Adrenal tumor characterization on magnetic resonance images

(1)

R E S E A R C H A R T I C L E

Adrenal tumor characterization on magnetic resonance images

Mucahid Barstugan

1

| Rahime Ceylan

1

| Semih Asoglu

2

| Hakan Cebeci

2

|

Mustafa Koplay

2

1

Electrical and Electronics Engineering, Konya Technical University, Konya, Turkey

2

Radiology Department, Medicine Faculty, Selcuk University, Konya, Turkey Correspondence

Mucahid Barstugan, Electrical and Electronics Engineering, Konya Technical University, Konya 42250, Turkey. Email: mbarstugan@ktun.edu.tr

Abstract

Adrenal tumors occur on adrenal glands and are generally detected on abdominal area scans. Adrenal tumors, which are incidentally detected, release vital hormones. These types of tumors that can be malignant affect body metabolism. Both of benign and malign adrenal tumors can have a similar size, intensity, and shape, this situation may lead to wrong decision during diagnosis and characterization of tumors. Thus, biopsy is done to confirm diagnosis of tumor types. In this study, adrenal tumor characteriza-tion is handled by using magnetic resonance images. In this way, it is wanted that patient can be disentangled from one or more imaging modalities (some of them can includes X-ray) and biopsy. An adrenal tumor image set, which includes five types of adrenal tumors and has 112 benign tumors and 10 malign tumors, was used in this study. Two data sets were created from the adrenal tumor image set by manu-ally/semiautomatically segmented adrenal tumors and feature sets of these data sets are constituted by different methods. Two-dimensional gray-level co-occurrence matrix (2D-GLCM), gray-level run-length matrix (GLRLM), and two-dimensional discrete wavelet transform (2D-DWT) methods were analyzed to reveal the most effective fea-tures on adrenal tumor characterization. Feature sets were classified in two ways: beni-gn/malign (binary classification) and type characterization (multiclass classification). Support vector machine and artificial neural network classified feature sets. The best performance on benign/malign classification was obtained by the 2D-GLCM feature set. The best results were assessed with sensitivity, specificity, accuracy, precision, and F-score metrics and they were 99.17%, 90%, 98.4%, 99.17%, and 99.13%, respec-tively. The highest classification performance on type characterization was obtained by the 2D-DWT feature set as 59.62%, 96.17%, 93.19%, 54.69%, and 54.94% for sen-sitivity, specificity, accuracy, precision, and F-score metrics, respectively.

K E Y W O R D S

adrenal glands, adrenal tumor classification, feature extraction, MR images, segmentation

1 | I N T R O D U C T I O N

Adrenal glands consist of the cortex and medulla. A hormone unbalance occurs after excessive hormone release from the medulla or cortex. This causes hormone-based clinical symp-toms. At this stage, imaging methods are used to detect the

adrenal tumor character. Some of the imaging techniques that are used in diagnosis are computed tomography (CT), magnetic

resonance (MR), and ultrasound.1Adrenal tumors can occur at

all ages. In the United States, between 200 and 500 people get this disease each year. Malign adrenal tumors are detected in adults, mostly. This disease occurs more in men than in DOI: 10.1002/ima.22358

(2)

women. If an adrenal tumor occurred on the adrenal gland only, without spreading to neighboring organs, the 5-year survival rate was determined to be 65%; if it has spread to neighboring organs, the 5-year survival rate was determined to be 44%. If an adrenal tumor has spread into far regions of the body, the

5-year survival rate was determined to be 7%.2In Konya city of

Turkey, 570 people consulted Selçuk University with a possi-ble adrenal tumor during 2011 to 2018. MR and CT scans were done on 478 and 92 people, respectively. Adrenal tumors were detected on 114 people who were scanned by MR machine. All of the people who were scanned by the CT machine had an adrenal tumor. During the CT scan, the patient is exposed to high radiation. However, the patient is not exposed to radiation during MR scan. Therefore, usage of MRI during the tumor detection and diagnosis is healthier for the patient. This study aims to implement a diagnosis system by using MR images for patients who have adrenal tumors.

In the literature, there are a few studies that have

consid-ered adrenal tumor classification. Saiprasad et al3

implemented ROI (region of interest) selection with the aim of extracting adrenal glands, which have adenomas, on CT images. Adrenal glands were manually segmented on the selected ROI, and histograms of each adenoma were calcu-lated. The mean classification accuracy was obtained as

87%. Biehl et al4proposed a machine learning system which

separates benign adrenocortical adenoma and malign adre-nocortical tumors. In their study, the classification process was done by using urine excretion values of steroid metabo-lites of 102 benign and 45 malign tumors. Generalized learn-ing vector quantization was used to separate the classes. In malign adrenal detection, high sensitivity and specificity

values were obtained. Biehl5 analyzed steroid metabolites

that represent the urine excretion of 32 biomarker patients during 24 hours, by the matrix relevance learning method. Tumors were classified as benign or malign. The obtained results showed that the proposed method can be an effective

diagnosis method. Saiprasad et al6 used the random forest

algorithm for the automatic detection of adrenal abnormality. In their method, CT images were separated into three classes by pixel-based classification: adrenal, left adrenal, and back-ground. Then, the adrenal abnormality was detected by his-togram analysis. The mean sensitivity and specificity values on 20 images were 80% and 90%, respectively. Koyuncu

and Ceylan7 studied the adrenal tumor classification of

dynamic CT images. Benign (24) and malign (8) adrenal tumor images were used in their study. ROIs, which have adrenal tumors, were selected, and the feature extraction pro-cess was implemented on the selected ROI by gray-level co-occurrence matrix (GLCM) and second-degree statistical methods. The extracted features were classified by the bounded PSO-artificial neural network (ANN) classifier. This method achieved a classification accuracy of 78.95%.

