The imaging diagnosis and prognostication of different degrees of traumatic brain injury (TBI) is very important for early care and clinical treatment. Especially, the exact recognition of mild TBI is the bottleneck for current label-free imaging technologies in neurosurgery. Here, we report an automatic evaluation method for TBI recognition with terahertz (THz) continuous-wave (CW) transmission imaging based on machine learning (ML). We propose a new feature extraction method for biological THz images combined with the transmittance distribution features in spatial domain and statistical distribution features in normalized gray histogram. Based on the extracted feature database, ML algorithms are performed for the classification of different degrees of TBI by feature selection and parameter optimization. The highest classification accuracy is up to 87.5%. The area under the curve (AUC) scores of the receiver operating characteristics (ROC) curve are all higher than 0.9, which shows this evaluation method has a good generalization ability. Furthermore, the excellent performance of the proposed system in the recognition of mild TBI is analyzed by different methodological parameters and diagnostic criteria. The system can be extensible to various diseases and will be a powerful tool in automatic biomedical diagnostics.
© 2018 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
Traumatic brain injury (TBI) is defined as the damage to brain caused by an external mechanic force exceeding its protective capacity, such as that caused by crush, falls, blast waves or penetration by a projectile. TBI is the most common disease with high mortality and disability. Based on severity, the level of TBI is graded as mild, moderate, and severe . It is estimated that, 10 million people suffer from deaths and hospitalizations directly attributed to TBI annually and 57 million people have experienced TBI worldwide, with 75% to 85% of these injuries categorized as mild . Thus, the diagnosis and prognostication of different degrees of TBI are very important for early care and clinical treatment.
The pathophysiology of TBI is complicated with diverse manifestations that results from both immediate and delayed mechanisms . The immediate primary injury triggers a cascade of molecular and biochemical events like cerebral blood flow changes, axonal shearing, and metabolic imbalance. The delayed second phase of damage may lead to neuronal and glial damage including brain edema, neuro-inflammation, and delayed neuronal death. Currently, the advanced imaging technologies for TBI mainly include diffusion tensor magnetic resonance imaging (DTMRI), positron emission tomography (PET), laser-induced photoacoustic imaging (PAI), and fluorescence molecular imaging (FMI). The DTMRI of TBI relies on differentiating diffusion of water molecules in tissues and has the ability of noninvasive, longitudinal assessment of hemodynamics. But there are still some technical issues concerning scanning protocols, analysis techniques, and different contrast agents . The 15O PET defines the distribution of oxygen extraction fraction to show the TBI-induced ischemic burden, but its statistical accuracy is poor in low blood flow regions and the modified control data is essential . The PAI system can be used for the noninvasive monitoring of TBI and post-traumatic rehabilitation, but the wave-front distortions will be induced by acoustic speed aberrations . The image quality of PAI are also limited by the instability of laser energy and the motions of animals. Intravenous injection of fluorophores is prerequisite for FMI and the detection sensitivity is limited to fluorescence quantum yield, photon scattering, and hepatic clearance . Of further note, the conventional imaging technologies cannot adequately depict brain injury in mild TBI because they are not sensitive to detect diffuse axonal injuries (DAI). Especially, mild TBI often appears quite normal on conventional computed tomography (CT) and MRI scans . Hence, the rapid and label-free imaging technology of distinguishing different degrees of TBI remains challenging in neurosurgery, especially for the exact recognition of mild TBI.
Terahertz (THz) imaging, as one of the most popular and advanced imaging technologies in recent years, has attracted great attentions in biological imaging and sensing applications due to its properties of unique physical characteristics of fingerprint spectrum, safety, and water sensitivity. To date, the research works on THz biological imaging have been extended to more than ten kinds of diseases, including various cancers and tumors [9–14], skin burns , arthritis , flap viability assessment , etc. In neurosurgery field, the THz spectroscopy and imaging detections have been studied to detect brain gliomas  and cerebral ischemia . Recently, we have demonstrated that the different degrees of TBI were distinguishable by THz imaging . In this paper we focus on exploring the feasibility of this imaging technique to discriminate among TBI tissue with different degrees.
