Investigation of random lasing as a feedback mechanism for tissue differentiation during laser surgery

: Laser surgery provides clean, fast and accurate cutting of tissue. However, it is diﬃcult to detect what kind of tissue is being cut. Therefore, a wrong cut may lead to iatrogenic damage of structures. A feedback system should automatically stop the cutting process when a nerve is reached or accidentally being cut to prevent its damage. This could increase the applicability and safety of using a laser scalpel in surgical procedures. In this study, random lasing (RL) is used to diﬀerentiate between skin, fat, muscle and nerve tissue. Among these tissue types, a special emphasis is made on the diﬀerentiation of nerve from the rest of the tissues, especially fat since nerve is covered by a fatty layer. The diﬀerentiation is done for ex-vivo tissues of a pig animal model. The results show that random lasing can be used to diﬀerentiate these tissue types also under room light conditions in open air.


Introduction
In the recent decades, there has been a rising usage of lasers for surgery due to their non-contact operational procedure, low or minimal invasiveness, high flexibility of the cutting geometry and their hemostatic and antiseptic effects [1,2]. For this reason, the laser has been applied as a surgical tool in many fields so far [3]. The surgical application cases involve the field ophthalmology [4], maxillofacial surgery [5], laser dentistry [6] and many others. However despite the advantages of laser surgery, there is no tactile feedback to the clinician performing the surgery. Therefore, there are only visible information of the surface available. Thus, there is a risk of iatrogenic damage or destruction of anatomical structures like peripheral nerves.
To overcome this issue, different approaches have been investigated as a feedback mechanism during laser surgery. These mechanisms are mainly categorized into imaging and non-imaging techniques. For imaging techniques, optical coherence tomography [7] and confocal microscopy [7] are some examples. However, they have the disadvantage of being too slow to provide feedback before damage is done to the tissue. Therefore, non-imaging techniques are preferable. So far, fluorescence spectroscopy [8], Doppler anemometry [9], acoustic emissions from the cutting process [10], diffuse reflectance spectroscopy [11], and laser induced breakdown spectroscopy (LIBS) [3] are among the investigated ones.
Doppler anemometry is not suited for fast detection due to the fact that the low frequencies below 20 kHz are the most important ones for differentiation of the tissue [9]. Therefore, this method would be too slow. Fluorescence spectroscopy and diffuse reflection spectroscopy have the disadvantage that the optical signal can be disrupted by room-light, and resultantly either a contact-based alignment upon the to be cut tissue or a dark room is essential during the laser surgery operation. Using the acoustic signal generated during the cutting process provides feedback, however only after the damage is done. Therefore, it is not suitable.
Comparing LIBS to all previous methods, it can be very fast, works under room light conditions and has a very high accuracy for tissue differentiation even between very similar tissue types like nerve and gland [12]. Additionally to LIBS, a random laser (RL) has the potential to provide information of a volume of the tissue on the surface as opposed to LIBS which can detect only the surface tissue type. However, the depth extent of detection of RL is not the purpose of this study due to the fact that the basic tissue discrimination has to be investigated first. Furthermore, LIBS measurements normally need a fine resolution and a large spectral bandwidth for its measurement. This requirement is expected to be lower for the RL. This limits the readout time of the spectrometer. With a RL it might be possible to detect the tissue type within or less than a micro second, it might be possible to stop an otherwise nerve damaging laser pulse. Moreover, a RL has the advantage that the RL can be generated easily with ultra-short pulse lasers [13].
A normal laser consists of a gain medium which amplifies light through stimulated emission and a cavity which provides a resonant feedback. For a normal laser, scattering would remove the photons from the beam path inside the cavity and decrease the laser performance. However, it is possible to use the scattering medium as a resonator. In this medium, light can be scattered multiple times and amplified at the same time to produce stimulated emission and lasing. This effect was first proposed in 1966 by Ambartsumyan et al. [14] when one mirror of a Fabry-Perot cavity was replaced by a scattering surface. Already two years later, Letokhov [15] predicted that it is possible that scattering can provide a feedback for stimulated emission without any mirrors. In 1995, this effect was named random laser (RL) [16]. A RL is a laser which can be described by the following properties [17]: 1. light is scattered multiple times and amplified by stimulated emission 2. there is a laser threshold which is the minimum pump intensity for having a positive gain. 3. the mean free path should be lower than the size of the system RL became widely investigated [17][18][19] due to the fact it is very easy to set up. A RL can be generated by combining scatterers with a laser dye. With optical pumping, the random laser is built. Out of the many studies done with RLs, the most interesting one is the study from Polson et al. [13] in which it could be shown that random lasers can be created from soft biological tissue by soaking it in Rhodamin 6G as fluorescence dye. They could show that RLs can be made from vegetables (potatoes), animals (chicken meat) and human organs (colon and kidney). Not only from soft tissue but also from hard tissue such as bone, a RL can be generated [20]. It could also be shown that RLs are extremely sensitive to small structural changes even at nanoscales [21]. Therefore, random lasing can be used to detect the bone deformation with small tensile forces of 8.9 N [22]. Due to this, nano scale bone deformation could be detected before any damage of the bone tissue occurs.
One of the most interesting studies is again from Polson et al. [23]. They showed that tissue mapping with RLs is possible. At a single patient, they differentiated between healthy and cancerous ex-vivo tissue from human intestines where the frozen tissue was soaked in Rhodamin 6G. It could be shown that the different scattering inside healthy and cancerous intestines leads to different spectral emissions. However, only one patient was used for the analysis. Therefore in this study, it is shown that tissue differentiation for at least between cancerous and healthy tissue might be possible with RL.