Li et al8extracted the features of benign and malign tumors

by using GLCM on CT images. Then, the extracted features were classified by the spatial Bayesian modeling method. There were 121 benign tumors and 109 malign tumors. These tumors were classified and the classification accuracy

was 80%. Koyuncu et al9performed an adrenal tumor

classi-fication on 114 dynamic CT images that had 90 benign and 24 malign tumors. After the feature extraction process, the extracted features were classified by different classifiers. Five types of tumors were classified as benign/malign. The

classification performance was 80.7%. Romeo et al10

per-formed an adrenal tumor classification by using MR images. Features of the manually segmented adrenal tumors were extracted and classified by the J48 algorithm. In their study, 60 MR images that have two different types of benign tumors (lipid-rich adenomas, lipid-poor adenomas) and one type of malign tumors (non-adenomas) were used, and the classification accuracy was 80%.

Literature findings show that mostly CT images have been used for adrenal tumor classification. The studies performed

beni-gn/malign classification. There is one study,10which performed a

tumor classification on MR images. In Reference 10, type charac-terization process was implemented on only 60 images.

In this study, 114 MR images, which have five different tumor types and consist of 112 benign tumors and 10 malign tumors, are classified as benign/malign adrenal tumors. In addition, the type characterization process of adrenal tumors is performed on the data set. Feature extraction methods and classification techniques are used to classify the adrenal tumor images. This study is the most detailed study in the literature on adrenal tumor classification on MR images.

This article is organized as follows. Section 2 analyzes the images statistically and visually. Section 3 briefly explains the manual segmentation method, the used feature extraction, and the classification techniques. Section 4 pre-sents the classification results. Sections 5 and 6 are the dis-cussion and conclusion, respectively.

2 | M A T E R I A L S

2.1 | Statistical features of data set

The data consist of 114 T1 and T2 phase MR abdominal images from the Radiology Department of the Medicine Faculty at Selçuk University. MR images were procured by SIEMENS AREA 1.5 2013 tool. Three experienced radiolo-gist in the Radiology Department identified and labeled the adrenal tumors and tumor types on the MR images.

In the data set, the radiologists detected that 112 benign

tumors have four types: adenoma, cyst, lipoma, and

lymphangioma. Table 1 shows the number of images and the classes and types of adrenal tumors. Nine images that have a

(3)

malign tumor were defined as metastasis. Some images (106) have a unilateral tumor, and eight images have bilateral tumors. Bilateral tumors in the images were separated into two different images. Therefore, 122 adrenal tumor images were obtained for benign/malign classification. There were five types for a total of 122 adrenal tumors (112 benign tumors have four types and 10 malign tumors have one type) for benign/malign classifica-tion. As seen in Table 1, there are only two images which consist of lymphangioma type; therefore, lymphangioma type was not used during type characterization process. Thus, there were four types for a total of 120 adrenal tumors (110 benign tumors have three types and 10 malign tumors have one type) for type charac-terization process.

2.2 | Visual features of data set

The images in the data set have some similarities and differ-ences, which make the classification process difficult. The differences that belong to the same class and the similarities that belong to a different class prevent the classifier from generalizing. Figure 1 shows an adrenal tumor, which was encountered on an MR scan of an adrenal gland.

As shown in Figure 1, a healthy adrenal gland is seen in the left scan. However, in the right scan, an adrenal tumor is seen. If a tumor occurred in both adrenal glands, there would be a bilateral tumor situation.

Problems that were expressed above can affect the classi-fication performance. These problems are as follows:

1. The same type of tumor can have different shapes and gray-levels.

2. The same type of tumor can have homogeneous or het-erogeneous structures.

3. Different types of tumors can have similar shapes. 4. Different types of tumors can have similar gray-levels.

Figure 2 shows sample images of the problems men-tioned above that the adrenal tumors in the data set can have similar features such as shape, gray-level, and size. At this point, a detailed classification process is needed to classify each tumor type.

3 | M E T H O D S

This study performs an adrenal tumor classification in two different ways. First, the benign/malign classification

pro-cess is implemented. In this propro-cess, the 2D-GLCM,11

gray-level run-length matrix (GLRLM),12 and two-dimensional

discrete wavelet transform (2D-DWT)13methods are used to

extract the features from the data set which is formed by manual or semiautomatic segmentation. The three extracted feature sets were classified by an support vector machine

(SVM)14separately, and benign/malign classification results

for each feature set are obtained. Second, the type characteri-zation process is implemented on adrenal tumors segmented by manual or semiautomatic segmentation. In the type char-acterization process, the 2D-GLCM and 2D-DWT methods are used to extract the features. Then, the two extracted

T A B L E 1 The adrenal tumor image set used

Classification

process Class Type Number of images

Number of tumors

Benign/malign Classification

Benign Adenoma 88 (82 unilateral, 6 bilateral) 94 Cyst 6 (6 unilateral) 6 Lipoma 9 (8 unilateral, 1 bilateral) 10 Lymphangioma 2 (2 unilateral) 2 Malign Metastasis 9 (8 unilateral, 1

bilateral)

10

Total 114 122

Type Characterization Benign Adenoma 88 (82 unilateral, 6 bilateral) 94 Cyst 6 (6 unilateral) 6 Lipoma 9 (8 unilateral, 1 bilateral) 10

Malign Metastasis 9 (8 unilateral, 1 bilateral)

10

(4)

feature sets are classified by the ANN separately, and type characterization results for these two feature extraction tech-niques are obtained.

