Tissue biomolecular and microstructure profiles in optical colorectal cancer delineation

Our study included 47 patients undergoing bowel resection at the Mercy University Hospital, Cork, Ireland (patient demographics are shown in table 1). The study was approved by the Clinical Research Ethics Committee of the University College Cork. All methods were performed in accordance with the relevant guidelines/regulations. Informed consent was obtained from all participants of the study. Out of the 47 tumors investigated, five were in pT1 stage, seven were pT2, 26 were pT3 and nine were pT4. Twenty-eight patients had local lymph node metastases whereas 19 had not.

Data collection involved measuring about 15 sites of ex vivo mucosal tissues and 15 of tumor tissues on each colorectal specimen after surgical resection. Representation of the tissue heterogeneity was ensured by collecting spectral readings from a typical area of 100 cm². Devascularization periods (i.e. time between the restriction of blood supply to the specimen and its removal) typically varied between 1 and 2 h. Once the specimen was removed, it was transported to a measurement bench and opened to expose the colonic lumen for the DRS measurements. The specimen was gently rinsed with water and cleaned afterwards in order to remove the remaining stool and excess of blood. Next, the specimen was placed in a measurement grid showing the coordinates of the sites where DRS spectra was collected. Then, mucosa and tumor regions were identified by experienced surgeons so that DRS spectra could be correlate with the types of tissues measured. The average time between the specimen removal and the start of the data collection was 40 min. Data collection was performed within an average time of 1 h after surgical resection. The tissue moisture was kept by using a damp cloth in order to preserve physiological conditions as much as possible throughout the data collection. Upon finishing the DRS data acquisition, each specimen was returned to the Pathology Department for processing and analysis according to standard protocols. Histopathology analysis was the ground truth of our DRS measurements.

Our DRS system employed is illustrated in figure 1. A broadband light source (HL-2000-HP, Ocean Optics, Edinburgh, United Kingdom) with emission ranging from 350 nm to 2400 nm was employed. Low-OH silica fiber optic probes including a 630 µm SDD quadrifurcated probe (small probe or probe 1, BF46LS01 1-to-4 Fan-Out Bundle, Thorlabs, Munich, Germany) and a 2500 µm SDD trifurcated probe (large probe or probe 2, Fibertech Optica, Anjou, Canada) were used to deliver the light to and from the tissue sample. Two spectrometers detected the DRS signal, one in the visible/short wavelength NIR wavelength range between 350 and 1140 nm (QE-Pro, Ocean Optics, Edinburgh, United Kingdom) and one in the long wavelength NIR range between 1090 and 1920 nm (NIR-Quest, Ocean Optics, Edinburgh, United Kingdom). The overlapping wavelength region was used to merge the spectra into one broadband (350–1920 nm) spectrum, as described in the data preprocessing section (section 2.5). We used the 630 µm SDD probe (probe 1) to probe the superficial tissue at depths ranging from 0.3–1.1 mm (estimated from Monte Carlo (MC) simulations of light propagation into tissues, data not shown), whereas the 2500 µm SDD probe (probe 2) collected reflectance from depths ranging from 0.1 to 1.8 mm (data not shown). Probe 1 contained 600 µm core diameter fibers for both illumination and collection. Probe 2 had one central 600 µm core-diameter source fiber, and ten 200 µm core-diameter collection fibers surrounding it. At the spectrometer end of probe 2, each group of five collection fibers were positioned linearly to match the slit of the spectrometers to increase the detected signal.

Zoom In Zoom Out Reset image size

The procedure to investigate the tissue depth probed is described in detail in [33]. Briefly the tissue depth investigated by using probes 1 and 2 was estimated through forward MC simulations of steady-state light transport in multi-layered tissues (MCML) [34] accelerated by graphics processing unit (GPU) [35, 36] using optical properties obtained from the analysis of the spectral fitting algorithm described in section 2.5. Light propagation through a semi-infinite homogeneous medium with refractive index of the outer medium (η_out) of 1, tissue refractive index (η_rel) of 1.4, anisotropy factor (g) of 0.9 was simulated. By using 10 million photon packets, we generated 5 × 5 mm light fluence maps with 10 μm of radial and depth resolution in order to compute the photon hitting density maps used to estimate the probed depth.

First, we collected the background and reference signals. Both these measurements were taken by placing each probe on a specialized holder designed to fix the distance between the fiber optic probes and our reflectance standard (FWS-99-01c, Avian Technologies LLC, New London, USA). Signal contamination due to backscattered light and ambient light was avoided be a closed environment with blackened walls of the holder. The background and reflectance measurements were taken with our broadband light source turned off and on, respectively. In order to avoid probe contamination, we covered our probes with polyvinyl chloride film and cleaned them with ethanol 70% after each set of measurements. All measurements of this study were performed with the same probes, which were positioned as close to 90° from the tissue surface as possible. Upon completion of the data collection, the data was safely stored for analysis. In this study, we collected a total of 1363 spectra for the small SDD probe (630 µm SDD) and 1526 for the large SDD probe (2500 µm SDD).

