Adaptive Boosting (AdaBoost)-based multiwavelength spatial frequency domain imaging and characterization for ex vivo human colorectal tissue assessment

Abstract

The current gold standard diagnostic test for colorectal cancer remains histological inspections of endoluminal neoplasia in biopsy specimens. However, biopsy site selection requires visual inspection of the bowel, typically with a white-light endoscope. Therefore, this technique is poorly suited to detect small or innocuous-appearing lesions. We hypothesize that an alternative modality—multiwavelength spatial frequency domain imaging (SFDI)—would be able to differentiate various colorectal neoplasia from normal tissue. In this ex vivo study of human colorectal tissues, we report the optical absorption and scattering signatures of normal, adenomatous polyp and cancer specimens. An abnormal vs. normal adaptive boosting (AdaBoost) classifier is trained to dichotomize tissue based on SFDI imaging characteristics, and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.95 is achieved. We conclude that AdaBoost-based multiwavelength SFDI can differentiate abnormal from normal colorectal tissues, potentially improving endoluminal screening of the distal gastrointestinal tract in the future.

Graphical Abstract

1 |. INTRODUCTION

Colorectal cancer remains the second most common malignancy diagnosed globally and represents the second leading cause of cancer mortality worldwide.[1] The gold-standard endoluminal screening technique for colorectal malignancies remains visual endoscopy, which involves visual inspection of the mucosal lining of the colon and rectum with a white-light endoscope. Abnormal-appearing areas are then biopsied for histologic analysis. Yet endoscopy remains limited by several shortcomings. This technique relies on visual identification of abnormal tissue to guide biopsy site selection—which can result in missed identification of small or sessile subsurface lesions that are hard to see.[2–4] Due to lacking quantitative measurement in visual endoscopy, the interpretations of the images are subjective and have interreader variation. To improve screening and surveillance of colorectal cancers, more sensitive and quantitative imaging modalities and methods are needed.

Computer-aided diagnosis (CAD) can assist the analysis and classification of medical images with objective and quantitative interpolation. For example, Song et al. applied a deep-learning model on endoscopic images to classify different colorectal polyp types and achieved a diagnostic accuracy of 81.3% to 82.4%.[5] Combined with other optical imaging modalities such as optical coherence tomography, photoacoustic imaging and fluorescence imaging, CAD has demonstrated the ability to differentiate normal from malignant colorectal tissues. [6–9] Recently, Sreyankar et al. reported the application of spatial frequency domain imaging (SFDI) of three visible wavelengths of 460, 530 and 630 nm for differentiating normal from malignant colon using a logistic regression model.[10] It achieved an AUC of 0.902 based on quantitative absorption and scattering information of human colorectal tissues. These studies have established that neoplastic disruptions of the normal colonic wall structure produce altered absorption and/or scattering coefficient patterns in comparison to normal controls.

Among CAD algorithms, Adaptive Boosting (AdaBoost) is a relatively new nonlinear machine learning algorithm.[11] AdaBoost can be combined with many other types of machine learning algorithms and the outputs of individual learning algorithms (weak learners) are combined into a weighted sum and the final model will converge to a strong learner. [12, 13] Since AdaBoost is a boosting-based algorithm, like most other ensemble methods, the likelihood of overfitting is very low.[12, 14] AdaBoost has recently been adopted by many researchers in the medical imaging field to assist diagnosis. These applications include discrimination of breast tumors in ultrasonic images,[15] brain tumor classification in magnetic resonance imaging[16–18] and lung bronchovascular classification in computed tomography.[19]

We report in this study the classification of colorectal tissues including normal, adenomatous polyp and cancer specimens using AdaBoost-based multiwavelength SFDI in the spectral range of 660 to 930 nm. Freshly excised colorectal tissues were imaged ex vivo. Wide-field absorption and scattering maps were constructed for the samples using nine discrete imaging wavelengths. An AdaBoost classifier was trained based on the absorption and scattering features and its performance was further evaluated using the AUC. The performance was also compared with support vector machine (SVM) classifiers.

