- Email: [email protected]
Determination of the relationship between internal auditory canal nerves and tinnitus based on the findings of brain magnetic resonance imaging
Biomedical Signal Processing and Control 40 (2018) 214–219
Contents lists available at ScienceDirect
Biomedical Signal Processing and Control journal homepage: www.elsevier.com/locate/bspc
Research Paper
Determination of the relationship between internal auditory canal nerves and tinnitus based on the findings of brain magnetic resonance imaging Burhan Ergen a,∗ , Murat Baykara b , Cahit Polat c a
Department of Computer Engineering, Faculty of Engineering, Firat University, Elazig, Turkey Department of Radiology, Faculty of Medicine, Kahramanmaras Sutcu Imam University, Kahramanmaras, Turkey c Department of Otolaryngology, Elazig Training and Research Hospital, Elazig, Turkey b
a r t i c l e
i n f o
Article history: Received 17 April 2017 Received in revised form 12 September 2017 Accepted 30 September 2017 Keywords: Extreme learning machine Internal auditory canal Statistical significance Tinnitus
a b s t r a c t This experimental study aimed to investigate a relationship between tinnitus and thicknesses of internal auditory canal and nerves in it. It was performed on brain magnetic resonance images of patients who consulted the ear, nose, and throat clinic with tinnitus complaint. Statistical hypothesis tests and classification experiments were performed on these data to find out structural differences in internal auditory channel components in patients with tinnitus after obtaining cross-sectional areas of nerves as thicknesses. Both the hypothesis tests and classification results showed that the thicknesses of nerves in tinnitus cases were different from those in normal cases. In particular, the hypothesis tests for the superior vestibular nerve and internal auditory channel showed the highest significance, indicating the relationship with tinnitus. The classification results indicated the possibility of classification for tinnitus identification, establishing a computer-assisted diagnostic system to help physicians. © 2017 Elsevier Ltd. All rights reserved.
1. Introduction Tinnitus is a disorder commonly encountered at ear, nose, and throat (ENT) clinics. It is defined as the perception of noise-like ringing without an external source. The prevalence of tinnitus that can be extremely disturbing is 3%–32% in the general population. Although the examining physician and the patient can hear objective tinnitus such as the ringing or humming sound, the patient only hears subjective tinnitus [1,2]. Some of its symptoms are stress, anxiety, depression, insomnia, and irritability. Although the pathophysiological reasons of tinnitus are poorly understood, it is believed that it is associated with the sensorineural hearing loss of various origins [3–6]. If the peripheral auditory system is damaged, the central auditory neurons receive the weakened auditory input with the affected frequency range. The central auditory system tries to compensate these alterations. In this case, the reason of tinnitus can be considered as a phantom sound because of aberrant plastic reorganization in the auditory center of the brain [4,7].
∗ Corresponding author. E-mail addresses: ergen@firat.edu.tr (B. Ergen), [email protected] (M. Baykara), [email protected] (C. Polat). http://dx.doi.org/10.1016/j.bspc.2017.09.023 1746-8094/© 2017 Elsevier Ltd. All rights reserved.
Subjective tinnitus is more frequently associated with internal auditory canal (IAC) pathology, presbycusis, Meniere’s disease, acoustic trauma, ossicle system deformities, and labyrinthitis. In addition, aging or loud noise exposures, which may cause hearing loss, are also related to tinnitus. The pitch of the phantom sound generally corresponding to hearing-loss frequencies may indicate hearing loss. However, it is not always associated with hearing loss with tinnitus [4,8]. Some studies that employed positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) propose that tinnitus generation is in the central auditory system [7,9–11]. The study of tonotopic brain changes with magnetic resonance imaging (MRI) and PET reported no distortion in auditory cortical organization and limbic system responses [7]. Another study carried out with children proposes that MRI would be mandatory for investigating tinnitus [9]. The reason of tinnitus is traditionally considered to reside in the inner ear because tinnitus sound is commonly perceived in the ear [12–14]. Therefore, MRI is recommended to investigate the existence of pathologies inside the canal for identification [15]. The aim of this study was to investigate the internal acoustic canal and the nerves in it. Fig. 1 represents the location of an inner ear in the auditory system. Sound collected by the outer ear is transferred to the middle ear through the ear channel. The middle ear transforms the
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
215
It is assumed that the CSAs have important information about the condition of nerves. After the data collection phase, statistical measurements were used to discover the existence of any relationship between the thickness of nerves and tinnitus. In addition, if a statistical difference exists between healthy volunteers and patients with tinnitus with respect to CSAs, it is possible to establish a computer-aided diagnostic (CAD) system for feature examinations. The possible statistical significance allows a classification system and therefore a CAD system. The possible existence of a statistical significance allows a classification system and therefore to be able to establish a CAD system. 2. Material
Fig. 1. Three parts of the ear system: inner ear, middle ear, and outer ear [21].
Fig. 2. A schematic diagram of the right IAC: (a) perspective view; (b) cross-sectional view [22].
sound energy into internal vibrations and a compressional wave in the inner ear. Transformation of the internal vibrations into nerve impulses is performed in the inner ear for transferring them to the brain [16–18]. The IAC in the inner ear contains four main nerves carrying the impulses to the brain. A schematic diagram of a right internal acoustic meatus that shows the placement of the nerves in IAC is given in Fig. 2. The nerves that pass through the IAC are the facial nerve (FN), vestibular superior nerve (SVN), inferior vestibular nerve (IVN), and cochlear nerve (CN). Each of the nerves has a constant location in the area of the fundus of IAC [19]. This anatomical knowledge makes it possible to observe structural differentiations. Today, structural forms of IAC and the nerves in it can be visualized using high-resolution techniques. In particular, MRI is more convenient to visualize the structural form and differentiation because the nerves are soft tissues. MRI monitoring of the auditory pathway has helped obtain valuable findings of the ear origin disorders such as vertigo, maneuver, tinnitus, and hearing loss. In a previous study, important findings were obtained indicating that nerve diameters were related to benign paroxysmal positional vertigo (BPPV) [20]. It was determined that the thicknesses of nerves in IAC were differentiating in BPPV cases. The experience and methods in the previous study have benefited the present study. This study focused on finding out any relationship between the thicknesses of IAC nerves and tinnitus. The brain MRI images of patients and healthy subjects were used to investigate how the structures of nerves were affected in tinnitus cases. The crosssectional areas (CSA) of IAC nerves were particularly investigated to determine how the thickness of nerves was affected.
The study was performed on those who consulted the ENT clinic with a tinnitus complaint. An otolaryngologist in the ENT clinic diagnosed that the patients had tinnitus definitely. The volunteers for the control group were chosen among the patients who applied to the radiology department. The volunteers in the control group needed brain MRI images for various reasons. The control and patient groups consisted of 23 healthy subjects and 23 patients. The local ethics committee approved this study, which was carried out in accordance with the Helsinki Declaration. The patients with tinnitus, the volunteers in the control group, and the local ethics committee provided the necessary permission. The brain MRI images of the patient and control groups were used to investigate the IAC region and measure its thickness and the thickness of the nerve in it. Both of the groups had similar demographic characteristics with respect to smoking, hypertension, cholesterol, and so forth. The MR device to obtain the brain images had the power of 1.5 T and an eight-channel high-definition brain coil. An otolaryngologist and a radiologist investigated the IAC canal and the nerves in the MRI images, which were processed using digital image processing techniques. Fig. 3a represents a brain image including the IAC region and the nerves in it as a sample for the collected data. The red square in Fig. 3b shows an IAC region and the nerves. A resolution improvement was performed for detecting the structures adequately after selecting and cropping the IAC region. Fig. 4a shows the cropped original image. Lanczos-2 kernel was used for resolution enhancement [23]. Fig. 4b shows the gray-level image representing the IAC region after resolution enhancement. Then, the obtained gray-level image was converted into a binary image using Otsu’s thresholding method [24]. The binary form of the cropped image given in Fig. 4c shows the specific location of the nerves in the IAC region. The white area refers to IAC, and the black areas in the white region refer to the nerves SVN, FN, CN, and IVN from the left top in clockwise rotation. The CSAs of the IAC canal and the nerves could be measured by counting the pixels in each area because a pixel referred to a voxel size. The MRI device provided this information on their produced Digital Imaging and Communication in Medicine (DICOM) Digital images [25]. Therefore, the real size of a tissue in MRI images could be calculated by a simple estimation. 3. Methods 3.1. Statistical measurements for significant difference The common way to determine any differences among the group data is to use statistical measurements and apply some tests. Many researchers adopt various statistical methods and common-sense techniques to examine the difference between the control and patient groups. Most widely used methods depend on statistical hypothesis tests and correlation calculation. The hypothesis tests
216
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
Fig. 3. MRI images demonstrating the IAC region: (a) sagittal oblique brain image; (b) binary image of the IAC region.
