Colon Cancer Diagnosis on Digital Tissue Images

Download Full-Text PDF Cite this Publication

Text Only Version

Colon Cancer Diagnosis on Digital Tissue Images

NCICN-2015 Conference Proceedings

Dr. K. Sivakami Sundari M.Tech.,Ph.D M. Vanaselvi

K.Vishaka

  1. Vasanthi

    Professor of Chandy College of PG Scholar PG Scholar PG Scholar Engineering, Tuticorin Chandy College of Engg Chandy College of Engg Chandy College of Engg

    Abstract Thepurposeofthisprojectistodevelopasoftwarewhichcanbeanaidfordifficult coloncancerdiagnosisandusingthissystemthepatientscanbehelpedwithanea rlydiagnosis.Theaimcanbeachievedwithprocessingandanalysingmicroscopi ctissueimages.Thispapercontainsthebasicknowledgesrelatedtotheprojectan dthedescriptionofthedevelopedsystem.The implemented algorithms determines the locations and featuresofglandsandsavetheseinformationforthesubsequentdiagnosis.Oneo fthemostimportantalgorithminthisprojectis the Color StructureCode, which performsa color basedsegmentationandtheoutputisthestartingpointofthefurtherprocess.

    IndexTermsmedicalimageprocessing,coloncancerdiagno- sis,glanddetection,nucleusdetection

    1. INTRODUCTION First of all the structure of the colon had to be known.

      structurehasmultiplelayers,themostimportantofthemisthemucosallayer,b ecausethelesionslikeinfectionsandtumorsareevolvefromthere.Inthermso fprocesstheimportanthistologicalcomponentscanbeseenonFig.1.Thediff erencebetweenthehealthyandinfectedmucosalstructurecanbeseenbythele sionsofthehistologicalcomponents.Thecomponents are

      theglands,Thegobletcellsinsidetheglands,thegobletscellsontheepiderian dthenuclei

      Somesimilarprojectswereanalysed,whichhelpedustosetupastartingpoint andtheseprojectsdrewourattentionsforalotofdifficultiesthatwehavetoface throughthetissueanalysis.In[1]thealgorithmscheckedthesizeoftheinputi mage,andifitwastoolargethenitwassplitintomultiplesubpictures.Thesesu bpictureswereprocessedonmultiplethreadswithGPGPU,andthismethodc ouldacceleratedtheprocessingtime,whichwasreallylongindefault.Afterth etestingphase,itwasfoundthatthealgorithmsresultedmorematchesthanthe realnumberofthenuclei,thusitwasnotpossibletomakeanaccuratediagnosi swiththeproject.

      In[2]theprocessingisseparatedintomultipleparts.Thepreprocessingconta insthresholdingandnoisefiltering.Forthenucleisegmentationthewatershe d[3]algorithm wasusedwhichhadanoverflowproblemsotheyhad to usesomecorrectiontoadjustit.Forthediagnosistheyanalysedmanyfeatures whichhasareallylongprocessingtime.Thedisadvantageoftheprogramisth atseveralstepsrequiredmanualconfiguration,whichcouldslowdownthepr ocessingtime.

      Inpaper[4]thecircle-fitistheprimaryandfundamental procedure and theprimitiveobjectsdefinedbythisalgorithmwillbetheanalysedcompone nts.Withtheusage of LAB colorspace the classification and separation ofthehistologicalstructureswasmucheasier.TherebyweusedtheLABcol orspaceinoutprojectsfortheanalysisofthehistologicalstructures.

      Inpaper[5],[6]and[7]newapproachesarementionedwhichcanbeapplied asamoreefficientwayoffeaturedescriptorsasthepreviousones.

    2. OUR APPROACH Wedevelopedtwonecessarymoduleintheneedofdiag-

nosis,onefortheglanddetectionandtheotherforthenucleidetection.Forth eglanddetectionweusedcolorsegmentationwithHSVandLABcolorspac e,tresholding,thenusingtheconnectedcomponentsmethodwecouldident ifytheglandsasindependentobjects.Forthenucleidetectionandseparatio ntwoprocedures were implemented

In the Implementation phase the statements of [8], [9], [10] were taken into account. The structure of our algorithm can be seen on Fig. 2. In follow ingsubsections the different parts of the algorithm will be shown.

