J Med Syst (2016) 40:221 DOI 10.1007/s10916-016-0573-7
SYSTEMS-LEVEL QUALITY IMPROVEMENT
Investigating the Effectiveness of Wavelet Approximations in Resizing Images for Ultrasound Image Classification Umar Manzoor 1
&
Samia Nefti 2 & Milella Ferdinando 2
Received: 1 March 2016 / Accepted: 11 August 2016 # Springer Science+Business Media New York 2016
Abstract Images are difficult to classify and annotate but the availability of digital image databases creates a constant demand for tools that automatically analyze image content and describe it with either a category or a set of variables. Ultrasound Imaging is very popular and is widely used to see the internal organ(s) condition of the patient. The main target of this research is to develop a robust image processing techniques for a better and more accurate medical image retrieval and categorization. This paper looks at an alternative to feature extraction for image classification such as image resizing technique. A new mean for image resizing using wavelet transform is proposed. Results, using real medical images, have shown the effectiveness of the proposed technique for classification task comparing to bi-cubic interpolation and feature extraction.
Keywords Ultrasound classification . Feature extraction . Image processing . Image resizing . Wavelet transformation . Neural networks
This article is part of the Topical Collection on Systems-Level Quality Improvement * Umar Manzoor
[email protected] Samia Nefti
[email protected] 1
King Abdulaziz University, Jeddah, Saudi Arabia
2
School of Computing, Science and Engineering, The University of Salford, Greater Manchester, Salford, UK
Introduction The vast amount of medical and other digital images made available in the recent years demanded a better and more accurate image retrieval and categorization. Feature extraction as image processing technique has proven to be very popular especially with medical images as almost any new technique innovated for the analysis of ultrasound image uses feature extraction technique. In general, feature extraction involves dimensionality reduction (i.e. reducing the amount of variables required to represent large data). Feature extraction can be categorized into three broad categories 1) Low Level feature extraction 2) Shape based feature extraction and 3) Texture based feature extraction. Ultrasound Imaging is very popular and is widely used to see the internal organ(s) condition of the patient [1]. Feature extraction (such as edge detection, shape or texture) helps in autonomous diagnosis of various diseases; however, feature extraction has its own limitation due to the nature of ultrasound images [2–4]. The noise/ speckle sometimes makes feature extraction impossible to be effective [5–8] mainly when the images includes high level of speckle/noise, blurring and distortions. Hence, in some application such as image classification, categorization, etc., using the whole image rather than just choosing a specific feature can be more effective as processing the whole image with its background/ environment can carry very important information which can be used efficiently during the classification and may affect the performance of the classifier. However, processing the whole image also has its down side, consider an image with the dimension of 600 × 600, and processing the whole image means each pixel will be considered as parameter ending up with 360,000 parameters which can be computationally` expensive.
221
J Med Syst (2016) 40:221
Page 2 of 9
Raphaël Marée in [9] has proposed a method which operates directly on pixel, a simple local sub-window extraction technique with induction of ensembles of extremely randomized decision trees is proposed, and the image is resized to 32 × 32 before classification which leads to loss of information. In this paper, we propose a new mean for resizing an image with minimal loss of information using Wavelet Transform approximation. There are two domains in which wavelet have proven to be very resourceful and widely used which are image compression and image de-nosing. The wavelet transform allows dismantling of the image by filtering the original image into different frequency components using a filter bank. This filter bank divides each dimension of the image into a low-pass and a high-pass region. The wavelet decomposition is usually preformed on images to compress it or to carry out wavelet parametric characterization. The image is simply two-dimensional array MxN which requires the use of two-dimensional discrete wavelet transform, the most common / simplest arrangement is to use the same scaling in the horizontal and vertical directions. R. Recknagel et al. in [10] proposed an algorithm using Wavelet Transform to separate local defects from the surface and the noise (measurement error and surface roughness) with a given manufacturer’s risk for piecewise smooth surfaces. Sonja Gregic et al. in [11] reviewed the important features of wavelet transform, its usage in image compression and discussed the image quality degradation by the process of wavelet compression and decompression. In any Image processing technique, pre-processing is usually an important step where the region of interest is enhanced/ selected and area/pixels of no interest are eliminated from the input image. Cropping is one of the pre-processing operations performed on the image to 1) enhance the region of interest and 2) remove all unwanted labels / noise etc. Cropping can play an important role in removing noise (such as measurement label) from the ultrasound image. Image enhancement techniques (such as smoothing, edge enhancement, noise reduction, adaptive thresholding…) improve the detectability of the objects in the image. In this paper, we proposed image resizing technique for ultrasound images classification using wavelet transform and compared it with feature extraction and bicubic-interpolation. Neural Network [12–15] is used as classifier and the results are compared with the conventional technique used for resizing i.e. bicubic-interpolation and with the feature extraction. The paper is organized as follows. In Section II, image processing, enhancement and feature extraction is presented. The proposed image resizing technique using discrete wavelet transform is discussed in section III. Experimental results for the classification of the ultrasound images are presented in section IV. In the end, conclusion is drawn.