Fig. 4. A cropped region of interest of the IAC region: (a) cropped image; (b) resolution-improved image; (c) binarized image showing the components.
are accepted as key elements for comparing two or more groups of data [26]. Indeed, these methods can put forward a statistical significance between the two groups. In statistical hypothesis testing, the null hypothesis assumes that no statistically significant difference exists between two or more populations. The statistics such as the mean values of distributions are calculated, and their difference is assessed for significance.to test the null hypothesis The significance is meant to reject the null hypothesis. Under the assumption that the null hypothesis is true, the P value is the level of confidence probability that the population data is well represented by the observed result determined by sampling the population. For instance, if the null hypothesis is of equality between two mean values collected from two populations, and if the null hypothesis is confirmed by the empirical observation, the P value is the probability that the two mean values observed are actually truly representative of an equality existing in between the actual populations. A small P value expresses high probability that the two population have different characteristics. The t-test, Wilcoxon signed-rank test, and Pearson’s correlation were used to test the null hypothesis. The null hypothesis of the tests is that no relationship exists between tinnitus data and the thickness data belonging to the IAC and its nerves. This means that the thicknesses of IAC and its nerves do not cause or affect the generation of tinnitus. The aforementioned tests are commonly used to determine a statistical significance. Since the data in the nature are assumed to have a normal distribution, the t test is generally used to determine the statistical differences. The use of the t test is especially true for the observed two groups [27,28]. The statistical significance is considered if the P value is below 0.05. This means that differences exist between the two groups of data observed. That is, the two groups show differences depending on the observed data. For the CSA data, the mean values (), standard deviations (), and P values are given for the experiments conducted on the patient and the control groups. However, the t-test is the most widely test for hypotheses. It is noted that the use of a t-test has a risk of some errors [29,30].
Wilcoxon’s signed-rank test is another recommended method to test multiple data sets [31,32]. As an alternative to the t-test, the Wilcoxon signed-rank test is nonparametric. It ranks the differences in the data set and compares the ranks for the positive and the negative differences without considering the signs [33]. In hypothesis tests, it is considered that a statistical significance exists if the P values are extremely low, such as around or less than 0.05. In addition to the hypothesis, correlation coefficients can be used to assess the similarity between the two groups. The Pearson product-moment correlation coefficient is also used to determine the significant difference [34,35]. It is the most widely used method for determining the correlation coefficients. If the correlation coefficient is near to 1, the groups are considered to be similar to each other. That is, no significant difference exists between the two groups, depending on the observed data. The correlation coefficients of the measurements are given for all CSA in the experimental tables. 3.2. Extreme learning machines Although feedforward neural networks have been successful in pattern recognition problems, its slow learning is still the main disadvantage. The main reason for this slow learning is that the neural network has gradient-based learning algorithms and tunes them iteratively. Unlike conventional iterative learning algorithm, weights and thresholds can be randomly generated, and the numbers of neurons in the hidden layer can be adjusted. After learning, extreme learning machine (ELM) can find out the optimal solution without setting any other parameters. This method not only decreases data processing time but also increases generalized efficiency. The output of ELM can be given as [36]: Ok (x) =
L
woi G (ai , bi , xk )
(1)
i=1
where a, b, and L are the learning parameters and the bias of the i-th hidden-layer neuron, respectively. Here, the i-th hidden-layer
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219 Table 1 Acronyms for the CSAs of inner ear components. FN SVN IVN CN IAC VT
Table 2 Statistical values related to the CSAs of IAC and its nerves.
Facial nerve Superior vestibular nerve Inferior vestibular nerve Cochlear nerve Internal auditory canal Vestibular total (inferior + superior)
neuron is connected to the output layer with woi . G (.) refers to the output of the I-th hidden-layer neuron [37]. The ELM output can be expressed in the compact form as follows:
O = WoT H
T
⎢ H=⎢ ⎣
Measurements
FN
SVN
IVN
VT
CN
IAC
Control Control Tinnitus Tinnitus Overall Overall t-test Wilcoxon Pearson
0.4919 0.2384 0.4442 0.2335 0.4693 0.2459 0.4910 0.6051 0.8712
0.5278 0.2250 0.4891 0.1344 0.5070 0.1973 0.0590 0.0727 0.7149
0.3353 0.2464 0.2614 0.1305 0.2812 0.1878 0.39352 0.30108 0.06384
0.8631 0.3744 0.7506 0.1822 0.8282 0.3051 0.0749 0.0777 0.5319
0.6147 0.3032 0.5706 0.2076 0.6068 0.3017 0.3633 0.4115 0.1604
24.4775 5.91042 21.6135 5.73889 22.6335 6.67977 0.0697 0.0829 0.2430
(2)
where H is the matrix of the hidden-layer neurons, and a (ki) refers to G (ai , bi , xk ), which is the input in the k-th row and i-th column in the hidden-layer output matrix H. H and W can be given in a matrix form as follows:
⎡
217
H1 (1)
···
H1 (N)
.. .
..
.. .
HL (N)
···
.
HL1 (N)
⎤
⎡
b1
b1
···
b1
⎤
⎢ . .. .. .. ⎥ ⎢ ⎥ ⎥ . . . ⎥ (3) ⎥ andW = ⎢ .. ⎢ ⎥ ⎦ ⎣ w11 w11 . . . w11 ⎦ w11
w11
···
w11
The ELM tries to find the minimum training error and the norm of the output weight by calculating the cost function given as: CF = O − Y
(4)
where Y is the expected output matrix. The learning speed of ELM is quite fast because no parameters need to be tuned during the training phase. Kernels can be integrated in ELM because they are support vector machines [38–40]. In this study, both the linear kernel and the Radial Basis Function (RBF) kernel were employed in this study. 4. Experiments and results The brain MRI images of the patients with tinnitus and healthy subjects were examined to investigate any relationship between tinnitus and the structure of IAC components. Particularly, the thickness of nerves and the area of IAC were evaluated as a structure. It was assumed that the thicknesses were related to the CSAs of IAC components: the nerves and the IAC through the collected MRI images. Indeed, the anatomy of the inner ear was spread over an extremely small area and had a complicated structure to visualize it completely. The visualization of the inner ear components required the specific positioning of the patient. Setting the patient and maintaining her/him in this special position was quite difficult. One of the most critical steps in this study was to obtain MRI images at the right angle to visualize CSAs. The CSAs of the IAC and the nerves were observed and measured under the guidance of ENT and radiology experts, who worked collaboratively to ensure that the patient was in the right position. The collected data was investigated in two ways to determine statistical significance, which referred to a relationship between tinnitus and the structure of IAC components. The first way was an approach to evaluate CSAs directly, while the latter way evaluated them proportionally. Therefore, statistical measurements and classification methods were applied on both direct data and proportional data. Table 1 shows the acronyms for the direct and proportional data used in Tables 2 and 3. They present the mean value and standard deviation of control data, tinnitus data, and whole data.
