Primary study of identification of parathyroid gland based on laser-induced breakdown spectroscopy

: The identification and preservation of parathyroid glands (PGs) is a major issue in thyroidectomy. The PG is particularly difficult to distinguish from the surrounding tissues. Accidental damage or removal of the PG may result in temporary or permanent postoperative hypoparathyroidism and hypocalcemia. In this study, a novel method for identification of the PG was proposed based on laser-induced breakdown spectroscopy (LIBS) for the first time. LIBS spectra were collected from the smear samples of PG and non-parathyroid gland (NPG) tissues (thyroid and neck lymph node) of rabbits. The emission lines (related to K, Na, Ca, N, O, CN, C 2 , etc.) observed in LIBS spectra were ranked and selected based on the important weight calculated by random forest (RF). Three machine learning algorithms were used as classifiers to distinguish PGs from NPGs. The artificial neural network classifier provided the best classification performance. The results demonstrated that LIBS can be adopted to discriminate between smear samples of PG and NPG, and it has a potential in intra-operative identification of PGs.


Introduction
Thyroid cancer is one of the most common endocrine tumors. Major treatments for thyroid cancer are thyroidectomy and central lymph node dissection. The diseased glands and surrounding lymph nodes are generally resected while leaving the normal tissues intact [1][2][3][4]. One of the main challenges during thyroidectomy is the accidental damage or removal of the parathyroid gland (PG), which accounts for the serious complications, including temporary or permanent postoperative hypoparathyroidism and hypocalcemia [4][5][6]. According to statistics, the incidence of inadvertent removal of PGs during thyroidectomy ranges from 8% to 19%, temporary postoperative hypoparathyroidism observed in 20% to 35% of cases after thyroidectomy, and the incidence of permanent hyperparathyroidism ranges 1% to 5% [7][8][9]. Thus, intra-operative identification and preservation of PG are particularly important with the prevention of complications after surgery. PG can be difficult to visually distinguish from the surrounding tissues due to the facts that its size is too small and its appearance is often similar to the lymph node, fat, and occasionally thyroid nodule [4].
The generally identification of PG based on the visual inspection is subjective and the results depend greatly on the experience of surgeons [3]. Other methods based on biopsy, such as histopathological analysis of frozen section, may result in devascularization and destruction of the PG [10,11]. In addition, the nanocarbon negative development technique, which can stain the thyroid and lymph nodes in black [12], is relatively costly, and the injection of nanocarbon may contaminate the surgical area. It is therefore urgent to develop a reliable, real-time, rapid, and low-invasive intra-operative identification methods of PG.
Several new optical technologies have been proposed in clinical studies, they may become complementary to visual inspection and improve the identification of PG during thyroidectomy [7]. Near-infrared autofluorescence (NIR-AF) spectroscopy/imaging is the most widely reported technology for identification of PG. Extensive in vivo studies have confirmed the feasibility of intra-operative NIR-AF, and the reported accuracy are encouraging [13,14]. However, the acquisition of AF signal needs to meet some requirements: the surface of the PG should be exposed to the camera through surgery, and the lights in operating room need to be repeatedly turned off. Tummers et al. [15] and Takeushi et al. [16] have demonstrated that it is possible to identify normal PG in thyroidectomy by exogenous fluorescence with methylene blue (MB) and 5-aminolevulinic acid (5-ALA), respectively. But this technique requires injection of contrast agent or oral medication, the detection effect will be affected with administration time increasing. Meanwhile, the safety needs to be further evaluated because the MB has certain neurotoxicity and the metabolites of 5-ALA are phototoxic.
Optical coherence tomography (OCT) is considered to provide microscopic images of PG without contrast agent. Sommerey et al. [17] analyzed ex vivo OCT images of PG and surrounding tissues, achieving the correct classification rate of 96.15% by measuring the backscattering intensity of OCT images. They also attempted to perform OCT studies in vivo but sensitivity and specificity were only 69% and 66%, respectively [18]. The discrimination between lymph node and PG was poor. The unsatisfactory imaging effect of probe may hinder the clinical application of OCT.
Laser-induced breakdown spectroscopy (LIBS) is a promising element analysis technique and has an increasingly wide utilization in biomedical field [19,20]. LIBS has the advantages of rapid detection, real-time analysis and low-invasion, and it is available for human and animal tissue without sample pretreatment and injection of contrast agent. These features make LIBS suitable for the intra-operative detection of biological tissue samples. LIBS can characterize the elemental composition of biological tissues. The emission intensities of elements of several normal and malignant human tissue were measured with LIBS in the study from Ghasemi et al. [21]. The differences in element concentrations between normal and malignant tissues can be reflected in LIBS spectra, which is foundation of pathologic diagnosis based on LIBS. Ex vivo researches on the differentiation of human/animal malignant tissues by use of LIBS have been concerned. Combined with machine learning methods, LIBS has successfully been applied for the discrimination of cutaneous melanoma tissues [22], liver cancer tissues [23], cervical cancer tissues [24], infiltrative glioma boundary [25], etc. Additionally, LIBS has also been used to distinguish the several normal tissues with similar composition. The group of Rajesh has confirmed that the potential of LIBS to identify some normal tissues from pigs (such as fat, nerve, muscle, etc.) to avoid the accidental removal in laser surgery [26][27][28]. The normal nerve and fat tissues with highly similar elemental composition were successfully differentiated based on LIBS in the study from Mehari et al. [27]. PG, lymph node and thyroid tissue contains the stable and similar elemental composition (except for iodine), but the concentrations of elements are different. In particular, iodine is only present in the thyroid tissue rather than PG or lymph nodes. These provided a theoretical basis for the identification of parathyroid based on LIBS.
In this work, we proposed to introduce the LIBS as a potential tool for the identification of PG for the first time. Considering the small size of PG and lymph node tissues, as well as the comparison of histopathological results, the tissue smears of rabbits were prepared as experimental objects. The ability of LIBS to represent specific elemental composition of PG and the surrounding tissue was evaluated based on collected spectra. We discriminated between PG and NPG by means of LIBS combined with feature selection and three kinds of typical classifiers. The aim of this study was to validate the feasibility of identification of PG through LIBS combined with machine learning algorithm.