In order to obtain the tissue reflectance $R\left( \lambda \right)$, the intensity measurements (${I^{{\text{Tissue}}}}$) of each spectrometer had their background subtracted ($\,{I^{\text{Background}}}$) and the resulting intensity was divided by the reference measurements ${I^{{\text{Reference}}}}$ (with background subtraction):

$$\begin{equation}R\left( \lambda \right) = \frac{{{I^{\text{Tissue}}}\,\left( \lambda \right) - \,{I^{\text{Background}}}\,\left( \lambda \right)}}{{{I^{\text{Reference}}}\,\left( \lambda \right) - \,{I^{\text{Background}}}\,\left( \lambda \right)}}.\end{equation} \tag{ 1 }$$

To merge the reflectance spectra and correct for any slight mismatch between the two spectrometers, we used the overlapping spectral range between 1090 and 1140 nm to perform an interpolation by using the weighted sum:

$$\begin{equation}R\left( \lambda \right) = \,\mathop \sum \limits_{i = 0}^{100} \frac{{\left( {\left( {100 - i} \right)\times\,{R_{\text{vis}}}\, + \,i\times\,{R_{\text{NIR}}}} \right)}}{{100}}.\end{equation} \tag{ 2 }$$

The resulting reflectance spectra $R\left( \lambda \right)$ were used for the estimation of chromophore concentrations and scattering parameters.

A flowchart of the spectral fitting algorithm is shown in figure 2. Details of the steps of this algorithm can be found in the supplementary material (available online at stacks.iop.org/JPD/54/454002/mmedia). Briefly, the obtained reflectance spectra $R\left( \lambda \right)$ was used to extract the tissue absorption and scattering related parameters. The absorption-based parameters included are the concentrations of blood (total hemoglobin (THb) percentage and THC), lipid (f_lipid), water (f_water), bile (f_bile), met-hemoglobin (f_metHb), as well as oxygen saturation (StO₂), and average vessel diameter R. The scattering-based parameters, on the other hand, included the reduced scattering amplitude $\alpha {^{^{\prime}}}$, Mie scattering power b_Mie, and the percentage contribution of Rayleigh scattering f_Ray. The spectral fitting methodology is illustrated in figure 2 with a detailed description provided in the supplementary material. Briefly, the extraction was performed by using a spectral fitting based on the four types of inputs to the algorithm: the measured DR spectrum; the pure chromophore absorption spectra (${\mu _{{\text{a,chromophore}}}}\left( \lambda \right)$); the Rayleigh and Mie scattering equations yielding the scattering coefficient as a function of wavelength; as well as a look-up table (LUT) of DR (R(µ_a,µ_s)) values describing the light propagation in tissue as a function of absorption coefficients (µ_a) between 0.01 and 300 cm⁻¹, and scattering coefficients (µ_s) between 0.1 and 1000 cm⁻¹. The LUT was generated by using a forward MC model of light propagation into tissues, which was accelerated by GPU [35, 36]. In this model, tissue optical properties including µ_a and µ_s were used as an input to simulate the light transport inside a homogeneous semi-infinite medium and calculate the reflectance at different positions of the medium (tissue) surface. MC simulations were performed with 10 million photon packets, refractive index of the outer medium (η_out) of 1, tissue refractive index (η_rel) of 1.4, anisotropy factor (g) 0.9. The results were stored with a radial and depth resolution of 10 μm. The contribution of photons launched in the excitation fiber was accounted through convolution of the incident photon beam [37]. The modeled reflectance was obtained by integrating the reflectance values in the area of the core diameter of the collection fibers (detector area) located at the specified SDD. This integration was performed for a wide range of optical properties involving values typically found in most biological tissues [38, 39]. The result is a MC LUT associating the modeled reflectance R(µ_a,µ_s) with µ_a and µ_s values for the geometry of our fiber optic probe.