Table 2 shows the results of statistical measurements, tests, and correlation coefficients. In addition to the tests and correlation coefficients, mean values () and standard deviations () were also common to evaluate how the data varied. The standard deviations and the mean values were given in the three categories with respect to the collected data: patients with tinnitus, healthy subjects, and overall values, which denoted whole data. “Tinnitus” and “Control” labels denote the patients with tinnitus and the normal healthy subjects in tables, respectively. “Overall” denotes the evaluation of all cases including the patients with tinnitus and the healthy subjects. No significant differences were observed between the patient and control groups on considering the mean values and the standard deviations into consideration. However, it could be realized that the mean values and the standard deviation for the cases with tinnitus were slightly smaller compared with the healthy cases. Nevertheless, this inference with no discrimination was applicable to all of the nerves. The results of the hypothesis tests and the correlation coefficients according to the nerves are also given in Table 2. According to the hypothesis tests, some of the P values were quite close to the values accepted as statistically significant. The statistically significant limit was just a hypothetical value. The values close to this limit could be also considered as statistically significant more or less. Therefore, the P values for SVN, VT, and IAC indicated statistical significance. Both the t-test and Wilcoxon signed-rank test indicated a statistical significance for SVN, VT, and IAC with the value pairs (0.05903–0.07273, 0.07496–0.07772, and 0.069711–0.08298), respectively. The significant values in the table are shown in bold. Although the t-test and Wilcoxon signed-rank test point out SVN, VT, and IAC, Pearson correlation coefficient claim that FN and SVN are related to tinnitus. Both tests and correlation coefficients found that the thickness of SVN was related to tinnitus. Table 3 represents the statistical values of the proportional data. The proportional data were obtained as the ratio of CSAs to IAC and some important nerves (FN and VT). The column names refer to these ratios. If the P values were taken into consideration, none of hypotheses tests showed any significance. These results indicated that tinnitus cases were not related to the ratio between the nerves and the IAC. However, the correlation coefficient for the proportional data put forward that the IVN/IAC and the VT/FN had a reasonable relationship to tinnitus. Although the results of the statistical hypothesis tests and the correlation coefficients made a claim that the structures of SVN, FN, VT, and IAC were related to tinnitus, it was hard to decide which nerve was the most dominant factor for tinnitus complaint. If these values indicated a statistically significant difference, a classification should be possible with respect to the CSAs of nerves. Thus, it was possible to determine the most dominant nerve on tinnitus by comparing the correct classification rates. The ELM classifier was accepted as one of the most reliable classifiers used widely in recent works.
218
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
Table 3 Relational statistical values related to the CSAs of IAC and its nerves.
Control Control Tinnitus Tinnitus Overall Overall t-test Wilcoxon Pearson
FN/IAC
SVN/IAC
IVN/IAC
VT/IAC
CN/IAC
SVN/FN
IVN/FN
VT/FN
CN/FN
CN/VT
0.0202 0.0087 0.0223 0.0145 0.0208 0.0120 0.4105 0.8314 0.91984
0.0222 0.0103 0.0230 0.0046 0.0239 0.0085 0.4875 0.5633 0.72264
0.0131 0.0076 0.0132 0.0091 0.0122 0.0080 0.3852 0.7151 0.97215
0.0353 0.0136 0.0363 0.0111 0.0361 0.0125 0.9283 0.9031 0.51967
0.0246 0.0085 0.0284 0.0139 0.0269 0.0121 0.5026 0.8551 0.00466
1.3695 1.0690 1.4137 0.8377 1.4734 0.9219 0.6754 0.7151 0.82095
0.7194 0.4251 0.7441 0.5129 0.6908 0.4345 0.3961 0.6051 0.67959
2.0889 1.2321 2.1579 1.2306 2.1642 1.1780 0.9711 0.6482 0.94091
1.4802 0.8926 1.7185 1.4789 1.6441 1.2358 0.7288 0.9031 0.09324
0.8022 0.4197 0.7935 0.3279 0.7884 0.3725 0.9409 0.8791 0.01373
Table 4 Classification results using different training rates. Training rate (%) 50 60 70
Mean Max Mean Max Mean Max
FN
SVN
IVN
VT
CN
IAC
0.52 0.59 0.48 0.61 0.47 0.64
0.71 0.90 0.75 0.89 0.76 0.92
0.56 0.63 0.48 0.66 0.51 0.71
0.70 0.86 0.71 0.88 0.71 0.91
0.47 0.54 0.47 0.55 0.51 0.64
0.74 0.91 0.76 0.94 0.82 0.93
ies, which required medical imaging, only hypothesis tests were usually accepted as sufficient. However, a classification of the data was preferred to ensure that tinnitus was related to the thickness of inner ear nerves and show the possibility of constructing a CAD system to assist ENT physicians. The CRRs of the classification for proportional data is given in Table 6 indicated that tinnitus cases were not related to the ratio between the nerves and the IAC.
5. Discussion Table 5 Classification results for the combination of three nerves. Training rate (60%)
SVN + VT
SVN + IAC
VT + IAC
Mean Max
0.70 0.94
0.75 0.94
0.71 0.94
For the classification phase, some portion of the patient data and the control data were used for training and the rest were used for testing the classification. The samples for the training and the test were chosen randomly. The classification process was repeatedly performed 10 times to ensure the result of classification. The result of classification for the CSAs of nerves is given in Table 4. In addition, the classification was repeated 10 times for three different training rates. Table 1 presents the average and the maximum correct classification rates (CRR) obtained with the ELM classifier using a radial basis function. The k-fold test was avoided in this study due to insufficient data. Actually, it was extremely difficult to find the patients with tinnitus and obtain their brain MRI images. Another difficult task in this study was to visualize the inner ear region at right angle as stated earlier. If the visualization of inner ear was not achieved at right angle, it would lead to the inaccurate measurement of CSAs. Also not being able to visualize the inner ear in the right position prevents measuring the thickness of inner ear components. This might be the reason why the value of P, which is the result of hypothesis tests, is slightly more than 0.5. The classification results in the tables were compatible with the results of the hypothesis tests. The classification tests were repeated for three training ratios of 50%, 60%, and 70%. The CRRs of classification for SVN, VT, and IAC were achieved as 0.92. The mean performances of ELM classifier were 0.7, 0.71, and 0.82 for the training rate of 70%. According to the classification result, the IAC was the nerve more related to tinnitus. Table 5 represents the classification result for the combination of three nerves. Although the maximum CRR increased at the same value, the mean CRR was slightly higher for using SVN and IAC features. In such stud-
Tinnitus that may be more troublesome than marginal hearing loss has several reasons. It is difficult to determine the cause of this disease, which affects almost 70% adults. It has been reported by researchers that environmental noise, change in gray matter, pathological problems in the sound channel, and structural problems of the inner ear can cause tinnitus [7,10,15]. Particularly, a number of studies have been performed to determine the differences that may be related to tinnitus in audio vestibular tract. Many findings on MRI screening of the audio vestibular tract indicates the importance of MRI examination in patients with tinnitus [15]. These studies depend on the belief that the pathological or structural condition of audio vestibular tract is related to tinnitus [2,11]. The studies have reported that intracranial abnormalities are responsible for the sensorineural hearing loss in patients. Furthermore, MRI of the temporal bone, cerebral angle, and brain have shown abnormal results in patient with tinnitus [6]. Taking into consideration these studies, the present study was performed with the idea that the inner ear canal and its nerves might have different thicknesses in tinnitus cases. Therefore, the inner ear region was visualized using brain MRI images in a patient with tinnitus. The statistical analysis and the classification results demonstrated that the thicknesses of the nerves were affected in tinnitus cases. In particular, the cross sectional area of the superior vestibular nerve and internal auditory canal showed the highest significance, indicating their relationship with tinnitus. However, similar results were obtained for the sum of the cross-sectional areas of vestibular and inferior nerves, it is clear that the superior vestibular nerve is actually related to tinnitus. Both the statistical results and the classification results did not show a relationship between the cross sectional area of inferior vestibular nerve and tinnitus. In addition, tinnitus problem was also associated with the structure of internal auditory channel. This study examined the cross sectional areas of the internal auditory canal and the nerves in it, where only the nerves were visible in brain MRI images. Since the internal auditory canal had a complex
Table 6 Correct recognition rates using proportional data. Training rate (60%)
FN/IAC
SVN/IAC
IVN/IA
VT/IAC
CN/IAC
SVN/FN
IVN/FN
VT/FN
CN/FN
CN/VT
Mean Max
0.45 0.57
0.47 0.64
0.45 0.57
0.45 0.57
0.54 0.64
0.44 0.57
0.46 0.64
0.40 0.57
0.36 0.50
0.42 0.64
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
structure, a special examination of its structure was a more correct approach. The mean values and the standard deviations indicated that the cross sectional areas of nerves and internal auditory canal were slightly smaller in tinnitus cases. Although the pathology of internal ear is reported to be related to hearing loss, vertigo, and tinnitus [11], it is not possible to take a biopsy from patients and examine them. Perhaps combining pathological and MRI findings may explain why these nerves are thinner in tinnitus cases. The findings of this study showed that tinnitus was particularly related to the inner ear components.
[14]
[15]
[16] [17] [18] [19]
6. Conclusions The results of this study emphasized the importance of MRI visualization in patients with tinnitus to find out the differentiation in audio vestibular tract. The experiments also showed that tinnitus classification was possible by considering the thickness of nerves and the cross-sectional area of internal auditory canal. Furthermore, the possibility of classification for tinnitus identification may facilitate establishing a computer-assisted diagnostic system that will help physicians.