Related work Medical images are difficult to classify and annotate especially ultrasound images as these images are low contrast and usually contain high noise. According to Alison Noble et al., the classification accuracy of ultrasound image is dependent on ultrasound image quality because the contrast between areas of interest is very low [16]. Studies proposed in the last decade have proved that ultrasound imaging improves the diagnosis of internal organ (such as liver, breast, heart etc) various diseases [17, 18]. Xiaofang Hou et al. in [19] proposed the classification of cardiac ultrasound image based on sparse representation, GLCM based texture features are used to compute the sparse solution with coefficients, support vector machine and artificial neural network are used as classifiers for experimentation. According to the authors, the proposed technique using sparse representation outperforms other techniques in terms of classification accuracy and has obvious advantages over other techniques when applied to computeraided diagnosis. Karthik Kalyan et al. in [20] presented a comparative analysis of different texture features extracted from ultrasound image to diagnosis the disease condition. Multilayer perceptron artificial neural network is used for classification and to identify which extracted feature is best for classification. The classification performance is evaluated using confusion matrix and receiver operating characteristics curve. According to the authors, GLRLM and mixed feature set showed best classification efficiency as compared to other features. Suhuai Luo et al. in [21] proposed automatic liver segmentation using three different methods (i.e. gray level, structure based, and texture based). According to the authors, each segmentation method has its own pros and cons, for example, gray scale based methods are very effective in tumour segmentation if the image contains high level of gray scale, however, these methods lose their effectiveness when the gray level of the image changes. Similarly, structure based methods are very effective when the boundary of the liver is unclear, however, these methods requires extensive training. Texture based methods use textures in the image to determine the boundaries; these methods achieve better efficiency when the boundaries are not clear. Using wavelet techniques in image classification for ultrasound imaging is not a new concept, researcher in the last decade have proposed many efficient techniques using wavelet for ultrasound image analysis, compression, resizing, de-nosing and classification [22, 23]. Henrik Nicolay Finsberg in [24] studies the use of wavelet techniques in ultrasound imaging and proposed windowed scattering transform (wavelet-based) technique for ultrasound image classification. Ultrasound images of
J Med Syst (2016) 40:221
Page 3 of 9 221
Fig. 1 Original images of abdomen, cranium, and femur
a) Abdomen
b)Cranium
c) Femur
Fig. 2 Cropped images
a) Abdomen
b) Cranium
c) Femur
221
J Med Syst (2016) 40:221
Page 4 of 9
Fig. 3 Enhanced images
a) Abdomen
b) Cranium
c) Femur finger-joints are used in this research to evaluate the proposed technique. According to the authors, experimental results show that the efficiency / accuracy of the proposed technique depends heavily on the size of training set (the larger the training set—the better the accuracy). Jong-Woo Han et al. in [25] proposed content-aware image resizing technique based on wavelet analysis, the authors modified the energy function in wavelet decomposition and introduced local energy estimation by weighting wavelet subband components. According to
the authors, experimental results demonstrated that the proposed technique preserves the shape of the main objects in the image after the resizing operation. Dhrub Kumar et al. in [26] presented a comparative analysis of different methods for enhancing ultrasound images, according to the authors wavelet based techniques are very effective as compared to other techniques. The authors investigated different wavelet families (such as Haar, Daubechies, Coiflet and Symlet) effectiveness for enhancing ultrasound images. Different decomposition levels / threshold selection methods were used in
Fig. 4 Original images resized using Bi-cubic interpolation
a) Abdomen
b) Cranium
c) Femur
J Med Syst (2016) 40:221
Page 5 of 9 221
Fig. 5 Show the original image and the decomposition at level two using DWT
Original Image
experimentation; results demonstrated that different wavelet families perform differently on different decomposition levels [26].
