Centre for Computational and Numerical Studies, Institute of Advanced Study in Science and Technology 2. Dept. of Statistics, Gauhati University

Transcription

1 Cervix Cancer Diagnosis from Pap Smear Images Using Structure Based Segmentation and Shape Analysis 1 Lipi B. Mahanta, 2 Dilip Ch. Nath, 1 Chandan Kr. Nath 1 Centre for Computational and Numerical Studies, Institute of Advanced Study in Science and Technology 2 Dept. of Statistics, Gauhati University ABSTRACT This work presents an approach for analysis of PAP smear images of cervical region based on cell nuclei distribution and shape and size analysis. PAP smear test is an efficient and easy procedure to detect any abnormality in cervical cells. But human observation is not always satisfying and it is a tedious task to manually analyze a large number of PAP smear images. The purpose of this study is to automate the screening process and to provide specific statistical data which will be helpful for detecting abnormalities in cervical region. The proposed approach is implemented in MATLAB, a high level, interactive environment for data visualization/analysis/computation. The MATLAB Image Processing Toolbox was used to segment the digital images and calculate various statistical data. By comparing cell nuclei distribution and taking into account the shape and size features MATLAB can be programmed to distinguish normal cervical cell to questionable ones. Keywords: PAP smear, cervical cancer, image processing, MATLAB 1. INTRODUCTION The automated detection and segmentation of cell nuclei in PAP smear images is one of the most interesting fields in cytological image analysis as observed by Plissiti et. al. [1]. There is a high degree of cell overlap in such images, the presence of more than one nucleus in a cell and the lack of homogeneity in image intensity. These confront any method to overcome the complexity of conventional cervical cell images and to accomplish a correct segmentation. In addition, the nucleus is a very important structure within the cell and it presents significant changes when the cell is affected by a disease and thus the accurate definition of the nucleus boundary is a crucial task. The identification and quantification of these changes in the nucleus morphology and density contribute in the discrimination of normal and abnormal cells in PAP smear images. The segmentation of nuclei in cytological images has been studied by several researchers [2-7]. In this paper we present a method for analysis of PAP smear images based on histogram and structuring element based segmentation and shape and size analysis of the cell nuclei. As mentioned earlier human observation is not always satisfying and according to Marroquin et. al. [8], 61.5% manual screening of available pathological tests remains unclassified. Our method includes segmentation, calculation of cell nuclei distribution and shape and size analysis of the cell nuclei. The first order statistics of an image is its histogram that gives information about the distribution of the gray level value in its dynamic range [0, L-1], where L is the number of gray levels. The histogram gives information of an image as a whole and hence it can be considered as global statistics of the image. Morphological operations are based on shapes. In a morphological operation, the value of each pixel in the output image is based on a comparison of the corresponding pixel in the input image with its neighbours. Dilation and Erosion are the two most common morphological operations. Dilation adds pixels to the object boundary of an image and erosion removes pixels on the object boundary of an image. The number of pixels added or removed from the object boundary of an image depends in the size and shape of the structuring element. For an image, A and a structuring element, B, dilation and erosion are defined as A B and AӨB respectively. Another morphological operation is filling operation which is carried out on binary or grayscale images. For binary images, filling operation changes connected background pixels to foreground pixels. For grayscale images, filling operation brings the intensity values of dark areas that are surrounded by lighter areas up to the same intensity level as surrounding pixels. This work is based on Marroquin et. al. [8] and Hernandez et. al. [9]. Hernandez described a way to segment biopsy images of brain cell and calculate the cell distribution. Marroquin used a fuzzy classifier after extracting features related to texture, shape and size to classify the breast biopsy images. The aim of this study was to provide assistance to pathologists and to enhance the accuracy of statistical data. 2. METHODOLGY The images used in this work are conventional PAP stained cervical cell images, acquired through a CCD camera adapted to an optical microscope. We used a 40 magnification lens and images were stored in JPEG format having size pixels. With consultation of the pathologists, we selected the following features to be extracted 245

