Presentation is loading. Please wait.

Presentation is loading. Please wait.

Applications of Evolutionary Algorithms

Similar presentations


Presentation on theme: "Applications of Evolutionary Algorithms"— Presentation transcript:

1 Applications of Evolutionary Algorithms
Jordan Hedges Alexander Emery Belinda Edmunds Jye Dewar

2 What are Evolutionary Algorithms?
Evolutionary Algorithms are a population-based heuristic optimization algorithm They use mechanisms that are based in biological evolution such as reproduction, and mutation We will be focuses on the applications of Genetic Algorithms, as they are the more commonly used form of Evolutionary Algorithm

3 Applications Some applications of Evolutionary Algorithms are as follows: Computer Automated Design Drug Design Medical Imaging Game playing AI Scheduling Our presentation will focus on both Drug Design and the Medical Imaging applications

4 Application: Drug Design
Explored by Douguet, Thoreau and Grassy who were seeking to use genetic algorithms to design novel molecules. Exhaustive search was completely impractical due to the size of the search space Results were compared with known ‘Drug Databases’ in order to streamline the massive search that would be undertaken

5 Solution Encoding Molecules are solutions and represented using Simplified molecular-input line-entry system Captures bonding (single, double or triple), structure (such as benzene rings) and branching. Human readable, but hard for a machine to manipulate during mutation and crossover. Weininger, D., J. Chem. Inf. Comput. Sci., 30 (1990) 237.

6 Steps Initialise population with fragment
Evaluate fitness of each molecule Select one or two molecules as parents Crossover or Mutation Check syntax of child Evaluate children Repeat as necessary

7 Initialisation Initial population is composed of fragments/molecules either provided, randomly selected, or randomly selected from a group of elements, these are taken from a chemical structure library. Example: 10 random derivatives of methane

8 Fitness Fitness score was based on several criteria:
Taken from Douguet, D., Thoreau, E. & Grassy, G. J Fitness Fitness score was based on several criteria: Indicative constraints: Volume Lipophilicity Electronic properties Imperative constraints: Such as correct valency Scorei = ∑p Wp*Scoreip Score is 100% when Xip ∈[Xpmin, Xpmax], otherwise ranges between 1 to 99% decreasing the further the molecule is from the properties’ goal range.

9 Taken from Douguet, D., Thoreau, E. & Grassy, G. J

10 Selection Parents are selected using roulette wheel method, selection likelihood increases based on fitness of molecule. Interestingly, scores are then scaled with the explicit intent of increasing the probability of less fit molecules being selected, most likely to increase chance of exploration. This effect is then toned down as the generational limit increases. Note that more favourable results are achieved when the fittest molecule is guaranteed selection.

11 Taken from Douguet, D., Thoreau, E. & Grassy, G. J

12 Crossover and Mutation
Two probabilities pm and pc give the probabilities of mutation and crossover respectively with their sum equal to 1, can be initially fixed or vary according to a strategy. Chosen between with roulette wheel selection. Two crossover operations and fourteen mutation operations, randomly selected from once a class of operation is chosen.

13 Crossover operators Crossover 1: Remove single bond in A and B and combine A and B. Crossover 2: Split molecule A into three parts and replace the centre with part of molecule B. Taken from Douguet, D., Thoreau, E. & Grassy, G. J

14 Mutation operators Fourteen different operators, randomly selected from. If multiple options for an operation, select from these with the listed probability.

15 Check syntax Children may have been created that are not physically possible One example being the wrong valence SMILES format converted to mol2 format to check properties Child is discarded and a new random operation is attempted This will continue for all impossible children

16 Results Taken from Douguet, D., Thoreau, E. & Grassy, G. J
Across many different sample sizes, populations of gave the best results Choosing the best off spring (highest fitness) resulted in more consistent results, with a higher average fitness after many generations

17 Different Generations
Taken from Douguet, D., Thoreau, E. & Grassy, G. J Different Generations

18 Application: Medical Imaging
Applications in medicine include: Radiotherapy Radiology Oncology Pediatrics Cardiology Endocrinology Surgery Orthopedics Pulmonology Obstetrics & Gynaecology Infectious Diseases Rehabilitation medicine Nerology Pharmacotherapy Health Care Management

19 Application: Medical Imaging
Radiology imaging techniques generate large amounts of data Computer-aided detection and diagnosis used for: Detection Segmentation Classification X-rays, MRI, CT scan, ultrasound, mammograms Aids in early detection of diabetes, breast, lung & colon cancer, coronary artery disease, congenital heart defect & neurological diseases

20 Medical Imaging: Mammograms
Majority of applications of GAs in radiology were used in mammography Radiologists fail to detect between 10% - 30% of malignant lesions Computer-Aided Detection (CADe) & Computer-Aided Diagnosis (CADx) CADe automatically detects and segments suspicious lesions CADx characterizes and estimates probability of malignancy

21 Image Segmentation & Detection
Research by Pereira, Ramos & Zanchetta do Nascimento from focuses on CADe systems for mammograms Proposed GA procedure allowed detection and segmentation of suspicious areas with 95% accuracy for true positives and between 1-3% sensitivity for false positives Image data set from the Digital Database for Screening Mammography (DDSM)

22 Challenges Feature selection for the genetic algorithm is important to reduce time, cost & complexity which impacts on detection accuracy. Tumors frequently indistinguishable from adjacent tissue Very inconsistent in size, shape and density Poor in image contrast Highly connected to surrounding tissue density Tissue background has no uniform characteristics

23 Methodology Algorithm designed to remove background artifacts
Filtering in the wavelet domain applied minimizes the mean square error between estimated random process & desired process Combine GA and wavelet transform to determine threshold levels & segmentation values Apply post-processing procedure Source: Periera et al, 2014

24 Outcome a) Original image from DDSM with lateral and oblique views
b) GA Segmented views, note number of false positive regions c) Malignant tumor detection after post- processing Source: Periera et al, 2014

25 Region of interest identified by previous studies
Tumor regions marked by GA segmentation procedure Image segmentation from proposed GA method Each image pair shows the cranio-caudal (CC) and medio-lateral oblique (MLO) views of the same region Source: Pereira et al, 2014

26 Medical Image Segmentation
A grey level image can be described a function f(x,y) f is the value of the feature at (x,y) (x,y) are spatial coordinates. A grey level image can be described as a function f(x,y) where f is the value of the feature at (x,y) and (x,y) are spatial coordinates.

