Proceedings of the International Conference , “Computational Systems and Communication Technology”

5TH MAY 2010 - by Einstein College of Engineering,

Tirunelveli-Tamil Nadu,PIN-627 012,INDIA
A Fuzzy based Automatic Pap Screening System – ROI Detection
M. Mohideen Fatima alias Niraimathi1, S.P. Raja2
1
Associate Professor, National College of Engineering, Tirunelveli, Tamilnadu, India
1
[email protected]

Abstract— In this paper an efficient and fast Fuzzy-based Cytology screening test has a 15% “false negative” rate and a
Automatic Pap Screening Systems -FAPSS- algorithm, that can be 5% “false positive” rate. A “false positive” is mistakenly
useful for future ACSS. The algorithm detects areas of the smear calling a normal specimen as “positive”. This results in
wherein the cells are located. Identification of the best areas for unnecessary anguish for the patient and additional expense of
screening provides the importance degree for evaluation of each time and money to cover such follow-up testing as a repeated
field. The results obtained by the proposed algorithm have a high Pap smear or a costly colposcopic examination. Recently,
concordance degree with the cytotechnologist evaluation. This
efforts to improve Pap smear performance have focused on
approach is based on fuzzy techniques which aids in a good
processing of both vagueness and in determination characteristic of reducing the number of “false negative” smears, that is, cases
cytology cell images. The proposed efficient and very fast fuzzy- in which premalignant or malignant cells have been
based APSS algorithm for ROI detection, which simulates misdiagnosed as normal. Measures adopted to improve
technologists' process, simplifies their work, and provides a very laboratory performance on this point include re-screening of
high ROI detection concordance percentage with the 10-20% of slides initially evaluated as negative. Recently,
cytotechnologist. The results obtained by the proposed approach several technologies have been developed to optimize Pap test
show the effectiveness of fuzzy techniques in vagueness treatment. screening by reducing the “false negative” rate associated with
cervical cytological screening. The two major components to
Keywords— Fuzzy techniques, Computer Vision, Scene analysis,
this “false negative” rate are “false negative” related to
Cervical cancer screening, Pap
sampling error and false negatives related to detection error.
I. INTRODUCTION About two-thirds of “false negative” are a result of sampling
error and the remaining one-third a result of detection error.
Cytopathology is a branch of pathology that studies Each of the new technologies is directed at one of these
and diagnoses diseases on the cellular level. The most components of “false negative”. Thin-layer cytology aims
common use of cytopathology is the Pap smear, used to detect primarily to fix sampling error, whereas computerized re-
cervical cancer at an early treatable stage.Cytopathology is a screening targets detection error. This implies that neither
very useful tool for screening of cervical carcinoma. Despite technology will be able to reduce “false negative” beyond a
all of the technological advances which have occurred in certain threshold.
medicine in the last 100 years, the cervicovaginal smear, or This type of problem typically requires a two-stage
Pap smear, in honor of its developer, Dr. George process.
Papanicolaou, is still one of the few tests where automation 1. Region of Interest -ROI- detection, it means intelligently
has not been yet achieved [2],[3]. Diagnostic cytology has two eliminating all normal images “the hay”. Normal images can
main steps: 1. The cytotechnologist microscopically examines be eliminated and suspicious images saved for further
the morphologic features of the cells, relates these findings to processing.
the patient's clinical history, and renders a cytologic 2. Region of Interest classification, wherein the suspicious
impression. 2. Cytopathologists diagnose disease by analyzing images are analyzed using computer algorithms to obtain best
cells previously selected by cytotechnologist. classification.
The aim of the work here presented is to develop an
automate algorithm addressed to ROI detection, and suitable
Over the last 60 years, several modification of of interacting with the human technologist. To do it we
terminology for cytological diagnosis was used. Today, the propose a Fuzzy-based Automatic Pap Screening System
most accepted one is The Bethesda System [4] proposed -FAPSS-. The system is a pixel classification approach that