LIBS experimental setup
The schematic of LIBS experimental setup in this study is illustrated in Fig. 1. A flash-pumped Q-switched Nd: YAG laser with 1064 nm wavelength, operating at the pulse frequency of 1 Hz, pulse duration of 5 ns, was used to excite the plasma. It has the beam diameter of Ø6 mm and pulse energy of 40 mJ. The laser beam was propagated through three plane mirrors and focused on the surface of samples by a convex lens (with a focal length of 100 mm). A He-Ne laser with 632.8 nm wavelength was used to point out the position of the focus. The samples were fixed on a three-dimensional motorized stage. Through a lens with a focal length of 36 mm, the plasma emission was collected into a fiber with Ø 600 µm diameter connected to a two-channel spectrometer (AvaSpec 2048-2-USB2, Avantes, wavelength range of 190-1100 nm, resolution range of 0.2-0.3 nm) equipped with an CCD camera. A spectral acquisition delay time of 1.28 µs was used to reduce Bremsstrahlung radiation. A photodetector was used to detect the laser pulse signal and a digital delayer (SRS-DG535, Stanford Research System) was used to trigger spectrometer after a preset delay time. The spectrometer was operated at the integration time of 2 ms. The Avasoft 7.0 software was used to control the spectrometer for the LIBS spectra collection.

Sample preparation and measurement
Three types of tissues including PG and surrounding lymph nodes and thyroid tissues were taken from three rabbits. Once removed, one fresh tissue was immediately used to make a tissue smear. The tissue smears were prepared by gently smearing the mucus of excised tissue onto the glass slides [29]. The animal experiments were conducted in Beijing Medical Discovery Leader laboratory. All prepared tissues smears were stored in a refrigerator and measured within 24 hours. Figure 2 shows the preparation and measurement procedure of smears sample.
The thyroid tissues were obtained directly because they can be visually recognized. PG and lymph nodes were identified by experimenters according to the prior experience, which may be inconclusive. Thus, excised PG and lymph node tissues were fixed, embedded, sectioned and stained with hematoxylin and eosin (H&E). The types of tissues were verified through pathological examination. Examples of the pathology pictures of PG and lymph nodes are showed in Fig. 3. PG, lymph nodes and thyroid in rabbits are similar to those in human body. PG consists of parenchymal cells (chief cells and eosinophils), fat cells, and fibrovascular stroma, and parathyroid hormone consisting of eighty-four amino acids is secreted by PG to regulate the calcium metabolism. As human organs, these tissues have relatively consistent elemental composition, including basic elements (C, H, O, and N) and trace elements (Na, K, Ca, Mg, etc.). However, PG, lymph nodes and thyroid play different roles in physiological regulation, which may lead to the differences in concentration of elements, especially the trace elements. All animal experiments procedures were approved by the medical and animal experiments committee of Beijing Institute of technology. For convenience, the lymph nodes and thyroid tissues were called the non-parathyroid (NPG) in this study. LIBS spectra were collected from different random positions on each of the smear samples.

