Application of Contrast-Enhanced Ultrasound Combined with Elastic Imaging Technology in Differential Diagnosis of Salivary Gland Tumors.

This paper proposes the effect of contrast-enhanced ultrasound combined with real-time elastic imaging technology in the differential diagnosis of salivary gland tumors. 200 patients were selected, including 120 males and 60 females. The age ranged from 9 to 83 years, with an average of 55.4 years. Among the 200 cases, there were 90 cases of single parotid gland on the right, 77 cases of single parotid gland on the left, 2 cases of bilateral (single parotid gland on each side), 2 cases of multiple parotid gland on the right (2 lesions), 1 case of 2 lesions on the left and 1 lesion on the right, and 1 case of multiple parotid gland on the left (4 lesions). 135 cases were located in the superficial lobe (78%) and 38 cases (22%) in the deep lobe of parotid gland. The ARIETTA 70 color Doppler ultrasound diagnostic instrument is used. The equipment is equipped with high-frequency contrast probe, real-time elastic imaging technology, and related software. The results showed that the detection rate of salivary gland tumors by ultrasound was 100% and the diagnostic coincidence rate was 71% (123/173). Ultrasound can not only identify the tumors in and around the parotid gland but also identify the location, size, and internal structure of the tumors. Combined with CDFI, it can make qualitative diagnosis of most benign and malignant salivary gland tumors and provide help for clinical treatment and operation plan. It is proved that contrast-enhanced ultrasound and real-time elastic imaging technology have advantages over gray-scale ultrasound in differentiating benign and malignant superficial enlarged lymph nodes, and the combined use can effectively improve the diagnostic efficiency.

1. Introduction

Salivary gland tumor is a common head and neck tumor in clinic, accounting for 3%–5% of head and neck tumors and 80% of salivary gland tumors. The pathological classification is extremely complex, in which epithelial tumors account for the vast majority and mesenchymal tissue sources are less. The imaging manifestations and biological characteristics of some salivary gland tumors with different pathological types lack specificity, so it is difficult to make a qualitative diagnosis before operation. Parotid radiography can only indirectly reflect the situation of the tumor and has little significance in the differentiation of benign and malignant tumors [1]. Puncture biopsy of salivary gland tumor is easy to cause rupture of parotid gland capsule, resulting in implant metastasis and recurrence. In recent years, with the rapid development of medical imaging technology, there are many studies on the differentiation of benign and malignant parotid tumors. Studies at home and abroad have found that multislice spiral CT (MSCT) and MRI dynamic enhanced scanning are of great value in the diagnosis of salivary gland tumors, which is conducive to the clinical formulation of surgical plans and the judgment of patients' prognosis. However, many studies have shown that it is still difficult to distinguish benign pleomorphic adenoma and malignant tumor of parotid gland with basically similar enhancement mode only by MSCT and MRI dynamic enhancement scanning [2]. A large number of studies have shown that ultrasound can identify benign and malignant tumors through the morphology, boundary, internal echo, and posterior echo of salivary gland tumors and can further evaluate the blood flow signal in salivary gland tumors through color Doppler flow imaging (CDFI), so as to provide more basis for the qualitative diagnosis of tumors [3-5].

