KR101824691B1 - Method for differentiation of fat-poor angiomyolipoma from clear cell renal cell carcinoma in contrast-enhanced mdct images using quantitative feature classification - Google Patents

Abstract

A method for distinguishing an aneurysmal lipoma and a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detector computed tomography image according to an embodiment of the present invention includes: (a) Extracting histogram features, brightness value percent features, and texture features from the isolated kidney tumor region in multi-detected computed tomography (MDCT) training images; (b) applying the extracted features to a combination of a plurality of feature selection methods and a plurality of machine learning-based classification methods to classify features and classification methods that exhibit the best classification performance as optimal classification features and optimal machine learning classification Selecting as a method; (c) applying the selected classification features to the selected machine learning classification method to obtain a classification model; (d) extracting the optimal classification features from the isolated kidney tumor region in a contrast enhanced multi-detector computed tomography test image; And (e) applying the extracted best-fit classification features to the classification model to distinguish between low-fat, angiomatous lipoma (fp-AML) and clear cell renal cell carcinoma (ccRCC).

Description

METHOD FOR DIFFERENTIATION OF FAT-POOR ANGIOMYOLIPOMA FROM CELL CELL RENAL CELL CARCINOMA IN CONTRAST USING A QUANTITATIVE FEATURE CLASSIFICATION IN A CONTRASTED MULTIPLE DETECTED COMPUTED CT IMAGING -ENHANCED MDCT IMAGES USING QUANTITATIVE FEATURE CLASSIFICATION}

The present invention relates to a method for distinguishing between low-fat, angiomatous lipoma and clear cell renal cell carcinoma using quantitative feature classification in contrast-enhanced multi-detector computer tomography (CE) MDMT .

Renal cell carcinoma (RCC) is the 9th most prevalent cancer in the world. Diagnosis of this cancer is increasing with increasing diagnosis using computational tomography (CT). Despite the classic symptoms of RCC, including side pain, lateral tumors and hematuria, most tumors are found incidentally without symptoms due to the development of CT and ultrasound imaging. Furthermore, 85% of the RCC cases have been reported as unsympathetic tumors of less than 4 cm in diameter. In these small renal masses (SRM), referred to as renal tumors of less than 4 cm in diameter, 20% to 40% are known to be benign type tumors, including angiomyolipoma (AML) come. Therefore, it has been proposed to distinguish between malignant SRM and benign SRM before surgery to avoid unnecessary surgery. In general, most cases of AML can be visually distinguished from RCC by the presence of visible fat in CT images. However, this visible fat is not visible in CT images in about 5% of AML cases. Thus, this sub-type of AML, referred to as low fat AML (fp-AML), exhibits both homogeneous and extremely dense characteristics compared to normal kidney parenchyma, as shown in Figure 1, It is difficult to distinguish it from RCC because of the visual similarity between branch conditions.

A number of approaches have been investigated to differentiate fp-AML and RCC from renal CT images. Investigators have used CT images that are not normally enhanced in view of the fact that image features on nonenhanced CT can more accurately classify the two subtypes of tumors. Chaudhry et al. (2003) used histogram features extracted from nonenhanced CT images to distinguish fp-AML from clear cell renal cell carcinoma (ccRCC) and papillary renal cell carcinoma (pRCC) by statistical analysis The potential was investigated. Schieda et al. (Schieda et al.) Used a Fisher exact test to identify imaging patterns such as density ratios, mass calcification, and hypodense rim to separate fp-AML and three types of RCCs. , As well as patient information such as sex and age. Hodgdon et al. (Hodgdon et al.) Used gray-level co-occurrence matrices (GLCM) and gray-scale continuous distance matrix (GLRLM) to classify fp-AML and RCC in non- level run-length matrices, we have developed a support vector machine (SVM) classifier using histogram features such as histogram mean and variance with texture features.

Although unenhanced CT images are known to have advantages when used to distinguish between two types of tumors compared to phase CT images, most approaches are based on CT images that are not enhanced, And augmented (CE) CT images. Lee-Felker et al. (Lee-Felker et al.) Used four types of phase images to classify subtypes of RCC using non-enhancement phase, cortical water phase, renal contrast phase, and excretory phase Morphological features and enhancement patterns were evaluated. Sung et al. Reported that the degree of roundness and the presence of capsules in order to differentiate fp-AML from RCC by chi-square testing of nonenhanced and early excretory phase CT ≪ / RTI > and morphological features and enhancement patterns were evaluated. Ishigami et al. (Ishigami et al.) Used CTLs to characterize five types of renal tumors (fp-AML, ccRCC, pRCC, chRCC and oncocytoma) using tumor attenuation and enhancement patterns , Which is a non-enhancement phase, a cortical water quality phase, and a delayed phase. Kim et al. Used a two-tailed t-test with the enhancement patterns of the three phases of CT (unenhanced phase, cortical water phase and initial excreted phase) , ccRCC, pRCC and chRCC. Xie et al. Used a statistical analysis based on wash-in and washout characteristics to compare the four phases of the CT (unenhanced) to the fc-AML and ccRCC Phase, cortical water quality, renal contrast, and delayed phase). Using multi-phase images, including extended selection of features such as enhancement patterns between images of different phases, provides an advantage in feature-based classification. However, multi-phase image analysis also has a number of limitations. First, in order to evaluate the enhancement patterns between images of different phases, time-consuming processes such as non-rigid matching between multiple passive segmentation and multi-phase images of the same tumor in images of various phases, Is required. Moreover, most of the above studies evaluated quantitative image features manually recorded by an expert. For example, the position (exophytic, intra-parenchymal or endophytic), shape (round or non-round), augmentation pattern (initial washout (early washout, gradual or prolonged), and damping type (hype-, iso- or hyper-attenuation) were manually evaluated in images entered by a radiologist in a multi-selection format. The quantitative feature evaluation process is not only time consuming but also limited by inter-observer variability and intra-observer variability. Recently, Lee et al. (2004) used CE CT images and quantitative images based on histogram mean, skewness, entropy, and GLCM texture patterns, Type (cyst, fat-visible AML, and ccRCC).