As the spectroscopy and image information in THz frequency exploding on the way, the automatic diagnosis technology based on THz database will have a great potential in biomedical diagnostics. Most of THz imaging adopted THz time-domain spectroscopy (TDS) system and continuous-wave (CW) THz source. Because the spatial and temporal components can be recorded simultaneously, THz imaging based on TDS system can obtain rich information of biological tissues, such as absorption coefficient, refractive index, and phase, etc. Classification and recognition for these data sets have been investigated in recent years and the algorithms mainly include principal component analysis (PCA), hierarchical clustering analysis (HCA), spectroscopic integration technique, and support vector machine (SVM). The distributions of normalized integration factors have been used to analyze the features of THz images in metastatic lymph nodes detection . The PCA and HCA have been used to distinguish cancer and healthy tissues in a single THz spectroscopic image . Classification of THz imaging data from excised breast tissue has been performed by SVM . The THz images of tumor and normal breast tissue could be distinguished and the accuracy achieved 92%. However, the area under the curve (AUC) scores of the receiver operating characteristics (ROC) curve were lower than 0.8, so its generalization ability is limited. CW THz imaging systems adopt CW THz source and can only get the intensity information with the advantages of high power and high-speed data acquisition. Currently, the feature extraction and recognition methods for CW THz imaging are rarely studied. Therefore, effective feature extraction methods and classification algorithms with excellent generalization ability for various THz imaging are of great significance in the development of THz applications.
In this paper, we demonstrate an automatic evaluation method of different degrees of TBI in a rat model based on THz transmission CW imaging with machine learning (ML). A new feature extraction method is proposed for biological THz images combined with the transmittance distribution features in spatial domain and statistical distribution features in normalized gray histogram. We use ML algorithms to construct and train feature database. Then, the classification of different degrees of TBI is analyzed by feature selection and parameter optimization. Moreover, the excellent performance of the classifier is demonstrated in the mild TBI recognition by calculating the methodological parameters of precision, sensitivity, and specificity. Optimal algorithm selection is also discussed based on different diagnostic criteria. The results indicate that our proposed method can be extensible to other diseases detection with THz imaging.
2. Materials and methods
2.1 Biological terahertz imaging and automatic evaluation system
The schematic diagram of the biological THz imaging and automatic evaluation system is shown in Fig. 1, including a THz transmission imaging system and a classifier based on ML.
A THz transmission imaging system operating at 2.52 THz is designed to acquire THz images of biological tissues, which can realize rapid and label-free biological imaging. A high-power THz gas laser (FIRL100, Edinburgh Instruments Ltd.) with tunable continuous THz wave is used in the imaging system, which has a good beam quality and high signal to noise ratio (SNR). Considering the power fluctuation of the THz laser, the THz wave, modulated by a 50-Hz chopper, is separated into two sub-beams by a wire grid polarizer (Microtech Instruments, Inc.). One sub-beam is used as the intensity reference to improve the SNR of THz imaging, which is monitored by Golay cell (1#, Tydex Ltd, GC-1P). The other sub-beam is focused on the biological samples by a focusing system which consist of two Tsurupica lens (f = 30 mm). The detected signal is monitored by the other Golay cell (2#, Tydex Ltd, GC-1P). The sectioned biological sample is mounted on a computer-controlled linear motor stages (SIGMA KOKI CO., LTD.) for the raster scan imaging. The imaging speed is about 100 ms/pixel and the imaging resolution is 260 μm × 380 μm measured by knife-edge method. A visible CCD camera is fabricated to acquire the visible images of biological tissues. The THz images and corresponding visible images display on the computer simultaneously for assisting diagnostics. The intensity of acquired pixel in images indicate the THz transmittance of biological tissues. The proposed THz imaging system has a great potential in imaging of fresh biological tissues due to the high detection sensitivity and SNR.
