Abnormality Detection In Endoscopic Images

Published Date: 02 Nov 2017

THROAT CANCER BY MORPHOLOGICAL OPERATIONS

Deepti Shikha and B.V. Dhandra*

Department of P.G. Studies and Research in Computer Science

Gulbarga University, Gulbarga

Abstract

Mathematical morphology provides a number of important images processing

operations and become the foundation of biomedical computing. The data extracted

from images continues to be a fundamental technique for achieving scientific

progress in experimental, clinical, biomedical, and behavioural research. Image

segmentation is an important component of image analysis, which partitions the

whole image under study into various disjoint regions based on potential features

such as gray value, color, texture, etc. An algorithm was developed to perform the

segmentation, classification and analysis of medical images, especially the

endoscopic images for the identification of commonly occurring throat cancer

abnormalities. It was observed that the proposed segmentation generates larger

number of regions in the abnormal images as compared to normal. Further, it was

seen that a large number of segmented regions generated in normal images due to the

presence of noise such as lumen regions, bright spots generated by the reflection of

light sources, etc.Indian Streams Research Journal

Medical imaging has been undergoing a revolution in the past decade with the advent

of faster, more accurate, and less invasive devices. This has driven the need for

corresponding software development which in turn has provided a major impetus for

new algorithms in signal and image processing. Digital image processing is important

for many biomedical applications. The medical images analyzed, used as diagnostic

tools and quite often provide insight into the inner working of the process under

study. It allows one to enhance the image features of interest while attenuating detail

irrelevant to a given application, and then extracts useful information about the scene

from the enhanced image [1, 2]. Often the raw image is not directly suitable for this

purpose and must be pre-processed in a form suitable for processing. Image

segmentation is the fundamental and important component of image analysis, which

partitions the whole image under study into various disjoint regions based on

potential features such as gray value, color, texture, etc. The segmentation process

presents both the uniformity of features within the region and an edge evolution, in

both the cases, the result should be balanced between adherence to the possible noisy

and incomplete data. The smooth segmentation results help for further analysis [3].

Throat cancer is one of the cancers which occur in the throat. This is disease is found

with different names like vocal cord cancer, throat cancer, laryngeal cancer, cancer of

the glottis. This is the cancer which is found in both men and women. Throat cancer

is one of the major causes of death around the world.

Throat cancer affects more men than women. It affects more people aged over 50

years and the risk factor includes smoking of cigarettes, chewing tobacco and heavy

alcohol consumption. Throat cancer can start in the oesophagus (food pipe), larynx

(voice box), thyroid gland or cells lining of the throat (squamous cells). The early

detection and characterization of throat cancer helps reduce the need for therapeutic

treatment and minimizes pain and suffering [4, 5, 6]

Endoscopy provides images better than that of the other tests, and in many cases

endoscopy is superior to the other imaging techniques such as traditional x-rays. In

some cases, a physician may discover an apparent abnormality during examination

that requires further analysis. This analysis can help to determine the cause of the

abnormality such as inflammation, infection and cancer. The process of computerized

visualization, interpretation and analysis of endoscopic images assist the physician for

fast identification of the abnormality in the images. In this direction, research works

are being carried out for classifying the abnormal endoscopic images based on their

properties like color, texture, structural relationships between the image pixels, etc [3,

7, 8].

One of the most rewarding areas of image processing is Mathematical Morphology.

Set theory forms the substratum of Mathematical Morphology. The objects in an

image are analogous to the sets in Mathematical Morphology. The geometric relations

amidst the points of such sets serve as the crux for the morphological operations [9].

Some of the premier operations that are instrumental for diverse image processing

problems include erosion, dilation, opening and closing. However, the drawback of

the mathematical morphology technique is that a part of the noise still remains [10].

A detailed introductory explanation of mathematical morphology is provided in [11,

12, 13]. It is particularly useful in providing basic building blocks to more

sophisticated imaging applications [14]. Using mathematical morphology, image data

can be filtered to either preserve or remove features of interest, sizing transformations

can be constructed, and information relating to shape, form and size can be easily

applied.