Zoom In Zoom Out Reset image size

By using the MC LUT, chromophore absorption spectra and scattering equations, the modeled DRS spectrum was reconstructed. First, chromophore concentrations and scattering parameters are used to obtain the tissue absorption and scattering spectra. Next, for each wavelength of these spectra, combinations of absorption and scattering coefficients were associated to values of reflectance R(µ_a,µ_s) in order to reconstruct the modeled DR spectrum ${R_{{\text{modeled}}}}\left( \lambda \right)$. ${R_{{\text{modeled}}}}\left( \lambda \right)$ was multiplied by a calibration factor and compared to the experimental spectrum ${R_{{\text{experimental}}}}\left( \lambda \right)$ in a fitting algorithm. The outputs of the algorithm are a fitted DR spectrum, the values of the fitting parameters and a value for the fitting fidelity.

In the algorithm we implemented the same spectral fitting method as used by Nachabe et al [40, 41] (and is described in in more detail in the supplementary material). Validation studies of our fitting procedure were conducted in both phantoms and biological tissue samples and has been was presented elsewhere [42]. It is important to note that we refer to ${f_{\text{blood}}} = {f_{\text{Hb}}} + {f_{\text{HbO}_2}}$ as THb percentage in tissue, even though THb is a relative volume percentage of pure blood. We also emphasize that ${f_{{\text{bile}}}}$ may contain traces of bilirubin found in blood at ranges between 5.8% and 40% in humans, especially considering that bilirubin blood serum levels can vary between 10 and 50 mg dl⁻¹ [43] and hemoglobin blood levels vary between 121 and 172 mg dl⁻¹ [39]. With that in mind, ${f_{\text{bile}}}{ }$ would represent the bile concentration added to a contribution of bilirubin varying between 5.8% and 40% of THb. The use of bilirubin absorption spectrum in our fitting is limited by the narrow wavelength range reported in the literature, which is not sufficient to fit a spectrum between 450 and 1590 nm.

Significant statistical differences between spectral fitting parameters of normal mucosa and tumors were tested by using two-sample t-test and Wilcoxon rank sum test. To evaluate the adequate use of two-sample t-test, Anderson–Darling and Lilliefors normality tests were performed.

The tissue chromophore concentrations and scattering properties obtained with our spectral fitting model were used for tissue classification in mucosa and tumor based on a binary decision tree algorithm. Our decision tree was built by using the classification and regression tree (CART) algorithm implemented with default fitctree MATLAB function. The method uses a recursive partitioning method to generate a classification tree from thresholds of the tissue parameters obtained by the spectral fitting described in section 2.5. These thresholds are selected from a training set of spectral data and applied to spectra on the test set for every iteration of cross-validation. This study used five-fold cross-validation, which consists of randomly dividing the dataset into five sets of training and test sets containing 80% and 20% of the spectra in our dataset, respectively. Each of these pairs is used to build a tissue classification model where the model is calibrated in the training set and validated on the test set. For each pair, the classification model was assessed by performance parameters involving the sensitivity, specificity, accuracy and area under the receiver operating characteristic curve (AUC) for identification tumor and mucosa tissues. The performance of the classification model was given by the mean of performance parameters among the five sets.

In order to ensure our tissue classification model was not overfitted, we reported the mean and standard deviation of the output of the 20 iterations of five-fold cross-validation with random sampling. In addition, we have used THb, R, f_lipid, f_water, f_bile, α', f_Ray and b_Mie as relevant parameters for tissue classification. StO₂, and f_metHb were not included in the model due to their variations in ex vivo settings, as it was not possible to control the blood oxygenation and conversion of Hb and HbO₂ into met-Hb once blood of the specimen was exposed to air.

Our spectral fitting analysis yielded relevant absorption and scattering parameters to be used for the tissue differentiation. Based on this analysis, we selected the tissue scattering-based microstructure parameters (reduced scattering amplitude $\alpha {^{^{\prime}}}$, Mie scattering power b_Mie, and the percentage contribution f_Ray), and concentrations of blood (THb), lipid (f_lipid), water (f_water), bile (f_bile), as well as oxygen saturation (StO₂) and average vessel diameter R as relevant parameters to be monitored and discussed (figures 3 and 4). All fitted variables are described in table 2. Values of f_MetHb were not considered relevant for our analysis as these values can vary due to exposure to oxidizing agents which convert Hb into metHb, even though f_MetHb needed to be included in our spectral fitting in order to eliminate the metHb influence on THb, THC and StO₂. Figure 3 shows a boxplot of the fitted parameters for all mucosa and tumor tissue spectra using probe 1. These plots suggest that the parameters which best discriminate mucosa and tumor tissues are THb, f_lipid and b_Mie.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figure 4 suggests a similar behavior of chromophore concentrations and scattering properties differentiating mucosa and cancer tissues could be observed in both superficial tissue (figure 3 for probe 1) and deeper tissue layers (figure 4 for probe 2). The differentiation of parameters such as the scattering amplitude α', f_lipid and b_Mie increased by using probe 2. Yet, the trend observed for lower f_Ray in the tumor surface (figure 3) does not hold for deeper tissue layers in tumors.

