Evaluation of a Peak-Free Chemometric Laser-Induced Breakdown Spectroscopy Method for Direct Rapid Cancer Detection via Trace Metal Biomarkers in Tissue

The ability to perform direct rapid analysis in air and at atmospheric pressure is a remarkable attraction of laser-induced breakdown spectroscopy (LIBS) for the diagnostic quantification of disease biomarker metals in body tissue. However, accurate trace analysis is limited by matrix effects and a pronounced background that masks the subtle (peak-free) analyte signals because tissue plasma is dense and most lines are optically thick. In this work, a peak-free chemometric LIBS method based on a single-shot (for rapidity and nondestructiveness) and an artificial neural network multivariate calibration strategy with spectral feature selection was evaluated for its utility for direct trace quantitative analysis of copper (Cu), iron (Fe), manganese (Mg), magnesium (Mg), and zinc (Zn) in model soft body tissue. The spectral signatures corresponding to the biometals (so-called because the metals are intrinsic to tissue biochemistry) were generated by spiking their known human-body-representative concentrations in molten paraffin wax. The developed multivariate analytical model achieved ≥ 95% accuracy as determined from the analysis of oyster tissue-certified reference material. The analytical models were tested on the liver, breast, and abdominal tissue biopsies. The results of applying the model to the clinical tissues indicated the absence or presence (including severity) of cancer as either malignant or benign, in agreement with the pathological examination report.


Introduction
In laser-induced breakdown spectroscopy (LIBS), the microplasma that is formed when the laser ablates a sample is a "spectral ngerprint" of the sample matrix. erefore, LIBS has great potential in medicine, where it may be used to gain diagnostic information by analysing anatomical specimens such as body uids and tissue. e trace biometals thereby detected may be explored and used as disease biomarkers because they are crucial for biopathological processes (a quarter to a third of all proteins require them to carry out their functions).
A promising application is the elusive early cancer diagnosis which may be realized by providing noninvasive detection and identi cation of speci c trace metal biomarkers in tissue because the tedious and destructive sample preparations required by other methods are in the LIBS technique inapplicable. Unfortunately, LIBS spectra in air and atmospheric pressure are adversely a ected by matrix e ects, shot-to-shot uctuations, and self-absorption [1][2][3]. Furthermore, LIBS signals su er from intensity and reproducibility degradation due to the softness, moisture content, and heterogeneity of tissue samples [4,5]. Biological tissues contain trace biometals in very low concentrations [6]. erefore, LIBS spectra of trace biometals from soft tissue show only a small number of subtle spectral lines from the biometals amid pronounced background and noise. e technical, theoretical, and mathematical aspects of LIBS are well reported [7][8][9][10][11]. It remains to understand more about the dynamics of how laser light interacts with tissue materials in order to fully realize the utility of LIBS for disease diagnostics. LIBS analysis of trace metals in tissues has been reported in the past, but the concentration levels of the metals of interest were either near or below the detection limit (DL). e well-known limitation of LIBS, namely, the relatively high DL of some metals with respect to their physiological levels has been pointed out by Santos Jr. et al. [12]. However, Adamson and Rehse [13] demonstrated DL <1 ppm for trace aluminium (Al) embedded as nanoparticles in surrogate tissue using LIBS. To achieve this, the authors used a calibration curve normalized by the nearby calcium (Ca) II line at 393.366 nm. Most quantitative LIBS methods involve univariate calibration curve and/or calibration-free approaches [14]. In univariate analysis, the useful spectral lines are carefully selected according to prior knowledge of the elemental components and biochemical characteristics of specific tissues, a laborious procedure. Needless to reiterate that the analytical approaches often fail to obtain the desired results for complex matrix samples. Furthermore, it is difficult to get suitable matrix-matched standards, making LIBS at best a semiquantitative tool for trace metal analysis in biological matrices. e challenge is how to improve the accuracy and speed of trace analysis by extracting useful analytical information from the LIBS high dimensionality data. Human tissues are molecularly complex and therefore nonlinear. erefore, more robust approaches are needed for evaluation of the physiological levels of trace biometals by extracting characteristic spectral information while suppressing the spectral interference and noise.
Multivariate chemometrics analysis methods [15][16][17] are applicable in this respect because the LIBS spectral signal consists of a series of vector data with interdependent variables. Chemometrics methods take into account nearly all the variables in the spectra, remove unnecessary and correlated information, and extract the most relevant variables. e errors generated by random and various nontarget factors in the spectra are also reduced. Among the available methods of regression analysis in chemometrics, artificial neural networks (ANNs) are the most "intelligent" enough to learn, memorize, and create relationships among spectral data without the need for characteristic spectral information [18]. e feasibility of using low signal-to-noise ratio (SNR) analyte profiles (here called peak-free) in LIBS was demonstrated when arsenic (As), chromium (Cr), copper (Cu), lead (Pb), and titanium (Ti) were modelled for direct trace (quantitative) analysis using partial least squares (PLS) and artificial neural networks (ANNs) [19], where ANNs were noted to be more robust than PLS at modelling spectral nonlinearity and correcting matrix effects. In the biomedical applications of LIBS, a single biomarker approach is highly unlikely to yield results that have diagnostic accuracy; therefore, the idea of using a basket of biomarkers has been suggested [20,21]. In this work, a chemometric peak-free LIBS approach was evaluated for its utility for direct, rapid but accurate trace quantitative analysis of copper (Cu), iron (Fe), manganese (Mn), magnesium (Mg), and zinc (Zn) simultaneously in soft body tissue. Such analysis would be useful in disease diagnostics applications based on absolute concentrations as well as the multivariate correlations and alterations of the analysed biometals (as the disease biomarkers) in body tissue as opposed to the practice of exploiting relative increases or decreases in intensities of major lines. e emphasis on rapid (single-shot) analysis and weak signals is crucial for clinical applications of LIBS. , and potassium permanganate (KMnO 4 )) in ethanol. e representative concentrations were selected in the ranges in which they occur in the human soft body tissues [22]: Fe: 30-170 μg/g, Mg: 962-502 μg/g, Zn: 20-200 μg/g, Cu: 1-10 μg/g, and Mn: 1-30 μg/g. e spiking concentration ranges were distributed using a research randomizer.