3.1 | Manual tumor segmentation before

classification

The adrenal tumor data set consists of 122 adrenal tumor images, which are 942 × 1160. Adrenal tumors on adrenal glands were manually segmented on T1- and T2-weighted MR images by

three experienced radiologists, separately. Then, tumors were set to be the same size as the largest tumor. Each tumor is positioned in the center of the 501 × 501 image. Figure 3 shows the data preparation process on T1-weighted images.

3.2 | Semiautomatic tumor segmentation

before classification

In this study, semiautomatically segmented adrenal tumors,

which were obtained in our previous study,15were classified.

F I G U R E 1 Normal adrenal gland and a tumor occurrence [Color figure can be viewed at wileyonlinelibrary.com]

(5)

During the semiautomatic segmentation method, T1- and T2-weighted images were used together and adrenal tumors were segmented. The implemented semiautomatic segmenta-tion scheme is presented in Figure 4.

As seen in Figure 4, abdominal area, fat layer, liver, and spi-nal cord were segmented on tumor image. The reason was to clean the neighbor structures of the adrenal tumors. By doing so, adrenal tumors were segmented with a better accuracy.

3.3 | Classification process and methods used

The data set has 112 benign adrenal tumors and 10 malign adrenal tumors. In this study, we solve two classification problems: benign/malign classification and type characteri-zation. During the benign/malign classification process, fea-ture extraction was applied to the data set. The 2D-GLCM, GLRLM, and 2D-DWT methods were used to extract the

features of benign/malign adrenal tumors, so three different feature set were obtained. The SVM classifier classified these feature set separately. During the classification pro-cess, 2-, 5-, and 10-fold cross-validation methods were used. The mean classification results after cross-validations were obtained. Figure 5 shows the scheme of the classification process used in both of the classification problems.

The feature sets formed by using 2D-GLCM and 2D-DWT were used for type characterization of adrenal tumors. The GLRLM method was not used during type char-acterization, because the GLRLM method gave low classifi-cation performance during benign/malign classificlassifi-cation.

The ANN classifier was used to classify the extracted features, because the ANN is a strong multiclass classifier. The same cross-validation process was applied during type characterization. The methods used in this study are as follows:

F I G U R E 3 Data preparation process before classification of adrenal tumors [Color figure can be viewed at wileyonlinelibrary.com]

(6)

• 2D-GLCM

• Gray-level run-length matrix • 2D-discrete wavelet transform • Support vector machines • Artificial neural networks

3.3.1 | 2D-GLCM

Gray-level co-occurrence matrix is a feature extraction method to obtain the second-degree statistical features. GLCM includes the relationships of different angles between the pixels of an image. Let a co-occurrence matrix that is obtained from a

G image be represented as P = [p(i, j | d,θ)]. At this point, the

co-occurrence matrix is used to evaluate the ith pixel frequency features with the jth neighbor pixel frequency features by

con-sidering the d length andθ direction. In this study, d = 1 was

selected. And so theθ angle is taken as 0.9The angular

sec-ondary moment, contrast, correlation, sum of squares: vari-ance, inverse difference moment, sum average, sum varivari-ance, sum entropy, entropy, difference entropy, difference variance, information measures of correlation 1, information measures of correlation 2, autocorrelation, dissimilarity, cluster shade, cluster prominence, maximum probability, and the inverse

difference features were extracted from the data set by the

2D-GLCM method.11,16,17

3.3.2 | Gray-level run-length matrix

Gray-level run-length matrix is a method of texture feature extraction on a high level. Let G be the number of gray-levels, R is the longest run, and N is the number of pixels in

the image. A GLRLM matrix is GXR, and each p(i,j |θ)

ele-ment gives the number of occurrences in theθ direction with

i gray-level and j run-length. The short-run emphasis, long-run emphasis, gray-level nonuniformity, long-run-length non-uniformity, run percentage, low gray-level run emphasis, and high gray-level run emphasis features were extracted

from the data set by the GLRLM method.12,18

3.3.3 | 2D-discrete wavelet transform

Two dimensional-DWT is a filter bank, which separates the image into frequency sub-bands by using an h low-pass filter and g high-pass filter. Approximation coefficients (LL), hor-izontal details (LH), vertical details (HL), and diagonal details (HH) represent the lowest frequency, horizontal high frequencies, vertical high frequencies, and high frequencies

in both directions, respectively.19The feature set was created

F I G U R E 4 Semiautomatic segmentation method on adrenal tumor data set. MR, magnetic resonance [Color figure can be viewed at wileyonlinelibrary.com]

(7)

by LL, LH, HL, and HH coefficients after 2D-DWT. The mean, variance, kurtosis, skewness, entropy, and energy of features were calculated from the coefficients.

During the feature extraction process with 2D-DWT, 36 different wavelet types were used. These wavelets are

db1, db2, db3, db4, db5, db6, db7, db8, db9, coif1, coif2, coif3, coif4, coif5, sym1, sym2, sym3, sym4, sym5, sym6, sym7, sym8, bior1.3, bior1.5, bior2.2, bior2.4, bior2.6, bior2.8, bior3.1, bior3.3, bior3.5, bior3.7, bior3.9, bior4.4, bior5.5, and bior6.8. After each wavelet transform, six fea-tures were extracted from four different coefficients (LL, LH, HL, and HH). Therefore, 24 features were obtained for one wavelet type after 2D-DWT.

