Angiomyolipoma with Minimal Fat: Differentiation From Clear Cell Renal Cell Carcinoma and Papillary Renal Cell Carcinoma by Texture Analysis on CT Images Lifen Yan, MD, Zaiyi Liu, MD, PhD, Guangyi Wang, MD, PhD, Yanqi Huang, MD, Yubao Liu, MD, PhD, Yuanxin Yu, MD, Changhong Liang, MD, PhD Rationale and Objectives: To retrospectively evaluate the diagnostic performance of texture analysis (TA) for the discrimination of angiomyolipoma (AML) with minimal fat, clear cell renal cell cancer (ccRCC), and papillary renal cell cancer (pRCC) on computed tomography (CT) images and to determine the scanning phase, which contains the strongest discriminative power. Materials and Methods: Patients with pathologically proved AMLs (n = 18) lacking visible macroscopic fat at CT and patients with pathologically proved ccRCCs (n = 18) and pRCCs (n = 14) were included. All patients underwent CT scan with three phases (precontrast phase [PCP], corticomedullary phase [CMP], and nephrographic phase [NP]). The selected images were analyzed and classified with TA software (MaZda). Texture classification was performed for 1) minimal fat AML versus ccRCC, 2) minimal fat AML versus pRCC, and 3) ccRCC versus pRCC. The classification results were arbitrarily divided into several levels according to the misclassification rates: excellent (misclassification rates #10%), good (10%< misclassification rates #20%), moderate (20%< misclassification rates #30%), fair (30%< misclassification rates #40%), and poor (misclassification rates $40%). Results: Excellent classification results (error of 0.00%–9.30%) were obtained with nonlinear discriminant analysis for all the three groups, no matter which phase was used. On comparison of the three scanning phases, we observed a trend toward better lesion classification with PCP for minimal fat AML versus ccRCC, CMP, and NP images for ccRCC versus pRCC and found similar discriminative power for minimal fat AML versus pRCC. Conclusions: TA might be a reliable quantitative method for the discrimination of minimal fat AML, ccRCC, and pRCC. Key Words: Texture analysis; computed tomography; angiomyolipoma; renal cell carcinoma. ªAUR, 2015
A
ngiomyolipoma (AML) as the most common benign solid renal tumor is not difficult to be diagnosed when macroscopic fat is appeared, but diagnosis is challenging for AML with minimal fat (1,2). Approximately 10%–17% of benign renal tumors are surgically resected (3,4), and AMLs account for 18%–59% of the excised benign tumors (3–5). In this regard, accurate differential diagnosis of minimal fat AML from renal cell cancer (RCC) is crucial to avoid unnecessary surgery. In previous studies, investigators have described some imaging features that are highly suggestive of minimal fat AML, such as high attenuation at unenhanced computed tomography (CT) with homogeneous prolonged enhancement (6,7), a small renal mass with homogeneous low signal intensity (SI) on T2weighted images (2,8), and the presence of microscopic fat at in- and opposed-phase images (2,9); however, imaging
Acad Radiol 2015; 22:1115–1121 From the Department of Radiology, Guangdong General Hospital, Guangdong Academy of Medical Sciences, 106 Zhong Shan Er Lu, Guangzhou, Guangdong Province 510080, China. Received September 27, 2014; accepted April 17, 2015. Funding Source: This work was supported by National Scientific Foundation of China (81271654 and 81271569). Address correspondence to: C.-H. L. e-mail: [email protected] ªAUR, 2015 http://dx.doi.org/10.1016/j.acra.2015.04.004
characteristics can be variable while clear cell RCC (ccRCC) often contains microscopic fat with decreasing SI on opposedphase images compared to in-phase images (8); papillary RCC (pRCC) often shows low T2 SI and homogeneous and gradual enhancement at CT or magnetic resonance (MR) images (5,6,10). In a word, there are no reliable imaging features to differentiate minimal fat AML from RCC. Recently, some quantitative techniques have been developed to detect subtle changes in tissues. Attenuation measurement histogram analysis is one of the quantitative methods which is used to calculate the percentage of pixels with negative CT numbers in the regions of interest (ROI). The use of attenuation measurement histogram analysis is advocated by Kim et al. (11,12) who reported that attenuation measurement histogram analysis can be used to detect a small amount of fat, which is helpful in distinguishing minimal fat AML from RCC. But this point is disputed by Chaudhry et al. (13,14). Texture analysis (TA) as another quantitative technique can also be used to characterize the properties of tissues. TA allows the computation of hundreds of texture features based on the SI of pixels in the ROI (15,16), which may yield potential information. Successful applications of TA have been reported in discriminating pathologic stages of hepatic fibrosis (15) and glioma (17), differentiating invasive ductal carcinoma from 1115
YAN ET AL
invasive lobular carcinoma (18) and hepatic cysts from hemangiomas (19), evaluating the therapeutic efficacy of metastatic RCC (20), non-Hodgkin lymphoma (21), and mammary carcinomas (22). Moreover, TA can be performed with commercial software and does not require highly specialized computer knowledge. It is easy for radiologists to obtain analysis results. To the best of our knowledge, TA has not been performed in differentiating minimal fat AML from RCC. Thus, the purpose of our study was to retrospectively evaluate the diagnostic performance of TA for the discrimination of minimal fat AML, ccRCC, and pRCC on CT images and to determine the scanning phase which contains the strongest discriminative power.
