Veterinary Thermographic Image Analysis

Gait Analysis Project, Combined Sides

Preliminary Gabor Filter Results

Proposed LIVS Thermographic Image Clinical Application Software User Interface Design

Project Number 7-64878

Report Number 4878-25

Scott E Umbaugh, BSE, MSEE, PhD

Jiyuan Fu, BSEE, MSEE Candidate

Samrat Subedi, BSEE, MSEE Candidate

Computer Vision and Image Processing Laboratory

Southern Illinois University Edwardsville

May 13, 2014

Spring Semester 2014

Submitted to:

Dr. Catherine Loughin

Dr. Dominic Marino, Chief of Staff

Long Island Veterinary Specialists

TABLE OF CONTENTS

Executive Summary…………………………………….………….……...3

Experimental Results……………….…….……………………………..... 4

FEPC Experimental Parameters…………………………………….4

Image Groups…....………………………………………………….5

Result Overview………………………..……..………….…………7

Anterior Stifle Results………………………………….…………...8

Lateral Stifle Results………………………………….……….…...13
Posterior Stifle Results………………………………….…….…...18
Paw Results………………………………………………………...23

New Texture Features UsingGabor Filters ……………………..………. 28

LIVS Thermographic Image Clinical Application Software…………….. 31

Summary and Conclusions ………………………………………………. 33

References…………………………………………………..……………. 34

Executive Summary

Overview

Research and development was continued to investigate the efficacy of feature extraction and pattern classification with thermographic images for the gait analysis project. Experiments were performed with 28 dogs, 14 normal and 14 abnormal. The previous reports contained results from all separate body parts [1,3], and results from the paw experiments [2]. Here we wanted to compare results from combining the left and right side images into one class. The overall results indicate that better success rates are obtained if the left and right side classes are separated. In this report we also include an introduction to the Gabor filters for texture feature extraction and pattern classification, and preliminary results from using them in the gait analysis project (p. 28). An overview of our planned clinical application software user interface is also presented (p. 31).

Materials and Methods for Gait Analysis Experiments

Twenty sets of experiments were performed with the images, five per view. The left and right side were not separated. The views used here include: 1) Anterior, 2) Lateral, 3) Posterior, and 4) Paw. Four color normalization methods [4] were used along with the original RGB images. Three data normalization [4] were used on the feature data, soft-max, standard normal density as well as the original raw feature data. The experiments were performed with the CVIP-FEPC, with each set having either 1023 (original images, without mean) or 2047 (color normalized) permutations. They used Nearest Neighbor (NN) and KNN (K=5)as the classification methods; histogram, spectral and texture features with a pixel distance of six were used. The testing method used was the leave-one-out method.

Best result overview comparison:

Combination / Left / Right
Anterior / 83.93%(47/56) / 82.14%(23/28) / 96.42%(27/28)
Lateral / 83.93%(47/56) / 85.71%(24/28) / 89.29%(25/28)
Posterior / 82.14%(46/56) / 85.71%(24/28) / 96.43%(27/28)
Paw / 81.13%(43/53) / 92.31%(24/26) / 92.59%(25/27)

As can be seen in the table better success rates are obtained by separation of the left and right side.

Experimental Results

FEPC Experiment:

IMAGES and VIEWS (regions)

28 dogs total, 14 abnormal and 14 normal. The abnormal dogs have one abnormal side and one normal side.

Anterior stifle

Lateral stifle

Posterior stifle

Paw stifle

CLASSIFICATION METHOD AND DISTANCE METRIC

K-Nearest Neighbor with K = 5 and NN, distance metric: Euclidean

FEATURES

Histogram features: Mean, Standard deviation, Skew, Energy and Entropy

Texture features: Energy, Inertia, Correlation, Inverse difference, and Entropy. The pixel distance was 6.

New texture functions are used.

Spectral Features was used with Rings = 3 and Sectors = 3.