3.3.4 | Support vector machines

Support vector machine is a learning method, which gives high classification accuracy in many applications. An SVM is based on two ideas. The first idea is to map feature vectors to a high dimensional space with a nonlinear method and to use linear classifiers in this new space. The second idea is to find a hyperplane, which separates the data with a high mar-gin. This plane is the best plane, which can separate the data

as well as possible.14

In this study, the three feature sets, which were created by 2D-GLCM, GLRLM, and 2D-DWT methods, were clas-sified by an SVM in the benign/malign classification of adrenal tumors.

3.3.5 | Artificial neural networks

An ANN is a mathematical modeling method that can learn the behavior of a system by using inputs and outputs. An ANN can produce acceptable solutions to unimplemented or unlearned problems. An ANN has an input layer, a hidden layer, and an output layer. These parameters can change according to the problem and are determined by trial and

error.20

In this study, an ANN was used in the type characteri-zation process. The ANN classified the two feature sets, which were created by the 2D-GLCM and 2D-DWT methods.

4 | E X P E R I M E N T A L R E S U L T S

This study presents an adrenal tumor classification in two ways: benign/malign classification and type characterization. Adrenal tumors were manually/semiautomatic segmented on both T1- and T2-weighted MR images. Each tumor image was set to be in an ROI, which is 501 × 501. Three feature sets were created from ROIs of adrenal tumors by using

2D-GLCM, GLRLM, and 2D-DWT methods which

obtained 19, 7, and 24 features, respectively.

Five different evaluation metrics (Equations 1-5) were used to assess the proposed method. These metrics are sensi-tivity (SEN), specificity (SPE), accuracy (ACC), precision (PRE), and F-score.

F I G U R E 5 Scheme of the classification process [Color figure can be viewed at wileyonlinelibrary.com]

(8)

Sensitivity = TP= TP + FNð Þ ð1Þ

Specificity = TN= TN + FPð Þ ð2Þ

Accuracy = TP + TNð Þ= TP + TN + FN + FPð Þ ð3Þ

Precision = TP= TP + FPð Þ ð4Þ

F−score = 2*TPð Þ= 2*TP + FP + FNð Þ ð5Þ

TP, TN, FP, and FN values are the number of true positives,

true negatives, false positives, and false negatives, respectively.21

4.1 | Benign/malign classification results of

adrenal tumors manually segmented

In this study, first, adrenal tumors that were manually seg-mented from T1- and T2-weighted MR images separately

were classified by SVM as benign/malign. Table 2 presents the classification results of adrenal tumors, which were seg-mented from T1-weighted MR images.

As seen in Table 2, the best classification result was obtained by the 2D-GLCM feature set with 10-fold cross-validation. The highest specificity in the GLRLM feature set was 10%. This means that malign adrenal tumors could not be correctly classified by the GLRLM feature set. However, the specificity was 70% in the 2D-DWT feature set. There-fore, 7 in 10 tumors could be correctly classified. In addi-tion, the specificity value was 90% by the 2D-GLCM feature set with 5- and 10-fold cross-validation. The mean accuracy value was 98.4% by the 2D-GLCM feature set with 10-fold cross-validation. The cost (C) parameter of SVM algorithm was taken as 1, which is default value of the SVM algo-rithm. Table 3 shows the classification results of adrenal tumors, which were segmented from T2-weighted MR images.

T A B L E 2 Classification results of adrenal tumors, which were manually segmented from T1-weighted images

Cross-validation Feature extraction method Number of features Evaluation metrics

SEN (%) SPE (%) ACC (%) PRE (%) F-score (%)

2-Fold 2D-GLCM 19 99.11 ± 1.26 60 ± 28.28 95.9 ± 1.16 96.56 ± 2.33 97.8 ± 0.58 GLRLM 7 99.11 ± 1.26 0 90.98 ± 1.16 91.73 ± 0.09 95.28 ± 0.64 2D-DWT db4 24 99.11 ± 1.26 70 ± 42.43 96.72 ± 2.32 97.46 ± 3.6 98.25 ± 1.21 5-Fold 2D-GLCM 19 99.13 ± 1.94 90 ± 22.36 98.37 ± 2.24 99.13 ± 1.94 99.11 ± 1.22 GLRLM 7 99.13 ± 1.94 0 91 ± 1.68 91.73 ± 0.15 95.28 ± 0.93 2D-DWT bior5.5 24 100 70 ± 44.72 97.53 ± 3.71 97.5 ± 3.73 97.7 ± 1.94 10-Fold 2D-GLCM 19 99.17 ± 2.64 90 ± 31.62 98.4 ± 3.38 99.17 ± 2.64 99.13 ± 1.83 GLRLM 7 99.17 ± 2.64 10 ± 31.62 91.86 ± 3.64 92.56 ± 2.62 95.72 ± 1.97 2D-DWT sym7 24 99.17 ± 2.64 70 ± 48.3 96.73 ± 4.22 97.5 ± 4.03 98.26 ± 2.25

Abbreviations: ACC, accuracy; PRE, precision; SEN, sensitivity; SPE, specificity.

