Research Article  Open Access
Breast Tumor Detection and Classification Using Intravoxel Incoherent Motion Hyperspectral Imaging Techniques
Abstract
Breast cancer is a main cause of disease and death for women globally. Because of the limitations of traditional mammography and ultrasonography, magnetic resonance imaging (MRI) has gradually become an important radiological method for breast cancer assessment over the past decades. MRI is free of the problems related to radiation exposure and provides excellent image resolution and contrast. However, a disadvantage is the injection of contrast agent, which is toxic for some patients (such as patients with chronic renal disease or pregnant and lactating women). Recent findings of gadolinium deposits in the brain are also a concern. To address these issues, this paper develops an intravoxel incoherent motion (IVIM) MRIbased histogram analysis approach, which takes advantage of several hyperspectral techniques, such as the band expansion process (BEP), to expand a multispectral image to hyperspectral images and create an automatic target generation process (ATGP). After automatically finding suspected targets, further detection was attained by using kernel constrained energy minimization (KCEM). A decision tree and histogram analysis were applied to classify breast tissue via quantitative analysis for detected lesions, which were used to distinguish between three categories of breast tissue: malignant tumors (i.e., central and peripheral zone), cysts, and normal breast tissues. The experimental results demonstrated that the proposed IVIMMRIbased histogram analysis approach can effectively differentiate between these three breast tissue types.
1. Introduction
Breast cancer is the most prevalent cancer and a major cause of death in women worldwide [1, 2]. Early diagnosis is critical for increasing the survival rate among women with breast cancer. However, current basic imaging approaches (e.g., mammography and ultrasonography [US]) have their limitations, especially in patients with implants, postoperative scarring, and dense breast parenchyma. These factors occasionally result in improper management of breast cancer [3–7]. Recent studies demonstrate that dynamic contrastenhanced magnetic resonance imaging (DCEMRI) techniques are the most promising methods for detecting tumors and assessing therapeutic response [7–9]. Furthermore, the additional diagnostic value of breast MRI is that it detects more multifocal, multicentric, and contralateral diseases [10, 11]. The sensitivity of DCEMRI in the diagnosis of breast cancer is as high as 88%100% [12, 13]. However, the specificity of DCEMRI varies from 50% to 97% [11, 14–16]. This variability may result from different lesion criteria and tumor heterogeneity used among studies [15–17]. Thus, differentiating a benign lesion from a malignant lesion using DCEMRI remains a challenge. To increase breast MRI specificity, one of the latest advancements in MRI technology, the apparent diffusion coefficient (ADC), which is obtained from diffusionweighted imaging (DWI), was developed to differentiate between benign and malignant breast lesions with 91.5% sensitivity and 86.5% specificity [18, 19]. A typical illustration of the regions of the breast tissues is shown in Figure 1.
DWI is a nocontrastenhanced MRI technique. It is useful for characterizing cancerous tissue heterogeneity by directly reflecting the Brownian motion of water molecules in body tissues, particularly in breast cancer [19–21]. It can obtain the physiological characteristics of breast tissue with the assumption of the monoexponential decay to quantitate the analysis of the water molecules’ ADC [22]. In malignant tumors, the ADC value is usually lower than that of the normal tissue because of the restriction of water diffusion resulting from increased cellularity of the tissue and reduced extracellular space in the tumor. Based on the findings of a metaanalysis [23], the specificity of DWI is higher than contrastenhanced MRI, although there is a wide range in threshold values from 0.90 mm^{2}/s to 1.60 x 10^{−3} mm^{2}/s to account for the use of different b values. The specificity of DWI and DCEMRI is 84% and 72% [23].