2 1. Size of the cell nuclei 2. Shape of the cell nuclei elongation occurs when abnormality occurs. The degree of change in shape is a good measure to analyze the shape of the nuclei. We took into account two factors. 3. Cell nuclei distribution The preprocessing is arranged into three steps: simplification and image enhancement, segmentation and feature extraction. I) Simplification and image enhancement The digitized images are coloured in RGB mode. Matrix corresponding to colour image is three dimensional and hence difficult to process. The images in RGB mode have three colour components and it is a tedious task to segment using those colour component. So due to the fact that gray level image are easy to process we converted the images to gray level by using the rgb2gray unction in MATLAB. Then we enhance the contrast of the images using histogram analysis using imadjust function in MATLAB. II) Segmentation The cell nuclei are darker than the surrounding cytoplasm and all the cell nuclei belonging to a class tend to have same gray level. We created the histogram for each image to view he density distribution of the different shades of gray. Based on the fact that cell nuclei are darker, we filtered out the light areas and created a uniform background. Then using structuring element function, we created binary images containing only the cell nuclei. We considered only round shape for segmenting cell nuclei, but all the nuclei are not round. So, the binary images are again processed with dilation and filling function. As a result of applying dilation and erosion, extra parts which were not part of the nuclei were removed and the boundary of the cell nuclei became prominent. Then we apply filling operation to create a uniform intensity level inside the cell nuclei. III) Feature Extraction a) Cell nuclei distribution The number of normal and abnormal cell in a cytological image is good criteria for identifying abnormality. The size of the cell nuclei can be treated as factor for determining whether they are normal or not. We calculated the area of each nucleus. The area was calculated in pixels. We, then, calculated the cell nuclei distribution per images and presented the result in tabular from with one column having the area and the other having the number of nuclei with that area. First one is Compactness. It is a dimensionless shape feature which measure compactness. It is defined by Sheng et. al. [10] as C = P 2 /A where P is the perimeter and A is the area of the cell nuclei. The calculation was done in pixel. Second one is Eccentricity. The cell nuclei are round in shape in normal condition. The roundness is not perfect and the shape can be considered as ellipse. We calculate the semimajor and semiminor axes lengths and the eccentricity for each nucleus using the following formula ε = {(a 2 b 2 )/b 2 } 1/2 where a and b are the semimajor and semiminor axes of the ellipse respectively. The eccentricity value 0 corresponds to a circle and with an increasing value the deviation from circle becomes more significant as observed by Marroquin. 3. RESULTS AND DISCUSSION The pre-processing step excludes all the background and leaves for further processing the parts of the image which contain isolated cells or cell clusters and as such is a fast process which results in the reduction of the region of interest in the image. Our method has been applied in several PAP smear images defined by an expert observer. The step for the detection of the cell nucleus centroid has exposed that the resulted points of the image indicate the area of the nuclei, as it is confirmed by the expert observer. Two images are shown here, one showing cancerous cells in initial stage that is mild dysplasia and the other showing normal cervical cell. The statistics were then obtained and compared to confirm the differences. The results are shown below. 4. SIMPLIFICATION AND IMAGE ENHANCEMENT Figure 1-2 (a), (b) and (c) show the results after applying the techniques mentioned above for simplification and enhancement of two images. b) Shape and size analysis Shape is another important feature for classification of cervical cells in PAP smear images. Generally, the shape of cell nuclei are round and 246

3 6. FEATURE EXTRACTION 1) Cell nuclei distribution according to area Table 1- Cancerous Cells Number of nuclei Area in pixels 2 45 Figure 1 - (a) Original image showing cancerous cells in initial stage that is mild dysplasia., (b) Image simplified /converted to gray level and (c) Image after enhancement Table 2 - Normal Cells Number of nuclei Area in pixels Figure 2 - (a) Original image showing normal cervical cell (b) Image simplified/converted to gray level and (c) Image after enhancement SEGMENTATION: The images were then segmented for the region of interest, that is the cell nuclei. They are shown below: Figure 3,4 - Segmented image showing only the cell nuclei The distribution of cell nuclei is an indicative measure of the number of normal and abnormal nuclei in a single slide. When cells in a tissue begin to show abnormal growth, the cells deviates from the normal morphological characteristics one by one. Size of the cell nuclei is one of the features that increase during abnormal or cancerous growth. We calculated the size of the each cell nucleus in terms of area and presented them in tabular form. The cell nuclei distribution table clearly shows that the abnormal samples tend to have more number of nuclei with higher value of area. The normal samples show a lower range of distribution as in the normal sample in Table 1 which has distribution range from 13 to 26 pixels. Whereas the cancerous sample in Fig.(c) has a range from 25 to 45 pixels. We can say that more the number of bigger nuclei more the extent of abnormality. 247