T A B L E 3 The classification results of adrenal tumors, which were manually segmented from T2-weighted images

2-Fold 2D-GLCM 19 99.11 ± 1.26 70 ± 14.14 96.72 97.38 ± 1.18 98.23 ± 0.02 GLRLM 7 100 0 91.8 91.8 95.73 2D-DWT db8 24 99.11 ± 1.26 70 ± 14.14 96.72 ± 2.32 97.37 ± 1.24 98.23 ± 1.25 5-Fold 2D-GLCM 19 99.13 ± 1.94 80 ± 27.39 97.53 ± 2.25 98.26 ± 2.38 98.67 ± 1.22 GLRLM 7 100 0 91.8 ± 0.18 91.8 ± 0.18 95.72 ± 0,1 2D-DWT db9 24 100 60 ± 41.83 96.73 ± 3.37 96.66 ± 3.39 98.28 ± 1.76 10-Fold 2D-GLCM 19 99.17 ± 2.64 80 ± 42.16 97.56 ± 3.93 98.3 ± 3.51 98.7 ± 2.1 GLRLM 7 100 0 91.8 ± 0.27 91.8 ± 0.27 95.72 ± 0.15 2D-DWT bior3.7 24 99.09 ± 2.87 70 ± 48.3 96.73 ± 5.77 97.49 ± 4.06 98.26 ± 3.11

(9)

T A BLE 4 The type characterization results of adrenal tumors, which were manually segmented from T1-weighted images and ANN parameters Cross-validation Feature extraction method Number of features Evaluation Metrics Learning rate Number of hidden layers Number of iterations SEN (%) SPE (%) ACC (%) PRE (%) F-score (%) 2-Fold 2D-GLCM 19 54.26 ± 2.65 86.72 ± 0.33 92 ± 0.1 60.31 ± 8.52 54.32 ± 3.43 1.4 50 6000 2D-DWT db7 24 59.62 ± 9.38 96.17 ± 1.13 93.19 ± 3.74 54.69 ± 12.52 54.94 ± 9.14 0.1 70 5000 5-Fold 2D-GLCM 19 42.54 ± 6.15 89.59 ± 7.24 87.83 ± 6.09 50.61 ± 16.65 44.31 ± 8.36 1.7 50 5000 2D-DWT sym6 24 60.39 ± 8.21 93.91 ± 1.05 90.93 ± 4.14 62.22 ± 10.45 58.57 ± 9.65 0.5 60 5000 Abbreviations: ACC, accuracy; ANN, artificial neural network; PRE, precision; SEN, sensitivity; SPE, specificity. TABL E 5 Type characterization results of adrenal tumors, which were manually segmented from T2-weighted images, and ANN parameters Cross-validation Feature extraction method Number of features Evaluation Metrics Learning rate Number of hidden layers Number of iterations SEN (%) SPE (%) ACC (%) PRE (%) F-score (%) 2-Fold 2D-GLCM 19 63.38 ± 13.25 95.36 ± 2.65 91.8 ± 3.08 58.41 ± 25.15 55.87 ± 16.47 0.5 80 8000 2D-DWT db1 24 53.84 ± 9.64 91.75 ± 5.04 92.12 ± 0.81 56.13 ± 15.98 52.33 ± 5.81 1.1 60 9000 5-Fold 2D-GLCM 19 59 ± 10.69 91.64 ± 5.27 89.34 ± 1.97 68.68 ± 13.21 57.24 ± 5.8 0.1 70 10 000 2D-DWT db4 24 56.17 ± 8.72 92.11 ± 4.73 91.61 ± 3.82 64.49 ± 13.36 58.08 ± 12.39 1.9 100 9000 Abbreviations: ACC, accuracy; ANN, artificial neural network; PRE, precision; SEN, sensitivity; SPE, specificity.

(10)

As seen in Table 3, malign tumors could not be classified correctly by the GLRLM feature set. Malign tumors were classified by the 2D-DWT feature set with 70% specificity, at most. However, the sensitivity and specificity were 99.17% and 80% with the 2D-GLCM feature set. Tables 2 and 3 show that the classification error of T1-weighted images is lower than T2-weighted images.

4.2 | Type characterization results of adrenal

tumors manually segmented

Adrenal tumors were classified according to their types in this section. Adrenal tumors, which include four types as adenoma, cyst, lipoma, and metastasis, are seen in Table 1. The ANN was used as the classifier, which is a strong multi-class multi-classifier, to multi-classify the tumors. An ANN has different parameters such as learning rate, number of iterations, and number of hidden layers. These three parameters were chan-ged. The optimum numbers of parameters were obtained for each cross-validation method after experiments. The learn-ing rate was increased from 0.1 to 2 by 0.1; the number of hidden layers was increased from 50 to 100 by 10; the num-ber of iterations was increased from 5000 to 10 000 by 1000. Traingdx function was used as activation function dur-ing the traindur-ing stage.

Two different feature sets were used in type characteriza-tion. The GLRLM method was not used to extract features from adrenal tumor types. The reason for this is that the GLRLM feature set in the benign/malign classification pro-cess showed low performance. The 2D-GLCM feature set created one feature set, and the 2D-DWT feature set created 36 different feature sets that were obtained by db1, db2, db3, db4, db5, db6, db7, db8, db9, coif1, coif2, coif3, coif4, coif5, sym1, sym2, sym3, sym4, sym5, sym6, sym7, sym8, bior1.3, bior1.5, bior2.2, bior2.4, bior2.6, bior2.8, bior3.1, bior3.3,

bior3.5, bior3.7, bior3.9, bior4.4, bior5.5, and bior6.8 wavelets. In addition, during the classification process, only 2- and 5-fold cross-validations were used. The reason for not using 10-fold cross-validation is that the cyst type has only six images. The number of cyst images is not enough for 10-fold cross-validation. Table 4 shows the type characteri-zation results of adrenal tumors, which were segmented from T1-weighted MR images, and ANN parameters, giving the best results.