In vivo, microscopic motion of water molecules detected by DWI is influenced by the diffusion of water molecules in the tissue structure and by the microcirculation of blood in the capillary network [24]. Signal attenuation on monoexponential DWI therefore presents a linear relationship but does not truly reflect the microstructure changes in organization. The reason for this finding is that the ADC value is overestimated due to the fact that it contains two types of information: (1) microcirculation perfusion of blood in a capillary network and (2) water molecule diffusion of the tissue structure. Intravoxel incoherent motion (IVIM), which was first described by Li Bihan et al. in 1986 [25], separates microcirculation perfusion and water molecule diffusion by using a biexponential model [22, 25, 26]. Using this biexponential model, DWI is able to use signal intensity produced by multiple bvalues to analyze water/tissue diffusitvity. At low bvalues, the signal intensity reflects water diffusion in tissues as well as microcirculation in a capillary network. At high bvalues, the signal intensity reflects tissue diffusivity. Through IVIM, biomarkers have been resolved as the perfusion fraction (PF), water molecule diffusivity (D), and pseudodiffusion (D). Recent studies demonstrate that the IVIM model is helpful for the differential diagnosis of breast lesions [27–31]. In general, a MR image can be viewed as a multispectral remote sensing image, in which each spectral band image can be considered as an image acquired by a particular pulse sequence [32]. Hyperspectral imaging has become an emerging remote sensing technology that improves traditional multispectral imaging by using hundreds of contiguous spectral bands. With such fine spectral resolution, many unknown subtle substances can be revealed for subpixel detection and mixed pixel analysis [33]. The application of hyperspectral imaging techniques in MR brain imaging has also been investigated [33–35]. Many medical imaging techniques are generally supervised and require training samples provided a priori such as maximum likelihood classification (MLC), Knearestneighbor (KNN) classification [36], neural networks [36], support vector machine (SVM) [36], and Fisher’s linear discriminant analysis (FLDA) [37]. To avoid this issue, the training samples should be obtained directly from the data to be processed in an unsupervised method. In this instance, we take advantage of the automatic target generation process (ATGP) developed by Ren and Chang [38] in conjunction with the spectral angle mapper (SAM) method, which allows users to select training samples automatically. This procedure is then followed by a supervised hyperspectral subpixel target detection technique, called kernelbased constrained energy minimization (KCEM), developed by Jiao and Chang [39], to detect breast lesion tissues for quantitative analysis for IVIMDWI. This technique considers MR image slices acquired by pulse sequences and DWI with weighting factors of 12 bvalues as a 12band multispectral image cube in which each sequenceacquired image is viewed as a spectral band image. The data set to be used for experiments was collected from 25 patients for IVIM imaging examinations.
The purpose of this study was three folds: (1) to measure IVIM parametric maps (i.e., tissue diffusivity [D], perfusion fraction [PF], pseudodiffusion coefficient [D], and slope) obtained from multiple b values of normal breast tissue, cysts, and malignant tumors; (2) to determine whether these parameters can be used to distinguish between benign and malignant breast lesions; and (3) to calculate advanced diffusion MRI metrics from histogram analysis (i.e., mean and median), and heterogeneity.
2. Background
2.1. DiffusionWeighted Imaging
Diffusionweighted imaging (DWI) is sensitive to the thermally driven random motion of water molecules. Molecular diffusion is the random motion of molecules, called Brownian motion [25, 40]. The motion is modified by the local tissue environment and the cell membranes. The motion of water molecules is more restricted in tissues with a high cellular density than in tissues with low cellularity. Signal intensity in DWI is inversely proportional to the degree of water molecule diffusion. The histologic structure is influenced by diffusion. The change of water diffusion in tissues is useful for MR imaging and can be used as a multifaceted tool to characterize tissue structures and identify and differentiate disease processes. Furthermore, DWI can be used to calculate the ADC, which can quantitatively measure the degree of tissue diffusivity.
