2.1.

Hyperspectral Camera Specifications

Two different HS cameras based on different acquisition techniques have been used to capture in vivo brain images. The technical specifications of both cameras, sensors, and lenses are presented in Table 1.

Table 1

Sensor and camera optics specifications of the different HS cameras used. The last five parameters are fixed during the acquisition of images in the operating room.

On one hand, the snapshot HS camera has a complementary metal oxide semiconductor (CMOS) sensor holding a $5\times 5$-mosaic pattern with a pixel size of $5.5\mu \mathrm{m}$ (MQ022HG-IM-SM5X5-NIR, Ximea GmbH, Münster, Germany). The analog to digital converter (ADC) of the sensor provides images with a resolution of 8 bits. In addition, a long pass filter (FELH0650, Thorlabs, Inc., Newton, New Jersey, United States) with a 650-nm cut-on wavelength is placed in front of the lens to remove non-negligible secondary harmonics, which were determined by the manufacturer in the spectral response curves during sensor production. Although the sensor resolution is $2048\times 1088\text{pixels}$, the active area of the sensor with the built-in spectral filters has $2045\times 1085\text{pixels}$. This reduced area is called the active filter zone, which omits the last three rows and last three columns from the total sensor resolution. Each capture with this snapshot camera produces an image containing the spatial and spectral resolution due to the $5\times 5$-mosaic pattern. Each mosaic contains approximately the same spatial pixel at different 25 wavelength bands within the NIR spectrum, specifically in the 660 to 950 nm spectral range. In addition, these bands are spaced among each other with a mean and standard deviation value of $12.11\pm 2.64\mathrm{nm}$. Hence, the mosaic pattern reduces the 2D output spatial resolution by a factor of 5 to obtain a 3D HS cube. In particular, the image with $2045\times 1085\text{pixels}$ generated by the sensor is arranged into a $409\times 217\times 25$ HS cube to perform the spectral analysis. The main advantage of the snapshot camera is its capability for real-time solutions, understood as processing a sequence of HS images to provide a live video of the scene. On the other hand, the other camera is based on the pushbroom linescan technology (Micro-Hyperspec^® E-Series, HeadWall Photonics Inc., Bolton, Massachusetts, United States), which holds a scientific CMOS sensor with an ADC of 16 bits and a pixel size of $6.5\mu \mathrm{m}$. The sensor acquires a single-spatial line with 1600 pixels and 394 wavelengths of information. Thus, the camera needs to be moved with an actuator to scan an image with as many lines as desired. All in vivo brain captures used in this study were scanned with 500 lines, producing images with a spatial resolution of $1600\times 500\text{pixels}$ and 394 spectral bands. Besides, the exposure time and frame period were set to 150 and 160 ms, respectively. Although the sensor is sensitive in the 365 to 1004 nm spectral range, those wavelengths acquired outside the 400 to 1000 nm spectral range need to be removed, as specified by the manufacturer. Eliminating such bands results in 369 effective wavelength bands separated by $1.62\pm 0.00\mathrm{nm}$ from each other.

It is worth noting that captures from both cameras were cropped spatially to help neurosurgeons during the labeling processing. Therefore, the spatial resolution of the HS linescan or snapshot captures are smaller than $1600\times 500\text{pixels}$ or $409\times 217\text{pixels}$, respectively. Although the pusbroom linescan camera provides more spatial resolution and spectral information, it is not suitable for real-time solutions due to the scanning procedure. The time spent to scan a brain image requires, $\sim 1\min$ and 40 s, whereas an image captured with the HS snapshot camera takes 100 ms.

2.2.