The classifier procedure based on ML includes the preprocessing of THz images, features extractions, classifier building, and classifier assessment and optimization. Multifarious ML algorithms are used to construct and train feature database and the optimized classifier is selected for the evaluation of diseases. The proposed system is extensible to THz images of various diseases. In this work, the automatic evaluation of different degrees of TBI is demonstrated.
2.2 Sample preparation and measurement protocol
A reproducible measurement protocol is established by standardizing all experimental steps of sample preparations and THz imaging. Simultaneously, all animal experiments are performed in accordance with the China Animal Welfare Legislation and are approved by the Third Military Medical University Committee on Ethics in the Care and Use of Laboratory Animals.
The adult male Sprague-Dawley rats are subjected to TBI, which weight about 265 g and are purchased from the Animal Center of the Third Military Medical University. The animal TBI model is established by a modified version of Feeney's method . The head of the anesthetized rat is mounted on a stereotaxic frame. With the rat disinfected and fur shaved, the scalp is opened by a midline incision. A right parietal craniotomy is performed by a dental drill. A 30-g stainless steel rod (4.5 mm in diameter, 5 mm in height) is free fell on the exposed dura from the height of 15 cm, 25 cm, and 35 cm to establish mild, moderate, and severe brain trauma, respectively. As the comparison group, the same surgical procedures are also performed on a sham operation group. Brain damage can be visualized by staining brain slices with 2% 2,3,5-triphenyltetrazolium chloride (TTC) and the degree of TBI is confirmed . Typical TBI models and the corresponding TTC-stained sections are shown in Fig. 2. The sectioned position lies in the middle of trauma. The result shows that, the more severe injury is induced, the larger pale area emerges in the stained brain tissues, which have been marked by the black dotted line. During the preparations of sliced samples for THz imaging, the fresh brain tissues with TBI are cut into 40μm thickness by a microtome (Leica, CM1950) and then sandwiched between two 500-μm quartz slides. In order to decrease sample dehydration, the oleic acid is used to cover the tissues. This method we have reported in . In our experiment, there are total 64 rats with TBI and 16 rats in the sham operation group. The numbers of mild, moderate, and severe TBI samples are 18, 27, and 19, respectively. When the scanning step of imaging system is set as 250 µm in the experiment, the THz images of 80 samples are acquired as shown in Fig. 3, where each THz image is from 1 rat. Here, the THz image is normalized using the reference of the transmitted THz intensity of two quartz substrates. The THz images are displayed in 5-color images and different colors indicate the transmittance of THz wave.
2.3 Preprocessing of terahertz images
As shown in Fig. 4, the preprocessing of THz images consists of rotation rectification, the selection of the region of interest (ROI), and the standardization of ROI for feature extraction of ML.
Because the feature extraction method we used in ML is based on the differences of distribution in spatial domain (range and position), the rotation rectification of THz images in preprocessing is imperative. Obtained from the THz image database, the THz transmittance of sectioned sham and TBI tissues is less than 0.5. To recognize the area of brain, the acquired greyed image is binarized and the threshold is set as 0.8 in our experiment. The excircle of the area of brain in the THz image is extracted to calculate the angle of rotation rectification. Then, the greyed image after rotation rectification is binarized and the brain area is separated from the background as the ROI (the white area). Finally, the ROI is resized in proportion to make the ROI with the same length, which can ensure the same length of feature vector for ML.
2.4 Features extraction
A new feature extraction method of biological THz images is proposed including transmittance distribution features in spatial domain and statistical distribution features in normalized gray histogram. All features are calculated in the ROI of images that have been preprocessed above.
