C 2005) Journal of Medical Systems, Vol. 29, No. 2, April 2005 ( DOI: 10.1007/s10916-005-3005-7
Imaging System for Visualization and Numerical Analysis of Cancer at Stomach and Skin Tissues Sadık Kara,1,3 Mustafa Okandan,1 Fulya S¸ener,2 and Mustafa Yıldırım1
Digital imaging of cancerous cells is instrumental not only in determining the characteristic of the cancer but also monitoring the progress of the disease in the follow up of the patient and adapting the treatment, accordingly. Therefore, we have developed an imaging system to display and layout the characteristics of normal and cancerous cells in an automated way. Image processing techniques are performed on the digitized images of stomach and skin tissues, in order to derive the number of cells, area of an individual cell, and average area of the cells in a certain size of an image window. KEY WORDS: imaging; cancer; stomach; skin.
INTRODUCTION Cancer is characterized as the outgrowing number of cells more than necessary in an uncontrolled way in one part of the body and the spread of this to other parts of the body very fast.(1) Normal body cells grow, divide, and know when to stop growing. Over time, they also die. Unlike these normal cells, cancer cells just continue to grow and divide out of control and do not die. Cancer cells usually group or clump together to form tumors. A growing tumor becomes a lump of cancer cells that can destroy the normal cells around the tumor and damage the body’s healthy tissues. Sometimes cancer cells break away from the original tumor and travel to other areas of the body, where they keep growing and can go on to form new tumors. This is how cancer spreads. The spread of a tumor to a new place in the body is called metastasis. When cancer presents at a specific part of the body, there is always risk of causing metastasis at another organ. Metastasis usually leads to the death of the patient. In the fast progressing cancer cases, metastasis forms in the early phase. On the 1 Electronics
Engineering Department, Erciyes University, Kayseri, Turkey. of Biomedical Device Technology, Erciyes University, Kayseri, Turkey. 3 To whom correspondence should be addressed at Electronics Engineering Department, Erciyes University, 38039 Kayseri, Turkey; e-mail:
[email protected]. 2 Department
179 C 2005 Springer Science+Business Media, Inc. 0148-5598/05/0400-0179/0
180
Kara, Okandan, S¸ener, and Yıldırım
other hand, in the slow progressing and less harmful cancer cases, metastasis arises much later and then cancer starts.(2,3) Therefore, early diagnosis and determination of the stage of disease are very crucial in the prevention and control of the cancer. The biological structure of the cancerous cells is different from the normal healthy cells. Some of the differences include the size of the cell, shape, staining method used for visualization of the cell, existence of malign large cells, and increase in the cell concentration.(4−10) Digital imaging of cancerous cells is instrumental in not only determining the characteristic of the tumor (benign or malign) but also monitoring the progress of the disease in the follow up of the patient. Particularly, it is very helpful in adapting the treatment to the variation of cancer stages.(11) Therefore, we have developed an imaging system to layout the number of cells in a certain size of image window and average area of the cells for normal and cancerous cells in an automated way. This computerized system will both minimize human caused errors and help archiving the patient data.
MATERIALS AND METHODS Images of the normal and cancerous cells sampled from tissues such as stomach and skin are acquired with the help of a microscope, camera, image capture card, and a PC. The magnified images at the microscope are passed on to the Snazzi III capture card through a CCD camera. The card converts the analog reading into digital form so that processing of images can be accomplished in the computer environment. Image processing techniques are performed on the digitized card and images transferred to PC in order to learn about the number of cells, area of an individual cell, and average area of the cells. In the first step, the tissue samples are stained and then located under the microscope. “Capture the image” command button on the designed graphical user interface (GUI) acquires the image from the camera and helps to visualize the cell itself on the computer screen (Fig. 1(a)). The images are converted to 8-bit Grey scale color format from the 24-bit RGB (Red–Green–Blue) using Eq. (1). Figure 1(b) shows the first hand images after the conversion, which is achieved by Eq. (2). A threshold level, such as 95 as in Eq. (2), is selected manually to obtain the crispy black–white cell regions. The threshold adjustment is also made on GUI. Gray Color = 0.299 × Red + 0.587 × Green + 0.114 × Blue color(x, y) < 95, black color(x, y) = color(x, y) > 95, white
(1) (2)
In order to make cell nucleus and cell regions clearer and more evident, Filter 1, Filter 2, and Filter 3 command buttons on GUI are used for filtration and noise elimination purposes. In the first filter, the partial cells existing on the boundary of the image window are cleared out in order to minimize error in the calculation of number of cells in the image window and mean area of cells (Fig. 1(c)). In this process, the image is
Imaging System for Visualization and Numerical Analysis of Cancer
181
Fig. 1. (a) First-hand image of healthy stomach tissue. (b) Image after conversion to black-white. (c) Image after Filter 1. (d) Image after Filter 3. (e) Image after eliminating cells with area smaller than mean area minus standard deviation.
