try to emulate some of the processes observed in natural evolution.
Natural evolution operates on encoding of biological entities in the
form of a collection of genes called a chromosome. Similarly, GAs
operate on string representation of possible solutions (individuals/
chromosomes) containing the features.
Selection : obeys Darwinian survival of the fittest strategy
Nature acts as environment. Objective function plays the role of
environment.
Variation is introduced mainly through genetic operations like
recombination (crossover) and mutation.
What are genetic algorithms?
Population of individuals
Termination conditions
Replacement technique
Probabilities to perform genetic operations
Genetic operators (recombination/crossover, mutation)
Selection procedure
Objective function & associated fitness evaluation criterion
Encoding/decoding (of individuals) mechanism
Components of a GA
Coding
Converts parameter values into chromosomal representation
For the continuous valued parameters decimal to the binary conversion
used.
For example 13 == 01101 (for 5 bit representation )
For a parameter having categorical values a particular bit position in the
chromosomal representation is set to 1 if it comes from that category.
For example the parameter marital status can have values from {married,
unmarried, divorced, widow}. So, unmarried == 0100 widow==0001
These strings (representing the parameters of a problem) are concatenated
to form a chromosome.
Encoding/decoding mechanism
It is the reverse of encoding.
For continuous valued parameter the binary representation is converted to
continuous value by the following formula
01101==40+(13/31)*(60-40)=48.387
For categorical valued parameters the value is found by consulting the
range of the parameter.
0001== widow
0100== unmarried
i=0
bits used-1
biti*2i
2(bits used)-1Lower bound + *(Upper bound - Lower
bound)
Decoding
A measure of chromosome's performance. More suitable
strings should get high fitness values.
Evaluation and selection
Selection gives more chance to better fitted individuals
(Mimics natural selection procedure)
Popular selection techniques
Roulette wheel selection
Stochastic universal sampling
Linear normalization selection
Tournament selection
Roulette wheel selection
Sum the fitness of all the chromosomes of the population.
Call it total-fitness.
Generate a random number n in [0, total-fitness]
Return the first chromosome whose fitness when added to
the fitness of the preceding population members is
greater than equal to n.
Example:
Let there be five chromosomes with fitness 20, 10, 40, 3, 18
Then total-fitness=91.
Say, the random number drawn (n) is 45.
Select the 3rd chromosome (since 20+10+40 > 45).
Choose mating pairs (from the selected chromosomes).
Check (using pc) whether this pair should go for
recombination or not. If yes, interchange chromosome
segments. one point, two point, multi point, uniform,...
Recombination/crossover
One point crossover
parent1: xy xy x y xy child1: xy xy x b ab
parent2: ab ab a b ab child2: ab ab a y xy
before
crossoverafter crossover
Recombination (contd.)
Two point crossover
parent1: xy xy x y xy child1: xy ab a y xy
parent2: ab ab a b ab child2: ab xy x b ab
before
crossoverafter
crossover
Introduces diversity.
Bit mutation.
Check (using pm) whether this bit should be
mutated or not. If yes, flip the bit.
001000 000000
Probabilities to perform genetic operations
May be fixed or made variable.
pc : 0.6 to 0.9 pm : 0.001 to 0.01
Mutation
Execute for a fixed number of generations/iterations.
Until a string with a certain fitness value is located.
Until the population attains a certain degree of
homogeneity (most of the individuals become similar).
Elitism (optional)
Keeps track of /store the best solution obtained so far.
Terminating criterion
Applications
Numerical, combinatorial and constrained optimization
Scheduling, TSP, Graph Coloring
Automatic programming - evolves computer programs for
specific tasks (Genetic Programming)
Pattern recognition - classification, clustering, prediction
Image processing --- segmentation, enhancement
Data mining --- rule mining, clustering
Immune systems - discovery of multigene, multiagent
systems
Economics - financial prediction, bidding strategies
Rough Sets
Rough Sets (RS)
Major mathematical tool for managing uncertainty
that arises from granularity in the domain of
discourse, i.e., from the indiscernibility between
objects in a set;
proved to be useful in a variety of KDD
processes;
offers mathematical tools to discover hidden
patterns in data (relevant to DM).
Used for knowledge extraction and encoding of
knowledge-based networks.
Principles of RS
Discovers redundancies and dependenciesbetween the features of a problem to be classified.
Approximates a given concept from below and from above, using lower and upper approximations.
These approximations are defined by the angle from which the set is viewed, i.e. through the information provided by the selected attributes (features).
Rough sets can be used to extract a set of rules in IF-THEN form, from crude domain knowledge, using decision table.
Rough Sets
. x
Upper
Approximation BX
Set X
Lower
Approximation BX
[x]B (Granules)
[x]B = set of all points belonging to the same granule as of the point xin feature space WB.
[x]B is the set of all points which are indiscernible with point x
in terms of feature subset B.