Data preprocessing and spectra selection
All spectra were visually inspected and significantly outlier spectra with low signal-to-background ratio were removed. The number of spectra available for each smear sample ranged from 65 to 120 spectra. A total of 1525 original spectra (773 PG spectra and 752 NPG spectra) from 20 smear samples of three rabbits were used. The changes in properties of laser-matter interaction had a great influence on LIBS [30]. There were significant intensity fluctuations in obtained LIBS spectra due to the heterogeneity of biological samples, the nonuniformity of smear samples and the shot-to-shot variations of laser pulse, which may affect the classification results. It is therefore necessary to do data preprocessing prior to further analysis.
Firstly, each spectrum was normalized according to Root mean square (RMS) value. The full range of the spectrum divided by RMS intensity defined as the square root of the arithmetic mean of the squares of the values. The RMS normalization may maintain the raw spectra pattern and reduce the spectrum fluctuation effectively [31].
Some obtained spectra should be regarded as outliers because they are quite different from the other spectra in datasets. A filtering method based on principal component analysis (PCA) and mahalanobis distance was used to select spectra in order to improve the performance of classification. PCA considers linear relationship within the data and thus it is sensitive to outliers [30]. Gaussian distribution of data was assumed and the mahalanobis distance of one spectrum to the average spectrum in principal components space was regarded as metric. Spectra data were analyzed using PCA and the first several PCs with cumulative variance over 98% were used to evaluate the mahalanobis matrices between all spectra and the average spectrum. The median of all mahalanobis distances was selected as the threshold. The spectra with nearer distance than the threshold were considered to be similar to the average spectra in the principal component space. These spectra with better correlation and consistency were utilized for subsequent analysis. Every individual smear sample dataset was filtered and the visualization result based on a PG sample dataset is illustrated in Fig. 4. The red solid circles indicate the spectra selected by this method. After filtering, three spectra were average for the final analysis.  Table 1 list the number of smear samples, and the number of spectra before and after filtering, as well as after averaging. Finally, a total of 252 spectra were obtained and used in subsequent analysis described in this work. The most significant emission lines in spectra were selected. These emission lines were ranked based on Random Forests (RF) algorithm and the prominent feature lines were utilized for classification. These three sets of spectra collected from three rabbits, named R1, R2 and R3, were divided into two datasets in such a way that each time, the spectra collected from two rabbits were regarded as a training dataset to establish the classification model, the spectra from the third rabbit were used as a testing dataset. This procedure was repeated for three times until the spectra data from each of rabbit had been used as the training dataset and testing dataset. The flowchart of the data analysis procedure is shown in Fig. 5.

Classification methods
Several machine learning algorithms have been applied to the classification and identification of LIBS spectra collected from biological tissue, including linear discriminant analysis (LDA) [22,27,28], k-nearest neighbor (k-NN) [25,32], support vector machines (SVM) [24,25,32], partial least-squares discriminant analysis (PLS-DA) [33], and artificial neural network (ANN) [23,33], etc. Han et al. [22] successfully distinguished between the melanoma and dermis tissue by LDA. This algorithm was also widely used to achieve the classification of various normal tissues of pigs by Mehari et al. [27,28] and had demonstrated excellent classification ability. k-NN and SVM were applied to the discrimination of glioma and infiltrating boundary tissues [25], as well as the classification of five soft pork tissues [32]. The best performance were provided by SVM in these two studies. SVM was also used for the identification of the cervical cancer tissue [24], and had shown pretty good classification. Additionally, PLS-DA and ANN were used to classify several tissues from chicken organs by Yueh et al. [33] and the classification results of the latter were superior to those of former. ANN with scaled conjugate gradient back-propagation also successfully classified the liver cancer samples [23]. LDA, SVM and ANN, these typical supervised learning algorithms, have shown excellent performance for the classification of biological soft tissue. They were utilized to establish the model for identification of PG in this study (Fig. 5). The details of these algorithms have been presented in other publications, and only a brief description is given in this section.
LDA is flowing from fisher's linear discriminant and maximizes the ratio of between-class variance to the within-class variance resulting in maximal separability. It rapidly and simply handles problems with two or more classes. In this work, LDA was performed using the statistical tool box in Matlab R2019b (Mathworks, USA).
SVM is considered to be suitable for LIBS data classification because of its capability to represent non-linear features. The basic idea of SVM is to separate class with a hyperplane by maximizing the margin between them. SVM models were built in Matlab utilizing LIBSVM toolbox 3.24. The radial basis function (RBF) was chosen as kernel function. Penalty parameter C and kernel function parameter g were two important parameters for the RBF kernel SVM and they were optimized by ten-fold cross validation and a grid search based on training dataset. The optimal parameter combination leading to the best cross validation result was used to build the classification model. ANN shows extremely excellent ability of self-learning and can find hidden relations among various features [34]. The fundamental structure of ANN is a three-layer network, which includes input layer (the number of neurons equals to the number of input variables), hidden layer (the number of neurons should be set according to the specific situation), and output layer (the outputs of neuron represent prediction classes). Here, back propagation gradient descent algorithm was used to construct ANN model by use of pattern recognition neural network toolbox in Matlab. It is necessary to emphasize that all classification models are established based training datasets by ten-fold cross validation to prevent overfitting.