Table 2 shows the mean and standard deviation of mucosal and cancerous biochemical and microstructural parameters extracted with our DR spectral fitting algorithm. In average, tumors had larger THb, THC, f_water, α', b_Mie, and smaller average vessel radius, f_lipid compared to normal mucosa. For probe 1, tumor tissues also exhibited lower f_Ray. Also, R, f_lipid, and b_Mie tend to be higher for probe 2 compared to probe 1 in both mucosa and tumor tissues, whereas all other parameters tend to be lower for probe 2. The standard deviation of THb, THC, f_lipid, f_water for probe 2 was lower to those of probe 1, even when considering the standard deviation relative to the mean values (coefficient of variation). The p-values of statistical tests leading to the representation of significant statistical differences in the boxplots can be found in table 3. Even though we show the result of statistical tests for t-test and Wilcoxon rank sum test, we emphasize that the distribution of spectral fitting parameters of normal mucosa and tumors are normal according to both Anderson–Darling and Lilliefors normality tests.

Classification trees were built with spectral fitting parameters of probes 1 and 2 in order to categorize tumors based on their particular ranges of chromophore concentrations and scattering properties. Figures 5(a) and (b) show the two separate branches of a decision tree with parameters leading to most of the differentiation between mucosa and tumor by using probe 1. These parameters are THb, R, f_lipid, f_bile, α', f_Ray and b_Mie. By using the five-fold cross-validation method described previously, we achieved 81.8% ± 1.1% sensitivity, 88.2% ± 1.4% specificity, 84.8% ± 0.6% accuracy, 0.900 ± 0.007 AUC. Also, it is important to note that parameters appearing close to the top of the decision tree are the most important for tissue classification. With this in mind, the parameters which appear on the top two layers of splits for probe 1 (figures 5(a) and (b)), i.e. f_lipid, α', f_Ray and b_Mie, are the directly or indirectly related to tissue scattering (section 4.3).

Zoom In Zoom Out Reset image size

Similarly to probe 1, figures 6(a) and (b) show the two separate branches of a decision tree with parameters leading to most of the differentiation between mucosa and tumor by using probe 2. These parameters are THb, R, f_lipid, f_bile, α', and b_Mie. By using five-fold cross-validation, we achieved 87.1% ± 0.6% sensitivity, 89.7% ± 1.0% specificity, 88.3% ± 0.6% accuracy, 0.911 ± 0.006 AUC. Differently from probe 1, the parameters which appear on the top two layers of splits for probe 2 (THb, R, f_lipid, α', and b_Mie) include both absorption and scattering parameters. Still, f_lipid, α', and b_Mie appear as important parameters for both probes and may be explored in future studies for endoscopic CRC detection.

Zoom In Zoom Out Reset image size

The relevance of our work is highlighted by the limitations overcome with respect to prior research on CRC detection during colonoscopy using DRS. Compared to previous studies, we have collected a substantially larger (2889 DRS spectra) at the extended wavelength range between 350 and 1920 nm by using fiber optic probes with two SDDs in order to probe superficial and deeper tissue layers. By using a 630 µm SDD fiber optic probe, signals could be captured from superficial tissue at depths between 0.5 and 1 mm (>0.8 mm for most wavelengths), whereas a 2500 µm SDD probe was used to collect signals from deeper tissue layers with between 0.5 and 1.9 mm (>1.4 mm for most wavelengths) [33]. Our extraction of chromophore concentrations and scattering properties is not restricted by limitations of diffusion theory, as we use probabilistic MC models to build the LUT containing reflectance values to fit the DRS spectrum. Furthermore, our LUT covers a wider range of optical properties (scattering coefficient varying from 0.1 to 1000 cm⁻¹ and absorption coefficient from 0.01 to 300 cm⁻¹) and a greater resolution (7200 reflectance values) compared to past studies, to the best of our knowledge.

In general, DRS measurements can be analyzed by using either machine learning methods [18–20, 22–29] for direct tissue differentiation or by spectral fitting models to obtain information on the tissue microstructure and biochemistry [18, 19, 30–32]. Machine learning is typically employed to build models for automated tissue discrimination without interpretation of the sources of this discrimination or to identify variables contributing most to the differentiation of two or more tissue types [18–29, 44–49]. Spectral fitting is used to extract chromophore concentrations and tissue optical properties such as the absorption and reduced scattering coefficients. The accuracy of spectral fitting models is dependent on the type of light transport model used as well the scattering and absorption properties of the medium where the light propagates [31, 40, 41, 50–56].