Acquisition System

The acquisition system used to gather the in vivo brain images is presented in Fig. 1. Starting from the left, the HS $5\times 5$-mosaic snapshot camera is located inside a 3D-printed white case. Inside the case, there is a servomotor employed to help focus the camera. An external light source with a 150 W 21 V EKE halogen lamp (MI150, Dolan-Jenner, Boxborough, Massachusetts, United States) and a dual gooseneck fiber optic (EEG28, Dolan-Jenner) were used to illuminate the brains. While the housing of the light source does not appear in the image, the dual fiber optics are shown glowing in Fig. 1. The system has been utilized as a real-time augmented reality (AR) application through laser imaging detection and ranging (LiDAR) equipped with an RGB camera. This technology enables the collection of geometric data of the brain surface by extracting a point cloud from the scene using depth and RGB information. The manufacturer indicates that the depth accuracy of the system depends on reflectivity and light conditions and is effective within a range of 0.5 to 3.86 m. Specifically, the random error, measured as the standard deviation, is $\le 17\mathrm{mm}$, whereas the typical systematic error is $<11\mathrm{mm}$ plus 0.1% of the measured distance, provided there is no multi-path interference. By combining depth information from the point cloud, ML classification results on HS data, and the RGB from the LiDAR camera in an AR interface, the effectiveness of surgical field exploration and tumor delineation can be enhanced.¹⁹ In this study, LiDAR is specifically used to measure the distance between the acquisition system and the brain surface. This measurement helps to focus the system accurately by determining the distance to a single point on the brain surface. Given that the linescan camera has a fixed focusing distance of 60 cm, the measurement of the LiDAR is necessary to ensure focused captures with the linescan camera. Therefore, all images were taken at a distance of 60 cm to ensure a fair comparison between the two cameras, taking into account the limited focus distance of the linescan camera. Furthermore, the motorized linear stage (X-LRQ300HL-DE51, Zaber Technologies Inc., Vancouver, Canada) below the imaging sensors is mainly used to move the HS linescan camera. This movement allows the scanning of the brain image to compose a HS cube with the desired spatial resolution. The sensors and fiber optics are all aligned in the same plane, which is perpendicular to that of the linear stage holding them. Although not included in Fig. 1, two additional motorized linear stages are used to tilt and provide height to the stage.

Fig. 1

Front view of real acquisition system with main components highlighted and labeled.

2.3.

Data Processing

All HS images, regardless of the camera used, are pre-processed using almost the same procedures, which are presented in Table 2.

Table 2

Pre-processing steps performed on the images captured with the different HS cameras used.

The first step is to obtain the reflectance information of the material interrogated by calibrating the raw information obtained with the sensor. Equation (1) eliminates the effect of the HS sensor and the lighting conditions captured with the raw images to obtain the reflectance information $R$ from the sample,

The second step is related to the formation of the HS cubes. The snapshot sensor captures 2D images that need to be rearranged into a 3D cube, which has a spatial dimension five times smaller than the 2D images due to the mosaic pattern of the sensor described in Sec. 2.1. However, the pushbroom camera has a simpler process to conform an HS cube. Such a process is called line stitching, which requires a precise linear actuator to move the HS camera to join adjacent spatial lines while avoiding overlap. The spectral correction is the third step and only applies to the snapshot camera. This correction is necessary to correct the response curves of the HS snapshot sensor, which presents crosstalks between adjacent pixels of the sensor that vary with the angle of incident light to the sensor.²² To correct this effect, the manufacturer provides a spectral correction matrix that modifies the response of the sensor to obtain ideal Gaussian curves. The spectral correction process is described in Eq. (2)

2.4.

In Vivo Human Brain Images

For this study, we used 10 in vivo brains from adult patients who have provided informed consent prior to surgery. These 10 patients were selected because two or more sterilized rubber rings were placed by neurosurgeons prior to acquire the HS images. Different colors were used for tissue identification, with green and black being used for healthy and pathological tissues, respectively. These rubber rings can be seen in the pseudo RGBs of Fig. 2 for patients with ID 190. In addition, the same images for the rest of the patients are presented in Fig. S1 in the Supplementary Material. As indicated previously, images of patients who have suffered different pathologies have been used. More specifically, patients with ID 177, 183, 184, 190, 193, and 203 have GB with isocitrate dehydrogenase (IDH) non-mutated. Patient ID 185 has grade II astrocytoma with IDH mutation, patient ID 192 has a metastatic testicular tumor, and patient ID 194 has a metastatic lung carcinoma tumor. Finally, patient ID 201 exhibits pseudoprogression in a GB with IDH non-mutated, indicating apparent GB progression likely attributable to treatment effects rather than actual tumor growth.