  1. Gland detection

    The digital tissue image contains alotofnoise and homo-geneous area that hinders the success fulgl and detection and segmentation. In the preprocessing phase we use an algorithm that blurs the similar colorshadesinHSVcolorspace,sothesignificantlydifferentcolorscanbe separated in to independent objects. Thus the white interiors of the glands can be clearly identifiable.

    WiththeColorStructureCode(furthermoreCSC)[11]wecanblurthepixel swithsimilarcolors,soinmostcasestheboundariesofglandscanbeseparat edwellfromtheotherpartsofthemucosalasitcanbeseeninFig.3.

    Fig.1.Structureofthetissueimage.Thedifferenceamongnuclei,glands,go bletcellsandgobletcellsoftheepidemiscanbefoundeasily.

    Color segmentatio n

    NCICN-2015 Conference Proceedings

    RGB ! LAB

    Thresholding

    Border lines Detection

    Connected

    Components Detection

    RGB ! LAB

    Thresholding

    Border lines Detection

    Connected

    Components Detection

    Erosion,

    noise reduction

    Thresholding

    Miss-detected glandsremoval

    Fig. 2. The flowchart of our algorithm

    The CSC is based on the Hartmann-like hierarchical region growing method, and the so-called islands compose the hier- archical structure in different levels. (See Fig. 4.) The island on the very first level contains seven pixel (one selected in the middle and 6 neighbours). On the next level this hexagonal component does not consist of seven pixel, but the former islands build-up a higher level new island. The building of this structure continues to that point, where the whole image is involved in one structure.

    In the LAB color space the A and B components determine the color which is independent from the luminousness. The A parameter means the red-green and the B component the blue- yellow transition. The L component is the luminance. A threshold value is determined considering the ratio of the aver-age intensity of homogeneous regions and the inhomogeneous regions. The result is an image which contains the contour

    1. (b)

      Fig. 4. The structure of CSC [11]. (a) The splitting phase of the algorithm using hexagonal pixel structure. (b) Composing segments from the homoge-neous pixels. curves of glands. In our tests nearly fifty tissue images were used to determine the appropriate threshold value.

      We have found that in threshold value determination only the A component has to be taken into account, because this parameter differs with the same value among the background and gland boundary regions. The result is binarised towards of the further process. The result of thresholding can be seen in Fig. 5. Because the intensity of some pixels in the contour line is close to the intensity of background pixels, so the boundaries do not constitute closed figures. To generate the connected components the opening morphological operation was used.

      The generated boundaries of glands give the most glands in the image. The further task is the separation and identification. The sequential connected component analysis method assigns unique color coordinates for all connected pixel sett, and these coordinates are the bases of further identifications. For all components the area was calculated, and too big (backgrounds) and too small (noises) components do not consider in the following process.

  2. Boundary generation and matching

    The glands on the boundary of the tissue were not de-tected after the earlier presented algorithm. For detection of these glands further process was necessary. In the first step noise reduction, then edge detection and connected component analysis were applied. After skipping the small regions the boundary lnes of glands were detected. As the last step dilation was used to enhance the boundary lines.

    1. (b) (a) (b

      Fig. 3. The effect of the CSC segmentation algorithm. (a) The original image, (b) the same image after using CSC algorithm.

      Fig. 5. Result of thresholding in LAB color space. (a) The image in LAB color space, (b) the same image after thresholding

      TABLE I

      NCICN-2015 Conference Proceedings

      EXPERIMENTAL RESULT VALUES WITH 10 TISSUE SAMPLES.