2 |. METHODS AND MATERIALS

2.1 |. Colon specimen preparation

Patients undergoing extirpative colonic resection at Washington University School of Medicine were recruited. From these patients, freshly excised colorectal specimens were imaged using a nine wavelength SFDI device. This study was approved by the Institutional Review Board, and the informed consent was obtained from all patients. Diagnoses were ascertained by subsequent histologic examination.

2.2 |. SFDI System

A low-cost, hand-held SFDI probe was used in this study and was described in detail in our previous study.[20] Briefly, 9 LEDs (660, 740, 780, 810, 830, 850, 890, 935 and 950 nm) were placed on a custom-designed printed circuit board (PCB). A rotational stepper motor (PG20L-D20-HHC0, NMB Technologies) was used to rotate the PCB in order to switch the LED that was positioned on the optical axis of the lenses. Light from the LED was homogenized by a beam diffuser, and then collimated by the collimating lens (Thorlabs, AC254–050-B-ML). A linear stepper motor (19541–12-905, Ametek) drives a printed sinusoidal pattern to provide three phase-shifted patterns shining on the tissue (0, 2π/3 and 4π/3). Two polarizer plates (one on the illumination path and one on the detection path) are used to reject specular reflection. A CMOS camera (EO-0413M-GL, Edmund Optics) combined with a 25 mm Fixed Focal Length Lens (67–715, Edmund Optics) was used to collect the diffused light. The overall acquisition time for all nine wavelengths was around 2 minutes per specimen.

2.3 |. Absorption and scattering feature characterization

The image reconstruction algorithm is similar to the previously reported methods.[10, 21, 22] Briefly, for each wavelength, three phase-shifted images of the diffused reflected light were used to extract the DC (spatial frequency = 0 cm⁻¹) and AC (spatial frequency = 1 cm⁻¹) components using amplitude demodulation method.[21, 23] A reference phantom was used to calibrate the diffuse reflectance components of the sample under the same illumination condition. Absorption coefficient (μ_a) and reduced scattering coefficient (μ_s′) maps were calculated using the two measured diffuse reflectance from the tissues and the phantom. The reconstruction time for all wavelengths was about 1 min. Independent regions of interest (ROIs) with sizes of ~5 mm × 3.75 mm were then selected from the reconstructed absorption and reduced scattering coefficient maps. A total of 88 ROIs (44 normal areas, 14 adenomatous polyp areas, and 30 cancer areas) from 16 patients were selected and processed. For each specimen, the ROIs were selected far away from each other to avoid overlap. More detailed information about ROI selection and the study group of patients are given in Tables 1 and S1. The averaged absorption and reduced scattering coefficient of each ROI were extracted for further statistical analysis and classification.

TABLE 1.

Characteristics of the group of patients imaged and ROI selection

Histologic examination	Number of patients involved	Age (mean ± SD)	Sex (% male)	Total number of ROIs	Averaged ROIs per patient	Median ROIs per patient
Normal	12	69 ± 15	58%	44	3.7	4
Adenomatous polyp	5	63 ± 9	60%	14	2.8	3
Cancer	10	70 ± 16	55%	30	3	3

ALGORITHM 1.

AdaBoost classifier

Input: D(y = {−1, +1}), T

Initialize input weights: ${\omega }_1^{(0)},{\omega }_2^{(0)},\dots ,{\omega }_n^{(0)}|=\frac{1}{n}$

For t = 1, …, T

1. Train a weak learner, ht, by minimizing the weighted training error;

2. Compute the weighted training error of ht:

${ϵ}_t=\sum _{i=1}^n{\omega }_i^{(t-1)}〚h_t\left({\overrightarrow{x}}_i\right)\ne y_i〛$

3. If ∊t < threshold value, end iteration.

4. Compute the “importance” of ht:

${\alpha }_t=\frac{1}{2}\log \left(\frac{1-{ϵ}_t}{{ϵ}_t}\right)$

5. Update the weights:

${\omega }_i^{(t)}=\frac{{\omega }_i^{(t-1)}}{Z_t}\times \{\begin{array}{r}\hfill e^{-{\alpha }_t}if{h}_t\left({\overrightarrow{x}}_i\right)=y_i\\ \hfill e^{{\alpha }_t}if{h}_t\left({\overrightarrow{x}}_i\right)\ne y_i\end{array}$

AdaBoost Output:

$g_T(\overrightarrow{x})=sign\left(H_T(\overrightarrow{x})\right)=sign\left(\sum _{t=1}^T{\alpha }_t{h}_t(\overrightarrow{x})\right)$

2.4 |. Feature selection

A total of 18 features, including the averaged absorption and reduced scattering coefficients for nine different wavelengths, were extracted from all ROIs. To explore a minimum number of wavelengths needed to achieve the same performance of classification, we generated a total of 502 data sets with different numbers of features using nine wavelengths or a subset of nine wavelengths. To obtain both oxygenated hemoglobin (HbO₂) and deoxygenated hemoglobin (Hb) spectral information, a minimal of two wavelengths were used in the subset of nine wavelength study. These features were input into the AdaBoost classifier to characterize the classification of normal, adenomatous polyp and cancer tissues.

2.5 |. AdaBoost classifier

AdaBoost is a boosting technique that combines multiple weak classifiers into a single strong classifier. Detailed steps of AdaBoost is given in Algorithm 1 and Figure 1.[24] Briefly, in each iteration t, each weak classifier, h_t, is trained by minimizing the weighted training error, ∊_t, as given in step 2 of Algorithm 1. Then the significance of the h_t is evaluated by computing α_t. If ∊_t is 0.5, then α_t is zero and the classifier is not counted for final AdaBoost output. If ∊_t is less than 0.5, then α_t is positive and the corresponding h_t is contributing to the final AdaBoost output. Otherwise, the corresponding ht is negatively contributing to the final AdaBoost output. If ∊_t is smaller than a threshold value which shows a good fitting, then the iteration will be stopped. The error weight ${\omega }_i^{(t)}$ for each iteration of boosting is updated by either increasing $(h_t\left({\overrightarrow{x}}_i\right)\ne y_i)$ or decreasing $(h_t\left({\overrightarrow{x}}_i\right)=y_i)$ as given in step 5. The final AdaBoost output is the weighted summation of all weak classifiers. The AdaBoost classifier with decision trees as weak learners is a benchmark tree-based ensemble classification algorithm.[24] Therefore, a simple one layer decision tree was chosen to be the weak learner. An itereation number T of 50 and a threshold value of 0.02 were found to be optimal in this study and used for all computations. All classifiers were built under Python 3.7 environment.

FIGURE 1.

Flow chart for AdaBoost algorithm

Different data sets described in Feature selection session were inputted into the AdaBoost classifiers for identifying colorectal tissue types. To train the abnormal vs. normal classifier, 30 normal areas and 30 adenomatous polyp and cancer areas were included in the training cohort and the remaining in the testing cohort. Thousand repeated tests were performed and averaged to obtain the ROC curve. For each test, the training set was randomly selected, and the remaining ROIs were used for testing.

For comparison with the well-known SVM classifier, two abnormal vs. normal SVM classifiers with linear kernel or radial basis function (RBF) kernel were also trained using 18 features from all 9 wavelengths.[25] The training set size and number of tests were the same as the AdaBoost classifier.

Additionally, we trained AdaBoost adenomatous polyp vs. normal and adenomatous polyp vs. cancer classifiers individually to better differentiate performance in both adenomatous and malignant settings. For each classifier, 10 adenomatous polyp areas, and 10 normal or cancer areas were used for training and the remaining for testing. The number of tests was the same as the abnormal vs. normal classifier.

2.6 |. Statistical analysis

The student’s t-test was used to evaluate the statistical significance within each individual feature. A p-value less than .05 is considered statistically significant. And the ROC and AUC were used for evaluating the accuracy of the trained AdaBoost or SVM classification models.

3 |. RESULTS

3.1 |. Reconstructed absorption and reduced scattering coefficient maps

Figure 2 shows representative absorption and reduced scattering coefficient maps (at 660 and 950 nm), H&E stained histology results and photographs of cancer and normal region from one colon sample. It was observed that the cancer region had a significantly higher absorption coefficient compared to the normal area. In contrast, the cancer region had a lower reduced scatter coefficient compared to the normal area.

FIGURE 2.