Fig. 2

Pseudo-RGB with healthy and pathological ROIs for patient ID 190. (a) Patient ID190-Snapshot. (b) Patient ID190-Linescan.

The guidelines of the Declaration of Helsinki have been followed, and the acquisition of HS images has been approved by the Research Ethics Committee of Hospital Universitario 12 de Octubre, Madrid, Spain (protocol code 19/158, May 28, 2019). All patients shaved the area to be operated on prior to the scalp incision. Then, a high-speed drill was employed to make burr holes in the skull, which are used to insert a cranial drill to perform the craniotomy. This procedure extracts a bone flap to expose the dura of the patient, and then, the durotomy is performed by cutting with a knife the dura to uncover the brain surface. Then, both HS cameras proceeded to acquire the in vivo brain surface.

Using the rubber rings ensures that the data of both cameras comes from the same spatial location of the brain surface. Furthermore, this procedure is similar to the one employed by Mühle et al.,²³ who used a plastic cursor to delimit a bordered region of interest (ROI) over biological organs to compare different HS cameras. In addition, the biomedical experts selected similar areas of size $21\times 21\text{pixels}$ inside the plastic cursor to perform their analysis. In our case, the areas we selected inside the rubber rings have a size of $5\times 5\text{pixels}$ since bigger ROIs got pixels outside the rings in the snapshot images. Although in the previous study, Mühle et al. took 10 HS snapshot images to average them; in our study, it was not feasible since the in vivo human brain is in motion due to the heartbeat. Moreover, the linescan camera is inevitably affected by the motion of the brain during the scanning procedure, hindering the possibility of averaging multiple captures. For these reasons, one HS snapshot and one HS linescan image were taken for each patient. A single linescan image takes $\sim 1\min$ and 40 s, whereas a snapshot measurement takes 100 ms. Notice that specular reflections have not been removed, as neither the cameras nor the light source had polarizers. Although the patients exhibit diverse pathologies, the spectral comparisons are conducted on an intra-patient basis. This entails the comparison of the spectral signatures of both cameras for each patient individually.

2.5.

Spectral Similarity Metrics

A way to compare both HS cameras is to employ spectral similarity metrics, which can assess how different reflectances are related to each other. Agarla et al. have examined 14 frequently used measures and grouped them into five categories based on the type of error they evaluate,²⁴ including the mathematical definition and implementation of all metrics. Furthermore, as stated by Agarla et al., selecting one measure for each of the groups, they describe can be sufficient to assess the spectral similarity.²⁴ For that reason, we decided to use the root mean square error (RMSE), the goodness-of-fit coefficient (GFC),²⁵ and the spectral angle mapper (SAM)²⁶ metrics. On one hand, RMSE can range from 0 to 1 since the maximum value of our data is 1, indicating a value of RMSE equal to 0 perfect similarity. On the other hand, GFC values can be between the 0 and 1 range, where a value of 1 indicates complete similarity. SAM values range from 0 to 1, indicating values closer to 0 with high similarity and values closer to 1 with low similarity.

These metrics are chosen since they are applicable to the spectral domain, can be used as loss functions (except SAM since angular metrics is not used to measure losses), and do not need extra requirements to be computed. Furthermore, the spectral signatures of the HS linescan camera are considered as the reference spectra, whereas those obtained with the HS snapshot camera are considered as the arbitrary signal. The previous assumption is grounded on the fact that the HS linescan camera gathers greater spatial and spectral data compared with the snapshot camera. We compute and compare the mean spectral signatures as illustrated in Fig. 3. For each patient, we independently obtain the mean spectral signatures of the pixels in the ROIs located inside the rubber rings, as shown on the left side of Fig. 3.