Image processing, enhancement and feature extraction Preprocessing Ultrasound / mammogram images are considered to be very difficult to analyze which makes pre-processing stage an essential step. In addition to the complexity of ultrasound / mammograms images there is also the added noise when the images were taken, noises such as measurement label Fig. 1. To avoid enhancement of noises, the following preprocessing steps are carried out: i) Image Cropping ii) Image Cleaning and iii) Visual Enhancement. In the Image cropping step, an optimal 300 × 300 pixel area is chosen which includes the area of interest and discarding the rest as shown in Fig. 2. The purpose of this step is to eliminating any part in the image which could affect the region of interest and complicate the classification stage. The second step is performed only on cranium images to remove the marks around the cranium as shown in Fig. 1. These marks are placed by the nurse when the
Two- dimensional decomposition
image is taken to identify the size of the cranium. In this step, these marks are removed which leaves behind black empty marks, however, morphological dilation process is applied to fill these gaps based on the neighboring pixels. In visual enhancement the image intensity values are adjusted using adaptive thresholding. Adaptive thresholding approach computes an independent threshold for each pixel on a given local window whose center is the pixel which is being binarized. If the value of the pixel is greater than the threshold, the brightness is weighted toward brighter output values, and if the value of the pixel is less than the threshold, the brightness is weighted toward lower (darker) output values. After successfully applying the adaptive thresholding, a morphological operation known as ‘thinning’ is applied. Thinning can be used for many purposes, but for this case it is particularly used to tidy up the binary image by reducing most unwanted object to single pixel thickness. Figure 3 shows the images of abdomen, cranium, and femur after preprocessing stage. Feature extraction After successfully enhancing the image, segmentation is performed, for more information on image segmentation see [27]. Having converted the image to segments
Fig. 6 Original images resized using DWT
a ) Abdomen
b) Cranium
c ) Femur
221
J Med Syst (2016) 40:221
Page 6 of 9
it is now easy to extract the required objects based on set of properties. The extracted properties are: Bounding box, also known as bounding rectangle is a rectangle which encloses the object. It is used to identify the shape and location of the object. Area, the area is simply the total number of pixel in the object; it is used here to identify the size of the object. Eccentricity is used to identify the ratio of the distance between the foci of the ellipse and its major axis length. The value obtained will be between 0 and 1 (if the object is circle the value will be closer to 0 and if it is straight line then the value will be closer to 1). After extracting the information from the image, input table is created for classification.
a) Feature extraction
Proposed technique As an alternative to feature extraction we have researched different methods to prepare the data for classification. The alternative methods involve resizing the enhanced images from 300X300 to 17X17 giving 289 inputs for the classification.
b) Bi-cubic Interpolation
Standard bi-cubic interpolation resizing This simply involves estimates the color at a pixel in the destination image by a weighted average of 4 × 4 pixels surrounding the closest neighboring pixel in the source image. Figure 4 shows original images resized using Bi-cubic Interpolation. The calculation behind bi-cubic interpolation is: 4 4 X X
apj xp yp
p¼1 p¼1
c) Discrete Wavelet Transform Resizing using wavelet approximation
Fig. 7 Classification results produced by the three techniques
The wavelet transform allows the dismantling of the image information by filtering the original image in to different frequency components using a filter bank. This filter bank divides each dimension of the image into a low-pass and a highpass region [28]. The wavelet decomposition is usually preformed on images to compress it or to carry out wavelet parametric characterization of data. The images are simply two-dimensional array MxN which requires the use of twodimensional discrete wavelet transform. The most common and simplest arrangement is to use the same scaling in the horizontal and vertical directions [29]. For image or MxN array a two-dimensional decomposition is performed, this leads to a decomposition of approximation coefficients at any level in four components: the approximation, and the
details in three orientations (horizontal, vertical, and diagonal) [30] as shown in Fig. 5. Applying Discrete Wavelet Transform decomposition on the original images produces the images as shown in Fig. 6. Also the images resized by DWT have produced better visual appearances than the standard bi-cubic interpolation. Table 1