Tissue preparation
For the analysis, fat, muscle, nerve and skin are taken from 15 bisected ex-vivo domestic pig heads. From each pig head, one sample of each fat, muscle, nerve and skin is taken. For a realistic comparison all tissues are taken from the cheek region: 1. nerve from the nervus infraorbitalis 2. fat from the regio buccali 3. muscle from the musculus masseter 4. skin from the regio infraorbitalis.
The size of the samples is around 20 x 20 mm 2 except nerve which had a size of about 10 x 30 mm 2 . No further cleaning process of the tissue is done.
Before the measurement, the tissue samples were put in a Rhodamin 6G (R6G) solution with water with a concentration of 2 · 10 −4 g ml for 24 hours. Despite R6G not being bio-compatible, RL can be generated with fluorescent anticancer drugs [24]. However, due to its availability and price R6G is used for the prove of principle in this study.
Experimental set-up The set-up is shown in Fig. 1. The main set-up consists of a laser and a collection optic connected to a spectrometer. The laser is a frequency doubled Nd:YAG-Laser (Q-smart 450, Quantel, Neuilly-Sur-Seine, France). The pulse duration is 5 ns and the pulse energy in this study is 100 mJ. The laser is focussed to a spot size of 40 µm. The final energy per area which is used in this study exceeds the safety limits of 20 mJ cm 2 for other modalities such as photo-acoustics by far. However, this does not pose a problem due to the fact that our proposed idea is to use the technique within the context of laser surgery where tissue ablation is required. In this case, the RL should stop an damaging laser pulse in the case of long pulses. If longer microsecond pulses are used for ablation, it could be possible to interrupt the process. Alternatively, the a laser which is used in burst mode can be used with this approach. In this case the first, pulse might be used to stop the other pulses.
The signal from the RL is collected with a collection optic and coupled into a fibre. The collection optic also contains a notch filter for 532 nm for removing the light from the excitation laser. The signal is measured with a spectrometer (Mechelle Me 5000 Echelle, Andor, Belfast, UK). The spectrometer is equipped with an ICCD camera (A-DH334T-18F-03 USB iStar ICCD detector, Andor, Belfast, UK) and has a spectral resolving power ( λ ∆λ ) of 6000. The spectrometer is triggered by the laser and the measurement is done with a delay of 0 ns and a measuring time of 1000 ns. It should be noted that the 1000 ns measuring time is used without optimization. The same signals can also be measured within 200 ns and less with zero delay. This was later tested for other samples. The repetition rate of the experiments is 1 Hz.
From each sample, 100 measurements are collected from the same spot. Despite the limitations of the amount of intra-animal variation by this approach, it is chosen so due the the following reasons: 1. The position on which the RL is generated on the sample might play a role on the outcome of the RL due to geometrical effects. To compensate this, the RL is only generated on only one spot 2. Due to the fact that we operate far above the Maximum Permissible Exposure of 20 mJ cm 2 , tissue damaging is inevitable. Thus, by repetitive shooting at the same position, the resulting damage/alteration is taken into account and become part of the dataset.
3. Moreover, due to the fact that the expected variance from changing the animal and with this also changing the position, we expect to have a higher variance than we might get from pure intra animal variations. By targeting the intra and inter animal variations at the same time, it should target already a more difficult data set.