The early applications of medical imaging sought to diagnose simple pathology such as

bone fracture or foreign bodies. Today, medical imaging has become a discrete medical

discipline and an essential part of prevention, diagnosis, and high treatment standards

throughout the world, revolutionizing virtually every aspect of clinical medicine. A number

of imaging modalities are routinely employed not only to rule out overt disease or

injury but also to reveal anatomical abnormality and dysfunction of organs, often well

ahead of clinical manifestations. Thus, medical image analysis has gained importance [15].

Medical imaging refers to the techniques and processes used to create images of the human

body (or parts thereof) for clinical purposes (medical procedures seeking to reveal,

diagnose or examine disease) or medical science (including the study of

normal anatomy and physiology) and depends on the analysis medical experts detect

the cancerous region based on the color information obtained from the endoscopic

images as discernible to their eyes [16].

In the present investigation, we have developed the algorithms and methods to

perform the segmentation, classification and analysis of medical images, especially

the endoscopic images for the identification of commonly occurring throat cancer

abnormalities to assist the physician for further diagnosis and treatment. The

proposed worked is described in section 2, experimental results of our algorithm in

section 3 and conclusions are given in section 4.

Proposed method

Image preprocessing and segmentation

The endoscopic images of the throat cancer considered as input images. The RGB color

image contains data three times more than a gray-scale image. So, after the image pixels

resized to a fixed value 128 * 128, the resized RGB image was converted into a grayscale

image. In the process of the conversion, the hue and saturation information are eliminated,

retaining the intensity component of each pixel in the image. The grayscale image was

smoothed using the average filter with the mask size 3X3, thereby reducing the noise present

in the image. Further, the bright spots formed due to the reflection of light source present in

the endoscope were eliminated by replacing such pixels with the average intensity value.

The pre-processed images fill the bright spots with average intensity value; this increases the

contrast of the resultant image. The complement of the pre-processed image was computed

by subtracting every pixel of the image with 255 as the intensity value in a gray scale images

are ranging from 0 to 255. In the output image, dark areas become lighter and light areas

become darker. The extended minima transform applied on the complemented image,

extended minima transform was a kind of threshold operation which bring most of the

valleys present in the image terrain to zero. The imimposemin function also changed a

valley�s pixel values to zero, deepest possible valley for unit 8 images. The combined process

of extended minima transform and the image imposition controlled the excessive over

segmentation of the image.

Feature Extraction

When the input data to an algorithm was too large to be processed and it is suspected

to be notoriously redundant (much data, but not much information) then the input data

were transformed into a reduced representation set of features (also named features

vector). Transforming the input data into the set of features called as feature

extraction. If the features extracted are carefully chosen it was expected that the

features set will extract the relevant information from the input data in order to

perform the desired task using this reduced representation instead of the full size

input. Some of the following features were extracted from the resultant segmented

images:

Area: the actual number of pixels in the region.

Eccentricity: the eccentricity of the ellipse that has the same second-moments as the region.

The eccentricity is the ratio of the distance between the foci of the ellipse and its major axis

length. The value is between 0 and 1 (0 and 1 are degenerate cases; an ellipse whose

eccentricity is 0 is actually a circle, while an ellipse whose eccentricity is 1 is a line

segment). This property is supported only for 2-D input label matrices.

On Pixel Density: The on pixels density of an image is the number of pixels in the extracted

region by the size of the image.

On pixel density =

?? ?? ??????

???? ?? ??? ?????

Experimental results

The morphological image segmentation method for detection of throat cancer in

endoscopic images was proposed and implemented on 20 abnormal and 7 normal

endoscopic images. The 20 images of sized 128x128 pixels containing the cancerous

region of the throat are considerd, which were assumed to be free from lumen region and

are preprocessed. The preprocessing involves smoothing of color images using average

filter. The results depicted in Fig.1 shows the experimental results of segmentation of

abnormal regions detected in the below test image using the proposed method.