4 2) Compactness 3) Eccentricity Table 3 - Cancerous Cells Perimeter Area Compactness Table 5 - cancerous cells Cell nuclei Eccentricity Table 4 - Normal Cells Perimeter Area Compactness Table 6 - Normal Cells Cell nuclei Eccentricity Compactness is the dimensionless shape measure of the cell nuclei. A normal nucleus has a well-formed and a compact shape in normal condition. Cells with abnormality gradually deform and the compactness decreases. We found out the compactness of the cell nuclei and found that normal cells have higher value of compactness then that of the abnormal nuclei which is very obvious from the tables showing the compactness of the normal and abnormal cells. Eccentricity is the measure of roundness of the cell nuclei. In general, the eccentricity can be said to be calculated from the width and height of the cell nuclei. The normal nuclei have a minimal proportion between the width and height and thus have greater roundness. Uncontrolled growth of the nuclei does not keep this uniform proportion and as result their eccentricity deviated farther away from zero (0). The eccentricity table of the normal sample shows some of the nuclei having 0 eccentricity value, whereas the abnormal sample deviates away from CONCLUSION: An effective method to identify and classify cervical cancer is becoming increasingly needed due to the fact that early detection and a decision of correct therapy 248

5 may save the patient. Medical images have various limitations such as low quality, presence of noise and human error in interpretation. Digital image processing can help the pathologists to a great extent. The statistical data can be used as benchmark to flag normal or questionable sample while the pathologist looks at the slide under a microscope which will be highly time saving. Some ideas for future enhancement includes: to design an interactive system where a pathologist can feed his own grayscale threshold or to automate the process by computer using histogram or fuzzy logic. Another enhancement can be to establish cutoff values between normal and abnormal values and to classify the abnormal values according to stage of the cancer. The images processed are magnified and the calculations are done in pixels. So, if a relation among magnification, pixels and actual size is established, the analysis will be more efficient. [2] Bamford P., Lovell B., Unsupervised cell nucleus segmentation with active contours, Signal Processing 71(2), pp , [3] Bamford P., Lovell B., A water immersion algorithm for cytological image segmentation, Proceedings of the APRS Image segmantation workshop, pp , University of Technology Sydney, Sydney [4] Mouroutis T., Roberts S. J., Robust cell nuclei segmentation using statistical modelling, IOP Bioimaging, 6, pp , [5] Garrido A., Perez de la Blanca N., Applying deformable templates for cell image segmentation, Pattern Recognition 33, pp , [6] Lee K.M., Street W.N., Learning shapes for automatic image segmentation, Proc. INFORMS-KORMS Conference, pp , Seoul, Korea, June ACKNOWLEDGMENT We would like to offer our gratitude to Dr. J. D. Sharmah, Chief consultant, Pathology department and Dr.(Miss) Banani Sharmah, Pathologist of Dr. B. Baruah Cancer Research Institute for providing the images and helping us to prepare documentation. This paper has the consent of all co-authors and authorities of the institute, where this study has been carried out and there exists no conflict of interest anywhere. REFERENCES [1] Plissiti M.E., Charchanti A., Krikoni O. and Fotiadis D.I., Automated segmentation of cell nuclei in PAP smear images, ITAB Proceedings International Special Topic Conference on Information Technology in Biomedicine, Greece, Ionnia, 26-28, October [7] Begelman G., Gur E., Rivlin E., Rudzsky M., Zalevsky Z., Cell nuclei segmentation using fuzzy logic engine, International Conference on Image Processing, Vol. 5, pp , October [8] Marroquin E. Martinez, Vos C., Santamaria E., Jove X., Socoro J.C., Non Linear Image Analysis for Fuzzy Classification of Breast Cancer, IEEE Proceedings of International Conference on Image Processing, vol.2, , [9] Hernandez L., Gothreaux P., Shih L., Towards Realtime Biopsy Image Analysis and Cell Segmentation. In Proceedings of IPCV, pp.81-87, [10] Sheng L., Rangayyan R. M., Desautels L., Application of Shape analysis to Mammographic Calcification. IEEE Trans. On Medical Imaging, Vol. 13, NO. 2,