Open Access
Add to BibSonomy
Abstract
Complete recognition of necrotic areas during small bowel tissue resection remains challenging due to the lack of optimal intraoperative aid identification techniques. This research utilizes hyperspectral imaging techniques to automatically distinguish normal and necrotic areas of small intestinal tissue. Sample data were obtained from the animal model of small intestinal tissue of eight Japanese large-eared white rabbits developed by experienced physicians. A spectral library of normal and necrotic regions of small intestinal tissue was created and processed using six different supervised classification algorithms. The results show that hyperspectral imaging combined with supervised classification algorithms can be a suitable technique to automatically distinguish between normal and necrotic areas of small intestinal tissue. This new technique could aid physicians in objectively identify normal and necrotic areas of small intestinal tissue.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Acute mesenteric ischemia (AMI) is an often underestimated, life-threatening gastrointestinal and vascular disease that occurs at various ages [1,2]. AMI accounts for 1-2% of patients with acute abdominal pain. Mesenteric artery embolism, mesenteric artery thrombosis or mesenteric vein thrombosis can cause inadequate perfusion of the small intestine, leading to ischemic necrosis [3]. Ischemia of intestinal tissues can also be caused by mechanical impact on the intestine (e.g., distortion, volvulus, intussusception), which can lead to intestinal necrosis [4]. In neonates, acute intestinal ischemia can be caused by aortic thrombosis and intestinal torsion, but also by hypoplastic left heart syndrome [5]. AMI is the cause of 10% of patients over 70 years of age diagnosed with acute abdominal disease, risk factors for AMI include heart failure, arterial hypertension, coronary artery disease, and atrial fibrillation [6]. AMI has become a current concern as the population ages and the incidence of high comorbidity increases [6]. Despite improvements in medical technology, AMI remains a life-threatening medical and surgical emergency with a mortality rate of 50% to 70% and is associated with the development of necrosis of small bowel tissue [3]. AMI has no specific signs or conventional clinical examination [7]. The CT is gold standard for diagnosing arterial and venous occlusions if the physician suspects that the patient is suspected of mesenteric ischemia. Treatment consists mainly of mesenteric arterial revascularization. Accurate identification of necrotic small bowel tissue requiring resection to prevent further disease progression and the ability to accurately distinguish between necrotic and normal tissue is a major clinical challenge [8]. It’s difficult for surgeons to accurately distinguish between normal and ischemic necrotic small bowel tissue sites, because the necrotic bowel segments sometimes do not appear at the first time [9]. This difficulty can lead surgeons to perform massive resections of small bowel tissue, which can result in surviving patients being adjudicated with short bowel syndrome and experiencing a decreased quality of life after healing. In clinical surgery, doctors often use their eyes to observe the different colors of small intestine tissues to identify normal and necrotic areas based on their experience [7,10], then remove the necrotic tissue and preserve the normal tissue. Due to variances in each physician's empirical discernment, this discrimination can also be vulnerable to individual physician subjectivity. Different clinical scenarios may cause errors in the surgeon's judgment, and sometimes necrotic small intestinal tissue is difficult to be identified accurately from the visual field of the surgeon. Sometimes biopsy samples are collected for pathological section preparation and a pathological identification of small intestinal tissue is made by observing the sections under a microscope [11]. This often leads to a long clinical identification time for small bowel tissue, which affects the rapid treatment of patients’ diseases. Therefore, there is an urgent need for a rapid, objective, and multiple informational combined discrimination technique for the clinical identification of small intestinal pathology. In this paper, we used an advanced technique, hyperspectral imaging (HSI), as an aid to detect normal and necrotic tissue in the small intestine.
Since the interaction between the electromagnetic radiation of light and a particular material is unique, the measured spectra obtained are often referred to as spectral fingerprints. Different substances have different spectral characteristics. HSI techniques can obtain spectral information in hundreds of continuous bands while acquiring two-dimensional spatial information about the target object [12]. The analysis of spectral features allows different types of substances to be distinguished. The pathological changes in biological tissue are closely related to its spectral characteristics. There are distinguishable spectral features in different wavelength regions, which allow the pathological changes in biological tissues to be objectively identified by computers [13]. The absorption and scattering properties of light by water, melanin and hemoglobin in cells change during disease progression in biological tissues [14]. Thus, the interaction between electromagnetic radiation of light and biological tissues carries information related to histopathology [15].
The advantage of HSI over invasive diagnostic methods (e.g., endoscopy, puncture procedure, biopsy production) is that it is noninvasive in all spectral range. For these reasons, HSI is an emerging application in the medical field [16]. In recent years, many researchers have explored this technology as a diagnostic aid in different medical applications [14], including cancer tissue detection [17–20], skin trauma assessment [21,22], prostate disease identification [23,24], and blood cell identification [25]. There is a limited amount of literature available in the field of hyperspectral detection of small intestinal tissues. No studies have been reported on the use of supervised classification techniques to objectively assist physicians in distinguishing normal and necrotic areas of small intestinal tissue.
In the present study, we analyzed hyperspectral data of normal and necrotic areas in Japanese large-eared white rabbits one hour after blocking the superior mesenteric artery. The main objective was to use the supervised classification approach to determine whether it is possible to discriminate between normal and ischemic necrotic areas of small intestinal tissue by processing only their spectral information. For this purpose, the hyperspectral data were processed using a supervised classification framework. Six different classifiers (support vector machines, artificial neural networks, bagged trees, linear discriminant analysis, logistic regression, and naive bayes) were used to automatically distinguish between normal and ischemic necrotic sites using the trained model with spectral information of small intestinal tissue as features.
2. Materials and methods
In this study, custom hyperspectral acquisition equipment was employed to collect hyperspectral data from small intestine samples taken from eight Japanese large-eared white rabbits (SPF class). To classify the small intestinal tissues, the data were analyzed using six different machine learning classification techniques. The materials and procedures used to achieve the proposed goal of distinguishing small intestinal tissue samples based on their spectral properties are described in this section.