scanned along the four sides and when a black pixel is encountered on the boundary, conversion of black pixels in the neighborhood to white is performed in two cycles (Fig. 2). Noise elimination takes place in the second filter stage. As the whole image is scanned, individual area of each cell is calculated. It is found out in the investigation of the acquired images that minimum cell area is about 50 pixel2 . The black regions with area smaller than 50 pixel2 are assumed to be artifact and cleared out by converting from black to white. As seen in Fig. 3, the third filter does edge smoothing and fills intracellular white pixels. During image scanning, when white pixel next to a black one is encountered, the image is searched thoroughly, only along the “x” dimension (while “y” is kept constant) till a second black pixel is found in the same row. And then if these two black pixels are found to be connected via other black pixels in the upper rows, then white pixels in between are converted to black. The same operation
Fig. 2. Cycling in the first filtering stage.
182
Kara, Okandan, S¸ener, and Yıldırım
Fig. 3. Filter 3 stage: Edge smoothing and filling intracellular white pixels.
is carried on the “y” dimension as “x” is kept constant (same column). In case the second black pixel in the same row or column is not found, the first black pixel next to white cell is converted to white (Fig. 3). With this type of filtering, the cellular regions that are left out as white due to inadequate staining and rough threshold level selection for conversion of the colorful image to black–white at the very first step are painted back to black. Image resulted after this third filter is shown in Fig. 1(d). n Xi (3) X¯ = i=1 n n ¯ 2 i=1 (Xi − X) S= (4) (n − 1)
Table I. Calculations Performed on Cancerous Stomach Tissue Image Windows
Values before 2nd filtration
Values after 2nd filtration
Values after eliminating cells with area smaller than mean area minus standard deviation
Sample tissue #
Threshold level
Number of cells
Mean area
Number of cells
Mean area
Number of cells
Mean area
Stomach 1 Stomach 2 Stomach 3 Stomach 4 Stomach 5 Stomach 6 Stomach 7 Average
125 125 140 95 145 110 120
13 16 19 21 17 14 17
324 394 377 337 380 310 374 356
13 16 19 21 17 14 17
383 425 446 371 418 395 389 403
9 12 10 18 14 11 15
453 472 508 416 501 457 468 467
Imaging System for Visualization and Numerical Analysis of Cancer
183
Table II. Calculations Performed on Healthy Stomach Tissue Image Windows
Values before 2nd filtration
Values after 2nd filtration
Values after eliminating cells with area smaller than mean area minus standard deviation
Sample tissue #
Threshold level
Number of cells
Mean area
Number of cells
Mean area
Number of cells
Mean area
Stomach 8 Stomach 9 Stomach 10 Average
90 90 83
23 28 21
253 263 214 243
23 28 21
255 269 223 249
19 16 15
278 286 232 265
Using Eqs. 3 and 4, the mean area and standard deviation of the area of the cells are calculated and the cells with areas smaller than mean area value minus standard deviation are eliminated to find a more precise average cellular area. Hence, the cellular images that are not completely improved in the subsequent filtration stages are cleared to make a more precise mean area calculation (Fig. 1(e)). Following this cleaning, average area and standard deviation are calculated for a second time. The end-results for each sample tissue image are shown in the last two columns of Tables I, II, III, and IV. RESULTS The whole process is tried on 10 stomach and 8 skin tissues. Tables I and II present the number of cells and mean area value for seven cancerous and three healthy stomach tissue samples, respectively. As seen in Tables I and II, the distinction between the mean areas of the cancerous cells and healthy cells is improved from the first filtration stage to last stage of elimination of cells with areas less than mean area minus standard deviation. This distinction was 113 (113 = 356 − 243) after the first filter and was improved to 202 at the last stage (202 = 467 − 265). Overall, it is obvious that cancerous tissues have larger cells (malign cells) than healthy tissues. The cancerous stomach cell areas are between 416 and 508, while healthy cells have areas from 232 to 286. Table III. Calculations Performed on Cancerous Skin Tissue Image Windows