Absorption and reduced scattering coefficient maps (at 660 and 950 nm), H&E stained histology results and photographs of a T2 adenocarcinoma and corresponding normal tissue

Figure 3 shows representative absorption and reduced scattering coefficient maps at 660 and 950 nm, H&E stained histology results and photographs of one colon specimen with an adenomatous polyp and corresponding normal area. This polyp has an absorption coefficient that is close to the normal area, which is relatively low, but the reduced scattering coefficient is close to the cancer area with a relatively low value.

FIGURE 3.

Absorption and reduced scattering coefficient maps (at 660 and 950 nm), H&E stained histology results and photographs of an adenomatous polyp and corresponding normal tissue. The photographs were taken in vivo using white-light colonoscope

3.2 |. Absorption and scattering feature characterization

A comparison between lesions of mean absorption coefficient and reduced scattering coefficient of all regions of interest at all wavelengths is shown in Figure 4. According to the student’s t-test, μ_a and μ_s′ for all wavelengths are significantly different between cancer and normal areas (P < .001). For all wavelengths, μ_a in adenomatous polyp areas are significantly different from cancer areas (P < .001), but has no significant difference compared to normal areas for wavelengths between 780 and 935 nm (P > .05). As for μ_s′, adenomatous polyps at all wavelengths show statistical significance from normal areas (P < .001) and statistical significance (P < .05) than malignancy areas except at 935 nm (P = .06).

FIGURE 4.

Boxplot of averaged absorption coefficient, A, and reduced scattering coefficient, B, of cancer, normal and adenomatous polyp (adenoma) groups

3.3 |. Testing result of AdaBoost classifier and SVM classifiers using all wavelengths

The ROC curve of the testing results of the AdaBoost abnormal vs. normal classifier is shown in Figure 5A. An AUC of 0.953, optimal sensitivity and specificity of 89.1% and 85.7% were achieved. The linear SVM and SVM with RBF kernel yield AUC of 0.885 and 0.934, respectively, and the ROC curves are shown in Figure 5B,C. Besides improved AUC, it can also be seen that AdaBoost has a smaller standard deviation compared to SVM.

FIGURE 5.

Receiver operating characteristic curve for abnormal vs normal AdaBoost classifier, A, SVM classifier with linear kernel, B, and SVM classifier with RBF kernel, C. The shaded zone marks the standard deviation

The ROC curves of the testing results of the AdaBoost normal vs. adenomatous polyp and adenomatous polyp vs. cancer classifiers are shown in Figure 6. They yield AUC of 0.900 and 0.879, respectively.

FIGURE 6.

Receiver operating characteristic curve for adenomatous polyp vs normal AdaBoost classifier, A, and adenomatous polyp vs cancer AdaBoost classifier, B. The shaded zone marks the standard deviation

3.4 |. AdaBoost classifier performance for subset of wavelengths

Table 2 shows the representative results of abnormal vs. normal classifier with subsets of nine wavelengths in an effort to identify the optimal set of wavelengths. With the optimal sets of wavelengths, an AUC larger than 0.94 can be achieved even using only two wavelengths (660 and 890 nm). With an increasing number of wavelengths, the classifier performance gradually increases.

TABLE 2.

Abnormal vs normal AdaBoost classifier performance with subsets of wavelengths

Number of wavelengths				Best AUC				Best wavelengths
2				0.941				(1, 7)
3				0.945				(1, 6, 7)
4				0.947				(1, 5, 6, 7)
5				0.946				(1, 5, 6, 7, 8)
6				0.947				(1, 4, 5, 6, 7, 8)
7				0.948				(1, 3, 4, 5, 6, 7, 8)
8				0.952				(1, 2, 3, 4, 5, 6, 7, 8)
9				0.953				All wavelengths
Notations for wavelengths:
Notation	1	2	3	4	5	6	7	8	9
Wavelength (nm)	660	740	780	810	830	850	890	935	950

4 |. DISCUSSION

This is the first report using an AdaBoost-based multiwavelength SFDI system for the classification of colorectal tissues including normal, malignant and adenomatous polyp specimens. Wide-field absorption and scattering maps were estimated for colorectal specimens over nine wavelengths. Drastically different quantitative characteristics were captured by SFDI for these three types of specimens. Elevated absorption coefficient among tumors, potentially due to malignant angiogenesis, were uniformly noted along with decreased scattering coefficient as a contrast to normal specimens, which may result from structural disruption of highly organized colorectal tissues. Adenomatous polyps were found to have a lower absorption coefficient than tumors and lower scattering coefficient than normal colorectal tissues.