Fig. 3

Procedure to compute the spectral similarities acquired with both HS cameras. The comparison is made independently by patient and tissue, using the spectra shared by both cameras.

Since the HS snapshot camera captures less spatial resolution, less pixels are included in the green and red ROIs inside the rubber rings compared with those captured by the HS linescan. Specifically, the ROIs created for all snapshot captures have a $5\times 5$ spatial dimension. Therefore, to provide a fair comparison, we created bigger ROIs for the HS linescan but randomly selected 25 pixels. Then, with those 25 pixels for each tissue and camera, the mean spectral signature and standard deviation are computed as presented with the plots in the middle of Fig. 3, which include black dashed rectangles to indicate the part of the spectrum shared by both cameras and used to compare them with the spectral similarity metrics. Finally, the RMSE, GFC, and SAM metrics are computed for each tissue using the matched wavelengths of the cameras presented in Table S1 in the Supplementary Material.

2.6.

Analyzing Absorbance Measurements to Identify Chromophore Absorption Peaks

To analyze the absorbance ($A$) spectral signatures and look for patterns that might indicate the presence of certain chromophores, we will use the reflectance ($R$) measured by the cameras. $A$ is commonly derived,^27–29 for each wavelength, from $R$ using Eq. (4)

This expression is derived from the Beer-Lambert law,³⁰ which expresses $A$ as $A={\log }_{10}(I_o/I)$. Here, $I_o$ represents the incident light, and $I$ is the light that has passed through the sample. In our specific context, we are dealing with reflectance information as defined in Eq. (1). The maximum reflected light corresponds to $I_{\text{white}}$ (analogous to what would be $I_o$ in the Beer-Lambert law), and $I_{\text{raw}}$ represents the light that has passed through the brain (similar to $I$ in the Beer-Lambert law). To account for the noise of the sensor, we include the dark measure from Eq. (1) ($I_{\text{dark}}$). This allows us to mitigate sensor noise, and as a result, we arrive at Eq (4) $A={\log }_{10}((I_{\text{white}}-I_{\text{dark}})/(I_{\mathrm{raw}}-I_{\text{dark}}))={\log }_{10}(1/R)=-{\log }_{10}(R)$.

The chromophores we will attempt to identify are those that contribute the most to absorb light in the NIR region (from 800 to 2500 nm) in adult brains. As stated by Correia et al., these are hemoglobin, water, lipid, and the following cytochromes: cytochrome aa3 (Cyt aa3), cytochrome b (Cyt b), and cytochrome c (Cyt c).³¹ The absorption spectra of the previous cytochromes on their redox state between 400 and 1000 nm were obtained from the Biomedical Optics Research Laboratory (BORL) Github repository.³² It is worth noting that we converted the molar extinction coefficient for the Hb to the absorption coefficient as specified by Prahl et al.,³³ considering that for the whole blood, there is 150 g of Hb per liter. Although this assumption may be doubtful for the measurements taken, it allows us to visually compare the spectrum of all chromophores with the same units. Also, it is worth noting that the absorbance, $A$, and the absorption coefficient, ${\mu }_a$, represent different phenomena. On one hand, $A$ is a property of a material that measures the fraction of light that can pass through in terms of intensity. On the other hand, ${\mu }_a$ is a property of the material that describes its effectiveness in absorbing light. We know from the Beer-Lambert law³⁰ that $A$ and ${\mu }_a$ are related through Eq. (5):

3.1.