[20]
[21] [22] [23]
[24]
References [25] [1] T.A. Gabr, M.A. El-Hay, A. Badawy, Electrophysiological and psychological studies in tinnitus, Auris Nasus Larynx 38 (2011) 678–683. [2] M. Shweel, B. Hamdy, Diagnostic utility of magnetic resonance imaging and magnetic resonance angiography in the radiological evaluation of pulsatile tinnitus, Am. J. Otolaryngol. 34 (2013) 710–717. [3] C. Polat, M. Baykara, B. Ergen, Evaluation of internal auditory canal structures in tinnitus of unknown origin, Clin. Exp. Otorhinolaryngol. 7 (2014) 160. [4] X. Zhou, S. Henin, G.R. Long, L.C. Parra, Impaired cochlear function correlates with the presence of tinnitus and its estimated spectral profile, Hear. Res. 277 (2011) 107–116. [5] D. Távora-Vieira, R. Marino, A. Acharya, G.P. Rajan, The impact of cochlear implantation on speech understanding, subjective hearing performance, and tinnitus perception in patients with unilateral severe to profound hearing loss, Otol. Neurotol. 36 (2015) 430–436. [6] P.C. Weber, R.I. Zbar, B.J. Gantz, Appropriateness of magnetic resonance imaging in sudden sensorineural hearing loss, Otolaryngol. Head Neck Surg. 116 (1997) 153–156. [7] A.C. Guinchard, N. Ghazaleh, M. Saenz, E. Fornari, J.O. Prior, P. Maeder, S. Adib, R. Maire, Study of tonotopic brain changes with functional MRI and FDG-PET in a patient with unilateral objective cochlear tinnitus, Hear. Res. 341 (2016) 232–239. [8] K. Boyen, E. de Kleine, P. van Dijk, D.R. Langers, Tinnitus-related dissociation between cortical and subcortical neural activity in humans with mild to moderate sensorineural hearing loss, Hear. Res. 312 (2014) 48–59. [9] E. Levi, E.K. Bekhit, R.G. Berkowitz, Magnetic resonance imaging findings in children with tinnitus, annals of otology, Rhinol. Laryngol. 124 (2015) 126–131. [10] A.M. Leaver, T.K. Turesky, A. Seydell-Greenwald, S. Morgan, H.J. Kim, J.P. Rauschecker, Intrinsic network activity in tinnitus investigated using functional MRI, Hum. Brain Mapp. 37 (2016) 2717–2735. [11] X. Wu, K. Chen, L. Sun, Z. Yang, Y. Zhu, H. Jiang, Magnetic resonance imaging-detected inner ear hemorrhage as a potential cause of sudden sensorineural hearing loss, Am. J. Otolaryngol. 35 (2014) 318–323. [12] M. Smits, S. Kovacs, D. De Ridder, R.R. Peeters, P. Van Hecke, S. Sunaert, Lateralization of functional magnetic resonance imaging (fMRI) activation in the auditory pathway of patients with lateralized tinnitus, Neuroradiology 49 (2007) 669–679. [13] M. Schecklmann, M. Landgrebe, T.B. Poeppl, P. Kreuzer, P. Männer, J. Marienhagen, D.S. Wack, T. Kleinjung, G. Hajak, B. Langguth, Neural correlates
[26]
[27] [28]
[29] [30] [31] [32] [33]
[34]
[35] [36] [37]
[38] [39] [40]
219
of tinnitus duration and distress: a positron emission tomography study, Hum. Brain Mapp. 34 (2013) 233–240. J.-J. Song, D. De Ridder, P. Van de Heyning, S. Vanneste, Mapping tinnitus-related brain activation: an activation-likelihood estimation metaanalysis of PET studies, J. Nucl. Med. 53 (2012) 1550–1557. A.A. Aarnisalo, H. Suoranta, J. Ylikoski, Magnetic resonance imaging findings in the auditory pathway of patients with sudden deafness, Otol. Neurotol. 25 (2004) 245–249. W.A. Yost, Fundamentals of Hearing: An Introduction, ASA, 2001. A.J. Gulya, L.B. Minor, M.E. Glasscock, D. Poe, Glasscock-Shambaugh Surgery of the Ear, PMPH, USA, 2010. R. Romand, I. Varela-Nieto, Development of Auditory and Vestibular Systems, Academic Press, 2014. H. Ryu, T. Tanaka, S. Yamamoto, K. Uemura, Y. Takehara, H. Isoda, Magnetic resonance cisternography used to determine precise topography of the facial nerve and three components of the eighth cranial nerve in the internal auditory canal and cerebellopontine cistern, J. Neurosurg. 90 (1999) 624–634. B. Ergen, M. Baykara, C. Polat, An investigation on magnetic imaging findings of the inner ear: a relationship between the internal auditory canal, its nerves and benign paroxysmal positional vertigo, Biomed. Signal Process. Control 9 (2014) 14–18. J. Hass, An Acoustics Primer, Indiana University, 2003, http://www.indiana. edu/?emusic/acoustics/acoustics.htm. (Accessed 26 September 2012). L. Albert, J. Rhoton, The cerebellopontine angle and posterior fossa cranial nerves by retrosigmoid approach, Neurosurgery 47 (2000) 93–129. E.H.W. Meijering, W.J. Niessen, M.A. Viergever, Quantitative evaluation of convolution-based methods for medical image interpolation, Med. Image Anal. 5 (2001) 111–126. P.K. Sahoo, G. Arora, A thresholding method based on two-dimensional Renyi’s entropy, Pattern Recognit. 37 (2004) 1149–1161. P. Mildenberger, M. Eichelberg, E. Martin, Introduction to the DICOM standard, Eur. Radiol. 12 (2002) 920–927. J.L. Amaral, A.J. Lopes, A.C. Faria, P.L. Melo, Machine learning algorithms and forced oscillation measurements to categorise the airway obstruction severity in chronic obstructive pulmonary disease, Comput. Methods Progr. Biomed. 118 (2015) 186–197. E.H. Livingston, Who was student and why do we care so much about his t-test? J. Surg. Res. 118 (2004) 58–65. E.H. Livingston, L. Cassidy, Statistical power and estimation of the number of required subjects for a study based on the t-test: a surgeon’s primer, J. Surg. Res. 126 (2005) 149–159. T.G. Dietterich, Approximate statistical tests for comparing supervised classification learning algorithms, Neural Comput. 10 (1998) 1895–1923. T.E. Nichols, A.P. Holmes, Nonparametric permutation tests for functional neuroimaging: a primer with examples, Hum. Brain Mapp. 15 (2002) 1–25. J. Demˇsar, Statistical comparisons of classifiers over multiple data sets, J. Mach. Learn. Res. 7 (2006) 1–30. J.J. Higgins, Introduction to Modern Nonparametric Statistics, 2003. W.M. Azmy, N. El Gayar, A.F. Atiya, H. El-Shishiny, M.L.P. Gaussian Processes, Negative correlation learning for time series prediction, in: J.A. Benediktsson, J. Kittler, F. Roli (Eds.), Multiple Classifier Systems: 8th International Workshop, MCS 2009, Reykjavik, Iceland, June 10–12, 2009. Proceedings, Springer Berlin Heidelberg, Berlin,Heidelberg, 2009, pp. 428–437. B.G. Häne, K. Jäger, H.G. Drexler, The Pearson product-moment correlation coefficient is better suited for identification of DNA fingerprint profiles than band matching algorithms, Electrophoresis 14 (1993) 967–972. M.-T. Puth, M. Neuhäuser, G.D. Ruxton, Effective use of Pearson’s product–moment correlation coefficient, Anim. Behav. 93 (2014) 183–189. G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Extreme learning machine: theory and applications, Neurocomputing 70 (2006) 489–501. N.-Y. Liang, G.-B. Huang, P. Saratchandran, N. Sundararajan, A fast and accurate online sequential learning algorithm for feedforward networks, IEEE Trans. Neural Netw. 17 (2006) 1411–1423. X. Wang, M. Han, Online sequential extreme learning machine with kernels for nonstationary time series prediction, Neurocomputing 145 (2014) 90–97. X. Liu, L. Wang, G.-B. Huang, J. Zhang, J. Yin, Multiple kernel extreme learning machine, Neurocomputing 149 (2015) 253–264. R. Yuan, Z. Li, X. Guan, L. Xu, An SVM-based machine learning method for accurate internet traffic classification, Inform. Syst. Front. 12 (2010) 149–156.