LIBS spectra and elemental analysis
The normalized LIBS spectra collected from PG, lymph node and thyroid smear samples and empty slide are shown in Fig. 6. In the average spectrum of empty slide, the significant emission lines of Si, Ca, Na, N, H, and O can be seen. The spectra from smear samples are an average of the spectra from three rabbits. When observing the spectra with the naked eye, the difference between the PG and NPG spectra is very subtle. The major elements common to three tissues were attributable to the K, Na, Ca, Mg, H, O, N, and Fe. Molecular bands from CN and C 2 were also observed in spectra. The spectra of smear samples are quite different from that of the empty slide. All significant emission lines (forty-seven emission lines) related above eight elements and two molecular bands in the spectra of PG and NPG were selected (Table 2) based on the maximum value in a certain spectral range. The National Institute of Standards and Technology (NIST) atomic emission database was utilized for determining the elements responsible for these emission lines. K, Na, Ca and Mg are all the important microelements in PG, lymph node and thyroid, they play the key role in controlling and regulating proper biological function. Significant Na, K and Ca emission lines can be seen in LIBS spectra. Although the intensity was weak, Fe emission line was also found in the spectra. Fe is an indispensable trace element and has been proven to exist in PG and its surrounding tissues by Lahav et al. [35] They analyzed the trace-element in various bovine neck tissues (i.e. muscle, thyroid, parathyroid, lymph node, thymus and salivary gland) using X-ray fluorescence (XRF). The existence of Fe and K was confirmed based on XRF spectra. In addition, iodine is a micronutrient that is essential for production of thyroid hormones [35], and mainly found in the thyroid. Unfortunately, significant iodine emission lines fail to be observed in the LIBS spectra obtained from thyroid, which may be due to a fact that the plasma emission of halogen element is difficult to be excited under the current pulse energy.
N, H and O emission lines were confidently attributed to endogenous and exogenous elements. C, H, O and N are the prominent elements in biological cell. The ablation of air may affect these emission lines because the LIBS measurement is exposed to air. It was noteworthy that the carbon elements can be observable from the molecular emission bands of CN and C 2 instead of pure carbon emission lines. CN bands may originate from the ablation of the C-N bond of amino acids structures in cells, or from the recombination between C components excited from the sample and N 2 in the air [25,26]. The spectra of PG and NPG smears illustrated a high degree of conformity in emission lines, but there were slight differences between intensities of emission lines for different tissues. In order to realize the discrimination between PG and NPG, feature selection and classification analysis were performed on the obtained LIBS spectra.