Evaluation of the RL
Additionally, the development of the intensity for the first five and the last five pulses is tested. There are two opposing effects which might alter the intensity of the RL. First, the quenching of the R6G might decrease the signal intensity. Second, coagulation due to the high intensity might introduce higher scattering and with this an increase in the RL-intensity may be observed. To generate a reliable measure for testing the significance of the RL intensity, the first five and last five spectra are averaged. Afterwards they are summed up to generate a single value, describing the intensity. With an ANOVA, it is tested if there is a significant increase or decrease of the signal intensity. Additional, the full width half maximum (FWHM), the maximum position and the fluctuations in the spectral intensity are calculated to characterize the RL. The fluctuations in the spectral intensity [25] are calculated as follows: where γ, β = 1, 2, ..., 100 is the spectrum for each spot and tissue type and ∆ β (k) = I γ (k) −Ī(k) whereĪ(k) is the mean of the hundred spectra at each point for each tissue type. The mean is calculated from each tissue type and animal separately. However, the P(q) presented, shows the summed up fluctuations.

Statistical analysis
The statistical analysis is done with the leave one out strategy. All data from one head is excluded and the rest is used for training (training data). Afterwards, the trained classifiers are tested at the left out data (test data). This is done for all 15 possible combinations of left out animals. Before the training process, the data is normalized and a principle component analysis (PCA) is done for the training data. The first 24 components are used for training. In average, more than 99 % of the information can be represented by the first 24 components. Thus, this is valid and it decreases the time to train the classifier significantly. The coefficient matrix of the training data is also applied to the test data before the classifiers are tested. For the classification two classifications are done. First, all four tissue types are tested. This multi class classification is done with Support Vector Machine (SVM), Random Forest (RF) and Linear Discriminant Analysis (LDA). Second, the differentiation between fat and nerve is tested separately due to the fact that this classification is normally the most difficult one and nerve is often surrounded by fat tissue. Moreover, it is tested how well nerve can be found in any of the other four tissue types. In the last two cases, the same classifiers and RobustBoost (RB) are used. RB is used because it also shows good results for tissue differentiation [26]. For all cases, only single spectra are used for training and testing. No averaging is done.   Figure 2 shows example spectra for fat, nerve, skin and muscle. All single spectra are normalized to one. The spectra generated in this study are nearly completely incoherent RL. There are no narrow peaks from coherent RL present. The single spectrum of muscle consists only of a single peak. Fat normally shows a peak widening for lower intensities. Most likely there are other materials present in the tissue. For the two other tissue types, it can happen that they show one single peak or two dominant peaks (Fig. 2) similar as in the mean spectra, shown in Fig. 3. Especially, skin is a multilayer organ, so multiple peaks are expected. This might be the reason that skin has the widest average peak as shown in Fig. 3. Also the nerve tissue consists of a fat layer, so that fat related peaks might exist. In Fig. 2 (left) this behaviour can be seen. The right nerve peak is completely identical with the fat peak. Therefore in the nerve tissue, also its fat layer can be detected. This pinpoints also the ability of the RL to detect tissue types inside the tissue volume, not only at the surface compared to LIBS.