MATERIAL AND METHODS The institutional review board approved this retrospective study and waived the requirement for informed consent. Patient Selection
Through a computerized search of institutional pathology database from January 2008 to April 2014, we identified 66 patients with histologically confirmed renal AML obtained by biopsy or surgical resection. After reviewing the digital images, 29 examinations were excluded because of an absence of preoperative CT (n = 25), only unenhanced CT (n = 1) or monophasic contrast-enhanced CT (n = 3) was performed. Two experienced radiologists with 36 and 16 years experience in abdominal imaging diagnosis independently reviewed the unenhanced CT images to excluded masses with macroscopic fat (n = 21). This review resulted in 18 minimal fat AMLs (AMLs without macroscopic fat at the unenhanced CT images) in 16 patients (1 patient with multiple pathologically confirmed AML and 2 of them were classified as minimal fat AMLs; 1 patient with 2 lesions of minimal fat AML in the right kidney), which were included in our study (14 women and 2 men; mean age, 44.5 years; age range, 26–61 years). The mean maximum diameter of the masses was 28.47 mm (range, 8–51 mm). To establish a size- and number-matched ccRCC and pRCC cohorts, we did a reverse search for ccRCCs and pRCCs in the institutional pathology database beginning in April 2014. The inclusion criteria were as follows: 1) the maximum diameter of tumor was less than 51 mm; 2) the patients underwent CT scan with three phases (the precontrast phase [PCP], the corticomedullary phase [CMP], and the nephrographic phase [NP]) in our department. We identified 18 patients with ccRCC (mean age, 53.9 years; age range, 36–79 years; mean maximum diameter, 33.22 mm; range, 15–49 mm) during March 2013 to April 2014, but only 14 patients with pRCC (mean age, 57.6 years; age range, 34–77 years; mean maximum diameter, 33.09 mm; range, 14–51 mm) during January 2008 to April 2014. We did not continue to search the cases with pRCC because the CT digital data of 2008 years ago was not stored in our 1116
Academic Radiology, Vol 22, No 9, September 2015
Radiology information system/picture archiving and communication systems (RIS/PACS). CT Examination
All patients underwent CT scan on a 64-section multidetector CT unit (LightSpeed VCT, GE Healthcare). The scanning protocol included data acquisition in three phases: PCP, CMP (30-second delay after contrast injection), and NP (90-second delay after contrast injection). Ninety to one hundred milliliter contrast material (Ultravist 370, Bayer ScheringPharma AG, Germany; Iopamiro 370, Bracco, Italy) was administered into an antecubital vein with a power injector at a rate of 3.5 mL/s. The scanning parameters for all phases were as follows: tube voltage, 120 kV; tube current, a weight-based automated tube-current modulation; collimation, 128 0.625 mm; rotation time, 0.6 ms; pitch, 0.984:1; and field of view, 360 mm. Conventional Imaging Analysis
CT Characteristics Analysis. Two radiologists (Y.-X. Y. and Y.-B. L. with 14 and 16 years experience in abdominal imaging diagnosis, respectively) who were unaware of the pathologic diagnosis reviewed the CT images in consensus at our RIS/PACS (Carestream Health, China). They evaluated the tumor attenuation on all the three phases (PCP, CMP, and NP) and enhancement degree on CMP and NP images. The attenuation values and the degree of enhancement were determined by a ROI. The enhancement degree was calculated by the difference between unenhanced images and enhanced images (CMP and NP images). Subjective parameters including tumor attenuation in comparison to the surrounding renal parenchyma and enhancement pattern over time were also analyzed. Tumor attenuation was subjectively classified as hypoattenuation, isoattenuation, or hyperattenuation compared to that of the surrounding renal parenchyma. The enhancement pattern over time was classified as an early