2.1 Biological samples
The biological samples used in this research work were small intestinal tissue models of Japanese large-eared white rabbits developed by physicians. In this experiment, rabbits were chosen as the animal model of mesenteric ischemia in small intestinal tissues because the intestinal tube size of rabbits is closer to that of human beings than that of rats. Moreover, the mechanism of small intestinal tissue necrosis caused by mesenteric ischemia in rabbits is similar to that in humans. The control of small intestinal tissue necrosis discrimination was manual intervention. Half of the Japanese large-eared white rabbits, male and female, weighing 2-2.5 kg, were selected, and animals in all experimental groups fasted for 12 hours before the experiment. The animals were preinjected intramuscularly with 10ug/kg dextromethorphan and 4 mg/kg tramadol, and anesthesia was induced intravenously using alfaxalone(2 mg/kg) before dissection. The abdomen of the rabbits was then debulked and an incision was made along the median line of the abdominal wall for open surgery. After the abdomen was opened, the superior mesenteric artery was blocked to create an animal model of small intestinal tissue necrosis (Fig. 1(a) and (b)). To keep the ischemic necrotic site from being disturbed by the surrounding normal tissues, a blockage to prevent blood collateral circulation was made to isolate the ischemic necrotic tissue from the normal tissue site (Fig. 1(c)).
Fig. 1. Biological samples. (a) Abdominal hair removal. (b) Blocking a portion of the superior mesenteric artery. (c) Blocking the flow of blood from tiny arteries in the normal intestine to the necrotic intestine.
Hyperspectral images of small intestinal tissues of specimens collected on eight occasions were included in this study, and each sample data collection included hyperspectral information of small intestinal tissues in both normal and necrotic regions. These animal models were developed by pediatric surgeons at the Second Hospital of Wenzhou Medical University, located in Zhejiang Province, China, and experienced physicians indicated the normal and ischemic necrotic areas of small intestinal tissue regions of interest. The study protocol and consent procedures were approved by the hospital ethics committee. Approved by the ethical review body of Animal Experiment Center of Wenzhou Medical University. The name of the experiment is ‘'Research and application of image spectral fusion technology for rapid identification of intestinal necrosis'‘; the consent was given for the rabbit experiment.
2.2 Data acquisition
This section describes the acquisition system used to acquire hyperspectral data of small intestinal tissues from specimens in this study, and briefly describes the hyperspectral camera, light source, and darkroom of the acquisition system. It also introduces the characteristics of the hyperspectral data of small intestinal tissues and summarizes the dataset of normal and ischemic necrotic sites of small intestinal tissues for each rabbit sample.
2.2.1 Acquisition system
A customized hyperspectral acquisition system was developed to acquire hyperspectral images from the small intestinal tissue of Japanese large-eared white rabbits. The system consists of a hyperspectral camera, a halogen light source, and an optical darkroom (Fig. 2). The hyperspectral camera is a SOC710-VP series from Surface Optics, USA, with an internally integrated push and sweep mechanism and detector. The system operates in the spectral range of 376∼1038 nm with a spectral resolution of 4.69 nm and can acquire hyperspectral information with 128 spectral channels. This band range allows the identification of normal and ischemic necrotic tissues depending on the amount of oxyhemoglobin and deoxyhemoglobin in biological tissues and was widely used in the medical field [16]. The advantage of the hyperspectral camera in use is that it can automatically calibrate the dark current and automatically match the scanning speed. The data were acquired with the lens pointing vertically downward to capture hyperspectral information about the rabbit's small intestinal tissue. This system also includes two power tunable halogen light sources (LOWEL PRO, USA) placed on either side of the sample to be measured and illuminated at 45° to provide uniform and stable light to the small intestinal tissue region. To prevent interference from external stray light, the experiments conducted in this study were performed in a custom-made optical darkroom (approximate dimensions 2m×1.5m×2 m).
Since the hyperspectral camera uses a built-in pushing and sweeping method, the camera and the target object should be kept from moving when shooting. When shooting, the aperture and focal length of the lens are adjusted so that the target object is imaged clearly. The light intensity of the captured image should not reach saturation. The pixel size of the hyperspectral camera is 696 × 520, and the spatial resolution is < 40 microns, they can meet the requirement of clinical differentiation of necrotic tissue of small intestine [26–28]. In the spectral dimension, the hyperspectral camera used in this study has a range of 128 bands. The time required to acquire one hyperspectral cube data was approximately 6 seconds. The peristalsis of small intestinal tissue is intermittent, and hyperspectral data acquisition was performed when the small intestinal tissue was not peristaltic.
2.2.2 Hyperspectral database
The hyperspectral acquisition system described previously was used to obtain a hyperspectral database of small intestinal tissues from specimens. Each hyperspectral cube consisted of 128 spectral channels and 696 × 520 pixels. Based on the process of generating animal models and experience, professional physicians marked the normal and ischemic necrotic areas of small intestinal tissue for subsequent work such as classification model building and prediction of unknown samples. In the actual clinical setting, although physicians can make an identification of small bowel tissue necrosis sites based on visual observation and empirical judgment, some small bowel tissue areas are difficult to distinguish in numerous scenarios. This study aimed to perform supervised classification modeling based on the different spectral characteristics of normal and necrotic regions of small intestinal tissues, and then to automatically distinguish normal and necrotic parts of small intestinal tissues of unknown samples by computer in a noncontact manner.
A region of interest (ROI) was selected from the hyperspectral cube of each specimen to extract the spectral dataset, which is subsequently described in detail in the preprocessing section. In the present study, two different tissues were defined: normal and ischemic-necrotic tissues of the small intestine. For each specimen, we choose two ROIs, one belonging to the normal region and the other one belonging to the ischemic necrotic region. Total 16 ROIs for 8 specimens. The size of each ROI is 150 × 100. Table 1 summarizes the spectral data of the specimens in this study, intercepting the ROI region from each hypercube and counting the number of pixel points for each tissue type separately. Each ROI has more than 10,000 effective spectrum pixel points after ignoring the background. The spectral features of normal and ischemic necrotic sites of small intestinal tissue are shown in the right panel of Fig. 3. The red curve indicates the average spectrum of selected normal tissues, and the blue curve indicates the average spectrum of selected ischemic necrotic tissues. By the spectral curve of Fig. 3, we found that the spectral curve of normal tissue had a small absorption peak at 500-560 nm, and the ischemic necrotic tissue had a significant absorption peak at 700-780. This is caused by the different levels of oxyhemoglobin and deoxyhemoglobin in the two different types of tissues.
Fig. 3. Illustration of the spatial and spectral characteristics of a hyperspectral image. Left: A hyperspectral image data cube consists of two spatial dimensions (x, y) and one spectral dimension (λ), which can be regarded as a spectrally resolved image or a spatially resolved spectrum. The two enclosing boxes capture local pixels of normal small intestinal tissue (red) and ischemic necrotic small intestinal tissue (blue). Right: average spectral profile of the two selected localized regions.