Values before 2nd filtration
Values after 2nd filtration
Values after eliminating cells with area smaller than mean area minus standard deviation
Sample tissue #
Threshold level
Number of cells
Mean area
Number of cells
Mean area
Number of cells
Mean area
Skin 1 Skin 2 Skin 3 Skin 4 Skin 5 Average
95 95 100 85 120
35 51 23 19 16
215 318 223 265 255 255
35 51 23 19 16
229 330 235 276 284 271
31 46 20 18 14
250 354 256 291 301 290
184
Kara, Okandan, S¸ener, and Yıldırım Table IV. Calculations Performed on Healthy Skin Tissue Image Windows
Values before 2nd filtration
Values after 2nd filtration
Values after eliminating cells with area smaller than mean area minus standard deviation
Sample tissue #
Threshold level
Number of cells
Mean area
Number of cells
Mean area
Number of cells
Mean area
Skin 6 Skin 7 Skin 8 Average
110 93 110
19 26 12
186 173 168 176
19 26 12
196 191 201 196
18 22 11
211 214 230 218
We cannot really say same kind of improvement in the differentiation of cancerous skin tissues from healthy ones was observed (Tables III and IV). The distinction stayed around 72 pixel2 (72 = 290 − 218). However, still larger cancerous cells are observed, and areas of cancerous cells range between 250 and 354, while areas of healthy cells are only in the range of 211–230. Besides, large malign cells can be visualized for diagnosis of cancer in not only stomach but also skin tissues. Visualization and area calculation complement each other to provide accurate pathological test results.
DISCUSSION AND CONCLUSION In the traditional pathologic tests, the cells are visualized under the microscope and the changes in the number and average areas of the cells in the follow-up of the patient are determined manually with the help of the eye. This kind of primitive technique and analysis does not yield sensitive and accurate results. In this study, number and area of the cell and average area covered by the cell in the selected image window, which are all very commonly used parameters in the pathologic tests, are presented. Our built system eases the pathology professionals’ tasks and saves them time. Besides, this programmed image processing analysis has the potential to eliminate both different interpretations and human errors, which might rise in the pathologic tests.
REFERENCES 1. Leung, K., Samuel., S., Chan, W. K., and Felix Wong, W. S., An expert system for the detection of cervical cancer cells using knowledge-based image analyzer. Artif. Intell. Med. 8:67–90, 1996. 2. Pothur, R., Srinivas Barnett, S., Kramer, B. S., and Sudhir, S., Trends in biomarker research for cancer detection. Lancet Oncol. 2:698–704, 2001. 3. Rozengurt, E., Neuropeptides as growth factors for normal and cancerous cells. Trends Endocrinol. Metab. 13:128–134, 2002. 4. Vujovic, N., Bakic, P., and Bzrakovic, D., Detection of potentially cancerous signs by mammogram follow up. In Doi, K., Giger, M. L., Nishikawa, R. M., and Schmidt, R. A. (eds.), Digital Mammography, 96, Elsevier Science B.V., Chicago, 1996. 5. Castellanos, M. A., A stochastic model for the evolution of cancerous tumors. Rev. Mex. F’isica 42(2):236–249, 1996.
Imaging System for Visualization and Numerical Analysis of Cancer
185
6. Dhlngra, J. K., Perrault , D. F., Jr., McMillan, K., Rebeiz, E. E., Shapshay, S. M., Manoharan, R., Feld, M. S., and Kabani, S., Spectroscopic analysis of precancerous and cancerous lesions of the oral cavity. IEEE Lasers Electro-Optics Soc. Annu. Meet. 1:332–333, 1996. 7. Ladik, J., Towards a unified theory of cancer initiation in the cell. J. Mol. Struct. THEOCHEM 504(1–3, 12):1–11, 2000. 8. Gerald, P., Cell imaging and morphology: Application to studies of inherited purine metabolic disorders. Pharmacol. Ther. 90:267–281, 2001. 9. Rizzuto, R., Carrington, W., and Tuft, R. A., Digital imaging microscopy of living cells. Trends Cell Biol. 8:288–292, 1998. 10. Thiran, J. P., Morphological feature extraction for the classification of digital images of cancerous tissue. IEEE Trans. Biomed. Eng. 43(10):1011–1020, 1996. 11. Funk, J. O., Cancer cell cycle control. Anticancer Res. 19(6A):4772–4780, 1999.