washout pattern (a tumor showed peak enhancement in CMP and then demonstrated a washout of at least 20 HU in NP), a gradual enhancement pattern (the tumor attenuation value in NP was at least 20 HU greater than that in CMP), and a prolonged enhancement pattern (the difference of tumor attenuation between CMP and NP ranged from 20 HU to 20 HU). Statistical Analysis. All data analysis was performed using SPSS (version 19, SPSS Inc., Chicago, IL). A P value < .05 was considered to indicate a significant difference. The independent samples t test was used to compare tumor attenuation in PCP, CMP, and NP and enhancement degree in CMP and NP between minimal fat AML and ccRCC, minimal fat AML and pRCC, and ccRCC and pRCC. The chi-squared analysis was used to compare the subjective parameters (tumor attenuation in comparison to the surrounding renal parenchyma in PCP and enhancement pattern over time).
Academic Radiology, Vol 22, No 9, September 2015
DIFFERENTIATE AML FROM RCC BY TA
Figure 1. PCP (a), CMP (b), and NP (c) images of a patient with pathologically proved AML. A manually defined irregular ROI is drawn in the tumor area, the line is drawn carefully to maintain an approximate distance of 2–3 mm from the tumor margin. AML, angiomyolipoma; CMP, corticomedullary phase; NP, nephrographic phase; PCP, precontrast phase; ROI, regions of interest. (Color version of figure is available online.)
TA and Feature Selection
Image Selection. We chose PCP, CMP, and NP images for TA. The images were anonymized and stored in BMP format on our RIS/PACS (Carestream Health, China). One to four consecutive axial image slices of each phase were chosen on the basis of maximum diameter of tumors and optimal representation of the largest tumor area. The slice thickness for all examinations was 5 mm, and the window width and level were 360HU and 60HU, respectively. ROI Definition. Texture parameter calculation as the first stage of TA was performed with the software package MaZda (16,23) (version 4.6, available at http://www.eletel.p.lodz.pl/ mazda/). A manually defined irregular ROI was placed on all the selected images (Fig 1). To minimize volume averaging with surrounding renal parenchyma and fat, the line was carefully drawn while trying to maintain an approximate distance of 2–3 mm from the tumor margin by referring to contrastenhanced images. Texture Feature Calculation and Selection. The texture parameters were derived, respectively, from image histogram (information about the intensity of pixels and without any spatial relations between the pixels on the image), gradient (information about the image intensity distribution and describes the histogram of the absolute gradient values of 3 3 neighborhoods of pixels in the ROI), run-length matrix (information about pixel runs with the specified gray-level values in a given direction and describes intensity homogeneity in specific directions in the ROI), co-occurrence matrix (information about changes of SI with increasing distance, describes the gray-level value distribution of pixel pairs along all directions at different distances in the ROI), autoregressive model (description of texture based on the statistical correlation between neighboring pixels), and wavelet transform (information about the frequency of similar SIs and describes the wavelet transform of the pixels in the ROI) (16,18,21). To minimize the influence of contrast and brightness variation, image gray-level intensity normalization was performed using a method that normalizes image intensities in the range (m 3s, m + 3s; m, mean gray level value; s, stan-