1. Jonasch E, Gao J, Rathmell WK. Renal cell carcinoma. BMJ 2014; 349: g4797 2. Skinner DG, Colvin RB, Vermillion CD, Pfister RC, and Leadbetter WF: Diagnosis and management of renal cell carcinoma: a clinical and pathologic study of 309 cases. Cancer 28: 1165-1177,1971. 3. Jayson M, Sanders H. Increased incidence of serendipitally detected renal cell carcinoma. Urology 1998; 51: 203-205 4. Ball MW, Bezerra SM, Gorin MA, Cowan M, Pavlovich CP, Pierorazio PM, et al. Grade heterogeneity in small renal masses: potential implications for renal mass biopsy. J Urol 2015; 193: 36-40 5. Jason Hafron, James D. Fogarty, David M. Hoenig, Maomi Li, Robert Berkenblit, Reza Ghavamian, Imaging characteristics of minimal fat renal angiomyolipoma with histologic correlations, Urology, Volume 66, Issue 6, December 2005, Pages 1155-1159 6. Huang WC, Atoria CL, Bjurlin M, Pinheiro LC, Russo P, Lowrance WT, et al. Management of Small Kidney Cancers in the New Millennium: Contemporary Trends and Outcomes in a Population-Based Cohort. JAMA Surg 2015; 150: 664-672 7. H. Chaudhry et al. Histogram analysis of small solid renal masses: Differentiating minimal fat angiomyolipoma from renal cell carcinoma. AJR 2012. 8. N. Schieda et al. Unenhanced CT for the diagnosis of minimal-fat renal angiomyolipoma. AJR 2014. 9. T. Hodgdon et al. Can quantitative CT texture analysis is used to differentiate fat-poor renal angiomyolipoma from renal cell carcinoma on unenhanced CT images. Radiology 2015. 10. S. Lee-Felker et al. Qualitative and quantitative MDCT features for differentiating clear cell renal cell carcinoma from other solid renal cortical masses. AJR 2014. 11. C. Sung et al. Angiomyolipoma with minimal fat: Differentiation of morphological and enhancement features from renal cell carcinoma at CT imaging. ActaRadiol2015. 12. K. Ishigami et al. Characterization of renal cell carcinoma, oncocytoma, and lipid-poor angiomyolipoma by unenhanced, nephrographic, and delayed phase contrast-enhanced computed tomography. Clinical Imaging 2015. 13. S. Kim et al. Differentiation of clear cell renal cell carcinoma from other subtypes and fat-poor angiomyolipoma by use of quantitative enhancement measurement during three-phase MDCT. AJR 2016. 14. P. Xie et al. Lipid-poor renal angiomyolipoma: Differentiation from clear cell renal cell carcinoma using wash-in and washout characteristics on contrast-enhanced computed tomography. Onco. Lett. 2016. 15. S. Lee, H. Hong, D. C. Jung, S. H. Park, and J. Kim, "Quantitative CT texture analysis and classification of small renal masses on contrast-enhanced CT images," Intl. Symp. Biomed. Im. 2016. 16. J. W. Xu and K. Suzuki, "Max-AUC Feature Selection in Computer-Aided Detection of Polyps in CT Colonography," IEEE Journal of Biomedical and Health Informatics, vol. 18, no. 2, pp. 585-593, March 2014. 17. Kononenko, Igor et al. Overcoming the myopia of inductive learning algorithms with RELIEFF (1997), Applied Intelligence, 7 (1), p39-55 18. A. Malhi and R. X. Gao, "PCA-based feature selection scheme for machine defect classification," IEEE Transactions on Instrumentation and Measurement, vol. 53, no. 6, pp. 1517-1525, Dec. 2004. 19. P. Pierorazio et al. Multiphasic enhancement patterns of small renal masses on preoperative computed tomography: Utility for distinguishing subtypes of renal cell carcinoma, angiomyolipoma, and oncocytoma. Oncology 2013. 20. N. Takahashi et al. Small (<4 cm) renal masses: differentiation of angiomyolipoma without visible renal cell carcinoma using unenhanced and contrast-enhanced CT. AJR 2015. 21. L. Yan et al. Angiomyolipoma with minimal fat: Differentiation from clear cell renal cell carcinoma and papillary renal cell carcinoma by CT analysis. Acad. Radiol. 2015. 22. Y. Lee et al. Does computer-aided diagnosis permit differentiation of angiomyolipoma without visible fat from renal cell carcinoma on MDCT? AJR 2015. 23. C. Yang et al. Are there useful CT features to differentiate renal cell carcinoma from lipid-poor renal angiomyolipoma? AJR 2013. 24. M. Kim et al. MDCT-based scoring system for differentiating angiomyolipoma with minimal fat from renal cell carcinoma. ActaRadiol 2013. 25. K. Kim, K. Kim, K. S. Cho, "CT histogram analysis: Differentiation of angiomyolipoma without visible fat from renal cell carcinoma at CT imaging," Radiology 246, 472-479 (2008)

The present invention is directed to quantitative classification of histograms and texture patterns in contrast enhanced multi-detector computed tomography (CT) images to detect positive angiomatous lipoma (fp-AML) It is intended to provide a method for distinguishing between low-fat, angiomatous lipoma and clear cell renal cell carcinoma, using quantitative characterization in contrast enhanced multi-detector computed tomography images, which can distinguish cancer (ccRCC).

According to an aspect of the present invention, there is provided a method for distinguishing an aneurysmal lipoma and a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-

(a) extracting histogram features, brightness value percentage features, and texture features from a separate kidney tumor region in a plurality of contrast enhanced multiple detection computerized tomography (MDCT) training images;

(b) applying the extracted features to a combination of a plurality of feature selection methods and a plurality of machine learning-based classification methods to classify features and classification methods that exhibit the best classification performance as optimal classification features and optimal machine learning classification Selecting as a method;

(c) applying the selected classification features to the selected machine learning classification method to obtain a classification model;

(d) extracting the optimal classification features from the isolated kidney tumor region in a contrast enhanced multi-detector computed tomography test image; And

(e) applying the extracted best-fit classification features to the classification model to distinguish between low-fat aneurysmal lipoma (fp-AML) and clear cell renal cell carcinoma (ccRCC).

A method for distinguishing an aneurysmal lipoma from a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention, Wherein the selecting the optimal classification features comprises:

selecting 20 features according to an area-under curve (AUC) analysis, a ReliefF feature selection method, and a principal component analysis (PCA); And

(b-2) selecting features selected by the area-under curve analysis (AUC), features selected by the ReliefF feature selection method and selected by the principal component analysis (PCA) And selecting common features among the features as the optimal classification features.