      True positive

      False positive

      True negative

      False negative

      Precision

      Sensitivity

      Specificity

      Accuracy

      1. tissue

      74%

      14%

      11%

      0%

      0.84

      1.00

      0.86

      0.92

      2. tissue

      77%

      9%

      11%

      4%

      0.90

      0.95

      0.90

      0.93

      3. tissue

      79%

      15%

      0%

      10%

      0.84

      0.88

      0.88

      0.87

      4. tissue

      73%

      17%

      3%

      7%

      0.81

      0.91

      0.83

      0.87

      5. tissue

      61%

      17%

      20%

      3%

      0.78

      0.95

      0.82

      0.87

      6. tissue

      73%

      13%

      8%

      5%

      0.85

      0.93

      0.86

      0.89

      7. tissue

      70%

      14%

      11%

      5%

      0.83

      0.94

      0.85

      0.89

      8. tissue

      72%

      15%

      7%

      6%

      0.83

      0.93

      0.85

      0.88

      9. tissue

      85%

      8%

      5%

      2%

      0.91

      0.98

      0.92

      0.95

      10. tissue

      68%

      17%

      8%

      6%

      0.80

      0.92

      0.83

      0.87

      Using this method the glands on the border of the tissue was detected as it can be seen in Fig. 6.

  3. Removal of miss-detected glands

    Using the color based segmentation the real glands can be found very well, but unfortunately many non-glands are detected as well. For avoid miss-detection only those regions are considered as glands where few nuclei can be found inside the region. Due to this the image of nucleus regions is generated. Then the number of nuclei in a gland region is counted. If the fraction of the nuclei in a gland candidate region is too high, then this region is not be considered as a gland region in the following process. The result of this method can be seen in Fig. 7.

  4. Nucleus detection using the HSV color space

    In the first approach the nuclei were detected considering the size and the specific dark color of them. Since the color based calculations are not precise enough the images were converted into the HSV color space. In this color space the nuclei have significantly high satu-ration value, so using a well-determined threshold value these can be retrieved. An

  5. Nucleus detection applying color segmentation

    At object determination the segmentation algorithms are very useful. In two other papers of the authors [12], [13] the region growing algorithm was applied implemented in parallel environment. In another approach the Color Structure Code method was applied. Using this method around the dark nucleus black color segments were appeared. These segments can be removed with erosion. The obtained pixels can be used for the description of nucleus positions, but do not give information about the size and shape of them.

    The advantage of this approach is that the nucleus region can be determined with bigger precision, but some dark region was detected in the glands as well.

    Iv TESTS AND RESULTS

    The main function of our system is the gland detection. Our approach was tested on ten representative tissue images. For analysing the efficiency of our system the frequently used measures of information retrieval were used [14]. These are the true positive (TP), false positive (FP), true negative (TN) and false negative (FN). The inherited measures were used as

    adaptive algorithm was implemented, which search connected weel:

    components with small size.

    TP

    The measure of badly detected nuclei was very low, but Precision =

    TP + FP

    (1)

    adjacent nucleus generates a bigger size component and these

    TP

    are not retrieved in the size based detection process. Sensitivity

    =

    TP + FN

    (2)

    TN

    Specificity

    =

    TN + FP

    (3)

    1. (b)

(a) (b

Fig. 6. Glands detection at the tissue border. (a) The originally detected glands, (b) the glands using the boundary line detection

Fig. 7. The influence of miss-detected glands removal. (a) The result without using the removal of miss-detected glands. (b) The result using the removal of miss-detected glands.

Overall, declared that our implemented system has a great advantage compared to other implementations in higher detec- tion rate of true glands.

As an additional development possibility we wish to take into account the application of nuclei detection algorithm in [15], which provides a more accurate detection rate of nuclei. In addition higher hit ratio of nuclei our approach may serve more effective results as well.

Fig. 8. (a) and (c) presents the original tissue images; (b) and

(d) the respective images after applying our method

TP + TN

Accuracy =

TP + TN + FP + FN

(4)

In Table I the mentioned result values can be seen. The advantage of our system is that there are very few number of those glands which are not detected. But the number of false positive cases is rather high, it means there are relatively big numbers of miss-detected objects in the image. The accuracy feature has to enhanced from the inherited property values, it is the most important feature in our test. The average value of this feature is 0.89, so a rather good gland detection can be implemented with our software.

IV. CONCLUSIONS

In Fig. 8-9 original tissue images and the obtained results can be seen.

On the ground of the tests our algorithm acts accurately in the most part of tissue images. The fraction of non-detected glands is averagely 6.2%, even the ratio of miss-detected glands is 18.8%. The separated components outside the gland are considered as miss-detected glands, but the most of the can be easily excluded.