Diffuse Reflectance Standard Measurements

To validate the spectral performance characterization of the HS cameras and enhance the reliability of the subsequent analysis of the patient data, reference measurements were conducted on the Zenith Polymer Reflectance Standard, which exhibits nearly ideal Lambertian 99% of diffuse reflectance (SphereOptics GmbH, Herrsching am Ammersee, BY, Germany). The spectral response of the polymer reference is provided by the manufacturer and is illustrated with a gray line in Fig. 4. Moreover, the Pearson correlation has been used between the system measurements and the reference polymer signature to evaluate the performance of the HS cameras. Furthermore, a miniature spectrometer (Ocean Insight, Orlando, Florida) was employed to analyze the spectral responses of the polymer reference. This device is capable of measuring in the visible (VIS) and near-infrared (NIR) spectrum from 350 to 925 nm, with an ADC resolution of 16 bits. For the purposes of this study, measurements taken with the spectrometer from 400 to 925 nm are presented with a red dashed line in Fig. 4, as this includes the spectral range under analysis in the following sections. The correlation obtained when using 1913 bands from the spectrometer is 96.91%. In the same figure, the results of the measurements taken with both HS cameras are illustrated with green dashed-dotted lines and orange dotted lines for the linescan and snapshot cameras, respectively. The correlation obtained with the linescan camera using 369 bands is 95.58%, whereas for the snapshot camera, it is 68.19% using 25 bands. Although the correlation with the snapshot camera is lower than that obtained with the spectrometer or the other HS camera, the orange line clearly demonstrates that the spectral response is relatively similar between 660 and 866 nm, with a Pearson correlation value of 95.55% using the 17 bands in the aforementioned range.

Fig. 4

Normalized reflectance spectral responses when the zenith polymer reflectance standard was illuminated with different sensors. The spectrometer measurement was conducted with a fiber optic positioned orthogonally to the polymer reference at a distance of 5 cm. The HS camera measurements were obtained by capturing a scene of the polymer at $\sim 40.5\mathrm{cm}$ distance, utilizing an ROI of $25\times 25\text{pixels}$. The responses from the HS cameras represent the mean spectral signatures of the polymer pixels in the ROI, with their corresponding standard deviation, as indicated by the shaded color between the mean and the corresponding data points. The spectrometer employs 1913 bands and encompasses the 400 to 925 nm spectrum. The HS linescan VNIR camera covers the 400 to 1000 nm spectral range using 369 bands. Finally, the HS snapshot NIR camera spectrum range covers from 660 to 950 nm using 25 bands. The Pearson correlation coefficient is presented after comparing the measured bands of each sensor with the polymer reference response.

3.2.

Brain Tissue Measurements

The reflectance measurements obtained with both HS cameras from the 10 in vivo human brains are shown in the Supplementary Material, specifically in Fig. 5. The illustrated spectral signatures have been obtained, for every camera measurement, using 25 pixels located inside the rubber rings as shown in the images in Fig. 2. We decided to analyze the effect of normalizing the calibrated and denoised data to check how it would influence on the spectral similarity metrics. Hence, the two columns to the left in Fig. 5 are data that have been calibrated and denoised, whereas the two columns to the right are the same data that have been normalized using Eq. (3). The spectral range analyzed to address the comparison is between 659.95 and 951.42 nm using 25 bands detailed in Table S1 in the Supplementary Material for each camera. When looking at the unnormalized data, similar spectral signatures are measured with both HS cameras for eight out of the 10 in vivo human brains. Note how the spectral signatures of both cameras for patients 192 and 193 are less similar compared with other patients for either the healthy or pathological tissues. Furthermore, the healthy tissue measurements from the different patients have values between 0.2 and 0.6 in most cases, excluding patient 192, which has values between 0.4 and almost 1.0. Moreover, the reflectance values for pathological tissues range from 0.1 to 0.4 for most patients, excluding patient 185 whose values are between 0.5 and 0.8. The variations in reflectance or spectral signatures could result from the biological differences between the patients and their distinct brain tumor pathologies. Overall, the mean spectral signatures of both cameras differ more from each other once the normalization is applied. For instance, the healthy tissue measurements of patient 183 show a 0.3 reflectance difference between the spectral signatures of both cameras, whereas the unnormalized data for the same patient has a variation of less than 0.05. Such behavior can be seen for most patients and tissues except the healthy spectral signatures of patient 192 because it already exhibits a great difference in the unnormalized measurements.