DATA NORMALIZATION METHOD

Soft-max with r = 1

Standard Normal Density

None

METHOD

Leave One Out

OVERVIEW

20 sets of experiments were performed, including color normalization

5 group anterior stifle images

5 group lateral stifle images

5 group posterior stifle images

5 group paw images

4 sets had 1023 permutations and 16 sets had 2047 permutations.

Anterior stifle Images Group:

For the first experiment (Anterior Stifle, Original for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the second experiment (Anterior Stifle, Lum for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the third experiment (Anterior Stifle, NormGrey for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the fourth experiment (Anterior Stifle, NormRGB for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the fifth experiment (Anterior Stifle, NormRGB-lum for color normalization), 56 images include 42 Normal and 14 Abnormal).

Lateral Stifle Image Group:

For the sixth experiment (Lateral Stifle, Original for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the seventh experiment (Lateral Stifle, Lum for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the eighth experiment (Lateral Stifle, NormGrey for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the ninth experiment (Lateral Stifle, NormRGB for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the tenth experiment (Lateral Stifle, NormRGB-lum for color normalization), 56 images include 42 Normal and 14 Abnormal).

Posterior Stifle Image Group:

For the eleventh experiment (Posterior Stifle, Original for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the twelfth experiment (Posterior Stifle, Lum for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the thirteenth experiment (Posterior Stifle, NormGrey for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the fourteenth experiment (Posterior Stifle, NormRGB for color normalization), 56 images include 42 Normal and 14 Abnormal).

For the fifteenth experiment (Posterior Stifle, NormRGB-lum for color normalization), 56 images include 42 Normal and 14 Abnormal).

Paw Stifle Image Group:

For the sixteenth experiment (PawImage, Original for color normalization), 53 images include 42 Normal and 11 Abnormal).

For the seventeenth experiment (Paw Stifle, Lum for color normalization), 53 images include 42 Normal and 11 Abnormal).

For the thirteenth experiment (PawImage, NormGrey for color normalization), 53 images include 42 Normal and 11 Abnormal).

For the nineteenth experiment (PawImage, NormRGB for color normalization), 53 images include 42 Normal and 11 Abnormal).

For the twentieth experiment (PawImage, NormRGB-lum for color normalization), 53 images include 42 Normal and 11 Abnormal).

Result Overview

Anterior stifle Images Group:

The best result of experiment with Anterior Stifle original for color normalization used is 78.57%.

The best result of experiment with Anterior Stiflelum for color normalization used is 78.57%.

The best result of experiment with Anterior StiflenormGrey for color normalization used is 83.93%.

The best result of experiment with Anterior Stifle normRGBfor color normalization used is 78.57%.

The best result of experiment with Anterior Stifle for normRGB-lumcolor normalization used is 82.14%

Lateral stifle Images Group:

The best result of experiment with Lateral Stifle original for color normalization used is 78.57%.

The best result of experiment with Lateral Stiflelum for color normalization used is 78.57%.

The best result of experiment with Lateral StiflenormGrey for color normalization used is 83.93%.

The best result of experiment with Lateral Stifle normRGBfor color normalization used is 83.93%.

The best result of experiment with Lateral Stifle for normRGB-lumcolor normalization used is 82.14%

Posterior stifle Images Group:

The best result of experiment with Posterior Stifle original for color normalization used is 82.14%.

The best result of experiment with Posterior Stiflelum for color normalization used is 80.35%.

The best result of experiment with Posterior StiflenormGrey for color normalization used is 80.35%.

The best result of experiment with Posterior Stifle normRGBfor color normalization used is 80.35%.

The best result of experiment with Posterior Stifle for normRGB-lumcolor normalization used is 82.14%

Paw Images Group:

The best result of experiment with Paw Stifle original for color normalization used is 79.25 %.

The best result of experiment with Paw Stiflelum for color normalization used is 79.25%.

The best result of experiment with Paw StiflenormGrey for color normalization used is 77.36%.

The best result of experiment with Paw Stifle normRGBfor color normalization used is 79.25%.

The best result of experiment with Paw Stifle for normRGB-lumcolor normalization used is 81.13%

Anterior StifleResults:

Results from Experiment Set #1.(Anterior Stifle).