dard deviation). Feature selection algorithms (Fisher coefficient [Fisher], mutual information [MI], and classification error probability combined with average correlation coefficients [POE + ACC]) were used to identify texture features with the highest discriminative power for classification. Ten features were chosen by each method automatically. Tissue Classification. Texture data analysis and classification was carried out by B11 application (version 4.6) of MaZda software package. Raw data analysis (RDA), principal component analysis (PCA), linear discriminant analysis (LDA), and nonlinear discriminant analysis (NDA) were run for the best texture features selected with the combination of Fisher, MI, and POE + ACC. K-nearest-neighbor classifier was performed for features resulting from RDA, PCA, and LDA. The artificial neural network classifier was used for features resulting from NDA. Texture classification for two disease combinations was performed for 1) minimal fat AML versus ccRCC, 2) minimal fat AML versus pRCC, and 3) ccRCC versus pRCC. The classification results were arbitrarily divided into several levels according to the misclassification rates: excellent (misclassification rates #10%), good (10%< misclassification rates #20%), moderate (20%< misclassification rates #30%), fair (30%< misclassification rates #40%), and poor (misclassification rates $40%). RESULTS Conventional Imaging Analysis
The data regarding tumor characteristics (tumor attenuation in PCP, CMP, and NP, enhancement degree in CMP and NP, and enhancement pattern over time) are described in Tables 1 and 2. No significant difference of tumor attenuation in PCP was found for all three comparisons (minimal fat AML vs. ccRCC, minimal fat AML vs. pRCC, and ccRCC vs. pRCC). The attenuation values and enhancement degree of minimal fat AML and ccRCC in CMP and NP were significantly higher than those of pRCC, but no significant difference was found between minimal fat AML and ccRCC. 1117
YAN ET AL
Academic Radiology, Vol 22, No 9, September 2015
TABLE 1. Attenuation Values and Enhancement Degree of Minimal Fat AML, ccRCC, and pRCC P
Parameter PCP attenuation CMP attenuation NP attenuation Enhancement degree (CMP) Enhancement degree (NP)
pRCC (n = 14)
Minimal Fat AML versus ccRCC
Minimal Fat AML versus pRCC
ccRCC versus pRCC
32.11 6.99 129.22 47.54 110.44 29.48 97.11 45.77
34.00 7.59 53.86 16.18 70.86 12.57 19.86 14.62
.476 .260 .874 .431
.465 .008 .013 .002
.926 .000 .005 .000
78.33 27.75
36.86 10.25
.833
.004
.007
Minimal Fat AML (n = 18)
ccRCC (n = 18)
43.89 5.49 109.44 39.83 95.89 29.48 65.56 39.98 52.00 29.03
AML, angiomyolipoma; ccRCC, clear cell RCC; CMP, corticomedullary phase; NP, nephrographic phase; pRCC, papillary RCC; PCP, precontrast phase; RCC, renal cell cancer. Data are means standard deviations in Hounsfield units.
TABLE 2. Subjective Analysis of Tumor Attenuation in Comparison to the Surrounding Renal Parenchyma and Enhancement Pattern P Value
Parameter Attenuation Hypoattenuation Isoattenuation Hyperattenuation Enhancement pattern Early washout Gradual Prolonged
Minimal Fat AML (n = 18)
ccRCC (n = 18)
pRCC (n = 14)
0 (0.0) 3 (16.7) 15 (83.3)
6 (33.3) 8 (44.4) 4 (22.2)
4 (28.6) 3 (21.4) 7 (50.0)
8 (44.4) 3 (16.7) 7 (38.9)
11 (61.1) 3 (16.7) 4 (22.2)
0 (0.0) 8 (57.1) 6 (42.9)
Minimal Fat AML Versus ccRCC
Minimal Fat AML Versus pRCC
ccRCC Versus pRCC
A
ngiomyolipoma (AML) as the most common benign solid renal tumor is not difficult to be diagnosed when macroscopic fat is appeared, but diagnosis is challenging for AML with minimal fat (1,2). Approximately 10%–17% of benign renal tumors are surgically resected (3,4), and AMLs account for 18%–59% of the excised benign tumors (3–5). In this regard, accurate differential diagnosis of minimal fat AML from renal cell cancer (RCC) is crucial to avoid unnecessary surgery. In previous studies, investigators have described some imaging features that are highly suggestive of minimal fat AML, such as high attenuation at unenhanced computed tomography (CT) with homogeneous prolonged enhancement (6,7), a small renal mass with homogeneous low signal intensity (SI) on T2weighted images (2,8), and the presence of microscopic fat at in- and opposed-phase images (2,9); however, imaging