Table 1. Spectral signature labeled dataset summary
2.3 Processing framework
The proposed preprocessing framework is based on a supervised classification scheme. Different substances have different spectral fingerprint properties. In this study, we identified normal and necrotic regions of small intestinal tissues by their spectral properties. Therefore, the input to the supervised classifier is the spectral information of normal and ischemic necrotic tissues of the small intestine. An overview of the preprocessing framework used in this study is shown in Fig. 4. The first stage of the framework consists of a preprocessing chain designed to compensate for the effects of environmental conditions and for the sensor response of the acquisition system during the acquisition of the HSI dataset. Six different supervised classification algorithms were then used to distinguish between normal and necrotic sites of small intestinal tissue. Finally, the performance of the classifiers is evaluated using the criteria used to evaluate the classifiers.
2.3.1. Data preprocessing
The preprocessing session proposed in this study is based on five steps: 1) calibration; 2) background removal; 3) region of interest selection; 4) first and last band removal; and 5) smoothing and SNV processing. Each step of the preprocessing is explained in detail below.
- 1) Calibration: The first stage of the preprocessing chain is related to the calibration of the hyperspectral images. Through calibration, the acquired images are converted from radiance observations to reflectance observations. The acquisition of small intestinal tissue data will have an impact on the data due to light source inhomogeneities and requires a black and white plate calibration of the raw image. This is a necessary step in the hyperspectral image preprocessing process [29,30]. As shown in Eq. (1).$${I_{ref}}(x,y,\lambda ) = \frac{{{I_{raw}}(x,y,\lambda ) - {I_{dark}}\,(x,y,\lambda )}}{{{I_{white}}\,(x,y,\lambda ) - {I_{dark}}\,(x,y,\lambda )}}.$$Where the ${I_{raw}}(x,y,\lambda )$ is the raw intensity value of a sample pixel. The ${I_{white}}\,(x,y,\lambda )$ is the intensity value of the standard reflective white plate at the corresponding pixel acquired by the camera, and the ${I_{dark}}\,(x,y,\lambda )$ is the intensity value of the corresponding pixel affected by the device dark current after covering the camera lens. The ${I_{ref}}(x,y,\lambda )$ is the normalized reflectance value at the pixel location $(x,y)$ and the wavelength band $\lambda $ [31].
- 2) Background removal: Valid data of small intestinal tissue must be extracted from the whole hyperspectral cube of the acquisition, so the background information needs to be removed. The data size for each acquisition is 696 × 520 × 128 (696: image length; 520: image height; 128: number of bands of hyperspectral data), which includes all the information in the hyperspectral camera field of view. Other objects exist in the scene of each acquisition of hyperspectral data of small intestinal tissue, such as medical gauze, blood stains, the rabbit's belly and hair, and intestinal mesentery of small intestinal tissue. For subsequent data processing, these objects with complex features need to be removed from the overall hyperspectral cube, and only the hyperspectral information of the rabbit's small intestinal tissue is retained. In this study, the OCSVM (OneClass SVM) algorithm was adopted to build a single classification model of rabbit small intestinal tissue. The processing steps are as follows.
- ii. There may be some interference of mesenteric information in the image after removing the background, which can be removed by using morphological filtering. The image after morphological filtering is shown in Fig. 5(d).
The background of the whole hyperspectral image is removed after the OneClass model and morphological filtering is demonstrated in Fig. 5. Some locations are overexposed due to specular reflections caused by wetting of the small intestinal tissue surface. The spectral information of these exposed positions is distorted and has no value for the subsequent supervised classification. After OneClass model processing, the exposed positions of the hyperspectral image are also removed, allowing the true spectral information of the small intestinal tissue to be retained, which greatly improves the reliability of the data and the training model. On this basis, it could be more convenient to select ROIs of normal and necrotic parts of small intestinal tissue separately. The ROIs could be selected more quickly, and the selected site would not have the background information, and the contour features of the selected site's intestine could be preserved as much as possible.
- 2) ROI selection: Doctors label different tissue types based on the location of the blocked arterial blood supply to the small intestinal tissues and empirical judgment. Based on the marked position, the region of small intestinal tissue of interest was selected for study from the images with the background removed.
- 3) Band reduction: The broad range of hyperspectral bands carries rich material information, which also includes the noisy first and last bands of the spectrum due to sensor-induced noise. These bands need to be removed to reduce the computer data processing time and interference with subsequent data classification. In this study, the first and last noisy bands of the hyperspectral data of small intestinal tissue were removed. The band used for subsequent processing ranged from 400 nm to 1000 nm, as shown in Fig. 6.
- 4) Smoothing and SNV: To reduce the subtle variations in the spectra, all spectral images were smoothed using the Savtizky-Golay (S-G) algorithm [32] (window width = 7 and polynomial order = 2). After the smoothing process, the spectral data are subjected to the standard normal transform (SNV) process, which mainly eliminates the effect of scattering from the object surface on the diffuse reflectance spectrum [33,34]. The diffuse reflective nature of light propagation in tissues can interfere with the spectral information acquired by hyperspectral cameras. The effect of diffuse light reflection can be eliminated using SNV pre-processing [35]. The above S-G smoothing processing and SNV processing prepare the spectral images for supervised classification.
2.3.2. Supervised classification
The pixel-level classification of hyperspectral images relies on the spectral characteristics of the pixel points of the hyperspectral images. Currently, many hyperspectral studies exist on pixel-level classification [36]. The supervised classification-based algorithms used in this study were Support Vector Machines (SVMs), Artificial Neural Networks(ANNs), Bagged Trees(BTs), Linear Discriminant Analysis(LDA), Logistic Regression(LR), and Naive Bayes(NB). These six supervised classification algorithms were shown to have high computational efficiency, and widely used to train supervised classifiers. These classification algorithms were used to distinguish between normal and necrotic sites of small intestinal tissue and were implemented in MATLAB 2021a (The MathWorks, Natick, USA).
SVMs are kernel-based supervised classifiers that have been widely used in the classification of hyperspectral images [37,38]. Related literature shows that SVMs are applied in the field of hyperspectral classification with excellent performance. The advantages of SVMs applied in hyperspectral classification are a strong theoretical foundation, high generalization ability, low sensitivity to the curse of dimensionality, and global classification solving ability. In this study, a linear SVM was used for the supervised classification of hyperspectral images of small intestinal tissue.
There are also many studies related to the application of ANNs as classifiers for hyperspectral classification. Related literature shows that ANNs have better performance in hyperspectral classification in the medical field [39,40]. The neural network used in this research work is a med-layer neural network.