The present invention also provides a method for distinguishing an aneurysmal lipoma from a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention, Features include a maximum brightness value, a percentage of positive pixels above the threshold of 190 and 210, a percent brightness value of 75% and 95%, and a GLCM sum entropy and GLRLM long-run emphasis ).

In addition, a method for distinguishing an aneurysmal lipoma from a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention, The learning-based classification method may include a kNN (k nearest neighbor) classification method or a random-forest (RF) classification method.

The present invention also provides a method for distinguishing an aneurysmal lipoma from a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention, Enhanced multiple-detection computed tomography (MDCT) training images were obtained from patients with clear-cell renal cell carcinoma (ccRCC) and training images obtained from patients with low-fat angioplasticoma (fp-AML) Training images.

According to one embodiment of the present invention, in a method for distinguishing an aneurysmal lipoma and a clear cell renal cell cancer from a fat-poor aneurysm using quantitative feature classification in an enhanced multi-detection CT image, The histogram and texture features of low-fat lipoangioma (fp-AML) and clear cell renal cell carcinoma (ccRCC) were separated on a computed tomography image and the extracted maximum brightness value, positive pixels greater than the threshold of 190 and 210 Based on the seven characteristics and machine learning, including the percentage of percentile, 75% and 95% percentile brightness values, and GLCM sum entropy and GLRLM long-run emphasis, It is possible to distinguish between angiomyolipoma and clear cell renal cell carcinoma.

1 shows fp-AML (left arrow, arrow) and transparent cell RCC (right arrow, arrow) shown in a contrast enhanced kidney CT image.
FIG. 2A is a flow chart of a method for distinguishing an aneurysmal lipoma of low fat from a clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detector computed tomography image according to an embodiment of the present invention.
FIG. 2B is a conceptual diagram of a method for distinguishing between an aneurysmal lipoma of low fat and a clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention.
Figure 3 shows an example of ROI contours for fp-AML (upper row) and ccRCC (lower row).
Figure 4 shows a heat map of normalized features for fp-AML and ccRCC patients; The x-axis is the histogram and texture features and the y-axis is the clustered patient.
Figure 5 shows the average ReliefF weights (bottom) calculated from AUC (top) and ReliefF selection measured from AUC analysis.
Figure 6 is a characteristic heat map of PCA-transformed training features; (a) shows original enhanced training features (upper) and (b) shows PCA-converted training features (lower).
Figures 7a, 7b and 7c show ROC curves of a classification method performed with four classifier models, based on (a) AUC, (b) ReliefF and (c) feature selection of PCA.
Figures 8A, 8B and 8C are line graphs of the number of features for classifier models based on accuracy versus (a) AUC, (b) ReliefF, and (c) PCA feature selection methods.
Figures 9a-9e illustrate five classification performances including (a) sensitivity, (b) specificity, (c) PPV, (d) NPV and (e) AUC for accuracy versus classifier models and feature selection methods Line graph.

BRIEF DESCRIPTION OF THE DRAWINGS The objectives, specific advantages and novel features of the present invention will become more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which: FIG.

Prior to that, terms and words used in the present specification and claims should not be construed in a conventional and dictionary sense, and the inventor may properly define the concept of the term in order to best explain its invention Should be construed in accordance with the principles and the meanings and concepts consistent with the technical idea of the present invention.

It should be noted that, in the present specification, the reference numerals are added to the constituent elements of the drawings, and the same constituent elements are assigned the same number as much as possible even if they are displayed on different drawings.

Also, the terms "first", "second", "one side", "other side", etc. are used to distinguish one element from another, It is not.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In the following description of the present invention, a detailed description of known arts which may unnecessarily obscure the gist of the present invention will be omitted.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 2A is a flow chart of a method for distinguishing an aneurysmal lipoma of low fat from a clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention.

Referring to FIG. 2A, in step S200, histogram features, brightness value percentage features, and texture features are extracted from a separate kidney tumor region in a plurality of contrast enhanced multi-detection computerized tomography (MDCT) training images.

In step S202, the extracted features are applied to a combination of a plurality of feature selection methods and a plurality of machine learning-based classification methods, so that the features and classification methods exhibiting the best classification performance are respectively referred to as optimal classification features and optimal machine learning As the classification method.

In step S204, the selected classification features are applied to the selected machine learning classification method to obtain a classification model.

In step S206, the optimal classification features are extracted from the isolated kidney tumor area in the contrast enhanced multi-detection computerized tomography test image.

In step S208, the extracted optimal classification features are applied to the classification model to distinguish aneurysmal lipoma of low fat from clear cell renal cell carcinoma.

A method for distinguishing between a fat-poor angiomyolipoma and a clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention includes: (Fp-AML: fat-poor angiomyolipoma) and malignant clear-cell renal cell carcinoma (ccRCC) using quantitative characterization of histogram and texture patterns from imaging (CE MDCT) : clear-cell renal cell carcinoma).

A dataset containing 50 CE MDCT images of 25 fp-AML and 25 ccRCC patients was used. From these images, the tumors were manually separated by a radiologist to define the region of interest (ROI). Using histogram and texture features and machine learning classifiers, a feature classification method for separating two types of kidney tumors is proposed. First, histogram features including basic histogram characteristics, percentages of pixels above certain thresholds and percentile intensities, and GLCM, GLRLM, and local binary pattern (LBP) ), 64 quantitative image features were extracted from each ROI. A number of feature selection methods including AUC analysis, ReliefF selection and PCA conversion were applied to select useful features. Finally, features that include logistic regression (LR), k nearest neighbor (kNN), support vector machine (SVM), and random-forest (RF) Classifiers were trained on selected features to identify positive fp-AML and malignant ccRCC. Each combination of feature selection and classification methods were tested using the 5-fold cross validation method and were tested for accuracy, sensitivity, specificity, positive predictive value (PPV: positive predictive value, a negative predictive value (NPV), and an area under receiver operating characteristic curve (AUC).

In feature selection, commonly selected features were evaluated by different feature selection methods. Five histogram features including the maximum brightness value, a threshold 190 and a percentage of pixels over 210, and a percent brightness value at 75% and 95%, and GLCM sum entropy and GLRLM long-run emphasis long-run emphasis) were jointly selected as key features to distinguish between two types of kidney tumors. In the feature classification, kNN and RF classifiers with ReliefF feature selection showed the best performance among other choices of feature selection and classification methods and ReliefF + kNN and ReliefF + RF achieved AUC of 0.782 and 0.780, respectively.