Fig 1: Segmentation process for endoscopic image of throat cancer

Input color endoscopic image, (b) resized image to pixel 128 * 128, (c) converted grayscale image,

(d) Image after smoothing, (e) Inverse transform image,(f) Complemented Image, (g) Extended

minima transform, (h) Image after imposition process.

The Table 4.1 and 4.2 exhibits the above described features obtained after the

proposed segmentation process applied on the abnormal and normal images. The

results are shown for 20 abnormal images and 7 normal images.

Area Eccentricity

On Pixel

Density

Cancer01 1772 0.96005 0.10815

Cancer02 313 0.9135 0.019104

Cancer03 4326 0.88148 0.26404

Cancer04 962 0.92178 0.058716

Cancer05 665 0.84245 0.040588

Cancer06 4447 0.79415 0.27142

Cancer07 604 0.98703 0.036865

Cancer08 2020 0.78793 0.12329

Cancer09 2772 0.93687 0.16919

Cancer10 903 0.92426 0.055115

Cancer11 721 0.97731 0.044006

Cancer12 419 0.86558 0.025574

Cancer13 6130 0.28642 0.37415

Cancer14 213 0.95002 0.013

Cancer15 1592 0.74536 0.097168

Cancer16 1711 0.59126 0.10443

Cancer17 468 0.93967 0.028564

Cancer18 2273 0.53358 0.13873

Cancer19 2485 0.53038 0.15167

Cancer20 535 0.95113 0.032654

Table 4.1: Extracted features from the segmented abnormal images

Area Eccentricity

On Pixel

Density

Normal01 2405 0.5392 0.14679

Normal02 0 0 0

Normal03 0 0 0

Normal04 0 0 0

Normal05 1982 0.53458 0.12097

Normal06 0 0 0

Normal07 0 0 0

Table 4.2: Extracted features from the segmented normal images

Further it was seen that a large number of segmented regions generated in normal

images due to the presence of noise such as lumen regions, bright spots generated by

the reflection of light sources, etc.

The following given hypothesis was considered from the normal images.

Ho: � = (620, 0.10, 0.05) (1)

H1: � ? (620, 0.10, 0.05) (2)

Given as ?? = Area, ??=Eccentricity and ??= On Pixel Density, were measured and

obtained the following results.

1 = 1766.55

2 = 0.8160105

3 = 0.1078212

The covariance and inverse of covariance matrix is given as

Covariance matrix = (3)

Inverse of Covariance = (4)

The Hotelling�s ?

?

statistics was computed for testing the above hypothesis and was given by

T

2

= 20

= 15.30 (5)

Critical value was calculated by the following given formula.

Critical Value = F p.n-p (0.10) = 8.18 (6)

The observed T

2

= 15.30>8.18

Here 8.18 the null hypothesis of normal images being rejected at 10% level of

significance.

Conclusions

Medical image analysis are crucial for obtaining solutions for the problems like image guided

surgery, description of anatomical regions, deformation analysis and visualization of

anatomical and physiological processes. Hence there is necessity to carry out the potential

research work in this direction. Therefore, in this present work, an attempt was made for

segmentation and analysis of endoscopic images for the detection of possible presence of

abnormality, which may assist the physician for the fast diagnosis of diseases and quick

treatment.

The segmentation results were compared with the manual segmentation performed by the

medical experts. The experimental results showed good agreement with the manual

segmentation. Thus, the proposed segmentation method can be used for automatic detection

of cancerous region in endoscopic images, which assists the physician for faster and proper

diagnosis of the disease for immediate treatment.

As stated above, the proposed method depends on the endoscopic equipment. The change in

the endoscope may lead to changes in the color property of the image, which may not yield

the expected results and large number of segments would be generated because due to the

presence of noise such as lumen regions, bright spots generated by the reflection of light

sources, etc. The limitations of the algorithm examined so far are certainly prone to

segmentation errors if the objects portrayed in the color images are affected by highlights,

shadowing, and shadows. These phenomena will cause the uniform colored surfaces to

change drastically. This may lead to the over segmentation of the regions. The only way to

overcome this drawback is to analyze how light interact with colored materials and to

introduce models for this physical interaction in the segmentation algorithms.