27 Image segmentation equation
Rep(Ri) represents the gray level of some region Ri K is the number of regions This equation is difficult to solve because Number of regions (k) is not known beforehand Representing gray levels implicitly requires considering all Possible partitions Source: C. Lai & C. Chang, 2009 This equation is difficult to solve due to the fact the proper number of regions is not known beforehand and the decision of representing gray level implicitly requires considering all the possible partitions. The k-means algorithm is one of the most widely used methods for solving this, however disadvantages of this method are k-means is sensitive to noise, and the intial cluster centers, and to apply this method the number of cluster centers needs to be known before hand The solution? To employ hierarchical evolutionary algorithms to automatically search the number of regions as well as properly finding the representative gray level for each region.

28 Hierarchical Evolutionary Algorithms
A variant of Genetic Algorithms Variable chromosome length Control gene and Parametric gene Hierarchical evolutionary algorithms are a variant of genetic algorithms. They use chromosomes of variable length. These chromosome consist of two types of genes. Control gene and parametric gene. The control gene dictates whether a parametric gene is used or not and is normally coded as binary digits, whereas a parametric gene can be any data type Source: C. Lai & C. Chang, 2009

29 Steps Source: C. Lai & C. Chang, 2009

30 Solution Representation
Control gene is made of binary digits and parametric gene is made of integer numbers representing the gray levels Can be viewed as a concise representation of variable length chromosomes. Number of control genes is decided by a soft estimate of the upper bound of the number of regions Source: C. Lai & C. Chang, 2009

31 Initial Population Gray levels are randomly selected from [0,L]
Where L is the number of discrete grey levels Control genes are generated randomly from {0,1}

32 Fitness Function As the aim of image segmentation is to make pixels in the same region similar to each other, the fitness function can be defined as For each chromosome, the parametric genes are extracted and then a segmented image is obtained by assigning their pixels to the corresponding regions Source: C. Lai & C. Chang, 2009

33 Selection Tournament selection method is utilised
Given Population P, tournament size Ntour Source: C. Lai & C. Chang, 2009 For selection, the tournament selection method is utilised, because time complexity is low. The basic idea behind the tournament method is randomly select a positive number N tour of chromosomes from the population and copy the best fitted item from them into an intermediate population. This process is repeated P times where P is the population size.

34 Crossover Uses the uniform crossover approach.
Is applied to the control gene and parametric gene simultaneously Two chromosomes are randomly selected as parents Crossover creates the offsprings on a bitwise basis taking each control parametric pair with a probability p Uniform crossover approach Is applied to the control gene and parametric gene simultaenously. Two chromosomes are randomly selected as parents, cross over creates the offspring on a bitwise basis taking the each allele with a probability p An allele is two or more alternative forms of a gene that arise by mutation and are found at the same place on a chromosome. P1 and P2 are parents, C1 and C2 are offsprings. The ith allele Source: C. Lai & C. Chang, 2009

35 Mutation Used to explore new areas of the search space and avoid local optimas. Applies bit mutation to control genes. (Flips 1 to 0 and vice versa) Mutation is used to explore new areas of the search space and avoid local optimas. Applying bitmutation to the control genes results in 1s being flipped to 0 and 0 being flipped to 1. Either of these cases will change the number of regions. Source: C. Lai & C. Chang, 2009

36 Results: Initial conditions
Generation number was 500 Population size was 20 Probabilities of crossover were 0.8 Probabilities of mutation were 0.15 Soft estimate of the upper bound of the number of regions was 10

37 Results Misclassified rate = |A - D|/A x 100
A is the number of actual pixel D is the number of segmented pixels Source: C. Lai & C. Chang, 2009

38 Results CT Head image The original Dynamic thresholding method K-means
Fuzzy c-mean CHNN Hierarchical Evolutionary Algorithm Source: C. Lai & C. Chang, 2009

39 References Pereira, D., Ramos, R. and do Nascimento, M. (2014). Segmentation and detection of breast cancer in mammograms combining wavelet analysis and genetic algorithm. Computer Methods and Programs in Biomedicine, 114(1), pp Lai, C. and Chang, C. (2009). A hierarchical evolutionary algorithm for automatic medical image segmentation. Expert Systems with Applications, 36(1), pp Ghaheri, A., Shoar, S., Naderan, M. and Shahabuddin Hoseini, S. (2015). The Applications of Genetic Algorithms in Medicine. Oman Medical Journal, 30(6), pp En.wikipedia.org. (2017). Computer-aided diagnosis. [online] Available at: aided_diagnosis [Accessed 11 Sep. 2017]. Douguet, D., Thoreau, E. & Grassy, G. J Comput Aided Mol Des (2000) 14:

40 Thanks for listening… Questions?


Download ppt "Applications of Evolutionary Algorithms"

Similar presentations


Ads by Google