The bagging tree algorithm is essentially a self-help sampling method, that performs multiple put-back sampling to reduce variance. It is a commonly used decision tree classification method. In hyperspectral data classification, some literature used the bagging tree algorithm to achieve competitive results [41,42]. In this experiment, the bagging tree algorithm was used to investigate the classification effect on small intestinal tissue.
The idea of the linear discriminant classifier is to project the training samples onto a straight line so that the projection points of similar samples are as close as possible and the projection points of dissimilar samples are as far away as possible [43]. It is such a line that the computer has to learn. Then the sample to be measured is projected onto this line and its class is determined according to the position of the projected points. Linear discriminant analysis has been used to achieve good classification results for hyperspectral data classification [44,45]. This study used linear discriminant analysis to perform supervised classification of small intestinal tissue.
Logistic regression is commonly used for binary classification and is a classifier rather than a regression method. The classification task is to fit a probability distribution by fitting a straight line and thus a probability distribution. In recent years, logistic regression has been widely used for hyperspectral data classification [46,47]. In this study, a logistic regression algorithm was used to classify small intestinal tissues.
Naive Bayes is a classification method based on Bayes’ theorem with the assumption of conditional independence of features [48]. The field of HSI technology has gained attention in recent years. Several research studies have demonstrated that the Bayesian classifier has good results for hyperspectral image classification [49,50]. Gaussian Naive Bayes (GNB) is a classification technique based on probabilistic methods and Gaussian distribution for machine learning. In this study, we also used GNB to study the classification of normal and necrotic tissues of small intestinal tissue.
2.3.3 Evaluation metrics
The supervised classification results of the six classifiers used in this study were evaluated using standard sensitivity, specificity, and overall accuracy metrics [51,52]. Sensitivity is related to the ability to test for the correct identification of conditions, and it results from the number of true positives (TP) divided by the total number of true positives and false negatives (FN) in Eq. (2). Specificity is related to the ability to detect the correct exclusion of symptoms and it is obtained by dividing the number of true negatives (TN) by the total number of true negatives and false positives (FP) in Eq. (3). Finally, the overall precision was calculated by dividing the total number of successful results by the overall number in Eq. (4).
2.4. Experiment description
To validate the ability of the supervised classification algorithm to distinguish between normal and necrotic tissues of the small intestine, three different case studies (CSs) have been proposed. This approach varies depending on the sample being included in the study population. These scenarios are described as follows.
Case Study 1 (CS1): The purpose of this case study is to check whether available labeling data can be used to distinguish between normal and necrotic small intestinal tissue and to avoid variability between sample data. The data explored in this case study include both normal and necrotic portions of small intestinal tissue as datasets. To avoid inter-sample data variability, data from each sample were used independently to train and test the supervised classifiers.
Case Study 2 (CS2): Considering the inter-sample variability in this scenario, all available marker data were combined into a unified dataset, and all sample data for the eight rabbits were included in this CS.
Case Study 3 (CS3): This case study is the closest study to clinical identification. In this scenario, data from each rabbit sample are used individually as a test set for the classification algorithm. The classifier model is trained by using the data belonging to the remaining rabbit sample markers. This CS represents a realistic case. That is when diagnosing the small intestinal tissue of a new patient, the classifier used is trained from the data of the previous patient.
In this research work, CS1 and CS2 used 10-fold cross-validation as the model validation scheme. The dataset was randomly divided into 10 folds, and 9 folds were used each time to train the classifier, and the remaining 1-fold was used as a prediction set to evaluate the performance of the classifier. This process was repeated until each fold was used to predict the performance of the classifier. The performance of the final classifier is the average of the performance obtained in each iteration. In CS3, cross-validation cannot be used, so the model is evaluated using hold-out validation, where the test set is the spectral data of a single rabbit sample, and the classifier is trained using all available spectral data of the remaining rabbit samples.
3. Experimental results
This section presents the results obtained after applying the supervised classification framework described in Section 2 to the small intestinal tissue hyperspectral dataset. These results give an estimate of the performance of each classifier for each CS. In addition, the computational cost per classifier is shown as a measure of the time required to train each classifier's performance. The 3.5 GHz Intel Core i9-10920X computer was used to train the model in this study.
3.1. Case Study 1
As mentioned before, CS1 shows the classification of data belonging to a single sample only. Estimates of model performance were obtained by 10-fold cross-validation. Table 2 shows the classification results of each classifier in this CS for each sample.
Table 2. Supervised classification results in CS1
The results presented in Table 2 show that normal and necrotic sites of small intestinal tissue can be well distinguished using all six supervised classifiers. Except for the GNB classifier, the sensitivity and specificity of the other classifiers were higher than 94% for each of these eight rabbit samples. The results obtained using the ANN classifier have an overall accuracy and sensitivity above 97% and a specificity above 98% for each sample, with the best classification results among the six classifiers. The linear SVM, BT, LDA, and LR classifiers also had a better ability to distinguish normal and necrotic tissues of the small intestine, all with an average overall accuracy of more than 98%.
In the CS1 scenario, all other classification algorithms except GNB achieved significant classification results. These classifiers all behave very similarly, and their average metrics for overall accuracy, specificity, and sensitivity are close to 99%. The classification results also depend on the object of study. Among these five classifiers, the accuracy of sample 1 is lower than that of the other samples. Table 3 shows the computational cost of this CS. The ANN classifier took the most time, with an average time of 27.07 seconds. The LDA classifier took the least time, with an average time of 2.98 seconds. The computational cost is related not only to the classification model algorithm but also to the individual differences between samples. For example, the SVM classifier took 40.80 seconds to train the classification model of sample 1, while it took only 8.81 seconds to train the SVM classifier of sample 7.
Table 3. The computational cost of each classifier
3.2. Case Study 2
The CS combined data from all available small intestinal tissues of 8 rabbits and aimed to introduce some inter-sample variability in the classification task. Model evaluation was completed by 10 cross-validations. The results obtained for all classifiers of this CS are shown in Table 4. All supervised classifiers had a good ability to distinguish between normal and necrotic small intestinal tissues. Except for the GNB classifier, the overall accuracy, sensitivity, and specificity of the other five classifiers were higher than 90%. However, the overall accuracy of each classifier showed worse results compared to CS1. In this case, ANN has the most competitive classification results with 99% overall accuracy, sensitivity, and specificity. The worst results were obtained using the GNB classifier, with an overall accuracy of 84.55%. The LDA classifier and the LR classifier had similar overall accuracies of approximately 92%.