Analytical models provide a solution to the radiative transport equation often using the diffusion approximation (typically valid when the scattering coefficient $\mu _{\text{s},\,{\text{total}}}^{\prime}$ is much higher than the absorption coefficient ${\mu _{\text{a},\,{\text{total}}}}$) [52, 57–67] or use a semi-empirical solution (often valid for media with particular sets of chromophores at specific geometries), while probabilistic models often use MC simulations to compute average parameters based on the photon packet paths and processes (e.g. absorption) inside the medium [31, 40, 41, 50–56]. Even though probabilistic models are more time-consuming, they provide the most accurate solution for the present case, as diffusion models are not accurate for short SDDs and for all tissue optical properties [31, 40, 41, 50–56, 68–82]. Among probabilistic models, we can exploit forward MC simulations to generate reflectance LUTs (i.e. a reflectance database as a function of μ_a and μ_s) for producing theoretical reflectance spectra based on wavelength-dependent of optical properties (referenced at scattering spectra and chromophore absorption spectra equations described in the supplementary material).

Previous studies have used spectral fitting models based on diffusion theory [52, 57–67] and may not be accurate in estimating chromophore concentrations and scattering parameters from tissues with optical properties out of the range of validity of diffusion theory [31, 40, 41, 50–56, 68–82]. The spectral fitting model used in this study uses MC simulations for a variety of optical properties (μ_s ranging from 0.1 to 1000 cm⁻¹ and μ_a ranging from 0.01 to 300 cm⁻¹). Therefore, our estimation of chromophore concentrations should be more accurate than prior studies. Furthermore, we combined our determined ranges of chromophore concentrations and scattering parameters which discriminate mucosa and tumor tissues with more than 90% AUC and 84% accuracy for probes investigating tissue depth up to 1 mm (small probe) and 1.9 mm (large probe).

We combine our spectral fitting model with machine learning in order to select the thresholds of chromophore concentrations and scattering parameters for highest differentiation between mucosa and CRC tissues. The selection of these thresholds can be performed directly by using the CART algorithm with no assumptions on the data distribution (i.e. reflectance values at each wavelength). Since the decision tree resulting from CART contains the most important parameters for tissue differentiation on the top (i.e. 1st nodes) of the tree, the importance of individual parameters can be checked by the order they appear on the tree. In addition, since parameters and their thresholds can be determined by tissue differentiation, it is possible to associate them with categories of mucosa and tumors occurring in cancer patients and, in particular, to identify cases where mucosa and tumors can be differentiated from each other. Thus, the results of our analysis offer not only a way to localize tumor tissues in normal mucosa, but a way of characterizing tissue microstructure and biochemistry that may influence clinical decision-making once parameter thresholds can be assessed in future in vivo studies involving sufficient number of patients. Finally, by understanding the cancer microstructure and biochemistry, it may be possible to learn the origin of cancer and related diseases, as well as to identify cases leading to better patient prognosis.

In this study, the most important parameters for discrimination between mucosa and tumor tissues for both probes were THb, R, f_lipid, f_bile, α', and b_Mie (figures 5(a), (b), 6(a) and (b)). Particularly for the small probe (probe 1), f_Ray was considered an important parameter for that discrimination based on reflectance of superficial tissue (figures 5(a) and (b)). As reported in previous studies, THb is expected to be higher in tumors due to the associated angiogenesis process to increase the local blood supply and support tumor growth. According to our results (tables 2 and 3 and figures 3 and 4), the investigated tumors had blood vessels with smaller average radius and slightly higher StO₂ in deeper tissue layers (probe 2) compared to mucosa, which may indicate that smaller vessels are created to increase the local oxygen supply in adenocarcinomas and carcinomas. Most previous studies investigated DRS for superficial tissue layers and reported higher THb and lower StO₂ in cancer tissues in relation to uninvolved mucosa [18, 19, 30–32]. By exploiting the wavelength range between 350 and 700 nm with a 550 μm SDD probe, Zonios et al [30] studied normal and cancerous tissues of 13 patients. The authors observed, respectively, THC of 13.6 ± 8.8 mg dl⁻¹ and StO₂ of 59% ± 8% for normal tissues, whereas adenomatous polyps exhibited THC of 72.0 ± 29.2 mg dl⁻¹ and StO₂ of 63% ± 10%. A higher THC and a slight increase of StO₂ for potentially malignant tissues (adenomatous polyps) agrees with our results of probe 1 (630 μm SDD probe). In addition, the mucosal THC of our study (THC 9.2 ± 9.5 mg dl⁻¹) was comparable to that of Zonios et al, while that of adenomatous polyps was about six-fold different from the THC values found in tumors of the present study (12.0 ± 8.7 mg dl⁻¹). The difference becomes even more apparent when comparing values to findings on probe 2, where lower THC and StO₂ were found (THC of 5.6 ± 2.4 mg dl⁻¹ and StO₂ of 50% ± 17% for mucosa, and THC of 7.9 ± 3.5 mg dl⁻¹ and StO₂ of 56% ± 18% for tumors, respectively). However, even lower values of THC have been found by Wang et al [31], who studied DRS of the tumors and surrounding normal tissues of 27 patients (17 with adenomatous polyps, eight with adenocarcinoma, and two with only normal tissues) in the visible wavelength range between 600 and 800 nm. By using probes with SDDs of 0.6, 1.5, 2.5, and 4 mm, Wang et al obtained mucosal THC of 59.8 ± 19.0 µM and tumor THC of 153.8 ± 38.6 µM. Even though the THC values are comparable to those of probe 2 of our study, their mucosal THC was lower while their tumor THC is higher on average (table 2). Wang et al found a much lower StO₂ for tumors (26.4% ± 6.1%) compared to normal tissue (60.7% ± 7.7%), which does not agree with our study. Finally, the THC values for probe 2 of our study are comparable to results obtained by Nandy et al [83], who performed spatial frequency domain imaging to ex vivo colorectal specimens of nine patients from which 15 (eight malignant and seven normal) regions were imaged. Images were taken by using 460, 530 and 630 nm wavelengths and two spatial frequencies (0 and 1 cm⁻¹). The authors found that THC varied between 46 and 80 μM in normal mucosa and between 57 and 118 μM in tumors.