Results produced by holdout validation
Feature extraction Bi-cubic interpolation DWT
Abdomen
Cranium
Femur
MSE
86.636 89.306 90.409
87.569 92.628 93.551
96.66 89.160 97.021
0.279 0.286 0.170
J Med Syst (2016) 40:221 Table 2
Total
Page 7 of 9 221
Results using feature extraction
Table 4
Results using discrete wavelet transform
Abdomen
Cranium
Femur
MSE
Abdomen
Cranium
Femur
MSE
75.807 84.183 89.442 87.304 83.018 86.169 84.124 87.297 87.585 89.67 85.46
80.760 86.081 87.849 86.153 85.946 86.060 84.011 86.186 83.806 83.97 85.082
97.005 96.558 95.239 96.741 97.115 96.473 96.875 96.379 96.579 95.863 96.483
0.276 0.252 0.270 0.282 0.277 0.269 0.274 0.267 0.276 0.271 0.271
91.456 90.565 90.740 89.283 93.251 93.097 93.879 91.105 92.823 92.078 91.828
94.061 92.946 94.820 92.583 95.033 95.756 94.422 93.352 92.511 92.381 93.787
94.478 93.947 97.824 97.763 95.140 95.291 92.241 93.743 97.852 93.587 97.007
0.301 0.301 0.300 0.295 0.291 0.305 0.302 0.301 0.299 0.299 0.299
Total
Classification and experimental results
Experiments and results
For categorization of the images, feed-forward back-propagation neural network (NN) is chosen as classifier.
For the classification two techniques are used i) holdout validation and ii) k-fold cross validation. For the holdout validation test dataset is chosen randomly from the initial sample and the remaining observations are retained as the training data, normally third of the initial sample is used as validation dataset. In K-fold cross-validation, the original sample is partitioned into K sets, a single dataset is retained as the validation data for testing the model, and the remaining K − 1 set are used as training datasets. The cross-validation process is then repeated K times (the folds), with each of the dataset used exactly once as the validation dataset. The K results from the folds then can be averaged (or otherwise combined) to produce a single estimation.
Feed-forward back-propagation neural network A feed-forward network has a layered structure with no connections within a layer. Each layer consists of nodes which receive their input from nodes of the previous layers and sends the output to next layer [31]. When a learning pattern is secured, the activation values are propagated to the output units and compared with desired output, this produces an error in each of the output unit. Let’s call this error eofor a particular output unit and it needs to be minimized (ideally the minimized value should be very close to zero) [12, 13]. To accomplish the above, the connections of the network are amended in such way that eoreaches its minimum value for the particular pattern. A common sigmoid function (logistic function) which is widely applied is used as activation function in this research [32, 33]. Table 3
Total
Results using Bi-cubic interpolation Abdomen
Cranium
Femur
MSE
91.220 89.010 90.385 88.849 91.767 91.555 92.409 90.626 91.661 90.636 90.812
96.019 94.003 94.852 93.672 97.168 96.034 94.786 93.175 95.878 91.868 94.745
91.749 89.468 90.237 86.797 92.424 89.950 90.374 89.468 89.269 90.933 90.067
0.290 0.297 0.292 0.291 0.285 0.302 0.301 0.296 0.286 0.296 0.294
Holdout validation In this experimentation using the holdout validation method, 30 % of the dataset is used for validation and 70 % is used for training. Figure 7a shows that there is slight confusion between abdomen and cranium and hardly any confusion with femur when Feature Extraction method is used, whereas Fig. 7c shows that there is less confusion between images when using Discrete Wavelet Transform instead of bi-cubic interpolation Fig. 7b and feature extraction methods. As we can see from the results (Table 1 and Fig. 7), the images resized using DWT have produced the best results but Table 5
Averages of all the results
Feature extraction Bi-cubic interpolation DWT
Abdomen
Cranium
Femur
MSE
85.46 90.812 91.828
85.082 94.745 93.787
96.483 90.067 97.007
0.271 0.294 0.109
221
J Med Syst (2016) 40:221
Page 8 of 9
with Mean Square Error slightly higher than the rest. Although the feature extraction method has not produced the best result for abdomen and Cranium but it has for the Femur, this could due to the fact the properties between abdomen and cranium are very similar in terms of shape and size as shown in Fig. 3. K-fold cross validation In this experimentation using K-fold cross validation, the dataset is divided into ten sets (i.e. k = 10). Tables 2, 3 and 4 shows the results produced by K-fold cross validation using feature extraction, bi-cubic Interpolation and discrete wavelet transform respectively. Table 5 shows that Discrete Wavelet Transform has produced the best overall results as compared to feature extraction and bi-cubic interpolation.