Color normalization:Original for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Anterior StifleOriginal Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Spectral
Texture Inertia
Texture Entropy (Experiment 656) / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 78.57%.
Texture Entropy / None / Anterior / 42 Normal and 14Abnormal / 76.79%.
Texture InvDiff / None / Anterior / 42 Normal and 14Abnormal / 76.79%.

Highest success rate for this body part: Experiment 656(Soft-max, r = l)

Sensitivity: 14.29%

Specificity: 100.00%

Results from Experiment Set #2.(Anterior Stifle).

Color normalization:Lum for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Anterior StifleLum Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture InvDiff
Histogram Mean (Experiment 80) / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 78.57%.
Texture Entropy / None / Anterior / 42 Normal and 14Abnormal / 76.79%.
Texture Entropy
Histogram Energy / None / Anterior / 42 Normal and 14Abnormal / 76.79%.

Highest success rate for this body part: Experiment 80(Soft-max, r = l)

Sensitivity: 42.86%

Specificity: 90.48%

Results from Experiment Set #3.(Anterior Stifle).

Color normalization:NormGrey for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Anterior StifleNormGrey Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture Inertia
Histogram StdDev(Experiment 264) / None / Anterior / 42 Normal and 14Abnormal / 83.93%.*
Texture Inertia
Histogram StdDev
Histogram Energy / None / Anterior / 42 Normal and 14Abnormal / 83.93%.
Texture Inertia
Texture InvDiff
Histogram Energy / None / Anterior / 42 Normal and 14Abnormal / 83.93%.

*Highest success rate for this body part: Experiment 264(None)

Sensitivity: 71.43%

Specificity: 88.10%

Results from Experiment Set #4.(Anterior Stifle).

Color normalization:NormRGB for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Anterior StifleNormRGB Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture InvDiff
Histogram StdDev
Histogram Energy
(Experiment 74) / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 78.57%.
Texture Inertia
Texture Correlation
Histogram Skew / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 78.57%.
Texture Inertia
Texture Correlation
Texture InvDiff
Texture Entropy
Histogram StdDev
Histogram Energy
Histogram Entropy / Standard Normal Density / Anterior / 42 Normal and 14Abnormal / 78.57%.

Highest success rate for this body part: Experiment 74(Soft-max, r = l)

Sensitivity: 21.43%

Specificity: 97.62%

Results from Experiment Set #5.(Anterior Stifle).

Color normalization:NormRGB-lum for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Anterior StifleNormRGB-lum Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Histogram StdDev
Histogram Energy (Experiment 10) / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 82.14%.
Histogram StdDev
Histogram Entropy / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 80.36%.
Histogram StdDev
Histogram Energy
Histogram Entropy / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 80.36%.

Highest success rate for this body part: Experiment 10(Soft-max, r = l)

Sensitivity: 64.29%

Specificity: 88.10%

Lateral StifleResults:

Results from Experiment Set #6.(Lateral Stifle).

Color normalization:Original for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Lateral StifleOriginal Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Spectral
Texture InvDiff
Histogram StdDev
Histogram Entropy (Experiment 553) / Soft-max, r = l / Lateral / 42 Normal and 14Abnormal / 78.57%.
Spectral
Texture InvDiff
Histogram StdDev
Histogram Skew
Histogram Energy / Soft-max, r = l / Lateral / 42 Normal and 14Abnormal / 78.57%
Spectral
Texture InvDiff
Texture Entropy
Histogram StdDev
Histogram Entropy / Soft-max, r = l / Lateral / 42 Normal and 14Abnormal / 78.57%

Highest success rate for this body part: Experiment 553(Soft-max, r = l)

Sensitivity: 42.86%

Specificity: 90.48%

Results from Experiment Set #7.(Lateral Stifle).

Color normalization:Lum for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Lateral StifleLumExperiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture Correlation
Histogram Skew (Experiment 132) / None / Lateral / 42 Normal and 14Abnormal / 78.57%.
Texture Correlation
Histogram Skew Histogram Energy / None / Lateral / 42 Normal and 14Abnormal / 78.57%
Texture Correlation
Texture InvDiff
Histogram Skew / None / Lateral / 42 Normal and 14Abnormal / 78.57%

Highest success rate for this body part: Experiment 132(None)

Sensitivity: 28.57%

Specificity: 95.24%

Results from Experiment Set #8.(Lateral Stifle).