Four images are selected to demonstrate the transmittance distribution features of THz images in spatial domain. Figure 5(a) is the THz image of sham operation group and Fig. 5(b)–5(d) show the THz images of typical mild, moderate, and severe TBI tissues selected from the THz image database. The images show that the area with low transmittance appears at the injury position with different distribution range but does not emerge in the sham operation group. It could be explained by the different degrees of brain edema induced by TBI where the water has high absorption of THz wave. In order to extract such transmittance distribution features in spatial domain (range and position) of the THz images, the features extraction method we used is shown in Fig. 5(e). The minimum, mean, and maximum transmittance values of each column (di,min, di,mean, and di,max, i = 1,2,3…80) are calculated, which characterize the transmittance distribution features of each column. These values constitute the first 240 values of the feature vector F. Besides that, the minimum, mean, and maximum transmittance values of the whole ROI are also included in the feature vector, as the 241th to 243th values of the feature vector F.
Observed from Fig. 5(a)–5(d), we also find that, the size and the intensity of the area with low THz transmittance are different in the THz images of different degrees of TBI tissues. In order to extract such regional statistical features, the gray histograms of the ROI of Fig. 5(a)–5(d) are obtained and normalized as shown in Fig. 5(f). Then, they are normalized by the maximum image pixel numbers in gray histograms for enhancing the relative differences of distribution. The normalized gray histogram indicates the relative proportion of pixel numbers with different THz transmittance in ROI and it can characterize the transmittance features of THz images in statistics. It is seen from the normalized gray histograms that the THz images of different degrees of TBI tissues have significant differences in distribution range and relative intensity. Thus, the statistical distribution features in normalized gray histogram are extracted as shown in Fig. 5(f), where hi (i = 0,1,2…255) correspond to the relative proportion of pixel numbers with the different THz transmittance.
The transmittance distribution features in spatial domain and statistical distribution features in gray histogram are combined as the final feature vector F with the dimension of 499.
2.5 Classifier building
Based on the features extracted from the THz images, we construct classifiers to train and predict the degrees of TBI. In order to improve the predictive power of the classifiers, we follow the procedures of feature selection, parameter optimization, and performance assessment for the ML algorithm.
Feature selection based on ReliefF algorithm  is performed for feature database of multi-label ML. ReliefF is a general and successful attribute estimator, which can detect conditional dependencies between attributes and provides a unified view on the attribute estimation in regression and classification . In the proposed system, the features are ranked by weight captured from the ReliefF algorithm and different numbers of features are included for ML training. In order to avoid parameters overfitting of classifiers, a part of features are used in classifier building and the leave-one-out cross validation (LOOCV)  are performed by one sample left out each time in training. Then, the performance of the classifiers are analyzed. The classifiers include k-nearest neighbor (kNN), support vector machine (SVM), and random forest (RF). In kNN classification, an object is classified by a majority vote of its neighbors, with the object being assigned to the class most common among its k nearest neighbors. SVM constructs a hyperplane or set of hyperplanes in a high- or infinite-dimensional space for classification, making the separation easier in that space. RF classification is an ensemble learning method that operate by constructing a multitude of decision trees at training time and outputting the class by majority vote. The classifiers with different parameters are used to obtain the total accuracy with different numbers of features included. The best subset of features and optimal algorithm parameters are selected for the highest accuracy of the classifiers. Furthermore, the ROC curve and AUC can demonstrate the generalization ability of ML systems, so we adopt the ROC curve and AUC scores for different degrees of TBI to assess the classification performance. The ROC curve indicates how much true positive rate (TPR) are recognized conditioned on a given false positive rate (FPR) and AUC score indicates the area of the zone under ROC curve.