5.
6.
7.
8.
9.
10.
11.
12.
Conclusion In this paper, resizing images using Discrete Wavelet Transform is implemented as an alternative approach to feature extraction technique, experimental results shows that the proposed approach is effective and takes less time as compared to traditional approaches. After achieving the expected results from the Discrete Wavelet Transform we intend to use this technology on real RGB images and find a suitable / generic method to cluster them. It will involve image decomposition and representing each of the coefficients detail as Gaussian in addition to the RGB colors which will also be represented by Gaussian. At the end of the image representation we will have a mixture of Gaussian consisting of six Gaussians which will hopefully be clustered using a fuzzy clustering technique.
13.
14.
15.
16. 17.
18.
References 1.
2.
3.
4.
Gao, S., Peng, Y., Guo, H., Liu, W., Gao, T., Xu, Y., and Tang, X., Texture analysis and classification of ultrasound liver images. Biomed. Mater. Eng. 24(1):1209–1216, 2014. Wang, W., Qin, J., Zhu, L., Ni, D., Chui, Y.-P., and Heng, P.-A., Detection and measurement of fetal abdominal contour in ultrasound images via local phase information and iterative randomized Hough transform. Biomed. Mater. Eng. 24(1):1261–1267, 2014. Noh, D.K., Lee, J.J., and You, J.H., Diaphragm breathing movement measurement using ultrasound and radio-graphic imaging: A concurrent validity. Biomed. Mater. Eng. 24(1):947–952, 2014. Yang, H.S., Yoo, J.W., Lee, B.A., Choi, C.K., and You, C., KIntertester and intra-tester reliability of ultra-sound imaging measurements of abdominal muscles in adolescents with and without idiopathic scoliosis: A case-controlled study. Biomed. Mater. Eng. 24(1):453–458, 2014.
19.
20.
21.
22.
23.
Afsar Minhas, F.A., Sabih, D., and Hussain, M., Automated classification of liver disorders using ultrasound images. J. Med. Syst. 36(5):3163–3172, 2012. Işik, H., and Arslan, S., An artificial neural network classification approach for use the ultrasound in physiotherapy. J. Med. Syst. 35(6):1333–1341, 2011. Singh, M., Singh, S., and Gupta, S., An information fusion based method for liver classification using texture analysis of ultrasound images. Inf. Fusion. 19:91–96, 2014. Ma, H.-Y., Zhou, Z., Wu, S., and Wan, Y.-L., A computer-aided diagnosis scheme for detection of fatty liver in vivo based on ultrasound kurtosis imaging. J. Med. Syst. 40(1), 2016. Marée, R., Geurts, P., Piater, J., and Wehenkel, L., Biomedical image classification with random subwindows and decision trees. In Proceeding of ICCV workshop on Computer Vision for Biomedical Image Applications (CVIBA 2005). Lect. Notes Comput. Sci. 3765, 220–229. Recknagel, R.-J., Kowarschik, R., and Notni, G., High-resolution defect detection and noise reduction using wavelet methods for surface measurement. J. Opt. A Pure Appl. Opt. 2(6):538, 2000. Grgic, S., Grgic, M., and Zovko-Cihlar, B., Performance analysis of image compression using wavelets. IEEE Trans. Ind. Electron. 48(3):682–695, 2001. Ardestani, M.M., Chen, Z., Wang, L., Lian, Q., Liu, Y., He, J., Li, D., and Jin, Z., Feed forward artificial neural network to predict contact force at medial knee joint: Application to gait modification. Neurocomputing. 139(2):114–129, 2014. Srinivas, Y., Stanley Raj, A., Hudson Oliver, D., Muthuraj, D., and Chandrasekar, N., A robust behavior of Feed Forward Back propagation algorithm of Artificial Neural Networks in the application of vertical electrical sounding data inversion. Geosci. Front. 3(5): 729–736, 2012. Zhang, L., and Subbarayan, G., An evaluation of back-propagation neural networks for the optimal design of structural systems: Part I: Training procedures. Comput. Methods Appl. Mech. Eng. 191(25– 26):2873–2886, 2002. Lu, J., Zhao, J., and Cao, F., Extended feed forward neural networks with random weights for face recognition. Neurocomputing. 