Contents lists available at ScienceDirect
Biomedical Signal Processing and Control journal homepage: www.elsevier.com/locate/bspc
Research Paper
Determination of the relationship between internal auditory canal nerves and tinnitus based on the findings of brain magnetic resonance imaging Burhan Ergen a,∗ , Murat Baykara b , Cahit Polat c a
Department of Computer Engineering, Faculty of Engineering, Firat University, Elazig, Turkey Department of Radiology, Faculty of Medicine, Kahramanmaras Sutcu Imam University, Kahramanmaras, Turkey c Department of Otolaryngology, Elazig Training and Research Hospital, Elazig, Turkey b
a r t i c l e
i n f o
Article history: Received 17 April 2017 Received in revised form 12 September 2017 Accepted 30 September 2017 Keywords: Extreme learning machine Internal auditory canal Statistical significance Tinnitus
a b s t r a c t This experimental study aimed to investigate a relationship between tinnitus and thicknesses of internal auditory canal and nerves in it. It was performed on brain magnetic resonance images of patients who consulted the ear, nose, and throat clinic with tinnitus complaint. Statistical hypothesis tests and classification experiments were performed on these data to find out structural differences in internal auditory channel components in patients with tinnitus after obtaining cross-sectional areas of nerves as thicknesses. Both the hypothesis tests and classification results showed that the thicknesses of nerves in tinnitus cases were different from those in normal cases. In particular, the hypothesis tests for the superior vestibular nerve and internal auditory channel showed the highest significance, indicating the relationship with tinnitus. The classification results indicated the possibility of classification for tinnitus identification, establishing a computer-assisted diagnostic system to help physicians. © 2017 Elsevier Ltd. All rights reserved.
1. Introduction Tinnitus is a disorder commonly encountered at ear, nose, and throat (ENT) clinics. It is defined as the perception of noise-like ringing without an external source. The prevalence of tinnitus that can be extremely disturbing is 3%–32% in the general population. Although the examining physician and the patient can hear objective tinnitus such as the ringing or humming sound, the patient only hears subjective tinnitus [1,2]. Some of its symptoms are stress, anxiety, depression, insomnia, and irritability. Although the pathophysiological reasons of tinnitus are poorly understood, it is believed that it is associated with the sensorineural hearing loss of various origins [3–6]. If the peripheral auditory system is damaged, the central auditory neurons receive the weakened auditory input with the affected frequency range. The central auditory system tries to compensate these alterations. In this case, the reason of tinnitus can be considered as a phantom sound because of aberrant plastic reorganization in the auditory center of the brain [4,7].
∗ Corresponding author. E-mail addresses: ergen@firat.edu.tr (B. Ergen), [email protected] (M. Baykara), [email protected] (C. Polat). http://dx.doi.org/10.1016/j.bspc.2017.09.023 1746-8094/© 2017 Elsevier Ltd. All rights reserved.
Subjective tinnitus is more frequently associated with internal auditory canal (IAC) pathology, presbycusis, Meniere’s disease, acoustic trauma, ossicle system deformities, and labyrinthitis. In addition, aging or loud noise exposures, which may cause hearing loss, are also related to tinnitus. The pitch of the phantom sound generally corresponding to hearing-loss frequencies may indicate hearing loss. However, it is not always associated with hearing loss with tinnitus [4,8]. Some studies that employed positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) propose that tinnitus generation is in the central auditory system [7,9–11]. The study of tonotopic brain changes with magnetic resonance imaging (MRI) and PET reported no distortion in auditory cortical organization and limbic system responses [7]. Another study carried out with children proposes that MRI would be mandatory for investigating tinnitus [9]. The reason of tinnitus is traditionally considered to reside in the inner ear because tinnitus sound is commonly perceived in the ear [12–14]. Therefore, MRI is recommended to investigate the existence of pathologies inside the canal for identification [15]. The aim of this study was to investigate the internal acoustic canal and the nerves in it. Fig. 1 represents the location of an inner ear in the auditory system. Sound collected by the outer ear is transferred to the middle ear through the ear channel. The middle ear transforms the
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
215
It is assumed that the CSAs have important information about the condition of nerves. After the data collection phase, statistical measurements were used to discover the existence of any relationship between the thickness of nerves and tinnitus. In addition, if a statistical difference exists between healthy volunteers and patients with tinnitus with respect to CSAs, it is possible to establish a computer-aided diagnostic (CAD) system for feature examinations. The possible statistical significance allows a classification system and therefore a CAD system. The possible existence of a statistical significance allows a classification system and therefore to be able to establish a CAD system. 2. Material
Fig. 1. Three parts of the ear system: inner ear, middle ear, and outer ear [21].
Fig. 2. A schematic diagram of the right IAC: (a) perspective view; (b) cross-sectional view [22].
sound energy into internal vibrations and a compressional wave in the inner ear. Transformation of the internal vibrations into nerve impulses is performed in the inner ear for transferring them to the brain [16–18]. The IAC in the inner ear contains four main nerves carrying the impulses to the brain. A schematic diagram of a right internal acoustic meatus that shows the placement of the nerves in IAC is given in Fig. 2. The nerves that pass through the IAC are the facial nerve (FN), vestibular superior nerve (SVN), inferior vestibular nerve (IVN), and cochlear nerve (CN). Each of the nerves has a constant location in the area of the fundus of IAC [19]. This anatomical knowledge makes it possible to observe structural differentiations. Today, structural forms of IAC and the nerves in it can be visualized using high-resolution techniques. In particular, MRI is more convenient to visualize the structural form and differentiation because the nerves are soft tissues. MRI monitoring of the auditory pathway has helped obtain valuable findings of the ear origin disorders such as vertigo, maneuver, tinnitus, and hearing loss. In a previous study, important findings were obtained indicating that nerve diameters were related to benign paroxysmal positional vertigo (BPPV) [20]. It was determined that the thicknesses of nerves in IAC were differentiating in BPPV cases. The experience and methods in the previous study have benefited the present study. This study focused on finding out any relationship between the thicknesses of IAC nerves and tinnitus. The brain MRI images of patients and healthy subjects were used to investigate how the structures of nerves were affected in tinnitus cases. The crosssectional areas (CSA) of IAC nerves were particularly investigated to determine how the thickness of nerves was affected.
The study was performed on those who consulted the ENT clinic with a tinnitus complaint. An otolaryngologist in the ENT clinic diagnosed that the patients had tinnitus definitely. The volunteers for the control group were chosen among the patients who applied to the radiology department. The volunteers in the control group needed brain MRI images for various reasons. The control and patient groups consisted of 23 healthy subjects and 23 patients. The local ethics committee approved this study, which was carried out in accordance with the Helsinki Declaration. The patients with tinnitus, the volunteers in the control group, and the local ethics committee provided the necessary permission. The brain MRI images of the patient and control groups were used to investigate the IAC region and measure its thickness and the thickness of the nerve in it. Both of the groups had similar demographic characteristics with respect to smoking, hypertension, cholesterol, and so forth. The MR device to obtain the brain images had the power of 1.5 T and an eight-channel high-definition brain coil. An otolaryngologist and a radiologist investigated the IAC canal and the nerves in the MRI images, which were processed using digital image processing techniques. Fig. 3a represents a brain image including the IAC region and the nerves in it as a sample for the collected data. The red square in Fig. 3b shows an IAC region and the nerves. A resolution improvement was performed for detecting the structures adequately after selecting and cropping the IAC region. Fig. 4a shows the cropped original image. Lanczos-2 kernel was used for resolution enhancement [23]. Fig. 4b shows the gray-level image representing the IAC region after resolution enhancement. Then, the obtained gray-level image was converted into a binary image using Otsu’s thresholding method [24]. The binary form of the cropped image given in Fig. 4c shows the specific location of the nerves in the IAC region. The white area refers to IAC, and the black areas in the white region refer to the nerves SVN, FN, CN, and IVN from the left top in clockwise rotation. The CSAs of the IAC canal and the nerves could be measured by counting the pixels in each area because a pixel referred to a voxel size. The MRI device provided this information on their produced Digital Imaging and Communication in Medicine (DICOM) Digital images [25]. Therefore, the real size of a tissue in MRI images could be calculated by a simple estimation. 3. Methods 3.1. Statistical measurements for significant difference The common way to determine any differences among the group data is to use statistical measurements and apply some tests. Many researchers adopt various statistical methods and common-sense techniques to examine the difference between the control and patient groups. Most widely used methods depend on statistical hypothesis tests and correlation calculation. The hypothesis tests
216
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
Fig. 3. MRI images demonstrating the IAC region: (a) sagittal oblique brain image; (b) binary image of the IAC region.