Acad Radiol 2015; 22:1115–1121 From the Department of Radiology, Guangdong General Hospital, Guangdong Academy of Medical Sciences, 106 Zhong Shan Er Lu, Guangzhou, Guangdong Province 510080, China. Received September 27, 2014; accepted April 17, 2015. Funding Source: This work was supported by National Scientific Foundation of China (81271654 and 81271569). Address correspondence to: C.-H. L. e-mail: [email protected] ªAUR, 2015 http://dx.doi.org/10.1016/j.acra.2015.04.004
characteristics can be variable while clear cell RCC (ccRCC) often contains microscopic fat with decreasing SI on opposedphase images compared to in-phase images (8); papillary RCC (pRCC) often shows low T2 SI and homogeneous and gradual enhancement at CT or magnetic resonance (MR) images (5,6,10). In a word, there are no reliable imaging features to differentiate minimal fat AML from RCC. Recently, some quantitative techniques have been developed to detect subtle changes in tissues. Attenuation measurement histogram analysis is one of the quantitative methods which is used to calculate the percentage of pixels with negative CT numbers in the regions of interest (ROI). The use of attenuation measurement histogram analysis is advocated by Kim et al. (11,12) who reported that attenuation measurement histogram analysis can be used to detect a small amount of fat, which is helpful in distinguishing minimal fat AML from RCC. But this point is disputed by Chaudhry et al. (13,14). Texture analysis (TA) as another quantitative technique can also be used to characterize the properties of tissues. TA allows the computation of hundreds of texture features based on the SI of pixels in the ROI (15,16), which may yield potential information. Successful applications of TA have been reported in discriminating pathologic stages of hepatic fibrosis (15) and glioma (17), differentiating invasive ductal carcinoma from 1115
YAN ET AL
invasive lobular carcinoma (18) and hepatic cysts from hemangiomas (19), evaluating the therapeutic efficacy of metastatic RCC (20), non-Hodgkin lymphoma (21), and mammary carcinomas (22). Moreover, TA can be performed with commercial software and does not require highly specialized computer knowledge. It is easy for radiologists to obtain analysis results. To the best of our knowledge, TA has not been performed in differentiating minimal fat AML from RCC. Thus, the purpose of our study was to retrospectively evaluate the diagnostic performance of TA for the discrimination of minimal fat AML, ccRCC, and pRCC on CT images and to determine the scanning phase which contains the strongest discriminative power.
MATERIAL AND METHODS The institutional review board approved this retrospective study and waived the requirement for informed consent. Patient Selection
Through a computerized search of institutional pathology database from January 2008 to April 2014, we identified 66 patients with histologically confirmed renal AML obtained by biopsy or surgical resection. After reviewing the digital images, 29 examinations were excluded because of an absence of preoperative CT (n = 25), only unenhanced CT (n = 1) or monophasic contrast-enhanced CT (n = 3) was performed. Two experienced radiologists with 36 and 16 years experience in abdominal imaging diagnosis independently reviewed the unenhanced CT images to excluded masses with macroscopic fat (n = 21). This review resulted in 18 minimal fat AMLs (AMLs without macroscopic fat at the unenhanced CT images) in 16 patients (1 patient with multiple pathologically confirmed AML and 2 of them were classified as minimal fat AMLs; 1 patient with 2 lesions of minimal fat AML in the right kidney), which were included in our study (14 women and 2 men; mean age, 44.5 years; age range, 26–61 years). The mean maximum diameter of the masses was 28.47 mm (range, 8–51 mm). To establish a size- and number-matched ccRCC and pRCC cohorts, we did a reverse search for ccRCCs and pRCCs in the institutional pathology database beginning in April 2014. The inclusion criteria