摘要
膀胱癌是一种严重威胁人类健康的全球性疾病。美国癌症协会(American Cancer Society)发布的《Cancer facts & figures 2009》指出:在恶性肿瘤疾病中,膀胱癌位居美国男性第四位、女性第八位,占男性恶性肿瘤发病数的7%,这一数字在女性为3%。同时,该报告估计全美2009年膀胱肿瘤新发病例数为70,980例,预计因膀胱肿瘤而导致死亡的病例为14,330例。根据SEER(National Cancer Institute’s Surveillance Epidemiology and End Result Registry)的统计结果,美国自1975年以来,膀胱癌的发病率增加了约40%。中国男性膀胱癌发病率位居肿瘤疾病的第八位,而且部分城市近年来肿瘤发病率报告显示,其发病率有快速增高趋势。除了常见的症状和体征等基础诊断方法外,膀胱镜是目前膀胱癌诊断的金标准。在实行膀胱镜检查的同时也会进行影像学检查以判断肿瘤浸润周围组织的情况。影像图片的判读和读片医师的临床经验密切相关。尽管在图像上肿瘤组织的灰度和肿瘤周围正常组织的灰度有很大的不同,但直接通过肉眼读片获得诸如膀胱肿瘤浸润深度——这样关键的临床信息仍然是非常困难的。本研究的目的是找到可以在MRI图像上区分膀胱肿瘤组织和膀胱壁组织的纹理特征,进而获得膀胱肿瘤浸润膀胱壁平滑肌的深度。
本实验共收集22例膀胱肿瘤患者数据及23例正常志愿者数据。所有数据分为三组:组A为肿瘤区域的纹理特征值;组B为肿瘤患者膀胱壁区域的纹理特征值;组C为正常志愿者膀胱壁区域的纹理特征值。其中组B又分为B1组和B2组,B1组由Ta,T1和T2期肿瘤患者膀胱壁区域的特征值组成(13例),B2由T3和T4期肿瘤患者膀胱壁区域的特征值组成(7例)。实验选取五类共42种纹理特指。实验步骤如下:首先,为保证图像分割的准确性,我们将MRI图像中的ROI区域手工勾勒出来;第二,计算ROI区域内的纹理特征值;第三,将上述分组的纹理特征值进行统计分析,筛选出所需特征;第四,在上一步特征筛选的基础上,将纹理分析、SVM、肿瘤生长特性结合应用,在MRI图像上确定膀胱肿瘤的浸润深度。
有35个纹理特征在组A和组B间表现出统计学差异,包括均值、熵、均匀度、方差、平滑度、三阶距;向量的模(自协方差系数)、粗糙度、对比度、线性度、粗略度(Tamura)、大部分基于灰度-梯度共生矩阵的纹理特征以及部分基于灰度共生矩阵的纹理特征。尽管粗糙度和粗略度在统计分析中表现出显著性差异,但是这两个特征值对图像的大小很敏感,因此本研究将其舍去。接下来对组B和组C进行统计学分析,有9个纹理特征在两组间表现出统计学差异,它们是熵、均匀度、方向度、T5,T6,T9(GLGCM灰度-梯度共生矩阵);f1,f9,f12(GLCM灰度共生矩阵)。第三步,在组B1、B2、C之间进行方差分析,主要分析的纹理特征是在第一步统计分析中表现出显著性差异的33项(舍去粗糙度和粗略度)。15项纹理特征在组B2和组C间表现出显著性差异,包括熵、均匀度、标准差、线性度、T3~T5,T7~T9(GLGCM)、f1,f8~f11(GLCM)。这些特征与第二步t检验得出的纹理特征在很大程度上存重叠。26项纹理特征在组B1和组C间没有表现出统计学差异,它们是:均值、平滑度、均值、向量的模(自协方差系数)、对比度、线性度、T1~T4,T7,T8,T10,T11(GLGCM), f2~f8,f10,f11,f14~f16(GLCM)。有9项纹理特征在在B2组和C组间表现出统计学差异,而在B1和C组间没有表现出差异;而表现出完全相反统计结果的纹理特征只有三阶距一项。第四,利用上述筛选出的纹理特征,将其与识别分类器——支持向量机(SVM)相结合,在MRI图像上识别肿瘤区域。根据肿瘤区域的分布并结合肿瘤生长特性判断肿瘤的浸润深度。将上述浸润深度与病理切片获得的浸润深度相对比,16例病例中有12例判断准确。
实验的初步结果显示,有33项纹理特征在膀胱肿瘤组织和患者的膀胱壁组织间存在统计学差异。实验结果进一步说明早期肿瘤周围的膀胱壁纹理与志愿者的膀胱壁纹理类似,而晚期肿瘤患者的膀胱壁纹理与志愿者膀胱壁纹理存在差异。可能的原因是由于病情进展,膀胱壁内血管增生以及平滑肌纤维化导致膀胱壁纹理改变。本实验最重要的结论是:通过将纹理分析与SVM分类器相结合,为直接在MRI图像上确定膀胱肿瘤浸润膀胱壁的深度提供了一种可行手段。
Bladder cancer is a severe disease that seriously threatens the human health all over the world.《Cancer facts & figures 2009》indicates that bladder cancer is the fourth most common malignancy in men and eighth in woman, it accounts for 7% of all malignancy in men and 3% in woman all over American. This report also estimates that 70,980 new cases and 14,330 deaths are expected to occur in 2009 in American. In addition, according to the National Cancer Institute’s Surveillance Epidemiology and End Result Registry (SEER), there has been a rising trend in bladder cancer incidence by approximately 40% since 1975. Bladder cancer is the eighth most common malignancy in men