Feature selection based on RF
LIBS spectra contain abundant information but a lot of them may be irrelevant to classification. Previous studies have confirmed that it is necessary to select spectral lines highly correlated to the classification, which may facilitate the data analysis and improve the accuracy of classification [36][37][38]. In this work, the elements of the emission lines observed in the LIBS spectra were determined based on the average spectrum of each tissue and the most significant forty-seven emission lines were visually selected (Section 3.1). RF algorithm was further used to rank the selected feature emission lines.
The importance weight of selected emission line in the spectra was measured based on RF algorithm in order to evaluate the contribution of this line to the classification result. The average Gini index decreasing was defined as the important weight of emission line, and then the emission lines were ranked based on their own important weight values. The details of this method have discussed in previous study [25,36].
As an example of the RF analysis, the normalized important weights of forty-seven feature lines obtained based on the training dataset from R1 and R3 are illustrated in Fig. 7 in the wavelength order. The dominant feature lines are K at 769.71nm, 766.29nm and Na at 589.51nm in Fig. 7. They all have the high importance weights when ranking the feature lines based on the other two training datasets. The weights of molecular bands C 2 (517.89nm), CN (382.86, 383.42nm) are also in the forefront. The feature lines associated with O (777.32nm), N (746.75nm), and Ca (612.07nm, 443.14nm) also have higher prioritization rank. The feature lines related to K, Na, CN, C 2 , Ca, O and N are in the top ten. Although they were not completely consistent due to the individual differences of rabbits, ranking of feature lines showed a similar trend in three sets of training datasets. In order to optimize and determine the number of characteristic spectral lines, one to twenty-five feature lines were used as the inputs of subsequent classifiers.

LDA analysis
In LDA analysis, the feature lines selected based on RF methods were directly used as inputs of LDA classifier. The number of feature lines will affect the accuracy of classification model. Figure 8 shows the relationship between classification accuracy of RF-LDA model for testing datasets of R1, R2 and R3 and the number of feature lines. When the same number of feature lines were used as inputs, RF-LDA method showed the best performance for testing dataset R2. The accuracies of model for testing dataset R2 were higher than those for testing dataset R1 and R3. Using the six feature lines as inputs of the classifier, the RF-LDA method provided the highest average accuracy, which was only 77.9%.
The optimal classification results represented by the confusion matrix are shown in Fig. 9. Sensitivity and specificity of LDA were calculated from the confusion matrix. For testing dataset R2, RF-LDA achieved the highest classification accuracy at 97.5% with nine feature lines. Figure 9(b) shows the optimal discrimination results of the PG and NPG of R2. Almost all PG and NPG were completely differentiated. Only one PG was mistakenly identified and one thyroid was misclassified as PG. For testing dataset from R1, the most important six feature lines were used to construct the classification model, which only achieved a highest accuracy of 78.8%. 91% PGs were correctly classified. However, 39% NPGs (seven lymph nodes and six thyroid) were identified as PGs, which resulted in a low specificity of 61%. On the contrary, with three feature lines inputs, RF-LDA model provided a superior specificity of 100% for R3 but the sensitivity is merely 30%.

SVM analysis
Necessary parameters, including penalty parameter C and kernel function parameter g, were optimized based on training datasets by use of a combined approach of 10-fold cross validation and a grid search for the first step of SVM analysis. With different number of feature lines, final SVM models were established by using the optimal values of C and g. Figure 10 illustrates the classification accuracy of three sets of testing datasets. RF-SVM methods also showed the better classification accuracy in testing dataset R2 than those in R1 and R3 with the same number of features. As a whole, the highest average accuracy of RF-SVM model based on the five feature lines reached 78.2%, which is slightly higher than that of RF-LDA model (77.9%).
The three models achieved the optimal classification results based on different number of feature lines. When the number of feature lines is more than seven, the classification accuracy has reached more than 90%. Using twenty-two feature lines, the classification model correctly identified 94.94% of testing dataset R2 ((C=1.1487, g=0.6598)). The confusion matrices of optimal classification are shown in Fig. 11. One PG and three thyroids were misclassified, and all lymph nodes were correctly identified. Using six feature lines, SVM classifier can only achieve the highest accuracy at 76.25% when predicting the class memberships of R1 (C=1.1487, g=0.5). The sensitivity (94%) was excellent but the specificity (52%) was poor. For R3, RF-SVM model showed the highest accuracy at 78.49% with five feature lines (C=16, g=0.1436). Only two NPGs were mistakenly identified as PGs, However, 49% PGs were misclassified as NPGs. These results are very similar to those obtained by RF-LDA model.