were as follows: 1) the maximum diameter of tumor was less than 51 mm; 2) the patients underwent CT scan with three phases (the precontrast phase [PCP], the corticomedullary phase [CMP], and the nephrographic phase [NP]) in our department. We identified 18 patients with ccRCC (mean age, 53.9 years; age range, 36–79 years; mean maximum diameter, 33.22 mm; range, 15–49 mm) during March 2013 to April 2014, but only 14 patients with pRCC (mean age, 57.6 years; age range, 34–77 years; mean maximum diameter, 33.09 mm; range, 14–51 mm) during January 2008 to April 2014. We did not continue to search the cases with pRCC because the CT digital data of 2008 years ago was not stored in our 1116
Academic Radiology, Vol 22, No 9, September 2015
Radiology information system/picture archiving and communication systems (RIS/PACS). CT Examination
All patients underwent CT scan on a 64-section multidetector CT unit (LightSpeed VCT, GE Healthcare). The scanning protocol included data acquisition in three phases: PCP, CMP (30-second delay after contrast injection), and NP (90-second delay after contrast injection). Ninety to one hundred milliliter contrast material (Ultravist 370, Bayer ScheringPharma AG, Germany; Iopamiro 370, Bracco, Italy) was administered into an antecubital vein with a power injector at a rate of 3.5 mL/s. The scanning parameters for all phases were as follows: tube voltage, 120 kV; tube current, a weight-based automated tube-current modulation; collimation, 128 0.625 mm; rotation time, 0.6 ms; pitch, 0.984:1; and field of view, 360 mm. Conventional Imaging Analysis
CT Characteristics Analysis. Two radiologists (Y.-X. Y. and Y.-B. L. with 14 and 16 years experience in abdominal imaging diagnosis, respectively) who were unaware of the pathologic diagnosis reviewed the CT images in consensus at our RIS/PACS (Carestream Health, China). They evaluated the tumor attenuation on all the three phases (PCP, CMP, and NP) and enhancement degree on CMP and NP images. The attenuation values and the degree of enhancement were determined by a ROI. The enhancement degree was calculated by the difference between unenhanced images and enhanced images (CMP and NP images). Subjective parameters including tumor attenuation in comparison to the surrounding renal parenchyma and enhancement pattern over time were also analyzed. Tumor attenuation was subjectively classified as hypoattenuation, isoattenuation, or hyperattenuation compared to that of the surrounding renal parenchyma. The enhancement pattern over time was classified as an early washout pattern (a tumor showed peak enhancement in CMP and then demonstrated a washout of at least 20 HU in NP), a gradual enhancement pattern (the tumor attenuation value in NP was at least 20 HU greater than that in CMP), and a prolonged enhancement pattern (the difference of tumor attenuation between CMP and NP ranged from 20 HU to 20 HU). Statistical Analysis. All data analysis was performed using SPSS (version 19, SPSS Inc., Chicago, IL). A P value < .05 was considered to indicate a significant difference. The independent samples t test was used to compare tumor attenuation in PCP, CMP, and NP and enhancement degree in CMP and NP between minimal fat AML and ccRCC, minimal fat AML and pRCC, and ccRCC and pRCC. The chi-squared analysis was used to compare the subjective parameters (tumor attenuation in comparison to the surrounding renal parenchyma in PCP and enhancement pattern over time).
Academic Radiology, Vol 22, No 9, September 2015
DIFFERENTIATE AML FROM RCC BY TA
Figure 1. PCP (a), CMP (b), and NP (c) images of a patient with pathologically proved AML. A manually defined irregular ROI is drawn in the tumor area, the line is drawn carefully to maintain an approximate distance of 2–3 mm from the tumor margin. AML, angiomyolipoma; CMP, corticomedullary phase; NP, nephrographic phase; PCP, precontrast phase; ROI, regions of interest. (Color version of figure is available online.)