UB W
Knowledge extraction using RS
Create a knowledge base, classifying
objects and attributes within decision
tables.
Remove some undesirable attributes
(knowledge discovery).
Analyze data dependency in the reduced
database to find the minimal subset of
attributes called reduct.
Rough Clusters
A pattern can belong to at most one lower
approximation
A member in a lower approximation is also a member
in the corresponding upper approximation
If a pattern does not belong to any lower
approximation, then it must belong to two or more
upper approximations
Indian Statistical Institute, Kolkata 69
Basic Idea
Indian Statistical Institute, Kolkata 70
dmin
d1
IF d1-dmin < threshold
THEN upper approximation of U2 and U1
ELSE lower and upper approximation of U2
v2
v1 Cluster Center
Pattern
The objects (or patterns) are assigned to a lower or an upper
approximation of a cluster as follows.
Fuzzy vs Rough
Fuzzy sets: Uncertainty in overlapping
regions modeled by membership values,
based on similarity relations.
Rough sets: Uncertainty due to granularity
from indiscernibility modeled by lower and
upper approximations. Ambiguity due to
absence of complete information.
Indian Statistical Institute, Kolkata 71
Role in Biomedical Image
Analysis
7/8/2015 Machine Intelligence Unit, ISI 72
Natural computing, through nature-inspired
strategies, plays a major role in intelligent
decision-making systems – with reference to
segmentation, classification, feature
extraction & selection.
Radiographic imaging modalities – like CT,
MRI, PET – help in providing improved
diagnosis, prognosis and treatment planning
for improved tumor management.
7/8/2015 Machine Intelligence Unit, ISI 73
Imperfect knowledge in medicine
Incomplete understanding of biological
mechanisms
Imprecise test measurements
Uncertainty of normal ranges for test results
Simultaneous presence of more than one
disease condition
Missing information
7/8/2015 Machine Intelligence Unit, ISI 74
Quantitative imaging
Computed tomography (CT) scans produce
tomographic (cross-sectional) slices, horizontally
& vertically, and are more detailed than X-rays.
Positron emission tomography (PET) detects
functional changes in a tissue (by tracing
positron-emitting radioactive tracer FDG – with
cancer cells using glucose faster) before related
anatomical changes become visible.
Magnetic resonance imaging (MRI) uses strong
magnetic field to make body tissues emit their
own radio waves of differing intensities.
7/8/2015 Machine Intelligence Unit, ISI 75
CT scan image for non small cell lung cancer
7/8/2015 Machine Intelligence Unit, ISI 76
Rough c-Means: Basic Idea
Indian Statistical Institute, Kolkata 77
dmin
d1
IF d1-dmin < threshold
THEN upper approximation of U2 and U1
ELSE lower and upper approximation of U2
v2
v1 Cluster Center
Pattern
The objects (or patterns) are assigned to a lower or an upper
approximation of a cluster as follows.
Rough-Fuzzy:
incorporate
membership µ
(maximum)
Rough-Fuzzy c-MeansMitra S, Banka H, Pedrycz W, IEEE Trans. SMC, Part B, vol. 36, pp. 795-805, 2006
Assign each pattern to the lower approximation or upper approximations of cluster pairs by computing fuzzy membership difference from cluster centroid pairs. Sort in descending form
If the difference is less than threshold
then assign to both upper approximation pairs
else assign it to the ith lower approximation such thatis maximum over c clusters
Compute new mean
for each cluster and
repeat until
convergence
Indian Statistical Institute, Kolkata 78
kX iUB
ji UBUB , ji UU ,
kX
iku
otherwiseu
Xu
UBUBUBifu
Xu
UBUBUBifu
Xu
wu
Xu
w
ik
ik
iik
iik
iik
iik
ik
ik
UBX
m
ik
UBX
k
m
ik
iii
UBUBX
m
ik
UBUBX
k
m
ik
iii
UBUBX
m
ik
UBUBX
k
m
ik
up
UBX
m
ik
UBX
k
m
ik
low
i
,
,
)(
)(
)(
)(
ApplicationS. Mitra and B. Barman, LNAI Vol. 5009, pp. 300-307, 2008.
Novel Application of Rough-Fuzzy (RF)Clustering (For Synthetic as well as CT scanimages of the brain)
RF Clustering simultaneously handles overlapof clusters (Fuzzy) and uncertainty involved inclass boundary (Rough)
Number of clusters was optimized via clustervalidity indices
Main objective was the diagnosis of the extentof brain infarction in CT scan images
Indian Statistical Institute, Kolkata 79
CT Scan Image Segmentation
Segmentation – Process of partitioning animage into some non-overlappingmeaningful regions
Segmentation here via Pixel Clustering
Study consists of cases of VascularInfarction of the Human Brain
Partitioning into five regions – Gray matter(GM), White matter (WM), Infarctedregion, Skull and the backround
Indian Statistical Institute, Kolkata 80
Fresh case of Vascular Insult (Original Image)
Indian Statistical Institute, Kolkata 81
1. Infarction is on the
left side.