3. Results and discussions
As shown in Fig. 6(a), the weights of different serial number of features, with large assigned to important attributes, are captured by ReliefF algorithm. The red area corresponds to the transmittance distribution features in spatial domain and the green area corresponds to the statistical distribution features in normalized gray histogram. According to the weight distribution, the overall importance of attributes of statistical distribution features in normalized gray histogram is higher than that of transmittance distribution features in spatial domain. Then the features are ranked by the weights and different numbers of features are used for the training of kNN, SVM, and RF. At this stage, the kernel function of SVM uses linear, quadratic, and polynomial kernel, respectively. The number of neighbors (k) of kNN increases from 4 to 15. The aggregating ensemble method of RF is performed by Bagging . The number of decision trees in RF is set as 25, 50, 100, 200, 300, 400, and 500, respectively. The number of features in the training of each decision tree is set as the square root of the number of features selected from ReliefF. As the number of features increases from 1 to 499, the total classification accuracies are obtained by LOOCV. The results are shown in Fig. 6(b)–6(d). With more features included, the accuracy of kNN increases and the highest accuracy reaches 86.25% with k equal to12 and all features used. Whereas, the accuracy of SVM increases first and then decreases, and the highest accuracy reaches 83.75% when using linear kernel and 141 features. The highest accuracy of RF is 87.5% when using 50 decision trees and 161 features. Note that the accuracy of the proposed automatic evaluation method is always higher than 81.25% for the classification of TBI degree through optimized features selection. RF has the highest classification accuracy.
In order to further demonstrate the performance of the classifiers when they reach the highest accuracy, the ROC curve and AUC scores are analyzed for different degrees of TBI respectively, as shown in Fig. 7.When different degrees of TBI classified by different classifiers, the AUC scores are all higher than 0.9, which indicates the proposed system has a good classification performance and excellent generalization ability. For the sham, mild, and severe group, the recognition ability of SVM and RF significantly outperform kNN. For the moderate group, the three algorithms have a close performance. It is worthwhile to mention that, the SVM and RF both have excellent recognition abilities of sham and severe group whose AUC scores are all higher than 0.98. On the whole, the AUC scores of RF for different degrees of TBI are all higher than the other two algorithms. Thus, considering the total accuracy and recognition ability of different degrees of TBI, RF is the best algorithm.
The exact recognition of mild TBI is the bottleneck for current biological imaging technologies in neurosurgery. Here, we analyze the performance of the proposed system in recognition of mild TBI. The methodological parameters of precision, sensitivity, and specificity for the mild TBI recognition has been calculated when the three algorithms reach the highest accuracy. Especially, it is worthwhile mentioned that these parameters can provide the significant reference for the algorithm selection based on the medical diagnostic criteria. Generally, the improvement of specificity and sensitivity can contribute to the reducing the misdiagnosis rate and missed diagnosis rate, respectively. As shown in Fig. 8, kNN has a higher precision and specificity of 0.875 and 0.968, respectively. RF has a higher sensitivity of 0.889. SVM has the moderate sensitivity and specificity, but its precision is low. Therefore, for the mild TBI recognition, kNN has the higher accuracy rate and lower misdiagnosis rate, and RF has the lower missed diagnosis rate.
In this work, 80 samples are classified by the proposed evaluation method and the severity of TBI has been well recognized. Further studies may be necessary with more biosamples and different kinds of diseases. And in vivo imaging of TBI after craniotomy will be performed during surgery based on THz reflection imaging system due to its low penetration of fresh tissue. The improvement of image resolution is also urgent for the precise localization of TBI region, which is helpful for the debridement. However, we believe that THz image could become a useful supplementary means for the clinical application, and the proposed method for automated detection using ML algorithms with appropriate parameters is general and powerful in THz biological imaging applications.
In summary, a biological THz imaging and automatic evaluation method has been demonstrated. THz images of different degrees of TBI with a rat model are acquired using THz transmission CW imaging. A new feature extraction method for biological THz images is proposed combined with the transmittance distribution features in spatial domain and statistical distribution features in normalized gray histogram. The classifiers, including kNN, SVM, and RF, are used to the classification of different degrees of TBI by feature selection and parameter optimization. The result shows that the AUC scores of the ROC curve are all higher than 0.9. Considering the total classification accuracy and recognition ability of different degrees of TBI, RF is the best algorithm and the highest accuracy is 87.5%. Furthermore, optimal algorithm selection is discussed for the recognition of mild TBI by different methodological parameters and diagnostic criteria. The proposed method can be extensible to various diseases and will be a powerful tool in automatic biomedical diagnostics.