initially in 1989, and actualized in 2001. In this system, performs a fuzzy local feature analysis and is faithful to the
cytologies are divided into two classes: negatives and principle of least commitment [3] in that partial information is
pathological. In the later group there are, according to the accumulated before a final decision is made.
seventy of the lesion, four diagnostic groups: atypia, low
grade lesion, high grade lesion and carcinoma. The main II. PIXEL CLASSIFICATON
objective of screening is identifying these pathological groups, The proposed -FAPSS- pretends to incorporate, as much as,
from atypia until carcinoma. Most atypias can be resolved possible cytotechnologists knowledge and experience, but
spontaneously, but up to 30% can progress to a carcinoma, if avoiding their subjectivity. Moreover, the results provided by
they are not adequately controlled[5]. Also, most lesions the ROI detection have to be independent of human bias and
would progress to a carcinoma[2-4]. It is estimated that the
 Proceedings of the International Conference , “Computational Systems and Communication Technology”
5TH MAY 2010 - by Einstein College of Engineering,
Tirunelveli-Tamil Nadu,PIN-627 012,INDIA
error, but making easier its use on the part of the technician. Background L H
To manage to get these goals our system makes use of Back Contour L VS
segmentation by pixel classification wherein local features are Region 1 DM VH
analyzed using fuzzy techniques [5] and [7] . Moreover, the Region1contour DM VS
algorithm has been applied to gray-scale images, because of Region2 LM VH
they are less sensitive to variations of lighting conditions and
staining quality than those of color images, and experienced 3.2 Fuzzy Rules
human observers can easily differentiate cells in gray level Considering variability and vagueness within the images and
images. Following the usual structure of pixel classification linguistic descriptions of the elements, we represent the Fuzzy
systems, the identification of areas related with the cells To knowledge by the Fuzzy rule base:
implement these steps in a suitable way, vagueness factors RC: if G-L(pij) = LM & Hom(pij) = M
introduced by vaginal secretions, inflamatory cells, cel- Then pij is Cell.
clumbing , bloodstaining , air-drying and other cell sampling RB: if G-L(pij) = L & Hom. (pij) = H
and preparation artifacts and have to be taken into account. Then pij is Background.
So, considering the technician way of work, after an analysis RBC:if G-L(pij) = L & Hom. (pij) = VS
of the images focused on gray level and texture Then pij is Contour of Background.
characteristics, besides Cells and Background, two more RR1: if G-L(pij)=DM & Hom.(pij) =VH
regions, containing the other elements appearing within the Then pij is Region 1.
cell images, have to be considered by our pixel classification RR1C:if G-L(pij)=DM & Hom.(pij) = VS
system.These regions are: Region 1. Includes inflamatory Then pij is Contour of Region1.
cells, cel-clumbing and bloodstaining, Region 2. Includes RR2:if G-L(pij)=LM & Hom.(pij) = VH
vaginal secretions, air-drying and other cell sampling and Then pij is Region 2.
preparation artefacts. .So, considering the technician way of work, after an analysis
of the images focused on gray level and texture
III. SYSTEM DESCRIPTION characteristics, besides Cells and Background, two more
3.1 Regions detection regions, containing the other elements appearing within the
Certain features are considered to detect the regions appearing cell images, have to be considered by our pixel classification
within the cell images. The overall process is shown in the system. These regions are: Region 1. Includes inflamatory
following Figure 1. The regions based on GLL and DOH is cells, cel-clumbing and bloodstaining, Region 2. Includes
shown in Table 1. vaginal secretions, air-drying and other cell sampling and
preparation artefact.
The Contour regions which are transition areas, with
special characteristics, have to be kept in mind.The gray level
lightness and homogeneity degree are evaluated, respectively,
by means of the average and standard deviation of the gray
values measured over a 3x3 window.Then, membership
functions are obtained starting from the probability density
and distribution functions, of these values evaluated over the
training images following the process explained in [4]. The
fuzzy sets so obtained are aggregated by the minimum,
according to the set of rules previously defined, to get the
regions fuzzy sets determined by the membership functions.
This type of method used to detect Region of Interest is called
as Direct Method of cytology cells detection.