in China and the incidence is climbing up rapidly in some cities recently. Besides basic diagnosis based on symptom and physical exam, cystoscopy is currently the gold standard for bladder cancer diagnosis. Meanwhile, radiological imaging is often preformed in conjunction with the cystoscopy for the evaluation of malignant invasion into adjacent structures. Radiograph reading and interpretation process depends greatly on the experience of radiologists. Since the densities of image voxels inside carcinomatous tissue differ only subtly from those of the surrounding and the appearance of bladder cancer varies greatly, it is quite difficult to acquire some essential information such as degree of muscle invasion (stage of bladder cancer) directly from radiological images, which is of great importance to accurate and early diagnosis and surgical treatment planning. The aim of this study was to explore characteristic texture features that could distinguish bladder cancer form bladder wall tissue in MR images, which may help us determine the invasion depth of bladder cancer into bladder wall muscle automatically.
Twenty-two consecutive patients with confirmed bladder tumors and twenty-three volunteers with fine bladders were included in this study. All the MRI data was divided into three main groups. Group A: pixels within carcinoma tissue area of patients in MRI images; group B: pixels within bladder wall area of patients in MR images; group C: pixels within bladder wall area of volunteers in MRI images. Group B was further subdivided into two small groups, B1: MRI data from patients of bladder cancer at Ta、T1 or T2 stage, whose bladder wall is thin (thirteen cases); B2: MRI data from patients of bladder cancer at T3 or T4 stage, whose bladder wall is thick (seven cases). Forty-two texture features from five categories were employed in this study. The experiment consists of four parts. In order to extract exact areas of carcinoma and normal bladder wall in MRI images, the ROIs (region of interest) were first selected manually. Second, texture features demonstrated the textures of ROIs were calculated. Third, statistical analysis was applied to results of features to reflect their significance. Fourth, texture analysis based on texture features which filtered from third step was preformed in conjunction with SVM (Support Vector Machine) which recognized patterns (textures) in MRI images. The stage of bladder cancer can be acquired by combining texture analysis with the characteristics of growth of bladder cancer.