136(20):96–102, 2014. Noble, J.A., and Boukerroui, D., Ultrasound image segmentation: A survey. IEEE Trans. Med. Imaging. 25(8):987–1010, 2006. Hareendranathan, A.R., Mabee, M., Punithakumar, K., Noga, M., and Jaremko, J.L., Toward automated classification of acetabular shape in ultrasound for diagnosis of DDH: Contour alpha angle and the rounding index. Comput. Methods Prog. Biomed. 129:89–98, 2016. Shi, J., Zhou, S., Liu, X., Zhang, Q., Lu, M., and Wang, T., Stacked deep polynomial network based representation learning for tumor classification with small ultrasound image dataset. Neurocomputing. 194(19):87–94, 2016. Hou, X., Zhu, P., and Ma, Y., Classification of cardiac ultrasound image sequences based on sparse representation. J. Softw. 9(9): 2309–2314, 2014. Kalyan, K., Jakhia, B., Lele, R.D., Joshi, M., and Chowdhary, A., Artificial neural network application in the diagnosis of disease conditions with liver ultrasound images. Adv. Bioinforma. 2014: 708279, 2014. doi:10.1155/2014/708279.14 pages Luo, S., Li, X., and Li, J., Review on the methods of automatic liver segmentation from abdominal images. J. Comput. Commun. 2(2): 1–7, 2014. Asamwar, R.S., Bhurchandi, K.M., and Gandhi, A.S., Interpolation of images using discrete wavelet transform to simulate image resizing as in human vision. Int. J. Autom. Comput. 7(1):9–16, 2010. Deng, C., Lin, W., and Cai, J., Content-based image compression for arbitrary-resolution display devices. IEEE Trans. Multimed. 14(4):1127–1139, 2012.
J Med Syst (2016) 40:221 24.
Finsberg, H. N., Wavelet techniques in medical imaging. classification of ultrasound images using the windowed scattering transform. Available Online: http://www.diva-portal.org/smash/get/diva2 :733307/FULLTEXT01.pdf 25. Han, J.-W., Choi, K.-S., Wang, T.-S., Cheon, S.-H., and Ko, S.-J. Wavelet based seam carving for content-aware image resizing. 16th IEEE International Conference on Image Processing (ICIP), pp 345–348, Cairo, 2009. 26. Das, S., Garg, S., and Sahoo, G., Comparison of content based image retrieval systems using wavelet and curvelet transform. Int. J. Multimed. Appl. 4(4):137–154, 2012. 27. Halkias, X.C., and Maragos, P., Analysis of Kirlian images: Feature extraction and segmentation. In Proceeding of IEEE International Conference on Signal Processing, pp 765–768, Beijing, China, 2004. 28. Anoop Suraj, A., Francis, M., Kavya, T.S., and Nirmal, T.M., Discrete wavelet transform based image fusion and de-noising in FPGA. J. Electr. Syst. Inf. Technol. 1(1):72–81, 2014.
Page 9 of 9 221 29.
30.
31.
32.
33.
Etehadtavakol, M., Ng, E.Y.K., Chandran, V., and Rabbani, H., Separable and non-separable discrete wavelet transform based texture features and image classification of breast thermograms. Infrared Phys. Technol. 61:274–286, 2013. Wang, X.-Y., Yang, H.-Y., and Fu, Z.-K., A New Wavelet-based image denoising using undecimated discrete wavelet transform and least squares support vector machine. Expert Syst. Appl. 37(10): 7040–7049, 2010. Sun, Y., and Cheng, A.C., Machine learning on-a-chip: A highperformance low-power reusable neuron architecture for artificial neural networks in ECG classifications. Comput. Biol. Med. 42(7): 751–757, 2012. Botzheim, J., and Földesi, P., Novel calculation of fuzzy exponent in the sigmoid functions for fuzzy neural networks. Neurocomputing. 129(10):458–466, 2014. Nambiar, V.P., Khalil-Hani, M., Sahnoun, R., and Marsono, M.N., Hardware implementation of evolvable block-based neural networks utilizing a cost efficient sigmoid-like activation function. Neurocomputing. 140(22):228–241, 2014.