Fig. 5

Mean spectral signatures with standard deviation for every patient in the study. Data comes from the 25 pixels included inside the rubber rings captured with both HS cameras. Continuous curves are from the snapshot camera, whereas dashed line curves are from the linescan camera. Plots are in green and red to represent the healthy and pathological tissues, respectively. The two columns on the left are the spectral signatures when data has been calibrated and denoised, whereas in the two columns to the right, data has additionally been normalized using a min-max normalization.

In addition, we can see an increase in the standard deviation shown in shaded colors around the mean curves due to the normalization. It is worth noting that the location of each pixel on the curved brain surface implies light variations, leading to slightly different measured reflectances. This, in turn, increases the difference between pixels within the same ROI when normalizing the reflectance to a range between 0 and 1. While measurements between cameras display a consistent trend across patients and tissues, snapshot measurements show more noise after normalization compared with the smoother linescan camera measurements. This results in more pronounced peaks and valleys in the spectrum of the snapshot measurements. However, there is a common peak across most measurements at $\lambda \approx 850\mathrm{nm}$, which is influenced by the laser used by the LiDAR while capturing only with the snapshot camera. Such a peak can be seen in patient 203 measurements in Fig. 5. Variations of the amplitude of such peak are due to the different angles at which the patients were captured. By comparing measurements with and without normalization, it is worth noting that the aforementioned peak at $\lambda \approx 850\mathrm{nm}$ is more pronounced once the normalization is applied.

To get an overview of the measurements, we illustrate in Fig. 6, the averaged results of all patients for both tissues presented in Fig. 6. The circular marks indicate the closest wavelengths between cameras specified in Table S1 in the Supplementary Material. By zooming into the common spectral range of the cameras between 659.95 and 951.42 nm from the previous plots, we present subfigures (c) and (d) to address in further detail the difference between measurements. Note the influence of the infrared peak emitted by the LiDAR at $\lambda \approx 850\mathrm{nm}$, illustrated in the dashed rectangles. Such a peak is only present in the snapshot measurements since the LiDAR could only be turned off when capturing with the linescan. By normalizing the pixels from subfigures (c) and (d), we obtained the mean spectral signatures with standard deviation for both tissues in subfigures (e) and (f).

Fig. 6

Mean reflectance spectral signatures with a standard deviation of all pixels included inside the $5\times 5$ ROIs from both HS cameras. The shorter spectral signatures from 661.61 to 951.42 nm with continuous lines correspond to the snapshot camera in panels (a) and (b), whereas the longer spectral signature with dashed lines corresponds to the linescan camera. On one hand, plots in panels (c) and (d) are the same as panels (a) and (b), but with emphasis on the area common to both. On the other hand, panels (e) and (f) are the normalized spectra from panels (c) and (d). The dashed rectangles indicate the spectral bands of the HS snapshot camera influenced by the infrared of the depth camera. (a) Calibrated and denoised healthy tissue. (b) Calibrated and denoised pathological tissue. (c) Calibrated and denoised healthy tissue after band removal. (d) Calibrated and denoised pathological tissue after band removal. (e) Calibrated, denoised, and normalized healthy tissue after band removal. (f) Calibrated, denoised, and normalized pathological tissue after band removal.

The analysis comparing the absorbance measurements using both HS cameras with the absorption coefficient spectra of deoxy-hemoglobin (Hb), oxy-hemoglobin (${\mathrm{HbO}}_2$),³³ Cyt aa3, Cyt b, Cyt c,³² and mammalian fat³⁴ is illustrated in Figs. 7 and 8. In addition, a detailed analysis of the results obtained with respect to the absorbance measurements is provided in Sec. S4 in the Supplementary Material, including Fig. S2 with all relevant absorption peaks for the chromophores under analysis.