As seen in Table 4, the best classification results were obtained with 2-fold cross-validation by the 2D-DWT fea-ture set. The best feafea-ture set among 36 feafea-ture sets was

obtained with the“db7” wavelet. The highest classification

accuracy was 93.19%. Table 5 presents the type characteri-zation results on adrenal tumors, which were segmented from T2-weighted MR images, and ANN parameters, giving the best results.

Table 5 shows that the best classification result was obtained with 2-fold cross-validation. The best feature set

was obtained with the “db1” wavelet in the 2D-DWT

method. The highest classification accuracy was 92.12%. Tables 4 and 5 show that the classification error of T1-weighted images is lower than T2-weighted images.

4.3 | Benign/malign classification results of

adrenal tumors segmented as semiautomatic

In addition to classification of adrenal tumors segmented manually, this study classified adrenal tumors segmented by a semiautomatic segmentation method. The implemented pipeline for semiautomatic adrenal tumor segmentation was

presented in our previous study.15Table 6 presents the

beni-gn/malign classification results that is obtained by SVM on adrenal tumors, which were segmented in Reference 15.

T A B L E 6 The classification results of adrenal tumors, which were segmented in Reference 15

2-Fold 2D-GLCM 19 94.43 ± 2.53 10 ± 14.14 89.34 ± 1.16 92.32 ± 0.93 94.31 ± 0.73 GLRLM 7 100 0 91.8 91.8 95.73 2D-DWT db2 24 100 60 96.72 96.55 98.25 5-Fold 2D-GLCM 19 98.26 ± 3.89 50 ± 35.36 94.27 ± 2.22 95.76 ± 2.95 96.91 ± 1.24 GLRLM 7 100 0 91.8 ± 0.18 91.8 ± 0.18 95.72 ± 0.1 2D-DWT bior2.8 24 100 60 ± 41.83 96.73 ± 3.37 96.66 ± 3.39 98.28 ± 1.76 10-fold 2D-GLCM 19 98.26 ± 3.68 50 ± 52.7 94.23 ± 3.99 95.83 ± 4.39 96.92 ± 2.13 GLRLM 7 100 0 91.79 ± 0.27 91.79 ± 0.27 95.72 ± 0.15 2D-DWT sym2 24 99.17 ± 2.64 60 ± 51.64 96.03 ± 5.55 96.73 ± 4.22 97.89 ± 2.96

(11)

TAB L E 7 Type characterization results of adrenal tumors, which were segmented in Reference 15 Cross-validation Feature extraction method Number of features Evaluation Metrics Learning rate Number of hidden layers Number of iterations SEN (%) SPE (%) ACC (%) PRE (%) F-score (%) 2-Fold 2D-GLCM 19 53.66 ± 4.55 89.6 ± 2.88 85.15 ± 0.65 43.99 ± 3.98 45.85 ± 3.57 0.5 90 6000 2D-DWT bior3.1 24 50.79 ± 9.69 87.85 ± 5.22 89.3 ± 0.99 61.41 ± 1.34 52.54 ± 5.24 1.1 100 6000 5-Fold 2D-GLCM 19 37.92 ± 9.45 86.49 ± 6.51 84.46 ± 5.82 45.8 ± 11.67 39.48 ± 9.43 1.3 50 5000 2D-DWT bior2.4 24 48.2 ± 28.72 80.51 ± 13.47 86.59 ± 4.62 50.64 ± 25.35 46.83 ± 23.99 0.9 60 9000 Abbreviations: ACC, accuracy; PRE, precision; SEN, sensitivity; SPE, specificity. TAB L E 8 Comparison of the best classification results of adrenal tumors, which were segmented by manual and semiautomatic methods, in benign/malign classi fication Segmentation method Cross-validation Feature extraction method Number of features Evaluation metrics SEN (%) SPE (%) ACC (%) PRE (%) F-score (%) Manual 10-Fold 2D-GLCM 19 99.17 ± 2.64 90 ± 31.62 98.4 ± 3.38 99.17 ± 2.64 99.13 ± 1.83 Semiautomatic 5-Fold 2D-DWT bior2.8 24 100 60 ± 41.83 96.73 ± 3.37 96.66 ± 3.39 98.28 ± 1.76 Abbreviations: ACC, accuracy; PRE, precision; SEN, sensitivity; SPE, specificity.

(12)

was obtained with the “bior2.8” wavelet in the 2D-DWT

method. The highest classification accuracy was 96.73%. The specificity value was obtained as 60%, at most. Even if all metrics shows high performance, the specificity value shows that the proposed method detects the malign tumor with 60% performance on adrenal tumors segmented by semiautomatic method. The cost (C) parameter of SVM algorithm was taken as 1, which is default value of the SVM algorithm.