The National Basic Research Program of China (973) (2015CB755403, 2014CB339802); National Key Research and Development projects (2016YFC0101001); National Natural Science Foundation of China (NSFC) (61775160, 61771332, 61471257); China Postdoctoral Science Foundation (2016M602954); Postdoctoral Science Foundation of Chongqing (Xm2016021); Joint Incubation Project of Southwest Hospital (SWH2016LHJC04, SWH2016LHJC01).
References and links
2. C. J. Murray and A. D. Lopez, “Global health statistics: a compendium of incidence, prevalence and mortality estimates for over 200 conditions,” Boston: Harvard School of Public Health, 1996.
4. A. Wunder, K. Schoknecht, D. B. Stanimirovic, O. Prager, and Y. Chassidim, “Imaging blood-brain barrier dysfunction in animal disease models,” Epilepsia 53(Suppl 6), 14–21 (2012). [CrossRef] [PubMed]
5. J. P. Coles, T. D. Fryer, P. Smielewski, K. Rice, J. C. Clark, J. D. Pickard, and D. K. Menon, “Defining ischemic burden after traumatic brain injury using 15O PET imaging of cerebral physiology,” J. Cereb. Blood Flow Metab. 24(2), 191–201 (2004). [CrossRef] [PubMed]
6. S. Yang, D. Xing, Y. Lao, D. Yang, L. Zeng, L. Xiang, and W. R. Chen, “Noninvasive monitoring of traumatic brain injury and post-traumatic rehabilitation with laser-induced photoacoustic imaging,” Appl. Phys. Lett. 90(24), 243902 (2007). [CrossRef]
7. X.-D. Zhang, H. Wang, A. L. Antaris, L. Li, S. Diao, R. Ma, A. Nguyen, G. Hong, Z. Ma, J. Wang, S. Zhu, J. M. Castellano, T. Wyss-Coray, Y. Liang, J. Luo, and H. Dai, “Traumatic brain injury imaging in the second near-infrared window with a molecular fluorophore,” Adv. Mater. 28(32), 6872–6879 (2016). [CrossRef] [PubMed]
8. M. E. Shenton, H. M. Hamoda, J. S. Schneiderman, S. Bouix, O. Pasternak, Y. Rathi, M.-A. Vu, M. P. Purohit, K. Helmer, I. Koerte, A. P. Lin, C.-F. Westin, R. Kikinis, M. Kubicki, R. A. Stern, and R. Zafonte, “A review of magnetic resonance imaging and diffusion tensor imaging findings in mild traumatic brain injury,” Brain Imaging Behav. 6(2), 137–192 (2012). [CrossRef] [PubMed]
10. H. Chen, T.-H. Chen, T.-F. Tseng, J.-T. Lu, C.-C. Kuo, S.-C. Fu, W.-J. Lee, Y. Fu Tsai, Y. Y. Huang, E. Y. Chuang, Y. J. Hwang, and C. K. Sun, “High-sensitivity in vivo THz transmission imaging of early human breast cancer in a subcutaneous xenograft mouse model,” Opt. Express 19, 21552–21562 (2011). [CrossRef] [PubMed]
11. J. Y. Park, H. J. Choi, K.-S. Cho, K.-R. Kim, and J.-H. Son, “Terahertz spectroscopic imaging of a rabbit VX2 hepatoma model,” Appl. Phys. Lett. 109, 064704 (2011).