Fig. 7

Mean absorbance spectral signatures of the 25 pixels included inside the rubber ring for both HS cameras. Data is calibrated and denoised. The shorter spectral signatures from 659.95 to 950.64 nm with continuous lines correspond to the snapshot camera, whereas the longer spectral signature with dashed lines corresponds to the linescan camera. The spectra of Hb and ${\mathrm{HbO}}_2$³³ are shown in panels (a) and (b) with continuous and dashed magenta lines, respectively, whereas cytochromes b, c, and aa3³² are shown in panels (c) and (d) with different shades of oranges. Furthermore, continuous orange lines correspond to the absorption coefficient spectra of mammalian fat in panels (e) and (f).³⁴ Hb, ${\mathrm{HbO}}_2$, Cyt. b, Cyt. c, Cyt. aa3, and fat are in ${\mathrm{cm}}^{-1}$, whose scale is in the right $y$-axis, whereas absorbance data measured with both cameras have its scale in the left $y$-axis. The A, B, and C dashed rectangles are used to indicate different absorption peaks of Hb, ${\mathrm{HbO}}_2$, Cyt. b, Cyt. c, Cyt. aa3, or fat, whereas the D rectangle points out a water absorption peak at $\lambda =976\mathrm{nm}$. (a) Healthy tissue. (b) Pathological tissue. (c) Healthy tissue. (d) Pathological tissue. (e) Healthy tissue after band removal. (f) Pathological tissue after band removal.

Fig. 8

Mean absorbance spectral signatures of the 25 pixels included inside the rubber ring for both HS cameras. Data is calibrated, denoised, and normalized. The shorter spectral signatures from 659.95 to 950.64 nm with continuous lines correspond to the snapshot camera, whereas the longer spectral signature with dashed lines corresponds to the linescan camera. The spectra of Hb and ${\mathrm{HbO}}_2$³³ are shown in panels (a) and (b) with continuous and dashed magenta lines, respectively, whereas cytochromes b, c, and aa3³² are shown in panels (c) and (d) with different shades of oranges. Furthermore, continuous orange lines correspond to the absorption coefficient spectra of mammalian fat.³⁴ Hb, ${\mathrm{HbO}}_2$, Cyt. b, Cyt. c, Cyt. aa3, and fat are in ${\mathrm{cm}}^{-1}$, whose scale is in the right $y$-axis, whereas absorbance data measured with both cameras have its scale in the left $y$-axis. The A, B, and C dashed rectangles are used to indicate different absorption peaks of Hb, ${\mathrm{HbO}}_2$, Cyt. b, Cyt. c, Cyt. aa3, or fat, whereas the D rectangle points out a water absorption peak at $\lambda =976\mathrm{nm}$. (a) Healthy tissue. (b) Pathological tissue. (c) Healthy tissue. (d) Pathological tissue. (e) Healthy tissue after band removal. (f) Pathological tissue after band removal.

3.3.

Reflectance Spectral Similarities to Compare Both Cameras

The spectral similarities between the reflectances measured with both HS cameras are presented in Fig. 9. This figure shows raincloud plots³⁵ for the healthy and pathological tissues with green and red colors, respectively. Raincloud plots are a useful graphical representation that addresses the challenge of data obfuscation in the presentation of error bars or box plots. These visualizations combine various data elements to display raw data points, probability density through half violin plots, and key summary statistics such as median, first and third quartiles, outliers with black diamonds, and relevant confidence intervals via boxplots. This combination produces a visually appealing and adaptable representation with minimal repetition. In addition, these plots have a red dot inside each box plot to illustrate the mean value of each distribution.