2. Literature Review

Kim believes that the tumor has irregular shape, wide invasion of surrounding structures, and the characteristics of neurophilic growth. The nerve invasion is mainly manifested in the abnormal thickening and irregular strengthening of the affected nerve anterograde or retrograde, and denervated muscle atrophy is an important indirect sign [6]. Kim et al. believed that acinar cell carcinoma can occur in all ages, most common in 40∼60 years old. It is a low-grade malignant tumor. When the tumor is small, it has a complete capsule and shows expansive growth. It is difficult to distinguish it from benign tumors. When the diameter is >3 cm, it shows the characteristics of invasive growth and the incidence of lymph node metastasis is low [6]. Antonioni et al. determined the interpolation direction according to the directional characteristics of contrast-enhanced ultrasound combined elastic imaging image. Therefore, the proposed direction-based interpolation algorithm can effectively improve the resolution of contrast-enhanced ultrasound combined elastic imaging image [7]. Costache et al. proposed a shape interpolation algorithm based on conditional morphology and successfully realized the superresolution reconstruction of shoulder contrast-enhanced ultrasound combined with elastic imaging technology [8]. Barbera et al. developed a log-based image interpolation algorithm for ultrasound analysis with contrast agents combined with elastography. The process of implementing the algorithm is to first record adjacent slices on a contrast ultrasound image combined with elastography using a multidensity recording method and then to interpolate the image using a two-cube interpolation algorithm. From the above, it can be seen that the interpolation-based superdensity recovery algorithm is simple and has the advantage of high real-time performance. The disadvantages are that the previous information on the CEUS combined elastographic image cannot be fully utilized, the superaccurate reconstruction algorithm based on the CEUS combined elastographic image is clearly lacking, and the reconstructed results often do not meet clinical requirements [9]. Chantarojanasiri et al. believed that compared with parotid gland, tumor T1WI mostly shows slightly lower signal and the signal changes of T2WI are different among different types of tumors. Most of the signals are uneven, which can be caused by the diversity of tissue components (such as pleomorphic adenoma), tumor death, and cystic change. More than half of the adenolymphoma in this group showed isosignal on T2WI, which was considered to be related to the dense lymphoid nuclei in the tumor [10]. Iglesias-Garcia et al. proposed a superdensity recovery algorithm based on sparse imaging. Implementation process: learn to get a complete dictionary from a sample library with high resolution and low resolution. The low-resolution ultrasonic combined elastographic image is divided into many overlapping small image blocks as a whole, and a sparse representation of the low-resolution image blocks is obtained based on a low-density supercomplete dictionary. High-resolution image patch combined with high-resolution full-size dictionary is recreated. Numerous reconstructed high-resolution image patches were combined with elastographic images and punctured into ultrasound with full contrast [11]. Kyoung et al. introduced the algorithm based on sparse representation to the superresolution reconstruction of contrast-enhanced ultrasound combined with elastic imaging technology brain image. First, the brain contrast-enhanced ultrasound combined with elastic imaging technology image is divided into three parts: cerebrospinal fluid, gray matter, and white matter. Also, the image blocks of the adjacent areas between the three different brain tissues are defined as the training set; this dataset contains a wealth of high-frequency information. The disadvantage of this algorithm is that the construction of training dataset requires segmentation algorithm and takes a long time [12]. Lam et al. extended the Rueda algorithm to superresolution reconstruction of diffusion-weighted contrast-enhanced ultrasound combined with elastic imaging technology images in 2014 [13, 14]. Figure 1 represents the network structure of RE-CNN.

Figure 1

Network structure of RE-CNN.

Based on the current research, this paper proposes the effect of contrast-enhanced ultrasound combined with real-time elastic imaging technology in the differential diagnosis of salivary gland tumors. 200 patients were selected, including 120 males and 60 females. The age ranged from 9 to 83 years, with an average of 55.4 years. Among the 200 cases, there were 90 cases of single parotid gland on the right, 77 cases of single parotid gland on the left, 2 cases of bilateral (single parotid gland on each side), 2 cases of multiple parotid gland on the right (2 lesions), 1 case of 2 lesions on the left and 1 lesion on the right, and 1 case of multiple parotid gland on the left (4 lesions). 135 cases were located in the superficial lobe (78%) and 38 cases (22%) in the deep lobe of parotid gland. The ARIETTA 70 color Doppler ultrasound diagnostic instrument is used. The equipment is equipped with high-frequency contrast probe, real-time elastic imaging technology, and related software.

3. Image Superresolution Reconstruction Based on Deep Learning and Contrast-Enhanced Ultrasound Combined with Elastic Imaging Technology

3.1. Image Degradation Model

High-resolution image recovery is the use of low-resolution images to restore high-resolution images. Superprecision reconstruction is the opposite. Low-resolution images may not have high-resolution images, but can be compatible with many high-resolution images. Many factors contribute to MR imaging, such as blurring of motion, unevenness of the field, time of acquisition, and noise. We assume that the probability distribution range of the noise and degradation mapping function has not changed. The pattern of image degradation in the spatial region can be expressed bywhere I represents the real high-resolution image, I represents the observed low-resolution image, H represents the degradation function, and n represents noise. Among them, the degenerate function can be expressed as the superposition of a blur function G and a downsampling operation ↓.