There were significant differences between group A and group B on thirty-five features, including Mean, Entropy, Uniformity, Standard deviation, Smoothness, Third moment (DG); Norm of Vector(auto-covariance coefficient); Coarseness; Contrast, Line_likeness; Roughness (Tamura features); all GLGCM features expect T6; all GLCM features expect f12,f13. Although Coarseness and Roughness showed significant differences, we still decided to cancel them because the values of these two features were greatly influenced by the size of ROI. In next step, t-test was made between the group B and group C. There were significant difference between two groups on nine features, including Entropy, Uniformity; Directionality (Tamura features); T5, T6, T9 (GLGCM features); f1, f9, f12 (GLCM features). Thirdly, analysis of variance was made among group B1, B2 and C. Features that shown significant differences in first test were employed in this step (Thirty-three features). There were significant differences between group B2 and C on fifteen features, including Entropy, Uniformity, Standard deviation, Line_likeness, T3~T5,T7~T9, f1,f8~f11(most features were same as the results of second t-test), while twenty-six features didn’t showed significant differences between group B1 and C, including Mean, Smoothness, Standard deviation, Norm of Vector, Contrast, Line_likeness, T1~T4,T7,T8,T10,T11,f2~f8,f10,f11,f14~f16. There were significant differences between group B2 and C on nine features, including Standard deviation, Line_likeness, T3, T4, T7, T8, f8, f10, f11, while there were no significant differences between group B1 and C on these features. On the contrary, only one features, i.e. the Third moment, showed an opposite statistical result between the two comparisons. Fourthly, we utilized texture features filtered from the third step, SVM and the characteristic of growth of cancer to determine the depth of tumor invasion(stage) into the bladder wall., Twelve(75%) of sixteen stage of tumors determined by pathology examine recognized by our method.
The preliminary results obtained in this study indicates that there are statistically significant differences existed in thirty-three texture features extracted from MR images between bladder cancer tissue and bladder wall tissue of patients. Additionally, the statistical results demonstrate that bladder wall tissue of patients differ from the bladder wall tissue of volunteers. The results of analysis of variance indicate that tissue of bladder wall of early stage of bladder cancer is different from the tissue of advanced stage. Moreover, tissue of bladder wall of early stage of bladder cancer is more similar to the tissue of bladder wall of volunteers. These observations indicate that with the development of tumor angiogenesis and fibrosis in smooth muscle of bladder wall, the patterns of texture feature in MR images change in patients with advanced stage of bladder cancer. The most important conclusion is that the stage of bladder cancer is quite possibly be identified by using the proposed method describe in this study.
引文
1. American Cancer Society. Cancer Facts and Figures 2009[M]. Atlanta(GA): American Cancer Society, 2009:4-10.
2. American Cancer Society. Cancer Facts and Figures 2008[M]. Atlanta(GA): American Cancer Society, 2008:4-10.
3. Kirkali Z, Chan T, Murugesan M, et al. Bladder cancer: epidemiology, staging and grading, and diagnosis[J]. Urology 2005;66(Suppl 6A):4–34.
4. Matthew J. Hayat, Nadia howlader, Marsha E.Reichman, Brendak. E dwards. Cancer Statistics, Trends, and Multiple Primary Cancer Analyses from the Surveillance, Epidemiology, and End Results (SEER) Program[J]. The Oncologist, 2007,12:20-37.
5. Ries LA, Eisner MP, Kosary CL, et al. SEER Cancer Statistics Review,1975-2000[M]. Bethesda: National Cancer Institute, 2003.
6. Bladder Cancer Clinical Guideline Update Panel. Bladder Cancer: Guideline for the Management of Nonmuscle Invasive Bladder Cancer: (Stages Ta,T1, and Tis):2007 Update[M]. American Urological Association, 2007:3-4.
7. The top ten malignancy mortality rates of China (man) [EB/OL], http://61.49.18.65/publicfiles/business/htmlfiles/zwgkzt/ptjnj/year2008/9.htm.
8.中国泌尿外科疾病诊断治疗指南编辑委员会.中国泌尿外科疾病诊断治疗指南(2007版)[M].中华医学会泌尿外科学分会, 2007:90-104.
9. Marco Grasso. Bladder Cancer: A Major Public Health Issue[J]. European urology, 2008;7:510-515.
10. Ifeanyichukwu I. Megwalu , Anna Vlahiotis, Mohamed Radwan, et al. Prognostic Impact of Comorbidity in Patients with Bladder Cancer[J]. European urology, 2008; 53:581–589.
11. National Comprehensive Cancer Network. Bladder Cancer (Including Upper Tract Tumors and Urothelial Carcinoma of the Prostate) V.I. 2010 [M]. National Comprehensive Cancer Network,2010:1-3(MS).