Table 4. Supervised classification results in CS2
Due to the larger composition of the CS2 dataset (more than 160,000 spectral features), the computational cost is much higher than that of CS1. In this CS, the computational time required for SVM and ANN classifiers is much higher than the one required for all other four classifiers. The time required to train the LDA classifier was the lowest. The BT classifier and the ANN classifier have little difference in overall accuracy, sensitivity, and specificity, but the computational cost of the BT classifier is one-tenth of the time required by the ANN classifier.
3.3. Case Study 3
This scenario illustrates a real-life clinical pathology identification. When diagnosing a new patient's disease, physicians need to make empirical judgments based on information from previous patients. Likewise, gastroenterology physician needs to use information from previous patients’ small intestinal tissue when diagnosing normal and necrotic areas of small intestinal tissue in current patients. Based on this idea, in this CS, the evaluation of the model is performed according to the hold-out method. The data of each rabbit sample are used as the prediction set, and the model is trained using the data of other rabbit samples in the database to make predictions on the current rabbit data.
Table 5 shows the classification results for each classifier for each rabbit sample, as well as the computational time results for the hold-out process. It can be seen that this classification result is not as accurate as the above two case studies. In addition, it can be observed that the sensitivity and specificity are not as balanced as in the above two cases. Among the classifiers under this CS, the overall accuracies of the SVM classifier and LR classifier are similar, both approximately 76%. The overall accuracies of the ANN, BT, LDA, and GNB classifiers are all in the range of 80%-85%. Among them, the BT classifier has the highest average overall accuracy and sensitivity, 84.12% and 87.59%, respectively. The GNB classifier has the highest average specificity, 82.91%.
Table 5. Supervised classification results in CS3
Unlike the results of CS1, the prediction accuracy for CS3 varies greatly between samples. There were very successful subject samples in this study, such as samples S3, S4 and S7, and their overall accuracy was higher than 80% under all classifiers. Some samples even showed similar classification accuracy as CS1, for example sample S7 using SVM classifier and LDA classifier. However, these models are not sufficient to produce high-quality results for small intestinal tissue identification in other rabbit samples.
Regardless of which classifier was used, sample S3 obtained the best classification results with an overall accuracy higher than 90%. According to Fig. 7, it can be found that there may be a chance of cross-fertilization between different classifiers. For example, for sample S5, the overall accuracy obtained using the GNB classifier was 92.12%, and the overall accuracy obtained using the ANN classifier was 65.08%. In contrast, for sample S6, the overall accuracy using the GNB classifier was 69.21%, while the overall accuracy using the ANN classifier was 90.15%. This fact supports the integration of supervised classifiers, where the misclassification of one classifier is compensated by the correct classification of the other classifier. Table 6 shows the computational cost of this CS. The linear SVM classifier takes the most time to train, followed by the mid-level ANN classifier. The LDA and GNB classifiers are the least computationally expensive, taking only approximately 3 seconds to train the classifier model.
Table 6. The computational cost of each classifier
Figure 8 shows representative images of this CS. Based on the physician labeling, hyperspectral data of normal and necrotic sites were extracted from the small intestinal tissue of previous samples, as shown in Fig. 8(a), to train each of the six supervised classifiers used in this study. The site to be evaluated was obtained from the currently developed small intestine tissue model, shown in Fig. 8(b). Physician labeling information for the evaluation site is shown in Fig. 8(c), where the site above the yellow line is normal tissue and the site below the yellow line is ischemic necrotic tissue. The recognition effect of each classifier for this evaluation site is shown in Fig. 8(d). From the result graph, it can be found that each classifier could roughly determine the normal and necrotic sites in this test. For the ANN classifier, especially the GNB classifier, they identified some necrotic tissues as normal tissues more often. According to the previous discussion, if the integration of supervised classifiers was taken in this test, the misclassification of the ANN and GNB classifiers could be compensated by the correct identifications of the other four classifiers.
Fig. 8. Representative images. (a). Training data for normal and necrotic tissues. (b). Animal model of the tissue site to be predicted. (c). Physician labeling of evaluation sites. (d). Identification of evaluation sites by six classifiers
4. Discussion
In this study, six supervised classification methods were proposed and validated to obtain an automated HSI-based tool to assist gastroenterology physician in differentiating necrotic from normal small bowel tissue. For this objective, a customized HSI acquisition system, capable of capturing hyperspectral images of small intestinal tissues in the VNIR range (400 nm to 1000 nm) was used. A framework for labeling and classification processing of hyperspectral samples of small intestinal tissues was developed. A total of 8 normal and ischemic necrotic small intestinal tissue samples from Japanese large-eared white rabbits were included in this study. From these samples, 8 HSI cubes were obtained, labeling more than 160,000 spectral features of normal and necrotic small intestinal tissues. Using this labeled database, six different supervised classification algorithms (SVM, ANN, BT, LDA, LR, and GNB) were evaluated using three different case studies based on the samples included as study subjects.
It should be emphasized that automatic classification of tissues will be possible when a library of spectral features from different types of tissues is collected based on the identification of small intestinal tissues. This means that a significant amount of work needs to be completed before using hyperspectral images for automatic identification. The proposed method is just a step forward in the implementation of HSI for the automatic identification of small intestinal tissues.
The results of the CS1 and CS2 experiments showed that the SVM, ANN, BT, LDA, LR, and GNB classifiers all had competitive results (achieving an overall accuracy of more than 80%) in distinguishing normal and ischemic necrotic tissues of the small intestine. The best classification results were obtained using ANN algorithm. Comparing the results of these CSs, the effect of differences between individual rabbit samples is evident.
In the case of CS3, the classification results do not follow the same trend as for CS1 and CS2, indicating that there is a correlation between the spectral features and the individual characteristics of the samples. In half of the experimental samples, the results are promising. For the other samples, the error of prediction was high. The reduction in the number of rabbit samples in this study may have led to inaccurate results derived in CS3 because the classifier was constructed using information from only eight rabbit samples. For these reasons, the classifier may not have sufficient information to build a model with a high degree of generalization, and therefore the model is affected by individual differences in the samples. In future work, to avoid this effect, the number of samples will be increased. However, the final CS led to some promising results: it was possible to detect normal and ischemic necrotic areas of the current small intestinal tissue using spectral information from previous diagnostic small intestinal tissues. This provides a valuable clinical reference for the use of computers to objectively assist physicians in the identification of small intestinal tissue.
In terms of the computational cost of training different classifiers, SVM and ANN classifiers are higher, while it takes less time to train LDA and GNB classifiers.