4.4 | Type characterization results of adrenal

tumors segmented as semiautomatic

The same process in Section 4.2 was applied to adrenal tumors segmented by semiautomatic segmentation method. ANN algorithm was used to classify the adrenal tumors. Table 7 presents the type characterization results on adrenal tumors, which were segmented in Reference 15.

was obtained with the “bior3.1” wavelet in the 2D-DWT

method. The highest classification accuracy was 89.3%. In type characterization, the sensitivity values did not show very high performance as in other metrics. So, the proposed method did not have very high performance on type charac-terization of semiautomatic segmentation results.

5 | D I S C U S S I O N

The adrenal tumor data set consisted of different types, with problems occurring during the classification process. Even if the adrenal tumor types are different from each other, the gray-levels, sizes, and shapes of adrenal tumors can be simi-lar to each other, especially the types of benign tumors. Therefore, three feature extraction methods were utilized to find the feature set that separates the tumor types with a high accuracy. A total of 38 feature sets were created for beni-gn/malign adrenal tumor classification with the 2D-GLCM (1 set), GLRLM (1 set), and 2D-DWT (36 sets) methods. Tables 8 and 9 present the highest classification results of adrenal tumors, which were segmented by manual and semi-automatic methods.

As seen in Tables 8 and 9, the best classification perfor-mances were obtained by manual segmentation method. In

our previous study,15 the mean segmentation performance

was obtained around 60%. There is an information loss about adrenal tumors around 40%. Therefore, the classifica-tion performances of semiautomatic segmentaclassifica-tion results have lower performance than manual segmentation method.

TABL E 9 Comparison of the best classification results of adrenal tumors, which were segmented by manual and semi-automatic methods, in type characterizati on Segmentation method Cross-validation Feature extraction method Number of features Evaluation metric Learning rate Number of hidden layers Number of iterations SEN (%) SPE (%) ACC (%) PRE (%) F-score (%) Manual 2-Fold 2D-DWT db7 24 59.62 ± 9.38 96.17 ± 1.13 93.19 ± 3.74 54.69 ± 12.52 54.94 ± 9.14 0.1 70 5000 Semiautomatic 2-Fold 2D-DWT bior3.1 24 50.79 ± 9.69 87.85 ± 5.22 89.3 ± 0.99 61.41 ± 1.34 52.54 ± 5.24 1,1 100 6000 Abbreviations: ACC, accuracy; PRE, precision; SEN, sensitivity; SPE, specificity.

(13)

If the performance of semiautomatic segmentation can be increased, the better classification performance is obtained.

There are few studies on adrenal tumor classification in the literature. Table 10 presents a literature comparison of adrenal tumor classification.

Koyuncu and Ceylan7classified adrenal tumors by using

an ROI, which includes an adrenal tumor. They performed only benign/malign classification and had an accuracy of

78.95%. Koyuncu et al9used five adrenal tumor types, and a

binary classification process (benign/malign classification) was performed on CT images. The 2D-GLCM method was used to extract the features. Four different feature sets were

obtained by using 0, 45, 90, and 135 directions in the

2D-GLCM method. The best classification accuracy was

80.7% with 0 direction. Therefore, in our study, we used

only the 0 direction in the 2D-GLCM method. Romeo

et al10 classified only 60 MR images, which have three

tumor types (lipid-rich adenomas, lipid-poor adenomas, and non-adenomas), with the J48 algorithm, and they obtained an accuracy of 80%.

This study:

• Used 122 adrenal tumors, which have five different types. • Performed benign/malign classification and obtained an

accuracy of 98.4%.

• Performed type characterization and obtained an accuracy of 93.19%.

• Had more MR images than the studies currently available.

• Had a wider range of benign tumors.

• Achieved higher classification performance than the stud-ies currently available.

• Is the most detailed adrenal tumor classification study currently available.

6 | C O N C L U S I O N

Adrenal tumor classification studies mostly use CT images. Some studies use urine excretion values and some studies use MR images. Therefore, there are no easily available standardized data sets of adrenal tumors. Although this study has better classification performance than the literature stud-ies, the aim of this study is to make a detailed analysis of adrenal tumors, which includes benign/malign classification and type characterization.

Tables 8 and 9 show that the manual segmentation method has good classification performance. In addition, among the feature extraction methods, the 2D-GLCM fea-ture set and SVM classifier gave better results in beni-gn/malign classification; the 2D-DWT feature set and ANN classifier gave better results in type characterization.

TABLE 10 Literature comparison of adrenal tumor classification Study Method Data Number of tumor types Number of total tumors Classification type SEN (%) SPE (%) ACC (%) Area Under Curve (AUC) (%) 4 Admire-LVQ Urine Test 2 147 (102 benign, 45 malign) Benign/malign 90 90 -96.5 6 Random forest CT 1 2 0 (10 benign, 10 malign) Benign/malign 80 90 -8 Bayesian Modeling of GLCM CT -230 (121 benign, 109 malign) Benign/malign -80 -7 PSO-YSA Dynamic CT 3 3 2 (24 benign, 8 malign) Benign/malign -78.95 84.29 9 Bounded PSO-YSA Dynamic CT 8 114 (90 benign, 24 malign) Benign/malign 75 85.22 80.7 78.61 10 J48 MR 3 6 0 (20 lipid-rich adenoma, 20 lipid-poor adenoma, 20 non-adenoma) Type characterization 78.67 79.67 80 79.47 This study 2D-GLCM-SVM MR 9 122 (112 benign, 10 malign) Benign/malign 99.17 90 98.4 -This study 2D-DWT-ANN MR 9 9 4 adenoma, 6 cyst, 10 lipoma, 10 metastasized Type characterization 59.62 96.17 93.19 -Abbreviations: ACC, accuracy; CT, computed tomography; MR, magnetic resonance; SEN, sensitivity; SPE, specificity.