In the depth learning-based image superresolution reconstruction algorithm, a large number of high-low resolution image pairs are usually used, which act as the training dataset, and then the convolutional neural network is used to learn the nonlinear mapping from low-resolution image to high-resolution image. Usually, a high-resolution image I is artificially degraded into a low-resolution image I. In this process, bilinear interpolation or bicubic interpolation is usually used to simulate the blur function G and downsampling operation ↓, and the noise n mostly uses Gaussian white noise [15].

3.2. Performance Index of Image Superresolution Reconstruction

The subjective rating index is a subjective assessment of image quality according to people's perceptions. At present, objective image evaluation indicators mainly include peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), while subjective evaluation indicators mainly use average voting scores (MOS). These evaluation criteria are presented below [16].

3.2.1. Peak Signal-to-Noise Ratio

After processing the peak signal-to-noise ratio (PSNR), it can be used to assess image quality, which is common in industries such as image overresolution, image compression, and image recovery. In the ultrahigh-resolution recovery area, the PSNR is expressed as the logarithm of the ratio of the actual squared error of the reconstructed image to the actual image and the maximum possible value of the image pixel. The mean square error is a measure of the error between the reconstructed image and the actual image. The specific definitions are as follows:where f(i, j) and g(i, j) represent the reconstructed image with height H and width w and the truth image, respectively. The PSNR of the two images can be calculated through the MSE and the maximum possible pixel value of the two images. The definition of PSNR is as follows:where n represents the number of bits of image pixels and 2 − 1 represents the maximum possible pixel value of the image. The unit of PSNR is dB. The larger the PSNR, the better the reconstruction quality of the image. Normally, 2 − 1=255 and the PSNR is between 20 dB and 40 dB. When n is fixed, PSNR is only related to MSE, that is, it is only related to the difference of pixels at the same position of the two images. This evaluation method is usually inconsistent with the visual perception of human eyes, so there will be great differences between PSNR and people's subjective feelings [17].

3.2.2. Structural Similarity

Structural similarity (SSIM) was first proposed by the Image and Video Engineering Laboratory of the University of Texas at Austin to characterize the similarity of two images. Structural similarity calculates the similarity from three aspects of image contrast, brightness, and structure and compares the similarity between two images in a more comprehensive way. SSIM is composed of brightness contrast, contrast ratio, and structure contrast. For the images to be tested, the average gray level is used as the brightness contrast, and the average gray level is shown as follows:

The brightness contrast function is shown in the following equation:

Take the image gray standard deviation as the contrast measurement as shown in equation (6) and the contrast ratio function shown in equation (7):

The contrast ratio estimation and brightness estimation of the two images can be expressed by the structural contrast function, as shown in formula (8), where σ is expressed as formula (9):

In equations (5), (7), and (8), constants C1, C2, and C3 are usually used to avoid denominators being 0. Usually,where L is the maximum possible pixel value of the image, K1 is usually equal to 0.01, and K2 is usually equal to 0.03.

Finally, a complete SSIM formula can be obtained, as shown in the following equation:

Generally, set a=β=γ=1, and the simplified formula (12) is finally obtained:

3.2.3. Average Opinion Score

Mean opinion score (MOS) provides a subjective evaluation standard for image quality evaluation. As mentioned earlier, when PSNR is used as the evaluation standard, it is often inconsistent with people's visual perception. Therefore, some researchers propose to use MOS as the evaluation index of image superresolution reconstruction as a subjective evaluation index. MOS evaluation criteria can be expressed bywhere R indicates the score of the i-th evaluator on the reconstructed image. The score is usually between 1 and 5. 1 means “poor quality,” 2 means “inferior quality,” 3 means “average quality,” 4 means “good quality,” and 5 means “very good quality.” However, MOS is usually affected by the subjective bias of different evaluators. Different evaluators usually have great differences in the quality evaluation of the same image. It is very difficult to use the same batch of evaluators for each evaluation [16]. In addition, the number of evaluation personnel also has a certain impact on MOS results. Based on the above reasons, the effectiveness and stability of MOS have been questioned. Some superresolution reconstruction methods perform generally on the objective evaluation indexes PSNR and SSIM, but they are much better than other methods when measuring subjective perception with MOS.

3.3. Introduction of Artificial Neural Network in Deep Learning

Deep learning is a branch of machine learning. Compared with machine learning, it is closer to the original goal—artificial intelligence. In recent years, deep learning technology has been introduced in many fields, including computer vision, and positive progress has been made on the whole, which also promotes the development of artificial intelligence to a great extent.