In a method for distinguishing an aneurysmal lipoma and a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention, a histogram Features based on texture and heterogeneity of tumor internal heterogeneity were selected as key features for classifying fp-AML and ccRCC, and the best combination of feature selection and classification methods to distinguish benign and malignant tumors Performance. The present invention can be used to evaluate useful features associated with malignant tumors for kidney tumors in contrast enhanced CT images.

In the following, referring to Figs. 2 to 9, a quantitative feature classification is used to distinguish an aneurysmal lipoma and a clear cell renal cell cancer from a contrast enhanced multi-detection CT image according to an embodiment of the present invention Will be described in detail.

Data acquisition

The data set includes 50 SRMs from 50 patients, acquired from December 2013 through July 2015. These SRMs consist of 25 fp-AML and 25 ccRCCs. For fp-AML, the data set includes only positive cases in which the tumor did not grow during the two year follow-up. For ccRCC, the data set includes only malignant cases positively diagnosed in pathological reports after surgical removal. The age, sex, and tumor size for the patients are summarized in Table 1.

Fp-AML
( N = 25) ccRCC
( N = 25) Average age 51.8 55.9 Patient's sex Woman 80% (20) 32% (8) man 20% (5) 68% (17) Mean tumor size 3.46 cm ³ 9.36 cm ³

CT imaging for all cases was performed in a single institution. Images of 50 patients were analyzed using MDCT (SOMATOM Sensation 64, Siemens Healthcare, Forchheim, Germany; LightSpeed VCT, GE Healthcare, Milwaukee, WI, USA; GE Discovery CT 750 HD, GE Healthcare, Milwaukee, , Philips Healthcare, Best, The Netherlands; SOMATOM Definition Flash, Siemens Healthcare, Forchheim, Germany; Revolution CT, GE Healthcare, Milwaukee, Wis., USA). All images were on the axial surface with a slice thickness of 1.0-3.0 mm, a resolution of 512 x 512 pixels, and a pixel size in the range of 0.66 x 0.66 mm &lt; 2 &gt; to 0.77 x 0.77 mm &lt; 2 &gt;. In all cases, both contrast enhanced and contrast enhancement phase images were obtained, and only enhancement images in the slowest retarded phase were used. Furthermore, 2 ml / kg non-ionic contrast agent (Xenetics, Iobitridol, Seoul, Korea) was injected intravenously at a rate of 3 ml / s and contrast enhanced images were obtained with a delay between 100 s and 120 s.

Feature extraction

According to one embodiment of the present invention, fP-AML and ccRCC are used to distinguish between a fat-poor angiomyolipoma and a clear cell renal cell carcinoma using quantitative feature classification in a multi-detection CT scan image of the present invention. We provide classification methods for SRMs in CE MDCT images to distinguish them. A method for distinguishing between low fat aneurysmal lipoma and clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detector computed tomography image according to an embodiment of the present invention includes three main steps: (1) feature extraction, (2) feature selection, and (3) classification. A method for distinguishing between a fat-poor angiomyolipoma and a clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detector computed tomography image according to an embodiment of the present invention is summarized in FIG.

First, in step S210, features representing image features of SRMs are extracted from CE MDCT images. As a preprocessing step for feature extraction, tumor division was performed and region of interest (ROI) was defined. For each patient 3D image, the radiologist selected one representative slice showing the largest section of the SRM. In this representative slice, the SRM was manually segmented by a radiologist using ImageJ software (National Institutes of Health, Bethesda, MD, USA). ROI was determined by performing morphological operations, i.e. erosion, on the manually segmented outline to include the tumor-internal ROI without noise-bearing information around the tumor boundary. FIG. 3 is a graph showing the results of SRM ROIs used in a method for distinguishing between fat-less aneurysmal lipoma and transparent cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention FIG. The upper row of Figure 3 is an example of an ROI outline for fp-AML, and the lower row images are an example of ROI outline for ccRCC.

From the segmented SRM ROI, 64 histograms and texture features were automatically extracted. A list of features is provided in Table 2. First, eighteen histogram features represent the intensity distribution of the SRM, including the basic histogram characteristics, the percentage of pixels over a certain threshold, and the percent brightness value. The seven basic histogram features included the mean, standard deviation, minimum and maximum brightness values, skewness, kurtosis, and entropy of the brightness value histogram. In six features with a percentage of pixels above a certain threshold, the ratio of pixels whose intensity brightness values exceeded a predetermined threshold value was calculated. In the five features with percentile brightness values, a predetermined percentage of brightness values on the intensity histogram were extracted. Second, the 46 texture features represented texture patterns within the SRM, including GLCM, GLRLM, and LBP features. In the 14 GLCM and 22 GLRLM features, textures with four pixels in different directions, 0 °, 45 °, 90 ° and 135 ° were calculated with a pixel distance of 1 pixel and mean and standard deviations between them were used . With respect to LBP features, nine rotation-invariant uniform patterns, including at most two binary transitions, were selected from 256 binary patterns while the remaining non-uniform patterns were selected from a single bin ( bin). To extract noise robust texture patterns, GLCM, GLRLM, and LBP were computed for mid-level ROI images by adjusting histograms from 256 to 64 levels. All extracted features were used in the feature selection and classification process.

Categories (No. of features):
Features (feature numbers) Histogram features (7):
Histogram mean (1), standard deviation (2), minimum (3) and maximum (4) intensities, skewness, kurtosis, Percentages of pixels above thresholds (6):
8-bit intensity thresholds with 150 (8), 170 (9), 190 (10), 210 (11), 230 (12) Percentile intensities at (5):
5% (14), 25% (15), 50% (16), 75% (17), 95% (18) GLCM features (14):
(21,22), sum average (23,24), sum variance (25,26), sum entropy (27,28), entropy (29,20) , 30), and difference entropy (31,32) GLRLM features (22):
Run-length non-uniformity (39,40), run-percentage (35,36), gray-level non-uniformity Level run emphasis 43,44, high gray-level run emphasis 45,46, short run low gray-level emphasis 47,48, short run high gray-level emphasis 45,46, (49,50), long run low gray-level emphasis (51,52), long run high gray-level emphasis (53,54) LBP features (10):
10Uniform patterns in LBP histogram

Feature selection

In step S212, from the extracted features, several features were selected and grouped to enhance the performance of the classifier. In feature classification, the selection of suitable features plays an important role in improving the discriminatory power of the trained classifier. Three feature-selection methods, area-under curve (AUC) analysis, ReliefF method and principal component analysis (PCA) were used in the present invention.

