Colorbased Diagnosis: Clinical Images

1
Color-based Diagnosis Clinical Images

• Research Project Funded In Part by NIH

Yue (Iris) Cheng, Dr. Scott E Umbaugh _at_ Computer
Vision and Image Processing Research
Lab Electrical and Computer Engineering
Department Southern Illinois University
Edwardsville E-mail cheng_at_westar.com https//www.
ee.siue.edu/CVIPtools
2
Overview

• Skin tumors can be either malignant or benign
• Clinically difficult to differentiate the early
  stage of malignant melanoma and benign tumors due
  to the similarity in appearance
• Proper identification and classification of
  malignant melanoma is considered as the top
  priority because of cost function
• Classification of skin tumors using computer
  imaging and pattern recognition
• Previous texture feature algorithm successfully
  differentiate the deadly melanoma and benign
  tumor seborrhea kurtosis
• Relative color feature algorithm is explored in
  this research for differentiate melanoma and
  benign tumors, dysplastic nevi and nevus
• Successfully classify 86 of malignant melanoma
  using relative color features, compared to the
  clinical accuracy by dermatologists in detection
  of melanoma of approximately 75

3
Materials and Tools

• Image database
• Original tumor images
• 512x512 24-bit color images digitized from 35mm
  color photographic slides and photographs
• 160 melanoma, 42 dysplastic, and 80 nevus skin
  tumor images
• Border images
• Binary images drawn manually and reviewed by the
  dermatologist for accuracy
• Software
• CVIPtools
• Computer vision and image processing tools
  developed at our research lab
• Partek
• Statistical analysis tools

4
CVIPtools
5
Method Design

• Creation of relative color images
• Segmentation and morphological filtering
• Relative color feature extraction
• Design of tumor feature space and object feature
  space
• Establishing statistical models from relative
  color features

6
Create Relative Color Skin Tumor Images

• Purpose
• to equalize any variations caused by lighting,
  photography/printing or digitization process
• to equalize variations in normal skin color
  between individuals
• the human visual system works on a relative color
  system
• Algorithm
• Mask out non-skin part in the image to calculate
  the normal skin color
• Separate tumor from the image
• Remove the skin color from the tumor to get a
  relative color skin tumor image
• CVIPtools functions were used to create relative
  color skin tumor images

7
Calculate Skin Color
8
Tumor Image
9
Relative Color Tumor Image
10
Segmentation and Morphological Filtering

• Image segmentation was used to find regions that
  represent objects or meaningful parts of objects
• Morphological filtering was used to reduce the
  number of objects in the segmented image
• Easy to use CVIPtools for experimenting and
  analysis

11
Relative Color Feature Extraction

• Necessary to simplify the raw image data into
  higher level, meaningful information
• Feature vectors are a standard technique for
  classifying objects, where each object is defined
  by a set of attributes in a feature space.
• Totally 17 color features and binary features
  were extracted using CVIPtools
• The three largest objects, based on the binary
  feature area, were used in feature extraction
• Histogram features, that is, color features, were
  extracted in each color band from relative color
  image objects

12
17 Features

• Histogram features in R, G, B bands
• Mean
• Standard deviation
• Skewness
• Energy
• Entropy

• Binary features
• Area
• Thinness

13
17 Features (Cont.)
14
Design Two Feature Spaces

• Tumor feature space
• consists of 277 feature vectors correspond to 277
  skin tumor images.
• each feature vector has 51 feature elements,
  which are the total of 17 features of each three
  largest objects within the same tumor.
• Object feature space
• had 842 feature vectors corresponding to 842
  image objects
• each feature vector has 17 feature elements,
  which were the binary features and color features
  stated as above

15
Establishing Statistical Models

• Two feature spaces serve as two data models in
  order to maximize the possibility of success
• Two classification models, Discriminant Analysis
  and Multi-layer Perceptron, were developed for
  both data models
• The training and test paradigm is used in
  statistical analysis to report unbiased results
  of a particular algorithm
• due to small size of data set, 282 images, we
  used the leave x out method, with both one and
  ten for x
• Partek software was used
• to analyze the data representing the features
• to develop a model or rules for classifying the
  tumors

16
Quadratic Discriminant Analysis

• A statistical pattern recognition technique based
  on Bayesian theory, which classifies data based
  on the distribution of measurement data into
  predefined classes
• Normalization the feature data as preprocessing
• performed to maximize the potential of the
  features to separate classes and satisfy the
  requirement of the modeling tool such as
  Quadratic discriminant analysis for a Bayesian
  distribution of the input data
• Variable selection was used to choose dominant
  features.