3.3 RONI Detection:

To provide cytotechnologist with an accurate ROI
detection and location, the output of the system must give
information about all the best areas for screening.The
identification of the Regions of Interest can be performed
locating the ROI directly or finding and eliminating the region
of no interest RONI. The RONI membership function is
Fig. 1: Overall process
TABLE I
obtained implementing this rule as follows
REGIONS BASED ON GLL AND DOH µRONI ( pij ) = Max ( µB ( pij ). µR1 ( pij ). µR 2 ( pij ))
Regions Gray Level Homogeneity
Cells LM M
 Proceedings of the International Conference , “Computational Systems and Communication Technology”
5TH MAY 2010 - by Einstein College of Engineering,
Tirunelveli-Tamil Nadu,PIN-627 012,INDIA
Here a fuzzy Max operation is performed between the
membership values of background, region 1 and region2 is
used to eliminate the non region of interest. The complement
of RONI gives the Region of Interest. This is also referred as
Indirect Method of cytology cells detection.

3.4 Cells Detection:

e)Region fuzzy sets f) Prob distri image
For a given image I, after obtaining µROI(pij) and
µC(pij) to decide if the image has to be analyzed, and the
membership function with which this analysis has to be
performed, next rule is applied:
Rcells:If
∀p(ij ) ∈ I , µ ROI ( pij ) > µ C ( pij )
≈

then I is analysed, else I is rejected.

3.5 ROI Detection and Location:
To get the final ROI to be given to cytotechnologist, g)ROI image h)TROI image
a regional analysis is performed. A fuzzy morphological
structural element has been used to detect the areas of the
image in which there exist interesting cells in a high enough
number. A circular structural element having a 2.9 pixels
radius has been chosen and the morphological operations are
performed on the threshold ROI.

IV. IMPLEMENTATION RESULTS

The experimental outputs are as shown in Figure 2(a-k).

i) histogram of image

a)original image b) g-l-l image

j) histogram of homogeneity image

c)homogeneity image d) prob.density
image
 Proceedings of the International Conference , “Computational Systems and Communication Technology”
5TH MAY 2010 - by Einstein College of Engineering,
Tirunelveli-Tamil Nadu,PIN-627 012,INDIA
results. Reviewing the existing rules, and if required adding
new rules to the rule basewe can improve accuracy. Revising
shapes of the fuzzy sets can be done for an apt membership
function since most cases, fuzzy systems are highly tolerant of
a shape approximation.

REFERENCES

k) Rcells image [1] Caroline McNeil and Nicole Gottlieb "Bethesda 2001: A Revised System
for Reporting Pap Test Results ",April 24, 2002.
[2] Koss LG. "The Papanicolaou test for cervical cancer detection: a triumph
V. CONCLUSIONS and a tragedy" JAM4 1989; 261, 737-43.
An efficient and very fast fuzzy based FAPSS [3] Koss LG, Schreiber K, Elberg P, Mango L. "Evaluation of the PAPNET
algorithm for ROI detection, which simulates technologists' Cytologic system for quality control of cervical smears". Am JClin Pathol
1994; 101, pp. 220-229.
process, simplifies their work, and provides a very high ROI [4] Karnon.J., Peters.J., Platt.J., Chilcott.J., McGoogan .E., and Brewer.N.,
detection concordance percentage with the cytotechnologist "Liquid- based cytology in cervical screening: an updated rapid and
has been proposed. The results obtained by the proposed systematic review and economic analysis" Health Technology
approach show the effectiveness of fuzzy techniques in Assessment 2004; Vol. 8: No. 20.
[5] Keller .J, and Gader.P, "Fuzzy Logic and the Principle of Least
vagueness treatment, habituation, cytotechnologist fatigue, Commitment in Computer Vision", Proceedings, IEEE Conference on
incomplete screening of a specimen, and cytotechnologist Systems, Man, and Cybernetics, Vancouver, BC,October, 1995, pp. 4621-
shortage. All of these factors contribute to human error. 4625.
Automation of the process is challenging due to the [6] Mamdani.E.H., "Application of fuzzy logic to Approximate reasoning
using linguistic synthesis". IEEE Trans. Computers C-26, No. 12, 1997, pp.
tremendous amount of data to be processed. We have used 1182- 1191 .
manual method of membership function generation from the [7] Montseny.E., and Sobrevilla.P., "Application of fuzzy techniques to the
gray level lightness and degree of homogeneity histograms. design of algorithms in computer vision", Mathware &Soft Computing Vol.
The fuzzy systems can be tuned in the following ways to 2- 3,1998, pp. 223-230.
[8] Park.J., and Keller.J., "Fuzzy Patch Label Relaxation in Bone Marrow
achieve best results. Reviewing model input and output Cell Segmentation", Proceedings, IEEE International Conference on Systems,
variables, and if required redefine their ranges. We can define Man, and Cybernetics, Orlando, FL, October, 1997, pp. 1133-1138.
additional sets on the universe of discourse. The use of wide [9] Sobrevilla.P., Montseny .E.,and Keller.J., "White Blood Cell Detection in
fuzzy sets may cause the fuzzy system to perform roughly. So Bone Marrow Images", Proceedings, on the l8th Intr. Conf. Of the North
American Fuzzy Information Processing Sociely, New York, June, 1999, pp.
by providing sufficient overlap between neighbouring sets an 403-407.
accurate fuzzy set can be derived. It is suggested that triangle- [10] Sobrevilla.P., Keller.J., and Montseny.E., "Using a Fuzzy
to-triangle and trapezoid-to-triangle fuzzy sets should overlap Morphological Structural Element for Image Segmentation". Proc. of 19th
between 25% to 50% of their bases to get more accurate NAFIPS International Conference, Atalanta, GA, USA, July 2000, pp 95-99.