Color normalization:NormGrey for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Lateral StifleNormGrey Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture Energy
Texture Inertia
Texture Correlation
Texture InvDiff
Histogram Mean
Histogram Energy (Experiment 978) / Soft-max, r = l / Lateral / 42 Normal and 14Abnormal / 83.93%.*
Texture Correlation
Texture InvDiff
Histogram Mean
Histogram Energy / Soft-max, r = l / Lateral / 42 Normal and 14Abnormal / 82.14%
Texture Energy
Texture Correlation
Texture InvDiff
Histogram Mean
Histogram Energy / Soft-max, r = l / Lateral / 42 Normal and 14Abnormal / 82.14%

*Highest success rate for this body part: Experiment 978(Soft-max, r = l)

Sensitivity: 57.14%

Specificity: 92.86%

Results from Experiment Set #9.(Lateral Stifle).

Color normalization:NormRGB for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Lateral StifleNormRGBExperiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture InvDiff(Experiment 64) / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 83.93%*
Texture InvDiff / Standard Normal Density / Anterior / 42 Normal and 14Abnormal / 83.93%
Texture Energy
Texture Correlation
Texture Entropy
Histogram StdDev
Histogram Skew / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 80.35%

*Highest success rate: Experiment 64(Soft-max, r = l)

Sensitivity: 57.14%

Specificity: 92.86%

Results from Experiment Set #10.(Lateral Stifle).

Color normalization:NormRGB-lum for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Lateral StifleNormRGB-lumExperiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture InvDiff
Histogram StdDev
Histogram Entropy (Experiment 73) / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 82.14%
Spectral
Texture InvDiff
Histogram Mean
Histogram Entropy / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 82.14%
Texture Inertia
Texture InvDiff
Texture Entropy
Histogram StdDev
Histogram Entropy / Soft-max, r = l / Anterior / 42 Normal and 14Abnormal / 82.14%

Highest success rate: Experiment 73(Soft-max, r = l)

Sensitivity: 71.43%

Specificity: 85.71%

Posterior StifleResults:

Results from Experiment Set #11.(Posterior Stifle).

Color normalization:Original for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Posterior StifleOriginal Experiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture Energy
Texture InvDiff
Texture Entropy
Histogram StdDev
Histogram Skew (Experiment 316) / Soft-max, r = l / Posterior / 42 Normal and 14Abnormal / 82.14%
Texture InvDiff
Texture Entropy
Histogram Entropy / Soft-max, r = l / Posterior / 42 Normal and 14Abnormal / 80.35%
Texture Energy
Texture InvDiff
Histogram StdDev / Soft-max, r = l / Posterior / 42 Normal and 14Abnormal / 80.35%

Highest success rate: Experiment 316(Soft-max, r = l)

Sensitivity: 35.71%

Specificity: 97.62%

Results from Experiment Set #12.(PosteriorStifle).

Color normalization:Lum for color normalization

Images:Dog GaitImage

Classes:Normal and Abnormal.

Texture Function: texture2, Spectral Feature were used

K-Nearest Neighbor:KNN= 5 and NN

Note: for complete results see the Excel spreadsheet file Posterior StifleLumExperiment.

Features
(texture pixel dist=6) / Normalization method / Body Part / Number of images per class / Classification Success
Texture Entropy
Histogram Mean (Experiment 48) / Soft-max, r = l / Posterior / 42 Normal and 14Abnormal / 80.35%
Texture Entropy
Histogram Mean / Standard Normal Density / Posterior / 42 Normal and 14Abnormal / 80.35%
Texture Energy
Histogram Mean
Histogram Entropy / Soft-max, r = l / Posterior / 42 Normal and 14Abnormal / 80.35%

Highest success rate: Experiment 48(Soft-max, r = l)