12. Ercan Kocakoc, Adem Kiris, Irfan Orhan, et al. Detection of Bladder Tumors with 3-Dimensional Sonography and Virtual Sonographic Cystoscopy[J]. J Ultrasound Med, 2008; 27:45–53.
13. Kim JK, Ahn JH, Park T, et al. Cystoscopy of the contrast material-filled bladder in patients with gross hematuria[J]. AJR Am J Roentgenol, 2002;179: 763-768.
14. Halil Arslan, Kadir Ceylan, Mustafa Harman, et al. Virtual Computed Tomography Cystoscopy in Bladder Pathologies[J]. Clinical Urology,2006; 32(2): 147-154.
15.于甬华,王仁本,于金明等.计算机辅助诊断在医学影像学领域的研究进展[J],生物医学工程研究,2005;24(2):133-136..
16. Barbara Hermes, Felix Sp?ler, Amjad Naami, et al. Visualization of the Basement Membrane Zone of the Bladder by Optical Coherence Tomography: Feasibility of Noninvasive Evaluation of Tumor Invasion[J]. UROLOGY, 2008;72(3):677-681.
17. National Comprehensive Cancer Network. Bladder Cancer (Including Upper Tract Tumors and Urothelial Carcinoma of the Prostate) V.1.2007 [M]. National Comprehensive Cancer Network,2007:1-3(MS).
18. Sylvester RJ. Natural history, recurrence, and progression in superficial bladder cancer[J]. Scientific World J,2006;27:2617–25.
19. John T. Leppert , Oleg Shvarts , Kelly Kawaoka ,et al. Prevention of Bladder Cancer: A Review[J]. European urology, 2006;49: 226–234.
20. M.H.Horing. Txture feature coding method for texture classification[J]. Optical Engineering,2003;43(1): 228-237.
21.张勇,王瓛,郭秀花.基于医学图像纹理特征的疾病诊断方法及其应用现状[J].北京生物医学工程,2008;27(4):543-544.
22. Balaji Ganeshan, Kenneth A. Miles, Rupert C.D. Young, et al. Texture analysis in non-contrast enhanced CT: Impact of malignancy on texture in apparently disease-free areas of the liver[J]. Eur J Radiol,2009;70:101-110.
23. ROBERT M,HAWLICK. Statistical and Structural Approaches to Texture[J]. PROCEEDINGS OF THE IEEE, 1979;67(5): 786-804.
24.丁玲,王崇俊,杨育彬,等.基于图像能量谱直方图的纹理检索算法RAH[J].计算机科学,2005; 32(1):194-201.
25. Wen-Jie Wu, Woo Kyung Moon. Ultrasound Breast Tumor Image Computer-Aided Diagnosis With Texture and Morphological Features[J]. Acad Radiol, 2008;15:873–880.
26. A. Karahaliou, I. Boniatis, P. Sakellaropoulos, et al. Can texture of tissue surrounding microcalcifications in mammography be used for breast cancer diagnosis?[J]. Nuclear Instruments and Methods in Physics Research A, 2007; 580: 1071–1074.
27. H.S. Sheshadri , A. Kandaswamy. Experimental investigation on breast tissue classification based on statistical feature extraction of mammograms[J]. Computerized Medical Imaging and Graphics, 2007;31: 46–48.
28. Shao-Jer Chen , Kuo-Sheng Cheng, Yen-Ting Chen, et al. Quantitative correlation between sonographic textural feature and histopathologicalcomponents for breast cancer: preliminary results[J]. Clinical Imaging, 2008;32:93–102.
29. Radhika Sivaramakrishna, Kimerly A. Powell, Michael L. Lieber. Texture analysis of lesions in breast ultrasound images[J]. Computerized Medical Imaging and Graphics, 2002; 26: 303-307.
30. Hui Li, Maryellen L. Giger, Olufunmilayo I. Olopade, et al. Computerized texture analysis of mammographic parenchymal patterns of digitized mammograms[J]. International Congress Series, 2004;1268: 878–881.