Preparation of Model
About 2 mL of molten paraffin wax was poured into a mould into which 5 mL of the prepared mixture was added.
e mixture was then stirred to ensure homogeneity. Stirring was done while the mixture was being heated at 78°C to ensure that ethanol and acetone boiled off. e mould was covered with an embedding cassette and placed in a freezer to cool and form a block. e block was sliced to 2 cm thickness, each weighing ∼2 g for LIBS analysis. Oyster tissue (NIST 1566B) powder was placed in a hydraulic press to also form method reference pellets of ∼2 g each.

Preparation of Human Cancer Tissue Biopsy Samples.
One (1) breast, two (2) liver, and one (1) abdominal tissue needle biopsies, which had been extracted through a routine surgical operation and histopathologically examined, were donated by Kenyatta National Referral Hospital. e tissues were trimmed to 2 mm thickness and placed in 10% formalin in labelled bottles. ey were dehydrated by soaking in absolute alcohol three times successively for an hour each. e tissue samples were thereafter cleared of alcohol by soaking in 50 : 50 alcohol for an hour, followed by toluene in three stages lasting 30 minutes each. e tissues were dipped into a mould filled with molten paraffin wax, previously placed in an oven at 58°C, and brought back to the oven overnight for infiltration of wax to fill up the spaces left in the tissue. e following day, the tissues were embedded in fresh molten wax at 58°C in moulds and left to cool at room temperature.
e blocks that formed were labelled and trimmed on the surface until the tissue was visible. ey were finally processed and fixed in paraffin wax to make 2 cm thick blocks from which 3 μm thin sections were prepared on Mylar films for LIBS analysis.

LIBS Spectral Acquisition and
Processing.
e LIBS system that was used in this work is a pulsed ND: YAG laser with a maximum energy of 50 mJ operating at a fundamental wavelength of 1064 nm and a 9 ns pulse width. e laser is fired onto a sample, directed by a focusing lens (focal length of 10.16 cm), exciting it to produce a microplasma that is characteristic of the sample under analysis. e optical-tosample distance was maintained at 30 mm following optimization. A fibre optical cable of 0.22 numerical aperture and 101 mm focal length collects the emission from the plasma plume through a lens into a set of seven HR 2000 atomic emission spectrometers in the spectral range of 200-980 nm, which spectrally disperse the radiation. Spectral data are acquired simultaneously and displayed on the computer screen with the help of OOILIBS software. Each charge coupled device (CCD) detector has 2048 pixels and an optical resolution of 0.065 nm. Figure 1 shows examples of the LIBS spectral responses of Fe in typical soft body tissue as indicated by the detected lines whose intensity clearly steadily increases with the spiked concentration of Fe in the blank matrix. From this, it is easy to differentiate between self-absorbed lines and resonant lines, as well as those with good oscillator strength and those that are interference-free to be used as potential candidates for multivariate calibration using spectral feature selection. e detectability of Fe by LIBS is demonstrated in Figure 2 using a clinically acquired liver biopsy sample.
As LIBS spectral responses from biomatrix analytes suffer from matrix effects, the spectra require preprocessing [23] because the intensity of the emission lines observed is a function of both the concentrations of the elements of interest as well as the thermochemical properties of the matrix that contains them. Preprocessing of spectra is essential to reduce noise and matrix effects. For this purpose, denoising, smoothing, baseline correction, and mean-cantering techniques were employed. Smoothing was done using the      Sivitzky Golay technique to obtain clear spectral line profiles, while wavelet transforms were used for denoising. Mean cantering enabled all the data across the spectral region to be involved in the modelling.