13. S. J. Oh, S.-H. Kim, Y. B. Ji, K. Jeong, Y. Park, J. Yang, D. W. Park, S. K. Noh, S.-G. Kang, Y.-M. Huh, J.-H. Son, and J.-S. Suh, “Study of freshly excised brain tissues using terahertz imaging,” Biomed. Opt. Express 5(8), 2837–2842 (2014). [CrossRef] [PubMed]
14. J. Y. Park, H. J. Choi, H. Cheon, S. W. Cho, S. Lee, and J.-H. Son, “Terahertz imaging of metastatic lymph nodes using spectroscopic integration technique,” Biomed. Opt. Express 8(2), 1122–1129 (2017). [CrossRef] [PubMed]
15. Z. D. Taylor, R. S. Singh, M. O. Culjat, J. Y. Suen, W. S. Grundfest, H. Lee, and E. R. Brown, “Reflective terahertz imaging of porcine skin burns,” Opt. Lett. 33(11), 1258–1260 (2008). [CrossRef] [PubMed]
17. N. Bajwa, J. Au, R. Jarrahy, S. Sung, M. C. Fishbein, D. Riopelle, D. B. Ennis, T. Aghaloo, M. A. St John, W. S. Grundfest, and Z. D. Taylor, “Non-invasive terahertz imaging of tissue water content for flap viability assessment,” Biomed. Opt. Express 8(1), 460–474 (2016). [CrossRef] [PubMed]
18. J. Shi, Y. Wang, D. Xu, C. Yan, L. Tang, P. Duan, Y. He, H. Liu, T. Chen, H. Feng, and J. Yao, “Biomedical diagnosis of cerebral ischemia with continuous-wave THz imaging,” in41st International Conference on Infrared, Millimeter, and Terahertz Waves (IRMMW-THz) (2016). [CrossRef]
19. H. Zhao, Y. Wang, J. Shi, L. Chen, K. Ma, L. Tang, T. Chen, D. Xu, J. Yao, and H. Feng, “High-sensitivity terahertz imaging of traumatic brain injury in a rat model,” in The 17th International Conference on Brain Edema and Cellular Injury (2017).
20. S. Nakajima, H. Hoshina, M. Yamashita, C. Otani, and N. Miyoshi, “Terahertz imaging diagnostics of cancer tissues with a chemometrics technique,” Appl. Phys. Lett. 90(4), 041102 (2007). [CrossRef]
21. A. J. Fitzgerald, S. Pinder, A. D. Purushotham, P. O’Kelly, P. C. Ashworth, and V. P. Wallace, “Classification of terahertz-pulsed imaging data from excised breast tissue,” J. Biomed. Opt. 17(1), 016005 (2012). [CrossRef] [PubMed]
22. D. M. Feeney, M. G. Boyeson, R. T. Linn, H. M. Murray, and W. G. Dail, “Responses to cortical injury: I. Methodology and local effects of contusions in the rat,” Brain Res. 211(1), 67–77 (1981). [CrossRef] [PubMed]
23. S. Yu, Y. Kaneko, E. Bae, C. E. Stahl, Y. Wang, H. van Loveren, P. R. Sanberg, and C. V. Borlongan, “Severity of controlled cortical impact traumatic brain injury in rats and mice dictates degree of behavioral deficits,” Brain Res. 1287, 157–163 (2009). [CrossRef] [PubMed]
24. Y. Wang, H. Minamide, M. Tang, T. Notake, H. Ito, and H. Ito, “Study of water concentration measurement in thin tissues with terahertz-wave parametric source,” Opt. Express 18(15), 15504–15512 (2010). [CrossRef] [PubMed]
25. I. Kononenko, “Estimating attributes: Analysis and extensions of RELIEF,” in European Conference on Machine Learning on Machine Learning(1994), pp. 171–182. [CrossRef]
26. M. Robnik-Šikonja and I. Kononenko, “Theoretical and empirical analysis of ReliefF and RReliefF,” Mach. Learn. 53(1/2), 23–69 (2003). [CrossRef]
27. R. Kohavi, “A study of cross-validation and bootstrap for accuracy estimation and model selection,” in International Joint Conference on Artificial Intelligence (Morgan Kaufmann Publishers Inc., 1995), 1137–1143.
28. L. Breiman, “A Bagging predictors,” Mach. Learn. 24(2), 123–140 (1996). [CrossRef]