In terms of scattering properties, tumors exhibit lower fraction of Rayleigh scattering (f_Ray), which may occur during cell division, when a relatively lower number of structures smaller than a wavelength exist in cells in order to create conditions for duplication of cellular structures. These structures include mitochondrion membranes (as mitochondria are under fragmentation for their equivalent distribution to daughter cells on cell division [84]) and cell membranes, replicating structures required for cell division. However, it is also important to remember that the number of mitochondria under fragmentation tends to increase during cell division in tumors. Since tumors contain cells in different stages of cell division, their tissue microstructure tends to be more heterogeneous compared to normal tissues. This heterogeneity can be observed in scattering parameters (figures 3 and 4), especially for superficial tissue (figure 3). The increase in the number of mitochondria and other organelles [85] agrees with the higher Mie scattering power (b_Mie) obtained in tumors, as it suggests that cancer tissues contain overall smaller-size particles and structures compared to normal tissues. Since Mie scattering is typically generated at organelle (e.g. mitochondria), nuclei, and cell scales [86], the presence of smaller structures (in average) generating Mie scattering agrees with the higher cell proliferation expected to be found in tumors [87–89].

Nandy et al [83] investigated scattering parameters by assuming an approximation of the total tissue scattering by a fitting to Mie scattering, which allows retrieving the scattering amplitude α' and the Mie scattering power b_Mie based on the three wavelengths investigated in their study instead of a broadband wavelength range as the present study. With this in mind, the authors reported that normal mucosa had α' between 11 and 19 cm⁻¹, b_Mie between 0.78 and 1.2, whereas tumor tissue had α' between 6 and 12 cm⁻¹, b_Mie between 0.2 and 0.8. While the values of α' and b_Mie are within the same range of those found for probe 2 in our study, their trend was exactly the opposite of our findings (lower α' and b_Mie for tumors compared to mucosa). Other studies have reported values of scattering coefficient at NIR wavelengths, which could be related to α', as the wavelength-dependent scattering coefficient has a much lower slope at NIR and longer wavelengths. Zeng et al [90] investigated eight cancer, one pre-malignant polyp, and five normal mucosa regions of nine patients by using a swept source-optical coherence tomography (SS-OCT) system. This system used 1310 nm as the center wavelength and had 110 nm for the full width half maximum bandwidth. In their study, cancerous specimens exhibited mean scattering coefficients (per image) ranging from 4.54 to 9.17 mm⁻¹, whereas those of normal tissues ranged from 2.50 to 8.21 mm⁻¹. The slightly higher scattering coefficients found in tumors agrees with the results of our study, as α' was higher for cancer tissues (table 2). On the other hand, higher scattering coefficient values disagree with results reported by Zhang et al [91], who used a spectral-domain OCT system with center wavelength at 850 nm and spectral bandwidth of 40 nm to study 16 cases of malignant and normal colorectal tissues. The authors found that the scattering coefficient of 1.41 ± 0.18 mm⁻¹ for malignant tissues (variation between 0.25 and 2.69 mm⁻¹) and 2.29 ± 0.32 mm⁻¹ for normal tissues (values ranging from 1.09 to 5.41 mm⁻¹). Although scattering parameters have different trends among studies, it is important to consider that the wavelength range has a large influence on the probed depth and tissue layers investigated. Thus, microstructural parameters will differ, especially given the heterogeneity of tumor tissues. However, the range of α' and b_Mie obtained in this study (table 2 and figures 3 and 4) was still in a range, which is comparable to previous studies. Factors contributing to the difference between mucosa and cancer tissues were described by Backman and Roy [16] in terms of early increase in blood supply as well as nanostructural and microstructural changes such as alterations in the collagen fiber crosslinking, cytoskeleton, chromatin structure.