Artificial neural network is also called neural network or quasineural network. Inspired by human neural structure, it is a mathematical model and computational model that imitates the structure and function of biological neural network, usually as an approximate or estimation function. Artificial neural network is composed of a large number of nodes connected with each other. These nodes are called neural units, which are abstracted from biological nerve cells. Neural unit is a multi-input and single output unit, and its model structure is shown in Figure 2, where x represents the input feature, w is the weight value between the input feature x and the neural unit, b represents the bias term, and y represents the output of the neural unit [18].

Figure 2

Neuron model.

The relationship between neuron input and output can be expressed by equation (14), where σ represents the activation function:

According to formula (14), when the activation function σ is not used, the neuron is a linear model. The expression ability of linear model is often not enough. The introduction of activation function strengthens the expression ability of the model to a great extent. At present, the commonly used activation functions are as follows: tanh function, sigmoid function, ReLU function, and Softsign function.

The formula of ReLU activation function is shown as follows:

The ReLU activation function makes the output of some neurons 0, which enhances the sparsity of the network, suppresses the dependence of relevant parameters, and can effectively avoid the problem of overfitting [19].

The formula of tanh activation function is shown as follows:

The output mean of tanh activation function is 0, so the convergence speed is fast and the number of iterations in training can be reduced.

The formula of sigmoid activation function is shown as follows:

Sigmoid activation function is characterized by monotonicity and continuity, and its output is between 0 and 1. It is usually used as output layer.

The connection of multiple artificial neurons forms an artificial neural network. As shown in Figure 3, it is a typical three-layer artificial neural network.

Figure 3

Artificial neural network model.

The first layer of neural network is called input layer, the last layer is output layer, and the middle of input layer and output layer is collectively referred to as hidden layer. In artificial neural network, the output of each layer will be used as the input of the next layer, and each hidden layer has its own weight value w and bias term b. In artificial neural network, w and b are parameters that can be learned. From the perspective of mathematical theory, artificial neural network can fit any complex function as long as it has sufficient parameter quantity and depth. In the actual training, the preprocessed data are imported into the network, then the deviation between the actual label value and the network output is calculated, and then the backpropagation algorithm is used to iteratively update the sum in the network. The above is the most basic artificial neural network model and the prototype of deep learning algorithm.

4. Experimental Results and Analysis

4.1. Subjects

There were 200 patients in this group, including 120 males and 60 females. The age ranged from 9 to 83 years, with an average of 55.4 years. Among the 200 cases, there were 90 cases of single parotid gland on the right, 77 cases of single parotid gland on the left, 2 cases of bilateral (single parotid gland on each side), 2 cases of multiple parotid gland on the right (2 lesions), 1 case of 2 lesions on the left and 1 lesion on the right, and 1 case of multiple parotid gland on the left (4 lesions). 135 cases were located in the superficial lobe (78%) and 38 cases (22%) in the deep lobe of parotid gland [20].

4.2. Instruments and Methods

The ARIETTA 70 color Doppler ultrasound diagnostic instrument was used, which is equipped with high-frequency radiography probe, real-time elastic imaging technology, and related software. First, routine color Doppler ultrasonography was performed to observe the size, shape, long and short diameter, and boundary of lymph nodes and to observe the internal echo and blood flow type. Then, for the lymph nodes to be biopsied, switch to the real-time elastic imaging mode, the operator vibrates evenly and slightly, and the pressure display bar takes the green full grid as the standard; then, record the image and score it. Finally, switch to the contrast mode, rapidly inject SonoVue contrast agent 2.0 ml/time through the elbow vein and immediately flush the tube with 5 ml normal saline. Playback and observe the distribution and filling of the contrast agent.

Two senior doctors engaged in MRI diagnosis of the head and neck read the films without knowing the pathological results. The evaluation contents include the following: ① location of tumor: the tumor in parotid gland is divided into deep and shallow lobes according to the parotid segment and its branches of facial nerve; ② morphological indexes included tumor shape, boundary, signal intensity, uniformity, enhancement degree, and infiltration of adjacent tissues and cervical lymph nodes. The signal intensity takes the normal parotid gland as the equal signal and takes it as the reference to divide the tumor signal into low, equal, and high signals. The enhancement effect is divided into mild, moderate, and significant enhancement according to the enhancement degree of normal glands.