Fig. 4. A cropped region of interest of the IAC region: (a) cropped image; (b) resolution-improved image; (c) binarized image showing the components.
are accepted as key elements for comparing two or more groups of data [26]. Indeed, these methods can put forward a statistical significance between the two groups. In statistical hypothesis testing, the null hypothesis assumes that no statistically significant difference exists between two or more populations. The statistics such as the mean values of distributions are calculated, and their difference is assessed for significance.to test the null hypothesis The significance is meant to reject the null hypothesis. Under the assumption that the null hypothesis is true, the P value is the level of confidence probability that the population data is well represented by the observed result determined by sampling the population. For instance, if the null hypothesis is of equality between two mean values collected from two populations, and if the null hypothesis is confirmed by the empirical observation, the P value is the probability that the two mean values observed are actually truly representative of an equality existing in between the actual populations. A small P value expresses high probability that the two population have different characteristics. The t-test, Wilcoxon signed-rank test, and Pearson’s correlation were used to test the null hypothesis. The null hypothesis of the tests is that no relationship exists between tinnitus data and the thickness data belonging to the IAC and its nerves. This means that the thicknesses of IAC and its nerves do not cause or affect the generation of tinnitus. The aforementioned tests are commonly used to determine a statistical significance. Since the data in the nature are assumed to have a normal distribution, the t test is generally used to determine the statistical differences. The use of the t test is especially true for the observed two groups [27,28]. The statistical significance is considered if the P value is below 0.05. This means that differences exist between the two groups of data observed. That is, the two groups show differences depending on the observed data. For the CSA data, the mean values (), standard deviations (), and P values are given for the experiments conducted on the patient and the control groups. However, the t-test is the most widely test for hypotheses. It is noted that the use of a t-test has a risk of some errors [29,30].
Wilcoxon’s signed-rank test is another recommended method to test multiple data sets [31,32]. As an alternative to the t-test, the Wilcoxon signed-rank test is nonparametric. It ranks the differences in the data set and compares the ranks for the positive and the negative differences without considering the signs [33]. In hypothesis tests, it is considered that a statistical significance exists if the P values are extremely low, such as around or less than 0.05. In addition to the hypothesis, correlation coefficients can be used to assess the similarity between the two groups. The Pearson product-moment correlation coefficient is also used to determine the significant difference [34,35]. It is the most widely used method for determining the correlation coefficients. If the correlation coefficient is near to 1, the groups are considered to be similar to each other. That is, no significant difference exists between the two groups, depending on the observed data. The correlation coefficients of the measurements are given for all CSA in the experimental tables. 3.2. Extreme learning machines Although feedforward neural networks have been successful in pattern recognition problems, its slow learning is still the main disadvantage. The main reason for this slow learning is that the neural network has gradient-based learning algorithms and tunes them iteratively. Unlike conventional iterative learning algorithm, weights and thresholds can be randomly generated, and the numbers of neurons in the hidden layer can be adjusted. After learning, extreme learning machine (ELM) can find out the optimal solution without setting any other parameters. This method not only decreases data processing time but also increases generalized efficiency. The output of ELM can be given as [36]: Ok (x) =
L
woi G (ai , bi , xk )
(1)
i=1
where a, b, and L are the learning parameters and the bias of the i-th hidden-layer neuron, respectively. Here, the i-th hidden-layer
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219 Table 1 Acronyms for the CSAs of inner ear components. FN SVN IVN CN IAC VT
Table 2 Statistical values related to the CSAs of IAC and its nerves.
Facial nerve Superior vestibular nerve Inferior vestibular nerve Cochlear nerve Internal auditory canal Vestibular total (inferior + superior)
neuron is connected to the output layer with woi . G (.) refers to the output of the I-th hidden-layer neuron [37]. The ELM output can be expressed in the compact form as follows:
O = WoT H
T
⎢ H=⎢ ⎣
Measurements
FN
SVN
IVN
VT
CN
IAC
Control Control Tinnitus Tinnitus Overall Overall t-test Wilcoxon Pearson
0.4919 0.2384 0.4442 0.2335 0.4693 0.2459 0.4910 0.6051 0.8712
0.5278 0.2250 0.4891 0.1344 0.5070 0.1973 0.0590 0.0727 0.7149
0.3353 0.2464 0.2614 0.1305 0.2812 0.1878 0.39352 0.30108 0.06384
0.8631 0.3744 0.7506 0.1822 0.8282 0.3051 0.0749 0.0777 0.5319
0.6147 0.3032 0.5706 0.2076 0.6068 0.3017 0.3633 0.4115 0.1604
24.4775 5.91042 21.6135 5.73889 22.6335 6.67977 0.0697 0.0829 0.2430
(2)
where H is the matrix of the hidden-layer neurons, and a (ki) refers to G (ai , bi , xk ), which is the input in the k-th row and i-th column in the hidden-layer output matrix H. H and W can be given in a matrix form as follows:
⎡
217
H1 (1)
···
H1 (N)
.. .
..
.. .
HL (N)
···
.
HL1 (N)
⎤
⎡
b1
b1
···
b1
⎤
⎢ . .. .. .. ⎥ ⎢ ⎥ ⎥ . . . ⎥ (3) ⎥ andW = ⎢ .. ⎢ ⎥ ⎦ ⎣ w11 w11 . . . w11 ⎦ w11
w11
···
w11
The ELM tries to find the minimum training error and the norm of the output weight by calculating the cost function given as: CF = O − Y
(4)
where Y is the expected output matrix. The learning speed of ELM is quite fast because no parameters need to be tuned during the training phase. Kernels can be integrated in ELM because they are support vector machines [38–40]. In this study, both the linear kernel and the Radial Basis Function (RBF) kernel were employed in this study. 4. Experiments and results The brain MRI images of the patients with tinnitus and healthy subjects were examined to investigate any relationship between tinnitus and the structure of IAC components. Particularly, the thickness of nerves and the area of IAC were evaluated as a structure. It was assumed that the thicknesses were related to the CSAs of IAC components: the nerves and the IAC through the collected MRI images. Indeed, the anatomy of the inner ear was spread over an extremely small area and had a complicated structure to visualize it completely. The visualization of the inner ear components required the specific positioning of the patient. Setting the patient and maintaining her/him in this special position was quite difficult. One of the most critical steps in this study was to obtain MRI images at the right angle to visualize CSAs. The CSAs of the IAC and the nerves were observed and measured under the guidance of ENT and radiology experts, who worked collaboratively to ensure that the patient was in the right position. The collected data was investigated in two ways to determine statistical significance, which referred to a relationship between tinnitus and the structure of IAC components. The first way was an approach to evaluate CSAs directly, while the latter way evaluated them proportionally. Therefore, statistical measurements and classification methods were applied on both direct data and proportional data. Table 1 shows the acronyms for the direct and proportional data used in Tables 2 and 3. They present the mean value and standard deviation of control data, tinnitus data, and whole data.
Table 2 shows the results of statistical measurements, tests, and correlation coefficients. In addition to the tests and correlation coefficients, mean values () and standard deviations () were also common to evaluate how the data varied. The standard deviations and the mean values were given in the three categories with respect to the collected data: patients with tinnitus, healthy subjects, and overall values, which denoted whole data. “Tinnitus” and “Control” labels denote the patients with tinnitus and the normal healthy subjects in tables, respectively. “Overall” denotes the evaluation of all cases including the patients with tinnitus and the healthy subjects. No significant differences were observed between the patient and control groups on considering the mean values and the standard deviations into consideration. However, it could be realized that the mean values and the standard deviation for the cases with tinnitus were slightly smaller compared with the healthy cases. Nevertheless, this inference with no discrimination was applicable to all of the nerves. The results of the hypothesis tests and the correlation coefficients according to the nerves are also given in Table 2. According to the hypothesis tests, some of the P values were quite close to the values accepted as statistically significant. The statistically significant limit was just a hypothetical value. The values close to this limit could be also considered as statistically significant more or less. Therefore, the P values for SVN, VT, and IAC indicated statistical significance. Both the t-test and Wilcoxon signed-rank test indicated a statistical significance for SVN, VT, and IAC with the value pairs (0.05903–0.07273, 0.07496–0.07772, and 0.069711–0.08298), respectively. The significant values in the table are shown in bold. Although the t-test and Wilcoxon signed-rank test point out SVN, VT, and IAC, Pearson correlation coefficient claim that FN and SVN are related to tinnitus. Both tests and correlation coefficients found that the thickness of SVN was related to tinnitus. Table 3 represents the statistical values of the proportional data. The proportional data were obtained as the ratio of CSAs to IAC and some important nerves (FN and VT). The column names refer to these ratios. If the P values were taken into consideration, none of hypotheses tests showed any significance. These results indicated that tinnitus cases were not related to the ratio between the nerves and the IAC. However, the correlation coefficient for the proportional data put forward that the IVN/IAC and the VT/FN had a reasonable relationship to tinnitus. Although the results of the statistical hypothesis tests and the correlation coefficients made a claim that the structures of SVN, FN, VT, and IAC were related to tinnitus, it was hard to decide which nerve was the most dominant factor for tinnitus complaint. If these values indicated a statistically significant difference, a classification should be possible with respect to the CSAs of nerves. Thus, it was possible to determine the most dominant nerve on tinnitus by comparing the correct classification rates. The ELM classifier was accepted as one of the most reliable classifiers used widely in recent works.