Due to the diffuse reflective nature of light in the tissue, it is possible for individual pixels to exhibit spatial correlation with neighboring pixels. Next, we will explore the inverse Fourier transformation of the modulation transfer function of the camera in the spatial frequency domain to obtain the point spread function (PSF) to evaluate the extent of spatial correlation [53–55]. To enhance the performance of hyperspectral imaging technology applied to the detection of necrotic tissue in the small intestine.
5. Conclusion
In summary, HSI has been shown to be a suitable technique for the automatic identification of normal and necrotic small intestinal tissues. More detailed studies including more samples and higher dimensional spectral and spatial resolution of hyperspectral images are needed for more sophisticated classification schemes. However, this study yielded excellent results in differentiating normal and ischemic necrotic small intestinal tissues in animal model experiments. Furthermore, combining these spectral analyses with clinical morphological analyses will potentially improve the accuracy of the overall identification of small intestinal tissue. Soon, such tools could help clinicians diagnose small intestinal biopsies and speed up the surgical procedure for each patient.
Funding
Wenzhou Social Development (Medical and Health) Science and Technology Project (ZY2021027); National Natural Science Foundation of China (62105245, 61893096014).
Disclosures
The authors declare that there are no conflicts of interest related to this article.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. M. Bala, J. Kashuk, E. E. Moore, Y. Kluger, W. Biffl, C. A. Gomes, O. Ben-Ishay, C. Rubinstein, Z. J. Balogh, I. Civil, F. Coccolini, A. Leppaniemi, A. Peitzman, L. Ansaloni, M. Sugrue, M. Sartelli, S. Di Saverio, G. P. Fraga, and F. Catena, “Acute mesenteric ischemia: guidelines of the World Society of Emergency Surgery,” World J Emerg. Surg. 12, 1 (2017). [CrossRef]
2. M. A. A. Sayed, F. A. Murad, M. A. Hamad, and A. Taha, “Management of acute mesenteric ischemia in Assiut University Hospital: a clinical audit,” Egypt. J. Surg. 40, 342–346 (2021). [CrossRef]
3. F. Kuhn, T. S. Schiergens, and E. Klar, “Acute Mesenteric Ischemia,” Visc. Med. 36(4), 256–263 (2020). [CrossRef]
4. A. R. Blaser, S. Acosta, and Y. M. Arabi, “A clinical approach to acute mesenteric ischemia,” Curr. Opin. Crit. Care 27(2), 183–192 (2021). [CrossRef]
5. I. I. Jeican, G. Ichim, and D. Gheban, “Intestinal ischemia in neonates and children,” Clujul medical (1957) 89(3), 347–351 (2016). [CrossRef]
6. B. Luther, A. Mamopoulos, C. Lehmann, and E. Klar, “The Ongoing Challenge of Acute Mesenteric Ischemia,” Visc. Med. 34(3), 215–221 (2018). [CrossRef]
7. M. Mehdorn, S. Ebel, H. Kohler, I. Gockel, and B. Jansen-Winkeln, “Hyperspectral imaging and indocyanine green fluorescence angiography in acute mesenteric ischemia: A case report on how to visualize intestinal perfusion,” Int. J. Surg. Case Rep. 82, 105853 (2021). [CrossRef]
8. H. A. Abualhamayel, M. Y. Alsayed, F. M. Alkudaysi, A. A. Almarhabi, O. A. Ghanem, M. T. Alanazi, A. O. Alsaeeri, B. A. Alshammari, F. S. Binkattal, N. T. Aldosari, and W. J. Hommadi, “Mesenteric Ischemia: Review over Pathophysiology and Means of Treatments in Surgical Department,” Pharmacophore 12(5), 25–31 (2021). [CrossRef]
9. G. Piton, A. Le Gouge, J. Boisrame-Helms, N. Anguel, L. Argaud, P. Asfar, V. Botoc, A. Bretagnol, L. Brisard, H. N. Bui, E. Canet, D. Chatelier, L. Chauvelot, M. Darmon, V. Das, J. Devaquet, M. Djibre, F. Ganster, M. Garrouste-Orgeas, S. Gaudry, O. Gontier, S. Groyer, B. Guidet, J. E. Herbrecht, Y. Hourmant, J. C. Lacherade, P. Letocart, F. Martino, V. Maxime, E. Mercier, J. P. Mira, S. Nseir, J. P. Quenot, J. Richecoeur, J. P. Rigaud, D. Roux, D. Schnell, C. Schwebel, D. Silva, M. Sirodot, B. Souweine, N. Thieulot-Rolin, F. Tinturier, P. Tirot, D. Thevenin, G. Thiery, J. B. Lascarrou, J. Reignier, and S. Clinical Res Intensive Care, “Factors associated with acute mesenteric ischemia among critically ill ventilated patients with shock: a post hoc analysis of the NUTRIREA2 trial,” Intensive Care Med. 48(4), 458–466 (2022). [CrossRef]
10. L. Urbanavicius, P. Pattyn, D. V. de Putte, and D. Venskutonis, “How to assess intestinal viability during surgery: A review of techniques,” World journal of gastrointestinal surgery 3(5), 59–69 (2011). [CrossRef]
11. X. Y. Wang, C. N. Chu, S. L. Sun, T. Xie, Z. H. Duan, K. Wang, B. C. Liu, X. X. Fan, X. J. Wu, and W. W. Ding, “Outcomes and clinical characteristics of transmural intestinal necrosis in acute mesenteric ischemia,” Scand. J. Gastroenterol. 54(8), 953–959 (2019). [CrossRef]
12. S. Junaid, J. Tomko, M. P. Semtsiv, J. Kischkat, W. T. Masselink, C. Pedersen, and P. Tidemand-Lichtenberg, “Mid-infrared upconversion based hyperspectral imaging,” Opt. Express 26(3), 2203–2211 (2018). [CrossRef]
13. B. Jansen-Winkeln, M. Barberio, C. Chalopin, K. Schierle, M. Diana, H. Kohler, I. Gockel, and M. Maktabi, “Feedforward Artificial Neural Network-Based Colorectal Cancer Detection Using Hyperspectral Imaging: A Step towards Automatic Optical Biopsy,” Cancers 13(5), 967 (2021). [CrossRef]
14. G. L. Lu and B. W. Fei, “Medical hyperspectral imaging: a review,” J. Biomed. Opt. 19, 1 (2014). [CrossRef]
15. S. Ortega, M. Halicek, H. Fabelo, G. M. Callico, and B. W. Fei, “Hyperspectral and multispectral imaging in digital and computational pathology: a systematic review [Invited],” Biomed. Opt. Express 11(6), 3195–3233 (2020). [CrossRef]
16. X. J. Ma, R. Liu, C. X. Li, W. L. Chen, and K. X. Xu, “Hyperspectral Imaging of in vivo Tissues: A Review,” Laser Optoelectron. Progress 57(8), 080002 (2020).