4.3. Statistical Methods

SPSS21.0 software is used for data analysis, receiver operating characteristic (ROC) curve is drawn for comparison of diagnostic efficiency, count data are expressed as percentage, and P < 0.05 is expressed as statistically significant [21].

4.4. Results

4.4.1. Pathological Results

Among the 200 cases, 170 cases were benign, including 59 cases of mixed tumor, 57 cases of adenolymphoma, 16 cases of basal cell adenoma, 14 cases of cyst, 5 cases of lymphoproliferative tissue, and 19 other cases. The coincidence rate of ultrasonic diagnosis was 71.1%. There were 30 cases of malignant tumors, 4 cases of mucoepidermoid carcinoma, 3 cases of adenoid cystic carcinoma, 2 cases of adenocarcinoma, and 21 other cases. The coincidence rate of ultrasonic diagnosis was 71.4%.

4.4.2. Ultrasonic Manifestation

① 59 cases of mixed tumor: the shape is round, quasiround, or nodular, with clear boundary, partial complete capsule, and uneven or uniform low echo inside. 24 cases were solid, including 4 cases with calcification. 29 cases were cystic and solid, and small anechoic areas could be seen inside. 6 cases were cystic. The length and diameter of the tumor were 1.2∼6.0 cm, and the blood flow grade was grade I∼II. ② There were 57 cases of adenolymphoma. The shape is quasiround or nodular, with clear boundary and complete capsule. 11 cases showed uniform hypoechoic; 19 cases showed uneven low echo, with short linear strong echo, which was divided into “fine grid” structure, and small focal anechoic with consistent size could be seen in the grid. 27 cases showed mixed echo and “coarse grid” structure, with short linear, patchy strong echo and large nonecho, papillary structure and dense dot echo, and dot echo floating when the probe was pressurized. The length and diameter of the tumor were 1.9∼5.2 cm, and the blood flow grade was grade II∼III. ③ There were 16 cases of basal cell adenoma. The shape was regular, and the boundary was clear. 10 cases showed hypoechoic parenchyma with uniform distribution, and 6 cases showed liquid dark area with mixed echo. The echo behind the tumor was enhanced. The length and diameter of the tumor were 1.0∼3.7 cm, and the blood flow grade was grade II. ④ Cysts in 14 cases: the shape was regular, the capsule was intact, and the posterior echo was enhanced, of which 12 cases were anechoic and 2 cases were hypoechoic. The length and diameter were 1.0∼3.5 cm, and the blood flow grade was grade 0. ⑤ Lymphoid hyperplasia occurred in 5 cases. The lesions were single or multiple, with regular shape, round, or oval, clear boundary, hypoechoic, and enhanced posterior echo. The length and diameter are 0.5∼3.5 cm, and the blood flow grade is grade II∼III, showing a “dendritic” distribution. ⑥ 30 cases of cancer: irregular shape, unclear boundaries, no capsule, hypoechoic, or irregular echo inside. In 4 cases, the capsule boundary is clear; in 1 case, it has an irregular low echo; in the liquid anechogenic section, it is 0.5 cm in diameter; and in 4 cases, it is finely calcified. Nine cases had posterior echogenicity [22]. The tumor is 0.5–6.0 cm in length and diameter, and the blood flow is from II to III. There were 7 cases of metastatic metastasis of the cervical lymph nodes. Comparisons of the performance of ultrasound on benign and malignant tumors are shown in Tables 1–3 and Figures 4–6. There are significant differences between benign and malignant tumors in terms of morphology, boundaries, capsules, liquefaction, calcification, and blood flow. There is no statistically significant difference between toxic and nontoxic internal echoes and back echoes. PSV and RI in benign tumors are lower than in cancer.

Table 1

Characteristics of morphology, boundary, envelope, internal echo, and posterior echo of the lesion.

Pathological type		Benign (n = 170)	Malignant (n = 30)	P value
Morphology	Regular shape	127	8	<0.05
	Irregular shape	43	22
Boundary	Clear	150	12	<0.05
	Unclear	20	18
Envelope	Yes	131	8	<0.05
	No	39	22
Internal echo	Uniformity	50	8	>0.05
	Uneven	120	22
With liquefaction (cyst solid) with calcification		66	1	<0.05
		5	4	<0.05
Posterior echo	Enhance	145	19	>0.05
	Attenuation or no change	25	11

Table 2

Blood segmentation of lesions.