17
Multi-Layer Perceptron

• A feed forward neural network
• neural networks modeled after the nervous system
  in biological systems, based on the processing
  element the neuron
• widely used for pattern classification, since
  they learn how to transform a given data into a
  desired output.
• Principal Component Analysis (PCA) as
  preprocessing
• a popular multivariate technique, is to reduce
  dimensionality by extracting the smallest number
  components that account for most of the variation
  in the original multivariate data and to
  summarize the data with little loss of
  information
• the dispersion matrix selected for PCA in this
  project is correlation

18
Multi-Layer Perceptron (Cont.)

• Creation, training and testing of neural
  networks
• Creation a neural network involves selection of
  hidden and output neuron types and a random
  number generation.
• Four output neuron types Softmax, Gaussian,
  Linear and sigmoid
• Three hidden neuron types Sigmoid, Gaussian and
  Linear
• Scaled Conjugate Gradient algorithm is used for
  learning in this project.
• Automated and independent of user parameters
• Avoids time consuming
• Stopping criteria, sum-squared error, is selected
  to determine after how many iterations the
  training should be stopped
• The trained data is then tested on itself first
  to examine how far the neural network is able to
  classify the objects correctly.
• Leave x partition out method is used for testing
  the algorithm

19
Experiments and Analysis in Tumor Feature Space

• Discriminant Analysis
• 24 features selected for leave ten out method

• 10 features selected for leave one out method

20
Experiments and Analysis in Tumor Feature Space
(Cont.)

• Discriminant Analysis (Cont.)

21
Experiments and Analysis in Tumor Feature Space
(Cont.)

• Multi-layer Perceptron

• Best features, being in the first three
  components of the PCA projection data, were used
• Success percentages of melanoma as high as 77
  and nevus is as high as 68

22
Experiments and Analysis in Object Feature Space

• Discriminant Analysis
• 8, 9, 11 and 12 significant features were
  selected respectively for leave one out method

23
Experiments and Analysis in Object Feature Space
(Cont.)

• Discriminant Analysis (Cont.)

• Yield consistent results in classifying melanoma
  from other skin tumor with above 80 success rate

24
Experiments and Analysis inObject Feature Space
(Cont.)

• Multi-layer Perceptron (MLP)

• 5 out of 12 hidden-output layer neuron
  combinations gave better classification results
• Leave one out method
• Yield success percentage as high as 86 for
  classifying melanoma.
• MLP is more consistent in classifying melanoma as
  well as nevus

25
Conclusion

• Multi-Layer perceptron (MLP) with feature data
  preprocessed by Principal Component Analysis
  (PCA) gave better classification results for
  melonoma than Discriminant Analysis (DA)
• The best overall successful rate of 78, of which
  percentage correct of melanoma is 86, nevus is
  62 and dysplastic is 56.
• The best classification results are achieved with
  sigmoid used as the hidden and output layer
  neuron type for the MLP with PCA on Object
  Feature Space.
• The three largest tumor objects are
  representative for the whole skin tumor.

26
Conclusion (Cont.)

• However the small percentage of melanoma
  misclassification as well as the relatively low
  success rate for nevus and dysplastic nevi
  suggests that we may not have the complete data
  set for the experiments.
• In order to achieve better classification
  results, future experiments
• Needs more complete skin tumor image database.
• Should combine texture and color methods to get
  better results
• Will include dermoscopy images

27
Acknowledgement

• Dr. Scott E Umbaugh, SIUE
• Mr. Ragavendar Swamisai
• Ms. Subhashini K. Srinivasan
• Ms. Saritha Teegala
• Dr. William V. Stoecker, Dermatologist, UMR

28
Thank You!
Yue (Iris) Cheng Graduate Student _at_ Computer
Vision and Image Processing Research
Lab Electrical and Computer Engineering
Department Southern Illinois University
Edwardsville E-mail cheng_at_westar.com https//www.
ee.siue.edu/CVIPtools