17. C. L. Tsai, A. Mukundan, C. S. Chung, Y. H. Chen, Y. K. Wang, T. H. Chen, Y. S. Tseng, C. W. Huang, I. C. Wu, and H. C. Wang, “Hyperspectral Imaging Combined with Artificial Intelligence in the Early Detection of Esophageal Cancer,” Cancers 13(18), 4593 (2021). [CrossRef]
18. M. Ishikawa, C. Okamoto, K. Shinoda, H. Komagata, C. Iwamoto, K. Ohuchida, M. Hashizume, A. Shimizu, and N. Kobayashi, “Detection of pancreatic tumor cell nuclei via a hyperspectral analysis of pathological slides based on stain spectra,” Biomed. Opt. Express 10(9), 4568–4588 (2019). [CrossRef]
19. I. H. Aboughaleb, M. H. Aref, and Y. H. El-Sharkawy, “Hyperspectral imaging for diagnosis and detection of ex-vivo breast cancer,” Photodiagn. Photodyn. Ther. 31, 101922 (2020). [CrossRef]
20. L. A. Zherdeva, I. A. Bratchenko, M. V. Alonova, O. O. Myakinin, D. N. Artemyev, A. A. Moryatov, S. V. Kozlov, and V. P. Zakharov, “Hyperspectral imaging of skin and lung cancers,” Proc SPIE 9887, 98870S (2016). [CrossRef]
21. M. Dietrich, S. Marx, M. von der Forst, T. Bruckner, F. C. F. Schmitt, M. O. Fiedler, F. Nickel, A. Studier-Fischer, B. P. Muller-Stich, T. Hackert, T. Brenner, M. A. Weigand, F. Uhle, and K. Schmidt, “Hyperspectral imaging for perioperative monitoring of microcirculatory tissue oxygenation and tissue water content in pancreatic surgery - an observational clinical pilot study,” Perioper Med-London 10(1), 1 (2021). [CrossRef]
22. G. Daeschlein, I. Langner, T. Wild, S. von Podewils, C. Sicher, T. Kiefer, and M. Junger, “Hyperspectral imaging as a novel diagnostic tool in microcirculation of wounds,” Clin. Hemorheol. Microcirc. 67(3-4), 467–474 (2017). [CrossRef]
23. H. Sasaki, H. Inaba, T. Kimura, K. Morishita, and S. Egawa, “Hyperspectral Imaging Can Distinguish Morphology Patterns in Prostate Cancer,” J Urology 201, E80 (2019). [CrossRef]
24. H. Akbari, L. V. Halig, D. M. Schuster, A. Osunkoya, V. Master, P. T. Nieh, G. Z. Chen, and B. W. Fei, “Hyperspectral imaging and quantitative analysis for prostate cancer detection,” J. Biomed. Opt. 17(7), 0760051 (2012).
25. Q. L. Li, M. Zhou, H. Y. Liu, Y. T. Wang, and F. M. Guo, “Red Blood Cell Count Automation Using Microscopic Hyperspectral Imaging Technology,” Appl. Spectrosc. 69(12), 1372–1380 (2015). [CrossRef]
26. B. Jansen-Winkeln, N. Holfert, H. Kohler, Y. Moulla, J. P. Takoh, S. M. Rabe, M. Mehdorn, M. Barberio, C. Chalopin, T. Neumuth, and I. Gockel, “Determination of the transection margin during colorectal resection with hyperspectral imaging (HSI),” Int. J. Colorectal. Dis. 34(4), 731–739 (2019). [CrossRef]
27. M. Barberio, F. Longo, C. Fiorillo, B. Seeliger, P. Mascagni, V. Agnus, V. Lindner, B. Geny, A. L. Charles, I. Gockel, M. Worreth, A. Saadi, J. Marescaux, and M. Diana, “HYPerspectral Enhanced Reality (HYPER): a physiology-based surgical guidance tool,” Surg. Endosc. 34(4), 1736–1744 (2020). [CrossRef]
28. B. Jansen-Winkeln, I. Germann, H. Kohler, M. Mehdorn, M. Maktabi, R. Sucher, M. Barberio, C. Chalopin, M. Diana, Y. Moulla, and I. Gockel, “Comparison of hyperspectral imaging and fluorescence angiography for the determination of the transection margin in colorectal resections-a comparative study,” Int. J. Colorectal. Dis. 36(2), 283–291 (2021). [CrossRef]
29. G. L. Lu, J. V. Little, X. Wang, H. Z. Zhang, M. R. Patel, C. C. Griffith, M. W. El-Deiry, A. Y. Chen, and B. W. Fei, “Detection of Head and Neck Cancer in Surgical Specimens Using Quantitative Hyperspectral Imaging,” Clin. Cancer Res. 23(18), 5426–5436 (2017). [CrossRef]
30. M. H. Aref, I. H. Aboughaleb, and Y. H. El-Sharkawy, “Custom optical imaging system for ex-vivo breast cancer detection based on spectral signature,” Surg. Oncol. 35, 547–555 (2020). [CrossRef]
31. M. Halicek, J. D. Dormer, J. V. Little, A. Y. Chen, and B. W. Fei, “Tumor detection of the thyroid and salivary glands using hyperspectral imaging and deep learning,” Biomed. Opt. Express 11(3), 1383–1400 (2020). [CrossRef]
32. J. Steinier, Y. Termonia, and J. Deltour, “Smoothing and differentiation of data by simplified least square procedure,” Anal. Chem. 44(11), 1906–1909 (1972). [CrossRef]
33. Y. M. Bi, K. L. Yuan, W. Q. Xiao, J. Z. Wu, C. Y. Shi, J. Xia, G. H. Chu, G. X. Zhang, and G. J. Zhou, “A local pre-processing method for near-infrared spectra, combined with spectral segmentation and standard normal variate transformation,” Anal. Chim. Acta 909, 30–40 (2016). [CrossRef]
34. R. J. Barnes, M. S. Dhanoa, and S. J. Lister, “Standard Normal Variate Transformation and De-Trending of Near-Infrared Diffuse Reflectance Spectra,” Appl. Spectrosc. 43(5), 772–777 (1989). [CrossRef]
35. E. Kho, L. L. de Boer, K. K. Van de Vijver, F. van Duijnhoven, M. J. T. F. D. V. Peeters, H. J. C. M. Sterenborg, and T. J. M. Ruers, “Hyperspectral Imaging for Resection Margin Assessment during Cancer Surgery,” Clin. Cancer Res. 25(12), 3572–3580 (2019). [CrossRef]