Evaluation of the RL
The average intensity of the RL increases by about 20 % from the first five pulse to the last five pulses on the same spot. The increase is significant with a p-value smaller than 10 −5 . This might indicate that the effect of the coagulation is more dominant that the quenching effect. Figure 4 shows the most important properties of the RL for this study. The presented P(q) represents the spectral fluctuations. According to Gomes et al. [25], the centre of the q-value show the regime of the RL. Fat and nerve tissue are in the stable RL-regime. However, in this regime many fluctuations of the signal are still present. For skin and fat, there is no clear answer based on the P(q)-value possible. However, the lower plots of the FWHM show that sometimes the skin shows a RL behaviour and sometimes not. In general, pre-experiments also showed that skin is the most complicated to generate a RL. However, if only one tissue type does not generate properly random lasing, the classification becomes even easier. At the same time the FWHM seems to fluctuate strongly in some cases such as the muscle tissue of the ninth animal. Additionally, there is a change of the maximal position of the RL. These changes are less random than the changes of the FWHM. The maximum often seems to reach a plateau.

Classification results
First, the principle component (PC)-plot of every measurement plus the mean from each animal are shown for PC2 and PC3 in figure 5. In general, there is separation between the different tissue types. There is a strong overlap between skin tissue and the rest. Moreover, fat and nerve tissue show a rather strong overlap. The same effect is found for the mean values for each spot/animal for each tissue type. In summary, it is expected that skin shows the worst classification results. Table 1 shows the classification results for the case that all four tissue types are used for classification. The total accuracy in this case is 88 % for SVM, 87 % for RF and 82 % for LDA. In general, SVM and RF provide similar results. However, the results from SVM tend to the extremes. The reason might be that SVM tends more to over fit than RF which normally does not show signs of over fitting. Table 2 shows the confusion matrix of the analysis. For all three classifiers it can be seen that nerve gets classified the best while the accuracy for skin and muscle are fairly weak. With LDA and RF the most mixing occurs between skin and muscle. Fat and nerve can get separated best especially by SVM and RF. The lower accuracy from LDA compared to SVM and RF is mainly caused by the misclassification of fat as skin. Misclassifications happen especially often in a ways that a tissue gets classified as skin which is expected already from the PC-plot. This happens most likely due to the wider spectrum and the potential double peaks the skin tissue might show. Therefore, there is just a bigger overlap with all other tissues than for fat, nerve and muscle have.
If nerve and fat are classified against each other, table 3 shows the results for the binary classification. The accuracy is between 96 % and 99 % for RF and RB and it is 99.6 % for the SVM. Thus, fat and nerve can be reliably classified. Especially due to the fact that the comparably fast LDA provides also very good results. Table 4 shows the results to distinguish nerve from other tissue. The accuracy is above 96 % for all classifiers. Hence, if nerve might be cut by accident by the cutting laser it might be possible to detect the nerve in time. Moreover, the high detection rate of nerve also highlights that the RL targets structural information which are expected to make a stronger difference between nerve and the other tissue types than e.g. the absorbance of the tissue.

Conclusion
The results of this study show that it is possible to use random lasing for tissue differentiation even in the case of thermal tissue damaging. However, the random lasing is very close to the lasing threshold in some cases. Currently the amount of R6G is not optimized due to the fact that R6G will not be the final fluorophor to be used. Due to the lack of optimization, random lasing did not occur every time. Especially for skin, the generation of the RL proved to be difficult. Additionally, it is possible to replace R6G by a bio-compatible fluorescence dye. Lahoz et al. [24] showed that a combination of moiety, tamoxifen (Tx) and nitro-2-1,3-benzoxadiazol-4yl (a commercial dye) is possible. However, the biggest downside is the long soaking duration which is used to enable RL. However, if a combination with another drug might be used such as shown from Lahoz et al. [24], it might be possible to use the standard injection for the anticancer drug to also deliver the fluorophor.
In future, it might possible to gather information about the current tissue under investigation within less than a micro second. Therefore, RL might allow correction of currently happening tissue ablation process. However, to enable this the data analysis has to be speeded up. At the current state, it is the main bottle neck. This can be done by using a spectrometer with a lower resolution and smaller measured part of the spectrum. Most likely, the range from 550-650 nm with a resolution of 1 nm would be enough for the tissue classification. This reduces the amount of data-points from a few thousand to 100 or even less. These 100 data-points should be able to be processed at a micro second time. Moreover, the results pinpoint the ability of the RL that the tissue type below the surface can also be detected. Thus, RL might be a good tool for in-vivo tissue differentiation.