In the AUC analysis, a quantitative assessment of each feature used to classify fp-AML and ccRCC was performed. Receiver-operating-characteristic (ROC) curves and area-under curves (AUC) for each feature were calculated to evaluate the performance of distinguishing the two classes of SRM. The ROC curve and AUC were calculated while continuously changing the thresholds of the feature values and measuring the true positive and false positive ratios. The AUC analysis provides discriminatory power of each individual feature to classify different types of SRM. With the calculated AUC score, the features are ranked and the top N _F features are selected from the sorted list, where N _F is the number of features selected.

In the ReliefF selection method, weights related to feature importance were calculated and feature selection was performed by selecting features with higher weights. The ReliefF method is one of the most commonly used feature selection methods in the realm of binary classification due to its low-order polynomial time efficiency, good noise tolerance and robustness to feature interactions. In the ReliefF method, feature weights are calculated by repeatedly updating them, emphasizing in-class margins and penalizing for inter-class margins. In the ReliefF selection step, the discriminating power of the feature combination is reflected in the selection result, while the AUC analysis only evaluates the features individually.

In the PCA transformation method, features are transformed by projecting features onto a new coordinate system. To effectively reduce the size of features without loss of information, the PCA method computes a new coordinate system with two constraints. First, the coordinates are not orthogonal to each other and correlated. Second, the coordinates are determined by the constraint that the level of change of the projected feature points is maximized. In contrast to the selection methods described above, which maintain the values of the selected features, the PCA transforms the feature values by feature space projection to emphasize the importance of the features. The first N _F features were selected from the transformed features in one embodiment of the present invention, since the discriminatory power of the features as well as the PCA-transformed features are ranked.

Using various types of feature selection methods, core features ranked jointly by multiple selection methods were obtained and analyzed. Furthermore, performance capabilities associated with various combinations of feature selection and classification methods have been evaluated for in-depth analysis of the relationship between classifier behavior and disease characteristics.

Classification

Finally, in step S214, classification of the features selected through the machine-learning classifier has been performed. In one embodiment of the present invention, four classifier models, including a logistic regression analysis (LR), a k-nearest neighbor (kNN), a support vector machine (SVM), and a random- ccRCC. &lt; / RTI &gt; In LR, the output class has been considered as a weighted linear combination of input features, and the problem is to find the optimal set of weights in a linear combination. In kNN, the output class is determined by observing k neighbor data for given unseen data. In the experiment, the number of observed neighbors was experimentally determined as k = 50. In the SVM, the hyperplane for the feature space is determined to distinguish binary class data. In the experiment, a kernel with a radial basis function (RBF) was adopted. In RF, a random decision tree (RDT) is bagged to avoid overfitting. In the experiment, the number of RDTs to form an RF classifier was experimentally determined to be n _T = 50.

Classifier models were trained using a 5-fold cross validation method, and data from 50 patients in the verification method were randomly divided into five groups, one group being used as test data And the remaining four groups were used as training data. Because of the limited number of data instances (N = 50), classifier models can be greatly affected by data bias and over-sum in the training phase. To overcome this difficulty, a data increment technique has been applied in which new training data derived from the original training data set is added. In the experiment, the training data increment is (1) selecting the data at random, (2) rotating the data with an angle resolution of 0.1 degrees, and (3) randomly flipping the data along the vertical axis (flipping). The number of increased data was also experimentally determined as N _D = 2 N _C , where N _C = 500 is the magnitude of the increased data for a single disease class.

result

To verify the ability to distinguish between benign fp-AML and malignant ccRCC, a data set comprising 25 fp-AML patients and 25 ccRCC patients was subjected to contrast enhanced multi-detection computer tomography Methods were used to distinguish between low fat, aneurysmal lipoma and clear cell renal cell carcinoma, using quantitative characterization in radiographic images. Experiments were performed using five-fold crossover validation by randomly superimposing the data set into five sets. The random overlap was performed 50 times and the scores were averaged to avoid overlapping bias. In the feature selection approach, twenty features selected by each feature selection method were cataloged and analyzed. Commonly selected core features from the top 20 lists from all three selection methods were evaluated. In the feature classification, six evaluation criteria were determined from the classification results: accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) And AUC. Accuracy represents the overlap between predicted and actual cases, regardless of malignancy. Sensitivity and PPV are true positive ratios for actual fp-AML cases and predicted fp-AML cases, respectively. Specificity and NPV are correspondingly the ratio of true ccRCC cases and true negative for predicted ccRCC cases. Finally, AUC represents the probability of a classifier that accurately discriminates between benign fp-AML and malignant ccRCC. The performance of the four classifier models was evaluated and analyzed for the choice of feature selection method and the number of selected features (N _F ).

Figure 4 shows a hit map of 64 normalized original features extracted from 50 patients. It shows that the difference between distributions of image features for fp-AML and ccRCC is hardly noticeable. It is therefore challenging to distinguish between the two diseases using conventional classification methods with these characteristics, and statistical methods such as the Student t test have been performed on a single feature to evaluate discrimination performance in most previous studies. From the increased training features formed from these original features, three feature selection methods, AUC analysis, ReliefF selection and PCA transformation, were applied to rank features in terms of discrimination to distinguish between the two classes.

In the AUC feature selection, the discriminatory power of the individual features was measured by AUC as shown in the upper row of FIG. Many features, particularly histogram features and GLCM texture features, exhibited a very high AUC in separating the two SRM classes. A list of features with the top 20 AUCs is given in Table 3. The top 20 AUC lists include three basic histogram features, two histogram percent features, two histogram percentile brightness values, seven GLCM texture features, four GLRLM texture features, and two LBP features.

In the ReliefF feature selection, feature importance weights were calculated, as shown in the bottom row of FIG. It has been noted that even though some texture features also represent important weights to be selected, histogram features achieve high feature weights. A list of features with the top 20 ReliefF weights is summarized in Table 3. This table includes six basic histogram features, four histogram percent features, three histogram percentile brightness values, three GLCM texture features, three GLRLM texture features, and one LBP characteristics.