Multivariate Calibration for Quantitative Analysis.
In this work, spectral features corresponding to Cu, Mn, Mg, Zn, and Fe of the model tissue samples were used to train the ANN model for trace quantitative analysis using MATLAB software. A recent review [24] sheds good light on ANN-based LIBS. ANN is one of the computational ways of mapping nonlinear input data to a target space. e most common of the network architectures is the multilayer feed-forward system, in which the input data proceed forward only (to the hidden layer and then to the output layer) and never make loops, as opposed to other techniques like the recurrent neural network system [25]. e power of the network depends on the transfer function, the learning rule, and the network architecture [26]. During training of the network, the neurons are optimized until the error in prediction is minimized and the network attains the desired level of accuracy. e trained network can then be given new input data to predict the output [27]. In this work, the best conditions were 3 neurons and a feed-forward back propagation algorithm. e model was trained using 60% of the data, 20% was used for validation, while the remaining 20% was used for testing. e model was trained a number of times until the one with the root mean square error of calibration (RMSEC) and regression coefficient (R 2 ) value closest to 1 was achieved. e model regression curves for training, validation, testing, and the overall curve are shown in Figure 3. e regression curves for Cu, Fe, Zn, Mn, and Mg, showing predicted concentration versus known concentration, are shown in  e NIST 1566B standard was used to estimate the analytical accuracy of the developed calibration model. In Table 1, the predicted values are compared against the standard reference values. It is noted that the ANN model is suitable for the determination of the concentration of Fe, Mn, Mg, Zn, and Cu in soft body tissue-the highest percentage deviation was −3.89% for Fe. e accuracy of prediction shows that the multivariate chemometrics analytical models developed here would be useful in noninvasive cancer diagnostics utilizing biometals as the disease biomarkers because trace bimetals offer the potential for early detection, tracking progression and recurrence, as well as monitoring of treatment response intrinsically in tissue.

Method Application to Needle Tissue Biopsies.
To better understand the metallome in disease, the determination of actual samples is of great importance [28]. e developed analytical models were investigated preliminarily on the liver, breast, and abdominal needle biopsies described above, and the results are shown in Table 2. e concentration ranges determined for the tissues were Fe (51.2-137.2 μg/g), Cu (5-18.7 μg/g), Zn (36-56.8 μg/g), Mg (78.2-507.4 μg/g), and Mn (8.8-19.5 μg/g) for liver; Fe (87.7-113.9 μg/g), Cu (10.9-12.3 μg/g), Zn (49.3 μg/g-55.7 μg/g), Mg (194.3-242.3 μg/g), and Mn (14.5 μg/g-16.1 μg/g) for breast; and Fe (96.7-125.7 μg/g), Cu (6.7-7.5 μg/g), Zn (88.3-93.9 μg/g), Mg (467.5-583.1 μg/g), and Mn (9.5-10.5 μg/g) for abdominal, respectively. e biometal concentrations in healthy and diseased (cancer) tissues are clearly different, as previously reported [29,30]. From Table 2, liver tissue sample number 2 and the abdominal samples, which had been histologically classified as malignant, had higher concentrations of Fe and relatively low concentrations of Cu as compared to the other liver tissue, which had been classified as benign. is shows there is an increased need for Fe in proliferating tissues due to the constant demand for supply of nutrients. e results for malignant and benign liver tissues clearly indicate the presence of cancer based on the trace biometals, which is in agreement with their pathological examination. It was also observed by other workers that the concentrations of Ca, Fe, Cu, and Zn are higher in neoplastic tissues (malignant and    benign) when compared with normal tissues [31,32]. ese trace biometals can be considered tumour biomarkers because it is possible to classify different tissues as normal or neoplastic, as well as different types of cancer, based on their concentrations. Further, all the trace biometals were statistically correlated with well-known prognostic factors for breast cancer.

Conclusion
is work involved the evaluation of a rapid chemometric peak-free LIBS technique for direct rapid (diagnostic) analysis of trace biometals in soft body tissue. A multivariate chemometrics calibration model was developed using ANN and based on paraffin wax for the determination of the concentrations of Fe, Mn, Mg, Cu, and Zn in soft body tissue. e model was successfully validated using oyster tissue as a certified reference material (CRM). e predicted biometal concentrations were within a range of less than 5% error.
is work has demonstrated that LIBS can be a useful technique for rapidly and directly detecting trace amounts of biometals in soft body tissue in the context of spectral diagnostics of disease. e method was tested on malignant and benign liver tissues, and the results agreed with those of histopathological examination, which is based on the microscopic examination of tissue morphology. Although the sample numbers in this work were too few to accurately assess the deviation between healthy and diseased tissue based on the analysed biometals, the method is not only potentially accurate but is useful for rapid diagnostics of cancer in soft body tissues. In a typical application, the analysed biomarkers would be measured and monitored to yield specific signatures that can be used for detecting cancer early before morphological features become apparent.

Data Availability
e datasets used in the study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.

Authors' Contributions
OEA prepared the simulated samples, collected spectral data, performed statistical and chemometric analyses of the samples, and prepared the initial draft of the manuscript. AHK and DKA conceived the project. AHK and DKA supervised data collection and chemometric analyses and provided feedback and suggestions on the manuscript before submission. All authors read and agreed with the final version of the manuscript.