An indirectly related parameter with scattering is the lipid concentration, as lipid is present in structures which generate Rayleigh scattering such as membranes and lipid droplets [92]. Then, a lower f_lipid in tumors (tables 2 and 3 and figures 3 and 4) may mean less membranes and lipid droplets are present in the targeted cells/tissues, which leads to lower f_Ray (table 2). In both probes 1 and 2, the lipid content is lower in tumors (tables 2 and 3), which agrees with previous cancer studies [93, 94]. In a Raman spectroscopy study, Bergholt et al [93] found that colorectal tumors (adenomas and adenocarcinomas) had lower lipid concentration in relation to hyperplastic polyps and higher water content compared to normal tissues. A decrease in lipid concentration in cancer tissues is further confirmed with magnetic resonance spectroscopy in breast cancer [94]. Similar studies showed that tumors contain higher water-to-lipid ratio [95–99], which also agrees with the results of this study (table 2). Therefore, lipid and water content can be used individually or in combined parameters for cancer detection.

Finally, a combination of the parameters THb, R, f_lipid, f_bile, α', b_Mie, f_Ray was sufficient to provide classify normal mucosa and cancer tissues with 84.8% ± 0.6% and 88.3% ± 0.6% accuracy with probes 1 and 2, respectively (figures 5(a), (b), 6(a) and (b)). The parameter selection for classification of the decision trees (figures 5(a), (b), 6(a) and (b)) described nine subtypes of mucosa and tumors for probe 1 as well as ten subtypes of mucosa and tumors for probe 2. The number of tissue types shows that tissues interrogated with probe 2 tend to more heterogeneous compared to those interrogated with probe 1 (figures 5(a), (b), 6(a) and (b)). Even though the division into more tissue subtypes was expected from probing deeper tissue layers, it is important to note that the classification performed with probe 2 used less parameters (THb, R, f_lipid, f_bile, α', b_Mie) compared to probe 1 (THb, R, f_lipid, f_bile, α', b_Mie, f_Ray) and still obtained a higher tissue classification performance.

In our previous work [33], we showed that the probed depth for probe 1 ranges from 0.5 to 1 mm (0.8 to 0.9 mm for most wavelengths between 430 and 1590 nm), whereas that of probe 2 mostly between 0.5 and 1.9 mm (higher than 1.4 mm for most wavelengths between 430 and 1590 nm). Even though it is possible to quantify the probed depth, specifying which tissue layers are being probed in the colorectal wall still remains a challenge. A typical thickness of the total wall is about 2 mm, which may also be changed by a number of factors [100]. Individual thicknesses of the tissue layers may have a wide range of thicknesses. According to Castro-Poças et al, mucosa layers can vary from 0.3 to 1.2 mm, submucosa can vary between 0.2 and 1.8 mm and muscularis propria can vary from 0.3 to 2.4 mm [100]. This variance may contribute to the heterogeneity of the biochemical and microstructural parameters obtained in the present study (table 2). However, when comparing these parameters per probed depth, THb, THC, f_lipid, f_water extracted from DRS spectra of probe 2 exhibited a relatively lower coefficient of variation compared to those of probe 1. This result suggests that probing more tissue layers with probe 2 did not generate more inhomogeneity on absorption parameters such as THb, lipid and water concentrations. It is unclear how probing deeper tissue layers affects other parameters and it is expected that multilayered tissue structures influence the scattering parameters reported in this study.