Pathological type	Blood segmentation
	0	Level I	Level II	Level III
Benign (n = 170)	69	88	8	5
Malignant (n = 30)	2	6	12	10
P value	<0.05	<0.05	<0.05	<0.05

Table 3

Frequency spectrum characteristics of lesions.

Pathological type	PSV	RI
Benign (n = 170)	18.3 ± 4.2	0.46 ± 0.08
Malignant (n = 30)	41.5 ± 9.4	0.72 ± 0.14
P value	<0.05	<0.05

Figure 4

X2 value of morphology, boundary, envelope, internal echo, and posterior echo of the lesion.

Figure 5

X2 value of blood segment of lesions.

Figure 6

X2 value of spectrum characteristic of lesions.

Among the 30 cases of malignant tumors, 8 cases were misdiagnosed as benign, including 4 cases of mixed tumor, 2 cases of adenolymphoma, and 16 cases of lymphoproliferative tissue. In this group, 57 cases of adenolymphoma were misdiagnosed as mixed tumor by ultrasonography in 19 cases. Cysts were misdiagnosed as solid tumors in 2 cases.

It can be seen from Figure 7 that the detection rate of salivary gland tumors by ultrasound in this group is 100% and the diagnostic coincidence rate is 71% (123/173) [23]. Ultrasound can not only identify the tumors in and around the parotid gland but also identify the location, size, and internal structure of the tumors. Combined with CDFI, it can make qualitative diagnosis of the benign and malignant of most salivary gland tumors and provide help for clinical treatment and the formulation of surgical plan.

Figure 7

Differential diagnosis error of contrast-enhanced ultrasonography combined with elastic imaging technology in salivary gland tumors.

4.5. Discussion

Benign epithelial tumors of salivary gland often occur in the superficial lobe of parotid gland, accounting for 73%, of which adenolymphoma can occur frequently and has certain specificity. In this group, 6 cases of recurrent pleomorphic adenoma showed extensive nodules in unilateral parotid gland, with diameter ranging from less than 1 mm to 15 mm, which was consistent with the literature report. Most benign tumors are found with a diameter of 2.0∼3.0 cm. They are located in the deep lobe and across the deep and shallow lobe of the parotid gland. The volume can be large, the shape is mostly round or oval, a few are lobulated, the boundary is mostly clear, and there is a complete envelope. The main reason for unclear boundary is the local capsule and/or extracapsule invasion of the tumor. This sign is more common in pleomorphic adenoma in this group of cases. Pathologically, the tumor often infiltrates around like pseudopodia, which is also one of the reasons for its high recurrence rate. Salivary gland malignant tumors are relatively rare, with many types, different degrees of malignancy, and diversified growth modes and speeds. Primary malignant tumors are mostly single and relatively large lesions, which are considered to be related to the rapid growth of tumors. The common signs of tumors include irregular shape, unclear boundary, uneven signal, and infiltration into surrounding tissues. The main reason for unclear boundary is that the tumor itself has incomplete capsule or no capsule, which is different from benign tumors. In this group, adenoid cystic carcinoma is particularly obvious, while 1 case of well-differentiated acinar cell carcinoma has clear boundary and is difficult to distinguish from benign tumors. Secondary tumors of salivary gland (such as lymphoma and lymph node metastasis) can occur frequently, often involving bilateral parotid glands, showing multiple nonfused nodules, combined with multiple enlarged lymph nodes in the neck, and the nodule signal is consistent with the enlarged lymph nodes in the neck [24].

5. Conclusion

Through the comprehensive analysis of MRI signs of salivary gland tumors, MRI can reflect the respective characteristics of common benign and malignant salivary gland tumors. It is possible to make differential diagnosis for most cases. With the development of MR imaging technology, diffusion-weighted imaging, MR dynamic enhancement imaging, MR spectrum analysis, and diffusion-weighted imaging based on incoherent motion in voxels are gradually applied to the diagnosis of salivary gland tumors, which will play a more and more important role in the differential diagnosis and preoperative evaluation of tumors.