2. The left side is
compressing the right
side
3. Dilation of the blood
ventricles
4. Severe edema
5. Division of brain into
gray matter, white
matter and the
cerebrospinal fluid
(CSF)
6. The third ventricle is
not visible here due
to severe edema
from the right
ventricle side
7. Cause: Cholesterol
Deposit, Blockage
Segmentation Result (RFCM)
White Matter
Gray Matter
Cerebrospinal Fluid
Infarcted Region
Comparative Analysis
(HCM, FCM, RCM & RFCM)
Indian Statistical Institute, Kolkata 83
RFCM: Better segmentation
of White matter, Gray matter
and CSF.
HCM (Noisy)FCM (Noisy)
RCM
(Noisy)
Skull
Gray Matter
White Matter
CerebroSpinal Fluid
Chronic Infarction (Original Image)
Indian Statistical Institute, Kolkata 84
1. Patient suffering
from vascular insult
2. Right and left
should have been
symmetric (the most
definite metric for
comparison)
3. Right side is dark
because it has not
received blood
supply for a very
long time
4. Due to this the
blood ventricles
have dilated and
have undergone
liquefaction (water)
5. Parenchyma is
infarcted
6. Arteries were
blocked due to high
cholesterol levels
7. Happens due to
normal old age
Chronic Infarction (RFCM Segmentation
result)
Indian Statistical Institute, Kolkata 85
White Matter
Gray Matter
Cerebrospinal Fluid
Infarcted Region
Skull
Subtle Case of Infarction (Original Image)
Indian Statistical Institute, Kolkata 86
1. The third ventricle
has dilated
2. Edema from below
3. Blockage of arteries,
no blood supply from
a long time
4. Dilation of left and
right ventricles due to
this as passage from
below is blocked
5. Problem modeling
same. although the
infarction here is
petty difficult to
locate
6. Tough problem of
segmentation for
infarction
7. Cause: Cholesterol
deposit, Blockage
Subtle Case of Infarction (RFCM
Segmentation result)
Indian Statistical Institute, Kolkata 87
Cerebrospinal Fluid
White Matter Uniform
merging of Gray Matter and the Infarcted region
Remarks
In the absence of an accurate index to test theaccuracy of segmentation results in CT scanimagery, we resorted to expert domain knowledge
36 frames of each case of infarction was studiedand results verified by an experienced radiologist
RFCM produced the best result as verified by expertradiologist
Results promise to provide a helpful second opinionto radiologists in case of Computer-AidedDiagnostic (CAD)
Indian Statistical Institute, Kolkata 88
Radiomics
High throughput extraction and analysis of advanced quantitative imaging features from medical images.
In cancer, several studies have reported the statistical association between the advanced imaging features extracted from the tumor regions of the routine medical images and the tumor stage, metabolism, patient survival and underlying gene expression patterns.
To ensure the reliability of quantitative imaging features, accurate and robust tumor delineation is essential. Tumor segmentation is one of the main challenges of Radiomics, as manual delineation is prone to high inter-observer variability.
It has been shown that semiautomatic segmentation approaches efficiently reduce inter-observer variability as compared to the time consuming manual delineations.
Parmar C, Rios Velazquez E, Leijenaar R, Jermoumi M, Carvalho, S., Mak, R. H., Mitra, S.,Uma Shankar, B., Kikinis, R., Haibe-Kains, B., Lambin, P., Aerts, H. J. W. L. (2014) Robust Radiomics feature quantification using semiautomatic volumetric segmentation. PLoS ONE 9(7): e102107. doi:10.1371/journal.pone.0102107
http://www.plosone.org/article/info:doi/10.1371/journal.pone.0102107
Objective Parmar C, Rios Velazquez E, Leijenaar R, Jermoumi M, Carvalho, S., Mak, R. H., Mitra, S., Uma
Shankar, B., Kikinis, R., Haibe-Kains, B., Lambin, P., Aerts, H. J. W. L. (2014) Robust Radiomics
feature quantification using semiautomatic volumetric segmentation. PLoS ONE 9(7): e102107.
doi:10.1371/journal.pone.0102107
http://www.plosone.org/article/info:doi/10.1371/journal.pone.0102107
In this study, a semiautomatic volumetric segmentation algorithm, available in the
free and publicly available 3D-Slicer platform, was investigated in terms of its
robustness for Radiomics feature quantification.
We extracted fifty-six CT 3D-Radiomic features from 3D-Slicer segmentations made
by three independent observers, twice, and compared them to the features extracted
from manual delineations provided by five independent physicians.