In the PCA feature selection, features are converted to different coordinate systems, so that an upper selection list of original features can not be generated directly as summarized for the two selection methods, as shown in FIG. Instead, a linear combination for the first principal component (PC) was averaged to evaluate feature importance in the PCA feature selection method. Considering that the discriminative power of the features in the PCA-transformed features is maximized in the first PC shown in the leftmost column of FIG. 6B, the original features with higher weights for the first PC are significant Can be considered as features. A list of the top 20 average first PC weights is shown in Table 3. In this list, two basic histogram features, four histogram percent features, four histogram percentile brightness values, seven GLCM texture features and three GLRLM texture features are included.

Rank Top 20 AUC  features Top 20 ReliefFweighted  features Top 20 PCA  weighted features One GLCM difference entropy (mean) Histogram max value Percentages of positive pixels above threshold 190 2 GLCM sum entropy (mean) Histogram entropy Percentile intensity at 95% 3 GLCM entropy (mean) Histogram skewness GLRLM long run emphasis (mean) 4 GLCM sum average (mean) Histogram standard deviation Percentile intensity at 75% 5 Histogram entropy Histogram min value GLRLM long run low gray-level emphasis (mean) 6 GLRLM low gray-level run emphasis (mean) Percentile intensity at 5% Percentages of positive pixels above threshold 170 7 Histogram max value GLCM difference entropy (mean) Histogram max value 8 Percentile intensity at 95% Percentages of positive pixels above threshold 190 GLCM sum variance (mean) 9 Percentile intensity at 75% Percentile intensity at 95% GLCM sum average (mean) 10 Histogram standard deviation GLCM sum entropy (mean) Percentile intensity at 50% 11 GLRLM long run low gray-level emphasis (mean) Percentages of positive pixels above threshold 230 Histogram mean 12 GLRLM long run emphasis (mean) GLRLM run length non-uniformity (mean) Percentages of positive pixels above threshold 210 13 GLCM angular second moment (mean) Percentages of positive pixels above threshold 170 GLRLM long run high gray-level emphasis (mean) 14 LBP (9) GLCM entropy (mean) GLCM sum variance (std) 15 LBP (3) LBP (10) Percentile intensity at 25% 16 Percentages of positive pixels above threshold 190 Percentile intensity at 75% GLCM contrast (mean) 17 GLCM sum variance (mean) Percentages of positive pixels above threshold 210 GLCM sum average (std) 18 GLCM sum variance (std) GLRLM long run emphasis (mean) Percentages of positive pixels above threshold 230 19 Percentages of positive pixels above threshold 210 GLRLM run percentage (mean) GLCM contrast (std) 20 GLRLM gray-level non-uniformity (mean) Histogram mean GLCM sum entropy (mean)

From the three lists of features having the top 20 selection scores or weights in the three feature selection methods, finding features that occur in two lists or three lists can be useful when examining meaningful features for a classification task useful. In the AUC and ReliefF selections, eleven features of the twenty features, including seven histogram features and four texture features, are generally listed. In the AUC and PCA selections, eleven features of the twenty features, including five histogram features and six texture features, were selected in both cases. Between the ReliefF and PCA selections, ten features of the twenty features were commonly selected, including eight histogram features and two texture features. On all three lists, the five histogram features, the histogram maximum brightness value 4, the percentage of positive pixels above the threshold of 190 (10) and 210 (11), 75% (17) and 95% 18), and two texture features of GLCM sum entropy (27) and GLRLM long-run emphasis (35) were common to the list.

A method for distinguishing between low-fat aneurysmal lipoma and clear cell renal cell carcinoma using quantitative feature classification in contrast enhanced multidetector CT images according to an embodiment of the present invention distinguishes between fp-AML and ccRCC It was applied to selected features. Table 4 shows the performance of four different classifiers with three different feature selection methods, measured according to accuracy, sensitivity, specificity, PPV and NPV.

Methods Accuracy (%) Sensitivity (%) Specificity (%) PPV ( % ) NPV  ( % ) AUC  Feature Selection LR 65.8 ± 4.3 69.7 ± 5.7 61.8 ± 6.8 64.8 ± 4.4 67.2 ± 4.7 kNN 70.3 ± 4.4 71.7 ± 6.1 68.9 ± 7.5 70.0 ± 5.2 71.0 + - 4.6 SVM 67.9 ± 3.9 67.7 ± 5.6 68.1 ± 5.9 68.1 ± 4.4 67.9 ± 4.1 RF 66.3 ± 4.5 68.3 ± 6.0 64.2 ± 6.9 65.8 ± 4.7 67.1 ± 4.9 ReliefF  Feature Selection LR 67.6 ± 5.4 66.3 ± 7.4 68.9 ± 7.8 68.3 ± 6.2 67.3 ± 5.5 kNN 72.3 ± 4.6 73.0 ± 6.2 71.7 ± 6.8 72.3 ± 5.2 72.8 ± 5.0 SVM 72.1 ± 4.2 71.0 ± 5.1 73.2 ± 6.1 72.8 ± 4.9 71.7 ± 4.1 RF 69.6 ± 3.9 68.1 ± 6.1 71.0 ± 5.8 70.3 ± 4.5 69.2 ± 4.2 PCA  Feature Selection LR 69.3 ± 4.3 70.2 ± 6.5 68.4 ± 6.0 69.1 ± 4.4 69.8 ± 5.0 kNN 66.3 ± 4.5 71.9 ± 5.6 60.7 ± 6.8 64.8 ± 4.5 68.4 ± 5.1 SVM 67.7 ± 3.7 68.5 ± 5.2 66.9 ± 6.4 67.6 ± 4.3 68.1 ± 3.9 RF 70.2 ± 4.6 67.5 ± 5.6 73.0 ± 7.1 71.7 ± 5.8 69.2 ± 4.4

In the AUC selection, kNN showed the best overall performance, and others achieved similar scores. In ReliefF selection, kNN and RF showed stronger performance than others. In the PCA selection, LR and RF achieved higher scores than kNN and SVM. In terms of the classifier model, LR showed stronger performance using PCA selection than other feature selection methods. Other classifier models, kNN, SVM and RF, achieved best performance using ReliefF selection. From the overall results, kNN and ReliefF selection outperformed other classification methods.