218
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
Table 3 Relational statistical values related to the CSAs of IAC and its nerves.
Control Control Tinnitus Tinnitus Overall Overall t-test Wilcoxon Pearson
FN/IAC
SVN/IAC
IVN/IAC
VT/IAC
CN/IAC
SVN/FN
IVN/FN
VT/FN
CN/FN
CN/VT
0.0202 0.0087 0.0223 0.0145 0.0208 0.0120 0.4105 0.8314 0.91984
0.0222 0.0103 0.0230 0.0046 0.0239 0.0085 0.4875 0.5633 0.72264
0.0131 0.0076 0.0132 0.0091 0.0122 0.0080 0.3852 0.7151 0.97215
0.0353 0.0136 0.0363 0.0111 0.0361 0.0125 0.9283 0.9031 0.51967
0.0246 0.0085 0.0284 0.0139 0.0269 0.0121 0.5026 0.8551 0.00466
1.3695 1.0690 1.4137 0.8377 1.4734 0.9219 0.6754 0.7151 0.82095
0.7194 0.4251 0.7441 0.5129 0.6908 0.4345 0.3961 0.6051 0.67959
2.0889 1.2321 2.1579 1.2306 2.1642 1.1780 0.9711 0.6482 0.94091
1.4802 0.8926 1.7185 1.4789 1.6441 1.2358 0.7288 0.9031 0.09324
0.8022 0.4197 0.7935 0.3279 0.7884 0.3725 0.9409 0.8791 0.01373
Table 4 Classification results using different training rates. Training rate (%) 50 60 70
Mean Max Mean Max Mean Max
FN
SVN
IVN
VT
CN
IAC
0.52 0.59 0.48 0.61 0.47 0.64
0.71 0.90 0.75 0.89 0.76 0.92
0.56 0.63 0.48 0.66 0.51 0.71
0.70 0.86 0.71 0.88 0.71 0.91
0.47 0.54 0.47 0.55 0.51 0.64
0.74 0.91 0.76 0.94 0.82 0.93
ies, which required medical imaging, only hypothesis tests were usually accepted as sufficient. However, a classification of the data was preferred to ensure that tinnitus was related to the thickness of inner ear nerves and show the possibility of constructing a CAD system to assist ENT physicians. The CRRs of the classification for proportional data is given in Table 6 indicated that tinnitus cases were not related to the ratio between the nerves and the IAC.
5. Discussion Table 5 Classification results for the combination of three nerves. Training rate (60%)
SVN + VT
SVN + IAC
VT + IAC
Mean Max
0.70 0.94
0.75 0.94
0.71 0.94
For the classification phase, some portion of the patient data and the control data were used for training and the rest were used for testing the classification. The samples for the training and the test were chosen randomly. The classification process was repeatedly performed 10 times to ensure the result of classification. The result of classification for the CSAs of nerves is given in Table 4. In addition, the classification was repeated 10 times for three different training rates. Table 1 presents the average and the maximum correct classification rates (CRR) obtained with the ELM classifier using a radial basis function. The k-fold test was avoided in this study due to insufficient data. Actually, it was extremely difficult to find the patients with tinnitus and obtain their brain MRI images. Another difficult task in this study was to visualize the inner ear region at right angle as stated earlier. If the visualization of inner ear was not achieved at right angle, it would lead to the inaccurate measurement of CSAs. Also not being able to visualize the inner ear in the right position prevents measuring the thickness of inner ear components. This might be the reason why the value of P, which is the result of hypothesis tests, is slightly more than 0.5. The classification results in the tables were compatible with the results of the hypothesis tests. The classification tests were repeated for three training ratios of 50%, 60%, and 70%. The CRRs of classification for SVN, VT, and IAC were achieved as 0.92. The mean performances of ELM classifier were 0.7, 0.71, and 0.82 for the training rate of 70%. According to the classification result, the IAC was the nerve more related to tinnitus. Table 5 represents the classification result for the combination of three nerves. Although the maximum CRR increased at the same value, the mean CRR was slightly higher for using SVN and IAC features. In such stud-
Tinnitus that may be more troublesome than marginal hearing loss has several reasons. It is difficult to determine the cause of this disease, which affects almost 70% adults. It has been reported by researchers that environmental noise, change in gray matter, pathological problems in the sound channel, and structural problems of the inner ear can cause tinnitus [7,10,15]. Particularly, a number of studies have been performed to determine the differences that may be related to tinnitus in audio vestibular tract. Many findings on MRI screening of the audio vestibular tract indicates the importance of MRI examination in patients with tinnitus [15]. These studies depend on the belief that the pathological or structural condition of audio vestibular tract is related to tinnitus [2,11]. The studies have reported that intracranial abnormalities are responsible for the sensorineural hearing loss in patients. Furthermore, MRI of the temporal bone, cerebral angle, and brain have shown abnormal results in patient with tinnitus [6]. Taking into consideration these studies, the present study was performed with the idea that the inner ear canal and its nerves might have different thicknesses in tinnitus cases. Therefore, the inner ear region was visualized using brain MRI images in a patient with tinnitus. The statistical analysis and the classification results demonstrated that the thicknesses of the nerves were affected in tinnitus cases. In particular, the cross sectional area of the superior vestibular nerve and internal auditory canal showed the highest significance, indicating their relationship with tinnitus. However, similar results were obtained for the sum of the cross-sectional areas of vestibular and inferior nerves, it is clear that the superior vestibular nerve is actually related to tinnitus. Both the statistical results and the classification results did not show a relationship between the cross sectional area of inferior vestibular nerve and tinnitus. In addition, tinnitus problem was also associated with the structure of internal auditory channel. This study examined the cross sectional areas of the internal auditory canal and the nerves in it, where only the nerves were visible in brain MRI images. Since the internal auditory canal had a complex
Table 6 Correct recognition rates using proportional data. Training rate (60%)
FN/IAC
SVN/IAC
IVN/IA
VT/IAC
CN/IAC
SVN/FN
IVN/FN
VT/FN
CN/FN
CN/VT
Mean Max
0.45 0.57
0.47 0.64
0.45 0.57
0.45 0.57
0.54 0.64
0.44 0.57
0.46 0.64
0.40 0.57
0.36 0.50
0.42 0.64
B. Ergen et al. / Biomedical Signal Processing and Control 40 (2018) 214–219
structure, a special examination of its structure was a more correct approach. The mean values and the standard deviations indicated that the cross sectional areas of nerves and internal auditory canal were slightly smaller in tinnitus cases. Although the pathology of internal ear is reported to be related to hearing loss, vertigo, and tinnitus [11], it is not possible to take a biopsy from patients and examine them. Perhaps combining pathological and MRI findings may explain why these nerves are thinner in tinnitus cases. The findings of this study showed that tinnitus was particularly related to the inner ear components.
[14]
[15]
[16] [17] [18] [19]
6. Conclusions The results of this study emphasized the importance of MRI visualization in patients with tinnitus to find out the differentiation in audio vestibular tract. The experiments also showed that tinnitus classification was possible by considering the thickness of nerves and the cross-sectional area of internal auditory canal. Furthermore, the possibility of classification for tinnitus identification may facilitate establishing a computer-assisted diagnostic system that will help physicians.