An abnormal vs. normal AdaBoost classifier was trained based on absorption and scattering features and achieved good AUC, sensitivity and specificity with this limited dataset. It has been shown that linear SVM has a relatively low AUC for our data because spectral features of μ_a and μ_s′ are nonlinearly related to the wavelength. SVM with a nonlinear RBF kernel shows improved performance. AdaBoost further improves SVM with RBF kernel because Adaboost with decision trees is suitable for nonlinear structure and robust to over fitting by taking a weighted average of many weak learners.[12] Thus, AdaBoost is more suitable for our colorectal tissue classification based on multiwavelength optical properties. Besides, the AdaBoost classifiers also achieved good AUC in differentiating adenomatous polyps from normal or cancer tissue, which demonstrated its potential to apply to more complex or challenging situations.

It is worth mentioning that a similar AUC can be achieved with reduced wavelengths in distinguishing between the normal and abnormal colorectal tissues, which can provide guidance for selecting the optimal wavelengths for probing colorectal cancer to reduce the cost and complexity of the system. For example, we observed similar performance in distinguishing between normal and abnormal tissues even with two wavelengths (660 and 890 nm), as shown in Table 2. This is likely due to vascular contrast between normal and abnormal colorectal tissues, as 660 nm is highly sensitive to Hb absorption and 890 nm is highly sensitive to HbO₂ absorption in the spectral range investigated in the study.[26] In addition to hemoglobin, the contrast between normal and abnormal colorectal tissue may involve other components such as water and lipid. Thus, a wider range of wavelengths has achieved superior classification performance. These results suggest that multiwavelength SFDI may be possible to assist clinical decision-making in several aspects. First, with a suitable segmentation algorithm, SFDI may be able to assist surgeons with tumor margin detection during surgeries based on tissue spectral absorption and scattering features and morphology changes.[27, 28] Second, the system may provide real-time assistance in detecting early, hard-to-identify lesions from normal endoluminal mucosa—which could lead to earlier identification of cancers. Currently, the data acquisition is using LabView, all image reconstruction and statistical analysis are performed in MATLAB, and the diagnosis is done using Python. It took less than 5 minutes for all the steps for one specimen. To further improve data acquisition and processing speed, a single snapshot SFDI system and algorithm can be adopted to reduce the data acquisition time by reconstructing optical properties with one single image instead of three phase-shifted images.[29] A deep learning-based reconstruction algorithm may also have the potential to reduce computational time.[30]

In the above-described models, an optimal cut-off value was determined to obtain both high sensitivity and high specificity. However, in clinical practice, a high-sensitivity value is critical because we cannot afford to miss a malignant lesion.[31] Under this decision criterion, the AdaBoost classifier would have correctly detected cancer and adenomatous polyps as abnormality with 95% sensitivity and 61% specificity.

This study must be considered in the context of several limitations. First, all imaged specimens were imaged ex vivo. The human in vivo environment is likely more complex and may produce different results in actively perfused tissues. Besides, all ex vivo specimens had been identified as abnormal with white-light endoscopy prior to resection and imaging via SFDI. Within our sample set, we did not encounter any patients with incidentally found neoplasia. Further in vivo study is needed to test the ability of AdaBoost-based SFDI to identify abnormities otherwise overlooked by standard endoscopy. Additionally, the system was not tested in the setting of all colorectal abnormalities, such as inflammatory bowel disease or non-neoplastic polyps. In the future, more patients including these tissue types need to be included for a complete evaluation of the AdaBoost-based SFDI system. In the future, more patients including thorough tissue types need to be included for a complete evaluation of the AdaBoost-based SFDI system. Fine-tune of the classifier’s hyperparameters should also be performed as we are targeting 100% sensitivity with high specificity in clinical application.

Abbreviations:

AdaBoost

adaptive boosting

AUC

area under a curve

CAD

computer-aided diagnosis

RBF

radial basis function

ROC

receiver operating characteristic

SFDI

spatial frequency domain imaging

SVM

support vector machine