Figure 7 illustrates the ROC curve of the proposed classification method of the present invention using different classifier models and different feature selection methods. As analyzed in Table 4, kNN showed the steepest curve among the other models when choosing AUC. The kNN and RF curves are located above the LR and SVM when ReliefF is selected, and the LR and RF curves are above the rest of them when the PCA is selected. From these curves and AUC, it can be seen that kNN and RF with ReliefF selection showed the best performance among other choices of classifier models and feature selection methods.

To investigate the behavior of classifier models on feature selection, the classification performance capabilities of the proposed classification method of the present invention were measured with different numbers of selected features (N _F ). FIG. 8 shows a line graph of classification accuracy and N _F of four classifier models and three feature selection methods evaluated here. When AUC is selected, the accuracy of RF decreases with decreasing N _F , while the accuracy with respect to kNN and SVM is relatively stable. The accuracy of LR decreases rapidly when N _F > 40, and when N _F <40 . In the ReliefF selection process, the accuracy of the RF decreases again as N _F decreases, but increases for other models. In particular, the LR shows a sharp increase in accuracy when N _F <30. When PCA is selected, the accuracy of LR, kNN, and SVM increases with decreasing N _F , while the accuracy of RF remains relatively stable. Especially for PCA selection, the accuracy for SVM with full PCA-converted features (N _F = 64) is severely degraded, whereas for N _F = 35 the accuracy increases sharply to the expected level. From the classifier behavior with N _F , the ReliefF selection method is effective for most classifier models regardless of the choice of N _F , and the PCA selection is small N _F It is obvious that it works particularly well with values. Other performance measures of classifier models, in this case sensitivity, specificity, PPV, NPV and AUC levels, showed a tendency similar to the accuracy trend, as shown in FIG.

Argument

In one embodiment of the invention, a method has been proposed for distinguishing fp-AML from ccRCC in CE MDCT images using histogram features and texture features with machine-learning classifiers. Despite its clinical importance, it is known that distinguishing between two types of SRMs is challenging due to their similar appearance. To distinguish between the two types of SRMs, feature classification methods have been developed for CE MDCT images consisting of feature extraction, selection and classification steps. In feature extraction, 64 features for the histogram and texture patterns were extracted from the central slice of the tumor ROI. During the process of feature selection, a group of core features used in the separation of the SRMs of the two classes was selected by various selection methods. Finally, for feature classification, multiple classifier models were trained using selected features to classify invisible SRM data. In the experiment on the data set, the core features that were selected in common by the three selection methods were evaluated, and the performance evaluation related to the feature selection and the various selection of the classifier models was performed.

Methods for distinguishing between low-fat aneurysmal lipoma and clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detector computed tomography image according to an embodiment of the present invention provide many benefits. First, feature extraction, selection, and classification were performed only on CE CT images. In most previous studies, multi-phase images including both nonenhanced and enhanced CT (CE) CT images were used as input images and enhancement patterns between different phases were evaluated as image features. Computing the enhancement patterns from multi-phase images is time consuming because it requires multiple passive segmentation of the same tumor in images of different phases as well as non-rigid matching between images to match tumor ROIs . In one embodiment of the present invention, a feature classification method using CE MDCT images has been proposed that is relatively simple and does not require non-rigid matching between multiple ROI definitions or multi-phase images. Moreover, while most existing methods focus on multi-phase images, the present invention contributes to evaluating useful image features and classification models that are specific to contrast enhanced images when considering their capabilities.

Secondly, in a method for distinguishing an aneurysmal lipoma and a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention, only quantitative image features Are evaluated and analyzed. Most of the previous methods have been used to determine the location (exophytic, intra-parenchymal or endophytic), shape (round or non-round) We used qualitative image features such as early washout, gradual or prolonged, and attenuation type (hype-, iso- or hyper-attenuation). These qualitative features were manually evaluated and recorded by the radiologist in a multi-selection format. These features generally provide valuable information as acknowledged by experts, but the process of characterization is time consuming as well as the risk of intra-observer variability and intra-observer variability. In contrast, a method for distinguishing between low-fat aneurysmal lipoma and clear cell renal cell carcinoma using quantitative feature classification in contrast enhanced multi-detector computed tomography images according to an embodiment of the present invention is based on the tumor ROI Using only quantitative image features that are automatically calculated from the images.

In the experimental results, a number of observations were recorded using the classification method proposed in the present invention. First, in the feature selection process, as shown in Table 3, seven features were commonly selected in the top 20 feature lists selected by the three selection methods. In these seven key features, the five histogram features of the maximum brightness value, the percentage of pixels above thresholds 190 and 210, and the 75% and 95% percentile brightness values are the histogram features of areas with bright brightness values in the SRM . Previous methods have shown that histogram analysis of CT attenuation measurements can have an advantage in distinguishing fp-AML from RCCs of various subtypes due to the presence of fat in fp-AML, regardless of visibility. They hypothesized that despite the absence of visible fat in fp-AML, the presence of fat in the tumor results in a small but present difference in the damping histogram of the fp-AML and subtype RCCs. The results of the method for differentiating between fat-less aneurysmal lipoma and clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detector computed tomography image according to an embodiment of the present invention are as follows: Histograms also show similar results in that they may also be slightly different between fp-AML and RCC, which mainly results in the selection of histogram features associated with histogram characteristics of regions with bright brightness values. The remaining two texture features, GLCM sum entropy and GLRLM long-run emphasis, provide heterogeneous texture patterns within the tumor. A number of studies involving contrast-enhanced CT images have shown that although fp-AML exhibits a lower level of lesion heterogeneity than RCC, the difference is almost invisible and is only accessible by quantitative texture analysis . The degree of lesion heterogeneity can also be observed as an important feature for distinguishing between two types of SRM on CE CT images, as shown during the main process of selection of heterogeneity-related texture patterns.

