WBC-Classification

Data Samples

Collect the data set from kaggle

Class	Eosinophil	Lymphocyte	Monocyte	Neutrophol
Images
Size of train dataset	2500	2500	2500	2500

The basic differnce between all four classes is total number of nuclie in each cell.Monocyte and lyphote have single nuclie but shapes are different,Monocyte looks like bean and lyphocyte looks like a circle.Eosinophil and Neutrophil have 2 and 3 nuclie respectively.

Task

Class	Eosinophil	Lymphocyte	Monocyte	Neutrophol	Task
HSV Image					This image is obtained by Ihsv=rgb2hsv(Img) command
HSV saturation plane					It is a saturation plane(2nd plane of hsv) obtained by Ihsv(:,:,2)

Feature extraction

Now by using extractLBPFeaturesmatlab command,extracted features from the images and each feature size is 1 x 59.So like wise collected 600 features from 600 images of each class.All features are stored in 600 x 59 matrix of respective classes{featuresA1,featuresA2,featuresA3,featuresA4}.Average vectors(featureB1,featureB2,featureB3,featureB4} of each size 1 x 59 are obtained by using all four 600 x59 feature matrix respectivly.Thease average vectors are used as metric to clasifiy the test images

Class	feature vectors(600x59)	Avg feature vector(1x59)
Eosinophil	features_A1	feature_B1 obtained from column wise average of features_A1 by using matlab code mean(features_A1,1)
Lymphocyte	features_A2	feature_B2 obtained from column wise average of features_A2 by using matlab code mean(features_A2,1)
Monocyte	features_A3	feature_B3 obtained from column wise average of features_A3 by using matlab code mean(features_A3,1)
Neutrophol	features_A4	feature_B4 obtained from column wise average of features_A4 by using matlab code mean(features_A4,1)

NOTE

Out of several methods like GLCM,Entropy,contrast,and between different feature extraction methods local binarypatterns and Average vector method is giving good accuracy so choosen them.

Cosine Similarity Measure

To classify the image,here I used cosine simlarity as a measure,which basically returns maximum value for closer vectors. For more details about similarity measure look into the code

Results

Class	Eosinophil	Lymphocyte	Monocyte	Neutrophil
number of test Images	250	250	250	250
Accuracy	34	88.50	81.	30.90

Conclusions

The classification accuracy of Lymphocyte and monocyte is good.Accuracy of the classification majorly depends on the dataset also.This dataset is not perfectly balanced,Please find few samples from netrophil and monocyte datasets,the dont actually looks like their own domin.Not only that while labling single image is assigned with multi class in given label set.So this is also effects the accuracy.

class	Eosinophil	Monocyte	Netrophil
dummy Images