36. B. K. Mohan and A. Porwal, “Hyperspectral image processing and analysis,” Curr Sci India 108, 833–841 (2015).
37. G. Urbanos, A. Martin, G. Vazquez, M. Villanueva, M. Villa, L. Jimenez-Roldan, M. Chavarrias, A. Lagares, E. Juarez, and C. Sanz, “Supervised Machine Learning Methods and Hyperspectral Imaging Techniques Jointly Applied for Brain Cancer Classification,” Sensors 21(11), 3827 (2021). [CrossRef]
38. J. X. Ke, L. B. Rao, L. M. Zhou, X. Y. Chen, and Z. Q. Zhang, “Non-destructive determination of volatile oil and moisture content and discrimination of geographical origins of Zanthoxylum bungeanum Maxim. by hyperspectral imaging,” Infrared Phys. Technol. 105, 103185 (2020). [CrossRef]
39. V. Dremin, Z. Marcinkevics, E. Zherebtsov, A. Popov, A. Grabovskis, H. Kronberga, K. Geldnere, A. Doronin, I. Meglinski, and A. Bykov, “Skin Complications of Diabetes Mellitus Revealed by Polarized Hyperspectral Imaging and Machine Learning,” IEEE Trans. Med. Imaging 40(4), 1207–1216 (2021). [CrossRef]
40. Z. Fang, R. B. Wang, M. Y. Wang, S. Zhong, L. Q. Ding, and S. Y. Chen, “Effect of Reconstruction Algorithm on the Identification of 3D Printing Polymers Based on Hyperspectral CT Technology Combined with Artificial Neural Network,” Materials 13(8), 1963 (2020). [CrossRef]
41. H. Jafarzadeh, M. Mahdianpari, E. Gill, F. Mohammadimanesh, and S. Homayouni, “Bagging and Boosting Ensemble Classifiers for Classification of Multispectral, Hyperspectral and PolSAR Data: A Comparative Evaluation,” Remote Sens. 13(21), 4405 (2021). [CrossRef]
42. C. J. Legleiter and S. W. Hodges, “Mapping Benthic Algae and Cyanobacteria in River Channels from Aerial Photographs and Satellite Images: A Proof-of-Concept Investigation on the Buffalo National River, AR, USA,” Remote Sens. 14(4), 953 (2022). [CrossRef]
43. S. Wan and Y. P. Wang, “The Comparison of Density-Based Clustering Approach among Different Machine Learning Models on Paddy Rice Image Classification of Multispectral and Hyperspectral Image Data,” Agriculture 10(10), 465 (2020). [CrossRef]
44. G. L. Lu, D. S. Wang, X. L. Qin, S. Muller, X. Wang, A. Y. Chen, Z. G. Chen, and B. W. Fei, “Detection and delineation of squamous neoplasia with hyperspectral imaging in a mouse model of tongue carcinogenesis,” J. Biophotonics 11, 1 (2018). [CrossRef]
45. R. H. Furlanetto, T. Moriwaki, R. Falcioni, M. Pattaro, A. Vollmann, A. C. Sturion, W. C. Antunes, and M. R. Nanni, “Hyperspectral reflectance imaging to classify lettuce varieties by optimum selected wavelengths and linear discriminant analysis,” Remote Sens Appl 20, 100400 (2020).
46. B. Chung, J. Yu, L. Wang, N. H. Kim, B. H. Lee, S. Koh, and S. Lee, “Detection of Magnesite and Associated Gangue Minerals using Hyperspectral Remote Sensing-A Laboratory Approach,” Remote Sens. 12(8), 1325 (2020). [CrossRef]
47. Q. Zhou, W. Q. Huang, S. X. Fan, F. Zhao, D. Liang, and X. Tian, “Non-destructive discrimination of the variety of sweet maize seeds based on hyperspectral image coupled with wavelength selection algorithm,” Infrared Phys. Technol. 109, 103418 (2020). [CrossRef]
48. A. N. N. Azmi, S. Khairunniza-Bejo, M. Jahari, F. M. Muharram, and I. Yule, “Identification of a Suitable Machine Learning Model for Detection of Asymptomatic Ganoderma boninense Infection in Oil Palm Seedlings Using Hyperspectral Data,” Appl. Sci. 11(24), 11798 (2021). [CrossRef]
49. L. L. Xu and J. Li, “Bayesian Classification of Hyperspectral Imagery Based on Probabilistic Sparse Representation and Markov Random Field,” IEEE Geosci. Remote Sensing Lett. 11(4), 823–827 (2014). [CrossRef]
50. A. Siedliska, P. Baranowski, J. Pastuszka-Wozniak, M. Zubik, and J. Krzyszczak, “Identification of plant leaf phosphorus content at different growth stages based on hyperspectral reflectance,” BMC Plant Biol. 21(1), 28 (2021). [CrossRef]
51. M. Halicek, H. Fabelo, S. Ortega, J. V. Little, X. Wang, A. Y. Chen, G. M. Callico, L. Myers, B. D. Sumer, and B. W. Fei, “Hyperspectral imaging for head and neck cancer detection: specular glare and variance of the tumor margin in surgical specimens,” J. Med. Imag. 6(03), 1 (2019). [CrossRef]
52. S. Ortega, H. Fabelo, R. Camacho, M. D. Plaza, G. M. Callico, and R. Sarmiento, “Detecting brain tumor in pathological slides using hyperspectral imaging,” Biomed. Opt. Express 9(2), 818–831 (2018). [CrossRef]
53. D. J. Cuccia, F. Bevilacqua, A. J. Durkin, F. R. Ayers, and B. J. Tromberg, “Quantitation and mapping of tissue optical properties using modulated imaging,” J. Biomed. Opt. 14(2), 024012 (2009). [CrossRef]
54. V. Pera, K. Karrobi, S. Tabassum, F. Teng, and D. Roblyer, “Optical property uncertainty estimates for spatial frequency domain imaging,” Biomed. Opt. Express 9(2), 661–678 (2018). [CrossRef]
55. Y. Y. Zhao, A. Pilvar, A. Tank, H. Peterson, J. Jiang, J. C. Aster, J. P. Dumas, M. C. Pierce, and D. Roblyer, “Shortwave-infrared meso-patterned imaging enables label-free mapping of tissue water and lipid content,” Nat. Commun. 11(1), 1 (2020). [CrossRef]