Fig. 9

Raincloud plots with SAM, GFC, and RMSE similarity metrics. The left column plots [(a), (c), and (e)] are computed when pixels in the ROIs have been calibrated and denoised, whereas plots in the right column [(b), (d), and (f)] have additionally been normalized (after removing non-matched wavelengths of the HS linescan camera) using the min-max normalization from Eq. (3). Green distributions and dots indicate healthy tissue, whereas in red, they indicate pathological tissue. There is one dot in every distribution for every patient described in Sec. 2.4. Big red dots joined by a red line indicate the mean value of the distribution. (a) SAM metric for calibrated and denoised data. (b) SAM metric for calibrated, denoised, and normalized data. (c) GFC metric for calibrated and denoised data. (d) GFC metric for calibrated, denoised, and normalized data. (e) RMSE metric for calibrated and denoised data. (f) RMSE metric for calibrated, denoised, and normalized data.

The pair of red dots shown in every plot of Fig. 9 are also connected to each other with a red line to visually see which distribution has a higher mean value. Furthermore, in these charts, each dot located below the box plots represents a single value of the corresponding similarity metric, which refers to the comparison of the mean spectral signatures for a particular patient and tissue type. For example, the leftmost green dot in Fig. 9(a) represents the SAM value of 0.029 obtained when comparing the healthy tissue of both cameras for the patient with ID 193.

For comparison purposes, each row in Fig. 9 shows a pair of raincloud plots for every metric described in Sec. 2.5. Plots located to the left of the figure are the similarity results computed when data were calibrated and denoised [Figs. 9(a), 9(c), 9(e), and 9(g)], whereas those to the right are obtained when data were additionally normalized [Figs. 9(b), 9(d), 9(f), and 9(h)]. The obtained metrics for the comparison of each patient, including both tissues under analysis, are presented in Table S2 in the Supplementary Material for further analysis. To evaluate the dispersion of the distributions, we will look at the interquartile range (IQR) which is computed as $\mathrm{IQR}=Q3-Q1$, being Q1 and Q3 the first and third quartiles of the distribution, respectively. The IQR is a measure of the spread of data that is robust against extreme values. It provides valuable information about the variability of the central portion of a distribution and is helpful for identifying potential outliers. Besides, a comprehensive examination of the outcomes comparing the reflectance measurements from both cameras is presented in Sec. S5 in the Supplementary Material.

3.4.

Identification of Chromophores in Absorbance Measurements

After studying the similarity of the reflectance measurements between the HS cameras in Sec. 3.2, we now attempt to identify the presence of any of the chromophores mentioned in Sec. 2.6 within the measured absorbances. Table S3 in the Supplementary Material presents SAM values resulting from the comparison between mean absorbance spectral signatures acquired using both HS cameras and the absorption coefficient spectra of the chromophores discussed in Sec. 3.2. The wavelength of the peaks of interest, the respective analyzed spectral ranges, and the number of bands considered are detailed. Notably, the comparison is made with data obtained from the linescan cameras in the VNIR, using all 369 wavelengths measured by the camera, and NIR regions for the snapshot and linescan cameras as well. For this latter case, we use the 25 overlapping bands between the two cameras, indicated in Table S1 in the Supplementary Material. To obtain the SAM values, we extracted the peak wavelength of interest and the fifteen wavelengths on each side of it for the chromophore under analysis. This ensures that the extracted bands correspond to the wavelength of interest and its nearby range. We have chosen to search for a maximum of 15 wavelengths on each side of the peak, as this allows us to extract spectral signatures with sufficient information. We then find the closest corresponding wavelengths between the selected chromophore bands and those measurable by each camera. It is important to note that it will not always be possible to use all 31 bands from the chromophore to identify a peak. This is because the spectral resolution of the cameras is lower than that used to measure the chromophores; hence, there will generally be fewer camera bands than those measured for each chromophore. For example, if the chromophore peak has a very narrow bandwidth, such as that from Cyt. b in its reduced state at $\lambda =555\mathrm{nm}$, there will not be many camera bands capable of measuring it. This also applies to searching for the closest snapshot camera wavelengths in the chromophores. In addition, we have included the mean spectral signatures of $A$ measured by the cameras together with the ${\mu }_a$ spectral signatures of the chromophore peaks analyzed in Figs. S3–S18 in the Supplementary Material, allowing the reader to visually check the measurements.