ANN analysis
It is necessary to use a classifier that has stronger learning ability. ANN models were established based on the back propagation gradient descent algorithm. The parameter setting of the neural network model should be taken into consideration. The number of input layer neurons was equal to the number of input variables, which ranges from one to twenty-five in our case. The number of output layer neurons was two (equals the number of classes). The number of hidden layer neurons, l, is determined by the formula: where n is the number of neurons in input layer; m is the number of output layer neurons, and it equals two in this case; α is a constant in the range of one to ten. With the increase in the number of input feature lines, the classification accuracy of ANN classifier for three rabbits both increased first and then decreased as shown in Fig. 12. ANN-RF model indicated the good classification performance for three testing datasets with the same number of feature lines. When using the ten feature lines, three models achieved the 92% of highest average accuracy, which had a significant improvement compared to the SVM and LDA.
The three models achieved the best classification accuracy for testing dataset R1, R2 and R3 based on eight, four and six feature lines, respectively. PG and NPG were differentiated by means of RF-ANN with an accuracy of over 97% for testing dataset R2 and R3. The hidden layer neurons for R2 and R3 were twelve and eight, respectively. RF-ANN also achieved the accuracy at 88.8% with eight feature lines and twelve hidden layer neurons for R1 testing datasets. The optimal prediction results of the three testing datasets are shown in Fig. 13. Almost all PGs and NPGs were correctly classified in R2 and R3. Only one PG and one thyroid were misclassified. The RF-ANN model provides good results in the discrimination between PG and NPG of R2 and R3. The classification performance for R1 were slightly poorer, but it is still acceptable. Compared with LDA and SVM, ANN classifier can effectively improve the classification accuracy of three sets of testing datasets. The combination of RF and ANN classifiers shows excellent performance. A plot of the best classification results of the three classifiers is shown in Fig. 14. For R1, R2 and R3, ANN provided the best accuracy overall. As a liner classifier, LDA model fails to solve the linearly non-separable problems. SVM with RBF kernel function is considered to be superior over traditional linear approaches due to the capability to represent non-linear features within data [39]. But the discrimination ability of LDA and SVM classifier was remarkable only for R2, but poor for R1 and R3. These results indicate that there are some variations between the spectra of the same type of tissues due to the individual differences between each rabbit. One of the possible reasons for poor classification may be the training datasets do not fully cover the spectral variations result from these differences.
ANN can overcome this issue to some extent because of its outstanding learning ability. The RF-ANN model provides excellent classification results while requiring less than ten feature lines. K and Na which are related to the control and regulation of physiological functions, play an important role in the discrimination between PG and NPG. The feature lines associated with C 2 , CN, Ca, O and N also contribute to the classification. The predicting time of RF-ANN model were less than 0.5 s. The feature selection based on RF reduces the number of input variables and effectively shortens the time of identification and prediction.
The receiver operating characteristic (ROC) curve was further used to evaluate the performance of the classifier. ROC curves of classification models obtained based on the optimal number of features are shown in Fig. 15. AUC can be utilized to evaluate the performance of model and don't require balanced sample distribution. Clearly, RF-ANN model got the highest AUC, which further indicated it had the excellent performance. RF-ANN is a suitable method for identification of PG based on LIBS.

Conclusion
The study focused on the feasibility of LIBS technique in the identification of PG. Obvious emission lines related to eight elements and two molecular bands can be observed in LIBS spectra, which demonstrated that LIBS technique can well characterize the elemental composition of PG and NPG smear samples.
Three typical supervised algorithms for the identification of PG were evaluated. The results demonstrated that even though the elemental composition of PG and NPG was similar, ANN classifier can provide encouraging classification performance combined with feature selection based on RF. With the same number of feature lines, RF-ANN model achieved the highest average accuracy of 92%. The highest classification accuracy, specificity and sensitivity of testing datasets R2 and R3 based on optimal number of feature lines were all over 97%. The accuracy of 88.5% was also achieved for testing dataset R1.
It's worth noting that RF-LDA and RF-SVM model failed to discriminant between PG and NPG for testing spectra of R1 and R3. Even if using ANN model, the classification accuracy rate was slightly lower when spectra from R1 were used as the testing dataset. This may be due to a fact that there were differences between different individual rabbits. In order to further improve the identification accuracy of PG, more samples and spectra should be obtained to form the representative training datasets that can cover as more differences and variations as possible.
This ex vivo study based on tissue smears demonstrated that LIBS combined with machine learning algorithms was a fast and effective method with potential applications in identification of parathyroid gland. Note that six spectra were converted into one spectrum (by spectra filtering and averaging) as the representative spectrum of the target sample in this work. In future measurement in vivo, one spectrum obtained by preprocessing is representative of the spectra from the target tissue. This mode will facilitate the improvement of diagnosis performance but affect the diagnosis timeliness to a certain extent. Another mode is a single spectrum collected from the target tissue for the identification of the type of tissue in order to ensure the diagnosis timeliness. These may need to be taken into consideration in future applications.