Cervical cancer classification using gabor filters 1026

CERVICAL CANCER CLASSIFICATION USING

GABOR FILTERS

2011 First IEEE International Conference on Healthcare Informatics, Imaging and Systems

Biology

Advisor : Yin-Fu Huang

Student : Chen-Ju Lai

OUTLINE

INTRODUCTION DATA COLLECTION METHODOLOGY AND RESULT CONCLUSION

INTRODUCTION

Cervical cancer Biopsy test Cervical intraepithelial neoplasia (CIN)

Input : histology images Feature extraction : texture , using Gabor

filter Classification method : K-Means Clustering Output : Normal/CIN1/CIN2/CIN3/Malignant

Pre-cancer

DATA COLLECTION

Pathology anatomy laboratory of Saiful Anwar hospital

Biopsy images : resolution 4080 x 3072 pixels

(categorized by an expert pathologist) 475 labelled images are used in this study

Normal CIN1 CIN2 CIN3 Malignant

60 70 50 50 245

DATA COLLECTION

CANCER GRADING

METHODOLOGYAND

RESULT

GABOR FILTER

Gabor elementary function

2D Gaussian function

From (1) and (2), the Gabor elementary function can be rewritten as

Spatial domain

σx and σy are the spread of the Gaussian in x and y directions

centre frequency

x'=x cos θ +y sin θ and y'=-x sin θ+y cos θ.

GABOR FILTER

Assuming σx and σy are the sameFrequency domain

U

V

θ

U0

φ

u

v

u'=u cos θ +v sin θ and v'=u sin θ+v cos θ

(U,V) can decision (U0 ,θ)

GABOR FILTER

Sample

Original (a) f = 0.2,θ = 0 0 (b) f = 0.2,θ = 45 0

(c) f = 0.2,θ = 90 0 (d) f = 0.2,θ = 135 0

COMPARE TEMPLATE

Compare each pixel with the templates. Supervised Training : generated templates

24 distinctive Gabor filters are used to generate a feature vector for each pixel and its neighbors.

background

basal stroma normalcells

abnormal cells

500 pixels 500 pixels

500 pixels

500 pixels

500 pixels

average average average average average

SEGMENTED IMAGE & K-MEAN CLUSTERING

Segmentation After each pixel compare with the five feature

vector templates. blue : background , yellow : basal , white : stroma, green : normal cell , red : abnormal cell

K-Means Clustering Based on the color. Quantify the normal nuclei and abnormal nuclei.

SEGMENTED IMAGE & K-MEAN CLUSTERING

CALCULATE THE RATIO AND GRADING

How to classify the image into categories ? Use the ratio of number of normal and abnormal

cells.

Benign the number of abnormal cells < 5

CIN 1 ratio between abnormal and normal cells < 1/3

CIN 2 ratio between abnormal and normal cells between 1/3 ~ 2/3

CIN 3 ratio between abnormal and normal cells> 2/3 or full

Malignant

ratio between abnormal and normal cells > CIN 3

CALCULATE THE RATIO AND GRADING

Table 1 shows the sample of the ratio between abnormal and normal cell.

CALCULATE THE RATIO AND GRADING

Table 2 shows the confusion matrix of the Gabor filter hybrid with K-means clustering.

The sensitivity of normal is 87%, CIN 1 is 86%, CIN 2 82 %,CIN 3 84% and malignant is 89%.

The percentage of specificity of this system is 85%.

(52/60)=0.87(60/70)=0.86(41/50)=0.82(42/50)=0.84(219/245)=0.89

COMPARED WITH SERVAL METHOD

Gray level Features , color K-mean and incremental thresholding.

CONCLUSION

A methodology of Gabor filter bank with hybrid K-means clustering algorithm has been proposed.

Designing Gabor filter bank with the optimum selection parameters and different classification method can improve performance using this algorithm.