Our project is based on a color-based approach. This has that advantage the detection is invariant for the shape, position and size of glands, whereas our algorithms in order to separate different structures consider only the color differences relative to the different histological structure. The disadvantage of the approach is that it is not suitable to for the separation of different merged glands or glands merged by the background. Another disadvantage is that complex correction algorithms needed to achieve precise results.

Fig. 9. (a), (c) and (e) presents the original tissue images; (b),

(d) and (f) the respective images after applying our method.

REFERENCES

[1] A. Remenyi,´ S. Szen´asi,´ I. Bandi,´ Z. Vamossy,´ G. Valcz, P. Bogdanov,

S. Sergy´an, and M. Kozlovszky, Parallel biomedical image processing with GPGPU-s in cancer research, in 3rd IEEE International Symposium on Logistics and Industrial Informatics, August 25-27. 2011, pp. 245248.

[2]´ C. Demir and B. Yener, Automated cancer diagnosis based´ on histopatho-logical images: a systematic survey, Rens Selaer Polytechnic Institute, Department of Computer Science, Tech. Rep., September 2005.

  1. V. T. DeVita Jr, T. S. Lawrence, and S. A. Rosenberg, Cancer: Principles and Practice of Oncology. Lippincott Williams and Wilkins, 2010,

    vol. 1.

  2. C. Gunduz-Demir, M. Kandemir, A. B. Tosun, and C. Sokmensuer, Automatic segmentation of colon glands using object-graphs, Medical Image Analysis, vol. 14, pp. 112, 2010.

  3. A. Rovid¨ and T. Hashimoto, Improved high dynamic range image reproduction method, Acta Polytechnica Hungarica, vol. 4, no. 3, pp. 4959, 2007.

  4. A. Rovid,¨ L. Szeidl, and P. Varlaki,´ The HOSVD based domain and the related image processing

    techniques, International Journal of Applied Mathematics and Informatics, vol. 5, no. 3, pp. 157164,

    2011.

  5. G. Gyorok,¨¨ M. Mako,´ and J. Lakner, Combinatorics at electronic circuit realization, Acta Polytechnica Hungarica, vol. 6, no. 1, pp. 151160, 2009.

  6. J. Tick, Business process-based initial modeling at software develop-ment, in 11th International Symposium on Applied Machine Intelligence and Informatics, Herlany, Slovakia, 2013, pp. 141144.

  7. J. Tick and Z. Kovacs,´ P-graph based workflow synthesis, in 12th International Conference on Intelligent Engineering Systems, Miami, Florida, USA, February 25- 29 2008, pp. 249253.

  8. J. Tick, Potential application of p-graph-based workflow in logistics,

    Aspects of Computational Intelligence: Theory and Applications: Re-vised and Selected Papers from the 6th IEEE International Symposium on Applied Computational Intelligence and Informatics, vol. 2, pp. 71 82, 2012.

  9. V. Rehrmann and L. Priese, Fast and robust segmentation of natural color scenes, in Computer Vision – ACCV98, ser. Lecture Notes in Computer Science, R. Chin and T.-C. Pong, Eds. Springer Berlin Heidelberg, 1997, vol. 1351, pp. 598606.

  10. S. Szen´asi´ and Z. Vamossy,´ Implementation of distributed genetic algorithm for parameter optimization in cell nuclei detection project,

    Acta Polytechnica Hungarica, in press.

  11. , Evolutionary algorithm for optimizing parameters of GPGPU based image segmentation, Acta Polytechnica Hungarica, in press.

  12. G. Chowdhury, Introduction to Modern Information Retrieval (3rd Edition). Facet Publishing, 2010.

  13. S. Szen´asi,´ Z. Vamossy,´ and M. Kozlovszky, GPGPU-based data parallel region growing algorithm for cell nuclei detection, in 12th IEEE International Symposium on Computational Intelligence and Informat- ics, Budapest, Hungary, November 21-22. 2011, pp. 493 499

Leave a Reply

Your email address will not be published. Required fields are marked *