In addition, our results indicate that deeper tissue layers of mucosa and tumor tend to have higher average vessel radius (R), lipid concentration (f_lipid), and Mie scattering power (b_Mie), whereas all other parameters tend to be lower for probe 2 (table 2 and figures 3 and 4). For tumors, a slightly higher R in deeper tissue layers may happen due to probing relatively less thin blood vessels (compared to the total number of vessels in the probed volume) created in the tumor surface during angiogenesis. A similar conclusion may be drawn from lower THb and THC (table 2 and figures 3 and 4), as most of the blood volume would be present on the superficial tissue layers. Since less blood vessels occur in deeper tumor layers, it would be expected a slower cell division, larger-sized cells, and less fragmented mitochondria. These three expected features lead to less refractive index mismatches, which agrees with our decreased scattering amplitude α'. Assuming Mie scattering is mostly generated by mitochondria [86], a lower number of mitochondria (larger size) would mean lower b_Mie, as per our results. Also, lower cell division rates mean less fragmented mitochondria, which generates less refractive index mismatches due to water-membranes transitions and thus leads to lower f_Ray. Finally, f_water was expected to be lower in the tissue surface due to dehydration caused during our ex vivo measurements. On the other hand, our procedure of keeping the tissue hydrated with a damp cloth from times to times during our measurements may have affected the water content extracted with our spectral fitting algorithm.

Our study has several limitations due to variations caused in ex vivo tissues and contact measurements. First, the interrogated cancer tissues were from diverse stages with predominance of advance stage tumors (pT2 and above). The measured tumor sites were selected based on the palpation and naked eye determination by experienced surgeons who demarcated the tumor region, although all these regions were confirmed subsequently with histology. It is important to note that this determination does not decrease the validity of our dataset and biochemical/microstructural analysis, as our dataset is larger compared to previous studies in the field (1363 spectra for probe 1 and 1526 for probe 2) and contains DRS spectra with variations due to parameters discussed in this section. With this in mind, early-stage cancer changes are expected to occur over the same tissue biomolecules and microstructure parameters but with lower magnitude in relation to advanced cancer.

Since our measurements were performed in ex vivo colorectal specimens, the blood oxygenation was not controlled during our data collection. However, we confirmed there was no significant variations on average Hb and HbO₂ in mucosa and tumor regions in a pilot ex vivo observation of the DRS signal in three patients. Our observation was limited by seven mucosa and seven tumor sites during the 1st 15 min of our measurements (data not shown). We are also aware that THb and StO₂ may increase and f_water may decrease in ex vivo tissues compared to in vivo, as described by Baltussen et al [101] in a previous study investigating different types of colorectal tissues (fat, tumor, and healthy colorectal wall). It is also important to note that, although the authors used the same optical technique (DRS) to draw those conclusions, their probe has a different SDD (1.29 mm center-to-center distance) compared to those of this study and their spectral fitting model uses diffusion theory approximation to estimate THb, StO₂ and f_water. In addition, the behavior of StO₂ is still not well understood, as previously reported studies disagree on their StO₂ trend on ex vivo tissues. While Baltussen et al reported an increased StO₂ in ex vivo tissues within 1 h after resection in relation to in vivo tissues, Salomatina et al [102] reported a decrease in StO₂ in mouse ear tissues 5–10 min after excision (ex vivo) of tissue and after 24 and 72 h of storage. Even though StO₂ varies, it was not considered as a relevant parameter in the classification of our decision tree. Also, we attempted to reduce tissue dehydration as much as possible by keeping the tissue moisture with a water wet wipe once every seven spectral measurements in average during the measurements. We attempted to not overhydrate tissues as they were wiped by not using an overly wet wipe. As illustrated by our results, only a small deviation on water content was found in tumor measurements, which can be mainly attributed to the tissue heterogeneity expected from cancer. Furthermore, if overhydration due to uncontrolled wiping (relatively 'random' event) had influenced our measurements, higher standard deviations in water concentration should have been associated with higher water-concentration measurements (i.e. tumor measurements), which is not what we observed.

We acknowledge that probe contact pressure and sufficient optical contact with interrogated tissues may be challenging to accomplish during in vivo endoscopy. With this in mind, a reduced sensitivity to optical contact and probe positioning is expected in our probe 2 due to its larger SDD, which may be related to higher measurement repeatability and less standard deviation present in some parameters of this probe. However, larger SDDs also means larger probed volume, which may cause more variations upon probe contact pressure if this pressure allows collecting the signal from heterogeneities of deeper tissue (e.g. interface between tissue layers). This type of variation was previously reported [103] and needs to be taken into account in order to correct DRS spectra. One should also note that the incorporation of large SDD probes is feasible in colonoscope channels, as long as the probe fits the diameter of the instrument channel (typically 2.8–4.2 mm [104, 105]).

Finally, not including non-cancer pathology in our study may change the classification performance of our CART model. With this in mind, future studies will include data on non-cancer pathology. On the other hand, the lack of non-cancer pathology does not affect our tissue biochemical/microstructural parameters. Therefore, comparison of these parameters with other studies is valid, especially considering our dataset includes a robust sampling of intra- and inter-patient variations (1363 + 1526 spectra of 47 patients) in both superficial and deeper tissue layers.