31. Omar S. Al-Kadi, D. Watson. Texture Analysis of Aggressive and Nonaggressive Lung Tumor CE CT Images[J]. IEEE Transactions On Biomedical Engineering, 2008;55 (7): 1822-1830.
32. Balaji Ganeshan, Kenneth A. Miles, Rupert C.D. Young, et al. Texture analysis in non-contrast enhanced CT: Impact of malignancy on texture in apparently disease-free areas of the liver[J]. Eur J Radiol, 2009;70:101-110..
33. Balaji Ganeshan, Kenneth A. Miles, Rupert C.D. Young, et al. Hepatic Enhancement in Colorectal Cancer: Texture Analysis Correlates with Hepatic Hemodynamics and Patient Survival[J]. Acad Radiol 2007;14: 1520–1530.
34. Shao-Jer Chen, Sung-Nien Yu, Jeh-En Tzeng, et al. Characterization of the major histopathological components of Thyroid Nodules using sonographic textural features for clinical diagnosis and management[J]. Ultrasound Med Biol.2009;35:201-208.
35. I. El Naqa, P.W. Grigsby, A. Apte, et al. Exploring feature-based approaches in PET images for predicting cancer treatment outcomes. Pattern Recognition, 2009;42:1162–1171.
36. H. Tamura, S. Mori, and T. Yamawaki. Texture features corresponding tovisual perception[J].IEEE Trans. On Systems, Man, and Cybernetics, 1978;Smc-8(6):460-473.
37.王志芳,刘玉红,王颖,等.基于数字图像处理的人类视觉对比度分辨率限制测定[J].生物医学工程学杂志, 2008;25(5):998-1002.
38. H Ji-guang. Gray Level-gradient Co-occurrence Matrix Texture Analysis Method [J]. Acta Automatica Sinica, 1984;10(1):22-25.
39.胡小锋,赵辉. Visual C++/Matlab图像处理与识别实例案例精选[M].北京:人民邮电出版社,2004,四维科技.
40. Robert M.Haralick, K.Shanmugam, Its'hak Dinstein. Textural Features for Image Classification[J]. IEEE transactions on systems,man and cybernetics, 1973, SMC-3(6): PP.610-621.
41. R.W. Conners, M.M. Trivedi, C.A. Harlow.”Segmentation of a High-Resolution Urban Scene Using Texture Operators[J]. Computer Vision, Graphics, and Image Processing, 1984,25:273-310.
42.王天,张国鹏,刘欣,张军英,卢虹冰.虚拟结肠镜的计算机辅助检测研究[J].中国生物医学工程学报, 2008;27(1):146-151.
43. Colleen A. Lingley-Papadopoulos, Murray H. Loew, Jason M. Zara. Real-time bladder-layer recognition: an approach to optical biopsy[J]. ISBI,2007: 1279-1279.
44. William H. Nailon, Anthony T. Redpath , Duncan B. McLaren. Texture analysis of 3D bladder cancer CT images for improving radiotherapy planning[J]. ISBI 2008: 652-655.
45. 18. Victor e. Reuter. The pathology of bladder cancer[J]. UROLOGY, 2006,67 (Suppl 3A): 11–18.
46.李丙春,耿国华,周明全等.一个医学图像分类器的设计[J].计算机工程与应用, 2004,17:230-232.
47. Marti Hearst. Support vector machines[J]. IEEE INTELLIGENT SYSTEMS, 1998 ,JULY/AUGUST:18-28.
48.王浩初,许顺良,樊树峰. 3D- FIESTA和FRFSE T2WI对后颅窝神经成像效能的对比研究[J].中国临床医学影像杂志,2008(19),6:385-397.
49. Bram van Ginneken, Bart M. ter Haar Romeny, Max A. Viergever. Computer-Aided Diagnosis in Chest Radiography:A Survey[J]. IEEE TRANSACTIONS ON MEDICAL IMAGING, 2001;20:1228-1240.
50.辛民宣,史正星,崔光彬等.基于内容的医学图像检索中对Tamura纹理特征的算法改进[J].医疗卫生装备, 2009,31(2):32-35.
51.史正星,付强,张国鹏等.不规则图像灰度共生矩阵生成方式的比较[J].医疗卫生装备, 2009;31(4):