Second, in the feature classification, it has been confirmed that the proper combination of feature selection methods and classifier models is crucial for good classification performance, as shown in Table 4 and Figures 7-9. First, LR showed excellent performance with fewer (N _F ) selected features, showing the best performance with PCA conversion, regardless of the feature selection method used. Thus, considering the prediction of class values by calculating a weighted combination of feature values, reducing the number of redundant features has a critical impact on its performance. Thus, PCA transformation is the most appropriate selection method for LR, because it reduces the level of redundancy from the original feature data. Second, kNN showed a relatively consistent score with respect to N _F with AUC and ReliefF selection, and showed better performance than other classifier models. A simple kNN is known to show better performance than other classifiers, e.g., SVM and RF, when the classes are highly overlapped in the given feature data. Here too, the SRM of both classes is highly overlapped in the feature space due to its visual similarity and texture similarity: Thus, kNN shows the best performance among other classifier models. In contrast to kNN, SVM showed average performance among other classifier choices. Since the SVM classifies different classes by computing hyperplanes separating the feature space into different classes, its performance can be degraded by severe overlap between the two classes. Furthermore, the performance of SVMs using PCA transformations was observed to decrease dramatically when N _F > 40. It can be inferred that these redundant features severely degrade the SVM performance, since only a small number of main features show an emphasized distribution while the remaining features show a noise-like distribution. Finally, RF also showed excellent performance regardless of the feature selection method used: however, its performance capability decreased as N _F decreased, as opposed to other classifiers. Thus, RF generally showed better performance when these features were not selected. Because RF includes feature selection steps in image multi-tree bagging processes, RF does not require additional feature selection steps, and can be implemented in a way that if RF is more suitable for high dimensional feature data, It was confirmed that the performance deteriorated. As a result, ReliefF + kNN and RF showed the best performance to distinguish fp-AML from ccRCC in CE MDCT images due to the distribution of feature space and the behavior of classifier models.

conclusion

According to a method for distinguishing between a fat-poor angiomyolipoma and a clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention, in a renal CE MDCT image, -Histogram and texture features of AML and ccRCC can be separated. First, histogram and texture features were extracted from the tumor ROI, including basic histogram characteristics, percent of pixels over a certain threshold, percentile brightness values, GLCM textures, GLRLM textures, and LBP features. With these characteristics, various feature selection methods (AUC analysis, ReliefF selection and PCA conversion) according to various classification methods (LR, kNN, SVM and RF) have been applied to distinguish fp-AML from ccRCC. In experiments on clinical data sets, the maximum brightness values, the percentage of positive pixels above the threshold of 190 and 210, the percentile brightness values of 75% and 95%, and the GLCM sum entropy and GLRLM long- Seven features have been selected as key features to distinguish between the two types of SRMs, based on histogram properties and texture heterogeneity of light brightness value regions, including long-run emphasis. Among these feature selection and selection of classification methods, kNN and RF classification with ReliefF selection showed the best performance in classifying the histogram and texture patterns of two SRMs. The AUC values of ReliefF + kNN and ReliefF + RF were 0.782 0.780. Based on the experimental results, a method for discriminating between fat-less aneurysmal lipoma and transparent cell renal cell carcinoma using quantitative feature classification in contrast enhanced multi-detector computed tomography images according to an embodiment of the present invention is described as a positive SRM It can be used to evaluate quantitative image features to differentiate malignant SRMs.

A method for distinguishing an aneurysmal lipoma of low fat from a clear cell renal cell cancer using quantitative feature classification in a contrast enhanced multi-detection computed tomography image according to an embodiment of the present invention includes a controller and a memory May be performed by a control unit.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. It is clear that the present invention can be modified or improved.

It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

S200: Feature extraction step
S202: Step of selecting a feature and selecting a classification method
S204: Classification model acquisition step
S206: Classification feature extraction step in the test image
S208: Step for distinguishing fp-AML from ccRCC
S210: Feature extraction step
S212: Feature selection step
S214: Character classification step

Claims (5)

(a) extracting histogram features, brightness value percentage features, and texture features from a separate kidney tumor region in a plurality of contrast enhanced multiple detection computerized tomography (MDCT) training images;
(b) applying the extracted features to a combination of a plurality of feature selection methods and a plurality of machine learning-based classification methods to classify features and classification methods that exhibit the best classification performance as optimal classification features and optimal machine learning classification Selecting as a method;
(c) applying the selected classification features to the selected machine learning classification method to obtain a classification model;
(d) extracting the optimal classification features from the isolated kidney tumor region in a contrast enhanced multi-detector computed tomography test image; And
(e) applying the extracted best-fit classification features to the classification model to distinguish between low-fat, angiomatous lipoma (fp-AML) and clear cell renal cell carcinoma (ccRCC) METHODS FOR DIFFERENTIATING TRANSITIONAL CELL CARCINOMA WITH LIVER AMBULANCE LYMPHOCYTE USING QUALITATIVE FEATURE CLASSIFICATIONS IN COMPUTER TOMOGRAPHIC IMAGES. The method according to claim 1,
Wherein the selecting of the optimal classification features in step (b)
selecting 20 features according to an area-under curve (AUC) analysis, a ReliefF feature selection method, and a principal component analysis (PCA); And
(b-2) selecting features selected by the area-under curve analysis (AUC), features selected by the ReliefF feature selection method and selected by the principal component analysis (PCA) Discriminating an aneurysmal lipoma of low fat and clear cell renal cell carcinoma using quantitative characterization in a contrast enhanced multi-detector computed tomography image, comprising selecting common features among the features as the optimal classification features Lt; / RTI &gt; The method of claim 2,
The optimal classification features include,
A maximum brightness value, a percentage of positive pixels above the threshold of 190 and 210, a percentile brightness value of 75% and 95%, and a GLCM sum entropy and GLRLM long-run emphasis A method for differentiating between low fat, aneurysmal lipoma and clear cell renal cell carcinoma using quantitative characterization in contrast enhanced multisection computed tomography images. The method according to claim 1,
The optimal machine learning based classification method comprises:
using a quantitative feature classification on contrast enhanced multi-detector computed tomography images, including k-nearest neighbor classification (kNN) or random-forest (RF) classification methods, A method for distinguishing between lipoma and clear cell renal cell carcinoma. The method according to claim 1,
The plurality of contrast enhanced multi-detection computerized tomography (MDCT) training images are obtained from training images obtained from patients with low-fat angioplasticoma (fp-AML) and patients with clear cell renal cell carcinoma (ccRCC) A method for distinguishing between an adipocyte lipoma of low fat and clear cell renal cell carcinoma using quantitative feature classification in a contrast enhanced multi-detector computed tomography image, comprising training images obtained from the subject.

Images