Shahed K Mohammed Farah Deeba Francis M Bui and Khan A Wahid Electrical and Computer Engineering University of Saskatchewan 1 Presentation Outline 2 Wireless Capsule Endoscopy 60000 Frames per patient ID: 792178
Download The PPT/PDF document "Application of Modified Ant Colony Optim..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Application of Modified Ant Colony Optimization for Computer Aided Bleeding Detection System
Shahed K. Mohammed, Farah Deeba, Francis M. Bui, and Khan A. WahidElectrical and Computer Engineering, University of Saskatchewan
1
Slide2Presentation Outline
2
Slide3Wireless Capsule Endoscopy
60000 Frames per patient
No control over movement
Only Few frames may contain any diagnostic information
Patient Friendly
Small Bowel Visualization
3
Slide4Gastrointestinal Bleeding
Ulcers
Angiodysplasia
Cancer
4
Slide5Automatic Bleeding Detection
5
Distinct hue of Bleeding
Different Color-spaces have been employed showing varying results on different dataset
Slide6Can cascading the feature subset from different color space improve the performance of the classifier?
How to solve the curse of dimensionality? 6
Slide7Feature Selection
7
Slide8Exhaustive Search Method
combinations to search for n dimensional feature space
8
Number
of Features
Number of
Possible Combinations
5
31
10
1023
15
32,767
20
1,048,575
30
1,073,741,823
Slide9Ant Colony Optimization
9
Metaheuristic approach to solve optimization problem
Inspired by pheromone trail laying behavior of real ants
Slide10How Ants Find the Smallest Path?
10
Slide11Feature Decision Making Using ACO
11
Each Node is a feature
Path travelled by the ant represents the feature subset
Slide12Probability Function
12
Slide13Flowchart of ACO for Feature Selection
13
Initialize ACO Parameter and Probability Function
Generate Feature Subset (Path of Ants)
For each subset, determine the classification model and the optimized error
Find the local best subset using the evaluation function
Pheromone Evaporation and Update
Stopping Criteria Reached
Global Best Feature Subset
Yes
No
Generate New Feature Subset
Slide14Probability Function for ACO
14
Slide15Proposed Probability Function for Modified Ant Colony Optimization (MACO)
15
Slide16Features Used in Bleeding Detection
16
Red
Green
Blue
Hue
Saturation
Value
Mean
Std.
Dev.
Energy
Skew
Entropy
Slide17Experimental Results
17
Slide18Dataset
Training
Testing
Bleeding
Non-bleeding
Bleeding
Non-bleeding
150
150
294
301
18
Slide19Classifier
Support Vector Machine (SVM) was used for deriving the classification model. Five fold cross validation was performed to estimate the optimization error
19
Slide20Comparison of Proposed Heuristic Information: Accuracy
20
Slide21Comparison of Proposed Heuristic Information: Number of Features
21
Slide22Comparison
Exhaustive Search (ES)Sequential Forward Search (SFS)Sequential Backward Search (SBS)
Binary Genetic Algorithm (BGA)
Random Subset Feature Selection (RSFS)
22
Slide23Performance: RGB Classifier
23
Feature Selection
Number of Features
Sensitivity
Specificity
Accuracy
Computational time
Exhaustive Search
2
0.9798
0.9800
0.9799
100
MACO
2
0.9798
0.9800
0.9799
1.88
Slide24Performance: HSV Classifier
24
Feature Selection
Number of Features
Sensitivity
Specificity
Accuracy
Computational time
Exhaustive Search
3
0.9866
0.9800
0.9833
100
MACO
3
0.9797
0.9700
0.9748
1.85
Slide25Performance on RGB-HSV Classifier
25
Feature Selection
Sensitivity
Specificity
Accuracy
Computational Time (min)
MACO
0.9966
0.9801
0.9882
147
ACO
0.5059
1
0.5058
47
SFS
0.9864
0.9900
0.9882
133
SBS
0.0034
1
0.5075
362
BGA
0.9864
0.7475
0.8655
111
RSFS
0.0404
1
0.6162
44
Slide26Comparison Between The Classifiers
26
Slide27Conclusion and Future Work
We proposed an efficient optimum feature selection algorithm.Compared to state-of-the-art feature selection algorithms, the proposed algorithm can select the optimum feature subset in a significantly less computational time.In future, we will extend this algorithm for feature selection for other abnormality detection algorithm.
27
Slide28Acknowledgement
28