[20]
[21] [22] [23]
[24]
References [25] [1] T.A. Gabr, M.A. El-Hay, A. Badawy, Electrophysiological and psychological studies in tinnitus, Auris Nasus Larynx 38 (2011) 678–683. [2] M. Shweel, B. Hamdy, Diagnostic utility of magnetic resonance imaging and magnetic resonance angiography in the radiological evaluation of pulsatile tinnitus, Am. J. Otolaryngol. 34 (2013) 710–717. [3] C. Polat, M. Baykara, B. Ergen, Evaluation of internal auditory canal structures in tinnitus of unknown origin, Clin. Exp. Otorhinolaryngol. 7 (2014) 160. [4] X. Zhou, S. Henin, G.R. Long, L.C. Parra, Impaired cochlear function correlates with the presence of tinnitus and its estimated spectral profile, Hear. Res. 277 (2011) 107–116. [5] D. Távora-Vieira, R. Marino, A. Acharya, G.P. Rajan, The impact of cochlear implantation on speech understanding, subjective hearing performance, and tinnitus perception in patients with unilateral severe to profound hearing loss, Otol. Neurotol. 36 (2015) 430–436. [6] P.C. Weber, R.I. Zbar, B.J. Gantz, Appropriateness of magnetic resonance imaging in sudden sensorineural hearing loss, Otolaryngol. Head Neck Surg. 116 (1997) 153–156. [7] A.C. Guinchard, N. Ghazaleh, M. Saenz, E. Fornari, J.O. Prior, P. Maeder, S. Adib, R. Maire, Study of tonotopic brain changes with functional MRI and FDG-PET in a patient with unilateral objective cochlear tinnitus, Hear. Res. 341 (2016) 232–239. [8] K. Boyen, E. de Kleine, P. van Dijk, D.R. Langers, Tinnitus-related dissociation between cortical and subcortical neural activity in humans with mild to moderate sensorineural hearing loss, Hear. Res. 312 (2014) 48–59. [9] E. Levi, E.K. Bekhit, R.G. Berkowitz, Magnetic resonance imaging findings in children with tinnitus, annals of otology, Rhinol. Laryngol. 124 (2015) 126–131. [10] A.M. Leaver, T.K. Turesky, A. Seydell-Greenwald, S. Morgan, H.J. Kim, J.P. Rauschecker, Intrinsic network activity in tinnitus investigated using functional MRI, Hum. Brain Mapp. 37 (2016) 2717–2735. [11] X. Wu, K. Chen, L. Sun, Z. Yang, Y. Zhu, H. Jiang, Magnetic resonance imaging-detected inner ear hemorrhage as a potential cause of sudden sensorineural hearing loss, Am. J. Otolaryngol. 35 (2014) 318–323. [12] M. Smits, S. Kovacs, D. De Ridder, R.R. Peeters, P. Van Hecke, S. Sunaert, Lateralization of functional magnetic resonance imaging (fMRI) activation in the auditory pathway of patients with lateralized tinnitus, Neuroradiology 49 (2007) 669–679. [13] M. Schecklmann, M. Landgrebe, T.B. Poeppl, P. Kreuzer, P. Männer, J. Marienhagen, D.S. Wack, T. Kleinjung, G. Hajak, B. Langguth, Neural correlates
[26]
[27] [28]
[29] [30] [31] [32] [33]
[34]
[35] [36] [37]
[38] [39] [40]
219
of tinnitus duration and distress: a positron emission tomography study, Hum. Brain Mapp. 34 (2013) 233–240. J.-J. Song, D. De Ridder, P. Van de Heyning, S. Vanneste, Mapping tinnitus-related brain activation: an activation-likelihood estimation metaanalysis of PET studies, J. Nucl. Med. 53 (2012) 1550–1557. A.A. Aarnisalo, H. Suoranta, J. Ylikoski, Magnetic resonance imaging findings in the auditory pathway of patients with sudden deafness, Otol. Neurotol. 25 (2004) 245–249. W.A. Yost, Fundamentals of Hearing: An Introduction, ASA, 2001. A.J. Gulya, L.B. Minor, M.E. Glasscock, D. Poe, Glasscock-Shambaugh Surgery of the Ear, PMPH, USA, 2010. R. Romand, I. Varela-Nieto, Development of Auditory and Vestibular Systems, Academic Press, 2014. H. Ryu, T. Tanaka, S. Yamamoto, K. Uemura, Y. Takehara, H. Isoda, Magnetic resonance cisternography used to determine precise topography of the facial nerve and three components of the eighth cranial nerve in the internal auditory canal and cerebellopontine cistern, J. Neurosurg. 90 (1999) 624–634. B. Ergen, M. Baykara, C. Polat, An investigation on magnetic imaging findings of the inner ear: a relationship between the internal auditory canal, its nerves and benign paroxysmal positional vertigo, Biomed. Signal Process. Control 9 (2014) 14–18. J. Hass, An Acoustics Primer, Indiana University, 2003, http://www.indiana. edu/?emusic/acoustics/acoustics.htm. (Accessed 26 September 2012). L. Albert, J. Rhoton, The cerebellopontine angle and posterior fossa cranial nerves by retrosigmoid approach, Neurosurgery 47 (2000) 93–129. E.H.W. Meijering, W.J. Niessen, M.A. Viergever, Quantitative evaluation of convolution-based methods for medical image interpolation, Med. Image Anal. 5 (2001) 111–126. P.K. Sahoo, G. Arora, A thresholding method based on two-dimensional Renyi’s entropy, Pattern Recognit. 37 (2004) 1149–1161. P. Mildenberger, M. Eichelberg, E. Martin, Introduction to the DICOM standard, Eur. Radiol. 12 (2002) 920–927. J.L. Amaral, A.J. Lopes, A.C. Faria, P.L. Melo, Machine learning algorithms and forced oscillation measurements to categorise the airway obstruction severity in chronic obstructive pulmonary disease, Comput. Methods Progr. Biomed. 118 (2015) 186–197. E.H. Livingston, Who was student and why do we care so much about his t-test? J. Surg. Res. 118 (2004) 58–65. E.H. Livingston, L. Cassidy, Statistical power and estimation of the number of required subjects for a study based on the t-test: a surgeon’s primer, J. Surg. Res. 126 (2005) 149–159. T.G. Dietterich, Approximate statistical tests for comparing supervised classification learning algorithms, Neural Comput. 10 (1998) 1895–1923. T.E. Nichols, A.P. Holmes, Nonparametric permutation tests for functional neuroimaging: a primer with examples, Hum. Brain Mapp. 15 (2002) 1–25. J. Demˇsar, Statistical comparisons of classifiers over multiple data sets, J. Mach. Learn. Res. 7 (2006) 1–30. J.J. Higgins, Introduction to Modern Nonparametric Statistics, 2003. W.M. Azmy, N. El Gayar, A.F. Atiya, H. El-Shishiny, M.L.P. Gaussian Processes, Negative correlation learning for time series prediction, in: J.A. Benediktsson, J. Kittler, F. Roli (Eds.), Multiple Classifier Systems: 8th International Workshop, MCS 2009, Reykjavik, Iceland, June 10–12, 2009. Proceedings, Springer Berlin Heidelberg, Berlin,Heidelberg, 2009, pp. 428–437. B.G. Häne, K. Jäger, H.G. Drexler, The Pearson product-moment correlation coefficient is better suited for identification of DNA fingerprint profiles than band matching algorithms, Electrophoresis 14 (1993) 967–972. M.-T. Puth, M. Neuhäuser, G.D. Ruxton, Effective use of Pearson’s product–moment correlation coefficient, Anim. Behav. 93 (2014) 183–189. G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Extreme learning machine: theory and applications, Neurocomputing 70 (2006) 489–501. N.-Y. Liang, G.-B. Huang, P. Saratchandran, N. Sundararajan, A fast and accurate online sequential learning algorithm for feedforward networks, IEEE Trans. Neural Netw. 17 (2006) 1411–1423. X. Wang, M. Han, Online sequential extreme learning machine with kernels for nonstationary time series prediction, Neurocomputing 145 (2014) 90–97. X. Liu, L. Wang, G.-B. Huang, J. Zhang, J. Yin, Multiple kernel extreme learning machine, Neurocomputing 149 (2015) 253–264. R. Yuan, Z. Li, X. Guan, L. Xu, An SVM-based machine learning method for accurate internet traffic classification, Inform. Syst. Front. 12 (2010) 149–156.