(14)

This study performed benign/malign classification and type characterization of adrenal tumors on MR images, pro-viding the most detailed adrenal tumor classification study currently available. T1- and T2-weighted MR images were used to classify adrenal tumors, and the best classification results were obtained on T1-weighted MR images.

The radiologists use different slices and scans to decide the type of adrenal tumors. Therefore, future works will examine 3D adrenal tumor data set, which will be created by using different slices and scans of MR imaging. In addition, more images will be taken from the Medicine Faculty at Selçuk University. Generalization of the classification struc-ture of type characterization will also be tried.

O R C I D

Mucahid Barstugan

https://orcid.org/0000-0001-9790-5890

R E F E R E N C E S

1. Kumar R, Anand V, Jana S. Adrenal lesions: Role of computed tomography, magnetic resonance imaging, 18F-fluorodeoxy-glucose-positron emission tomography, and positron emission tomography/computed tomography. Cancer Imaging. Elsevier; 2008:269-279.

2. Michigan University CCC. A Patient's Guide to Adrenocortical Cancer. 2012. http://www.cancer.med.umich.edu/files/adrenal-patient-handbook.pdf.

3. Saiprasad G, Saenz N, Chang C, Siegel E. Prototype decision sup-port system for evaluation of adrenal glands incorporated into rou-tine CT workflow. Radiological Society of North America 2010 Scientific Assembly and Annual Meeting; 2010.

4. Biehl M, Schneider P, Smith D, et al. Matrix Relevance LVQ in Steroid Metabolomics Based Classification of Adrenal Tumors. ESANN 2012 proceedings, European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learn-ing. Bruges (Belgium), 25-27 April 2012.

5. Biehl M. Admire LVQ—adaptive distance measures in relevance learning vector quantization. KI-Künstliche Intelligenz. 2012;26 (4):391-395.

6. Saiprasad G, Chang C-I, Safdar N, Saenz N, Siegel E. Adrenal gland abnormality detection using random forest classification. J Digit Imaging. 2013;26(5):891-897.

7. Koyuncu H, Ceylan R. Classification of adrenal lesions by bounded PSO-NN. Signal Processing and Communications Appli-cations Conference (SIU), 2017 25th; 2017: IEEE.

8. Li X, Guindani M, Ng C, Hobbs B, editors. Classification of adre-nal lesions through spatial Bayesian modeling of GLCM.

Biomedical Imaging (ISBI 2017), 2017 IEEE 14th International Symposium on; 2017: IEEE.

9. Koyuncu H, Ceylan R, Asoglu S, Cebeci H, Koplay M. An exten-sive study for binary characterisation of adrenal tumours. Med Biol Eng Comput. 2018;57:1-14.

10. Romeo V, Maurea S, Cuocolo R, et al. Characterization of adrenal lesions on unenhanced MRI using texture analysis: a machine-learning approach. J Magn Reson Imaging. 2018;48:198-204. 11. Haralick RM, Shanmugam K, Dinstein IH. Textural features for

image classification. IEEE Trans Syst Man Cybern. 1973;3(6): 610-621.

12. Sohail ASM, Bhattacharya P, Mudur SP, Krishnamurthy S. Local relative GLRLM-based texture feature extraction for classifying ultrasound medical images. Electrical and Computer Engineering (CCECE), 2011 24th Canadian Conference on; 2011: IEEE. 13. Shensa MJ. The discrete wavelet transform: wedding the a trous

and Mallat algorithms. IEEE Trans Signal Process. 1992;40(10): 2464-2482.

14. Kulkarni SR, Harman G. Statistical learning theory: a tutorial. Wiley Interdiscip Rev Comput Stat. 2011;3(6):543-556.

15. Barstugan M, Ceylan R, Asoglu S, Cebeci H, Koplay M. Adrenal tumor segmentation method for MR images. Comput Methods Programs Biomed. 2018;164:87-100.

16. Soh L-K, Tsatsoulis C. Texture analysis of SAR sea ice imagery using gray level co-occurrence matrices. CSE J Article. 1999;47: 780-795.

17. Clausi DA. An analysis of co-occurrence texture statistics as a func-tion of grey level quantizafunc-tion. Can J Remote Sens. 2002;28(1):45-62. 18. Chu A, Sehgal CM, Greenleaf JF. Use of gray value distribution of run lengths for texture analysis. Pattern Recogn Lett. 1990;11 (6):415-419.

19. Eltoukhy MM, Faye I, Samir BB. A statistical based feature extraction method for breast cancer diagnosis in digital mammo-gram using multiresolution representation. Comput Biol Med. 2012;42(1):123-128.

20. Fausett LV. Fundamentals of Neural Networks: Architectures, Algorithms, and Applications. Englewood Cliffs, NJ: Prentice-Hall; 1994.

21. Ruuska S, Hämäläinen W, Kajava S, Mughal M, Matilainen P, Mononen J. Evaluation of the confusion matrix method in the vali-dation of an automated system for measuring feeding behaviour of cattle. Behav Processes. 2018;148:56-62.

How to cite this article: Barstugan M, Ceylan R,

Asoglu S, Cebeci H, Koplay M. Adrenal tumor characterization on magnetic resonance images. Int J

Imaging Syst Technol. 2020;30:252–265.https://doi.