The assessed fifty-six 3D-radiomic features quantified I) tumor intensity, II) tumor
shape, and III) tumor texture. Intra-class correlation coefficient was used to quantify
the reproducibility of imaging features across observers.
It was observed that quantitative imaging features extracted from semi-automatically
segmented tumors have lower variability and are more robust as compared to
features extracted from manual tumor delineations.
Figure 1. Schematic diagram depicting the overview of the analysis.
Parmar C, Rios Velazquez E, Leijenaar R, Jermoumi M, et al. (2014) Robust Radiomics Feature Quantification Using Semiautomatic
Volumetric Segmentation. PLoS ONE 9(7): e102107. doi:10.1371/journal.pone.0102107
http://www.plosone.org/article/info:doi/10.1371/journal.pone.0102107
Semiautomatic segmentation in 3D slicer
GrowCut algorithm implemented in 3D-Slicer was used (www.slicer.org). GrowCut is
an interactive region growing segmentation strategy. Given an initial set of label
points the algorithm automatically segments the remaining image.
For N-class segmentation, the algorithm needs N initial sets of labeled pixels (one set
corresponding to each class) from the user. Based on these, it automatically
generates the region of interest (ROI), which is the convex hull of the user-labeled
pixels with an additional margin.
Next, it iteratively labels all the remaining pixels in the ROI utilizing user-supplied
pixel labels. Individual pixels are labeled by computing a weighted similarity metric of
a pixel with all its neighbors, where the weights correspond to the neighboring pixel's
strength.
The algorithm converges when all the pixels in the ROI have unchanged labels
across several iterations.
3D-Slicer provides a graphical user interface (GUI) as the front end and an efficient
algorithm as the back end for the GrowCut segmentation. After loading the patient
data, the process begins with the user initialization of the foreground and background
by manually marking the area inside and outside the tumor region. Next, the Growcut
automatic competing region-growing algorithm gets activated, and segments the ROI
into foreground and background regions. Thereafter, background and the surrounding
isolated foreground pixels are removed following visual inspection.
Intra-class correlation coefficient (ICC) was calculated in order to quantify the feature
reproducibility. The ICC is a statistical measure, ranging between 0 and 1, indicating
null and perfect reproducibility, respectively.
Since two 3D-Slicer segmentations from each of the three observers were considered
for the analysis, the six 3D-Slicer segmentations were divided in to two sets, each
having three segmentations (one from each observer).
Box-plot comparing intra- and inter-observer reproducibility (ICC) of radiomics features.
Parmar C, Rios Velazquez E, Leijenaar R, Jermoumi M, et al. (2014) Robust Radiomics Feature Quantification Using Semiautomatic
Volumetric Segmentation. PLoS ONE 9(7): e102107. doi:10.1371/journal.pone.0102107
http://www.plosone.org/article/info:doi/10.1371/journal.pone.0102107
Natural Computing in Medical
Image Analysis
S. Mitra and B. Uma Shankar,
``Medical image analysis for cancer
management in natural computing
framework", Information Sciences, Vol.
306, pp. 111-131, 2015.
• Segmentation: Fuzzy sets,
neural nets, rough sets, GA
• Feature extraction: Texture, co-
occurrence matrix, shape,
intensity gradient, statistical,
wavelets, color
• Classification: Neural nets,
fuzzy sets, GA, swarm
intelligence, support vector
machine
7/8/2015 Machine Intelligence Unit, ISI 95
Personalized Medicine
Offers RIGHT drug
To the RIGHT disease
At the RIGHT time
With the RIGHT dosage
7/8/2015 Machine Intelligence Unit, ISI 96
Radiogenomics Radiographic imaging features can serve as molecular
surrogates (substitutes) of gene expression patterns, thereby
contributing towards diagnosis, prognosis, and gene--
associated treatment response of various forms of cancer.
It is observed that tumors having greater genomic
heterogeneity are more likely to develop treatment resistance
and are prone to faster metastasis, thereby resulting in poorer
prognosis.
The genomic heterogeneity, translated into intra-tumoral
heterogeneity, can be potentially captured through medical
imaging in a non-invasive manner.
Such association has the potential for a patient-specific
personalized management of tumors
7/8/2015 Machine Intelligence Unit, ISI 97
Radiogenomics and Personalized MedicineS. Mitra and B. Uma Shankar, ''Integrating radio imaging with gene expressions towards a
personalized management of cancer", IEEE Transactions on Human-Machine Systems, 44, 664-677, 2014.
7/8/2015 Machine Intelligence Unit, ISI 98
Summary
Soft Computing: Fuzzy sets, Artificial neural
networks, genetic algorithms, rough sets
Role in Biomedical Image Analysis
Radiomics
Radiogenomics
Thank You
7/8/2015 Machine Intelligence Unit, ISI 99
7/8/2015 Machine Intelligence Unit, ISI 100
Top Related