TA and Feature Selection
Image Selection. We chose PCP, CMP, and NP images for TA. The images were anonymized and stored in BMP format on our RIS/PACS (Carestream Health, China). One to four consecutive axial image slices of each phase were chosen on the basis of maximum diameter of tumors and optimal representation of the largest tumor area. The slice thickness for all examinations was 5 mm, and the window width and level were 360HU and 60HU, respectively. ROI Definition. Texture parameter calculation as the first stage of TA was performed with the software package MaZda (16,23) (version 4.6, available at http://www.eletel.p.lodz.pl/ mazda/). A manually defined irregular ROI was placed on all the selected images (Fig 1). To minimize volume averaging with surrounding renal parenchyma and fat, the line was carefully drawn while trying to maintain an approximate distance of 2–3 mm from the tumor margin by referring to contrastenhanced images. Texture Feature Calculation and Selection. The texture parameters were derived, respectively, from image histogram (information about the intensity of pixels and without any spatial relations between the pixels on the image), gradient (information about the image intensity distribution and describes the histogram of the absolute gradient values of 3 3 neighborhoods of pixels in the ROI), run-length matrix (information about pixel runs with the specified gray-level values in a given direction and describes intensity homogeneity in specific directions in the ROI), co-occurrence matrix (information about changes of SI with increasing distance, describes the gray-level value distribution of pixel pairs along all directions at different distances in the ROI), autoregressive model (description of texture based on the statistical correlation between neighboring pixels), and wavelet transform (information about the frequency of similar SIs and describes the wavelet transform of the pixels in the ROI) (16,18,21). To minimize the influence of contrast and brightness variation, image gray-level intensity normalization was performed using a method that normalizes image intensities in the range (m 3s, m + 3s; m, mean gray level value; s, stan-
dard deviation). Feature selection algorithms (Fisher coefficient [Fisher], mutual information [MI], and classification error probability combined with average correlation coefficients [POE + ACC]) were used to identify texture features with the highest discriminative power for classification. Ten features were chosen by each method automatically. Tissue Classification. Texture data analysis and classification was carried out by B11 application (version 4.6) of MaZda software package. Raw data analysis (RDA), principal component analysis (PCA), linear discriminant analysis (LDA), and nonlinear discriminant analysis (NDA) were run for the best texture features selected with the combination of Fisher, MI, and POE + ACC. K-nearest-neighbor classifier was performed for features resulting from RDA, PCA, and LDA. The artificial neural network classifier was used for features resulting from NDA. Texture classification for two disease combinations was performed for 1) minimal fat AML versus ccRCC, 2) minimal fat AML versus pRCC, and 3) ccRCC versus pRCC. The classification results were arbitrarily divided into several levels according to the misclassification rates: excellent (misclassification rates #10%), good (10%< misclassification rates #20%), moderate (20%< misclassification rates #30%), fair (30%< misclassification rates #40%), and poor (misclassification rates $40%). RESULTS Conventional Imaging Analysis
The data regarding tumor characteristics (tumor attenuation in PCP, CMP, and NP, enhancement degree in CMP and NP, and enhancement pattern over time) are described in Tables 1 and 2. No significant difference of tumor attenuation in PCP was found for all three comparisons (minimal fat AML vs. ccRCC, minimal fat AML vs. pRCC, and ccRCC vs. pRCC). The attenuation values and enhancement degree of minimal fat AML and ccRCC in CMP and NP were significantly higher than those of pRCC, but no significant difference was found between minimal fat AML and ccRCC. 1117
YAN ET AL
Academic Radiology, Vol 22, No 9, September 2015
TABLE 1. Attenuation Values and Enhancement Degree of Minimal Fat AML, ccRCC, and pRCC P
Parameter PCP attenuation CMP attenuation NP attenuation Enhancement degree (CMP) Enhancement degree (NP)
pRCC (n = 14)
Minimal Fat AML versus ccRCC
Minimal Fat AML versus pRCC
ccRCC versus pRCC
32.11 6.99 129.22 47.54 110.44 29.48 97.11 45.77
34.00 7.59 53.86 16.18 70.86 12.57 19.86 14.62
.476 .260 .874 .431
.465 .008 .013 .002
.926 .000 .005 .000
78.33 27.75
36.86 10.25
.833
.004
.007
Minimal Fat AML (n = 18)
ccRCC (n = 18)
43.89 5.49 109.44 39.83 95.89 29.48 65.56 39.98 52.00 29.03
AML, angiomyolipoma; ccRCC, clear cell RCC; CMP, corticomedullary phase; NP, nephrographic phase; pRCC, papillary RCC; PCP, precontrast phase; RCC, renal cell cancer. Data are means standard deviations in Hounsfield units.
TABLE 2. Subjective Analysis of Tumor Attenuation in Comparison to the Surrounding Renal Parenchyma and Enhancement Pattern P Value
Parameter Attenuation Hypoattenuation Isoattenuation Hyperattenuation Enhancement pattern Early washout Gradual Prolonged
Minimal Fat AML (n = 18)
ccRCC (n = 18)
pRCC (n = 14)
0 (0.0) 3 (16.7) 15 (83.3)
6 (33.3) 8 (44.4) 4 (22.2)
4 (28.6) 3 (21.4) 7 (50.0)
8 (44.4) 3 (16.7) 7 (38.9)
11 (61.1) 3 (16.7) 4 (22.2)
0 (0.0) 8 (57.1) 6 (42.9)
Minimal Fat AML Versus ccRCC
Minimal Fat AML Versus pRCC
ccRCC Versus pRCC