Spinecho echoplanar DWI is the most popular clinical technique to produce diffusionweighted images in which diffusion is described by the following monoexponential equation [25]:within which S_{b} is DWI with diffusion attenuation at the b value and is DWI without diffusion attenuation, is the diffusion gradient, γ is the proton gyromagnetic ratio, and δ is the diffusion gradient duration. Equation (2) presents b value (expressed in seconds per square millimeter) and represents the strength of diffusion weighting in which S_{b} is the signal intensity with the diffusion gradient. S_{0} is the signal intensity without the diffusion gradient. The monoexponential model is traditionally used to calculate the decay of ADC in (2) and results from two b values, b_{1} and b_{2}, as follows:
2.2. Intravoxel Incoherent Motion Imaging
Using the value of ADC in (3) does not sufficiently describe tissue behavior. To address this issue, Le Bihan et al. [25] proposed a novel approach, called IVIM MRI, which demonstrated that the ADC values of data samples obtained at low b values can be used to measure two IVIM parameters: (1) pseudodiffusion (D) and (2) the perfusion fraction (PF) of pure molecular diffusion and microcirculation, or blood perfusion [25, 26]. In particular, D and PF can be used as quantitative biomarkers to describe changes in diffusivity and microcapillary perfusion of tissues. Increasing the bvalue can change the values of D and PF and therefore can reflect the diffusivity of different tissues and tissue microcapillary perfusion.
DWI is influenced by the diffusion of water molecules, which is affected by microcapillary blood in the capillary network in tissue structures. Both processes cause phase dispersion in DWI, and consequently signal attenuation. Le Bihan et al. [25] described the behavior of protons that displayed signal attenuation in DWI as IVIM imaging, which does not represent a linear relationship as shown in Figure 2. The biexponential model involves curve fitting for IVIM imaging to separate the estimation of tissue diffusivity from perfusion by using multiple b values, in the following equation:
where D is the pure diffusion coefficient, S_{b} is the signal intensity in the pixel with diffusion gradient, and S_{0} is the signal intensity in the pixel without a diffusion gradient. This biexponential analysis, based on the IVIM sequence, can be used to describe the mixture of perfusion and diffusion resulting from DWI in three parameters: pure diffusion coefficient [D], perfusion fraction [PF], and pseudodiffusion coefficient [D].
2.3. Band Expansion Process
Intravoxel incoherent images do not provide sufficient spectral information to be processed as hyperspectral images; therefore, the band expansion process (BEP) developed by Ren and Chang [41] is used to resolve this issue for IVIM analysis. This process produces new images from the original IVIM images via nonlinear functions. Its premise is similar to the moment generating function used by a random process that makes use of the statistics of all orders to describe probabilistic behaviors of a random process. In the BEP, only the statistics of the second order of IVIM images, autocorrelation, and crosscorrelation functions are used for this purpose. Such nonlinearly generated images provide nonlinear spectral information contained in IVIM images and can help improve data analysis.
The BEP presented in this section is a nonlinear process using correlation functions to generate new band images from the original set of multispectral images.
The BEP for IVIM images is as follows.
Step 1. The order IVIM images are obtained as follows: _ = the set of original IVIM images
Step 2. The order correlated IVIM images are obtained as follows: (i)= the set of autocorrelated IVIM images;(ii) = the set of crosscorrelated IVIM images.
In case a rescaling is needed, auto or crosscorrelated IVIM images can be normalized by variances in the IVIM images such as and .
Step 3. The order correlated IVIM images are obtained as follows:(i) = the set of autocorrelated IVIM images;(ii) = the set of two crosscorrelated IVIM images;(iii)= the set of three crosscorrelated IVIM images.
As in Step 2, auto or crosscorrelated IVIM images can be similar. Therefore, it may require normalization by the variances in the IVIM images. For example,
Step 4. Nonlinear correlated IVIM images are obtained as follows:(i) = the set of IVIM images stretched out by the square root;(ii) = the set of IVIM images stretched out by the logarithmic function.Figure 3 shows an example of IVIM images (a) ground truth (b) compared to the one in (c) without using BEP.
(a)
(b)
(c)
2.4. Automatic Target Generation Process
Once IVIM images are nonlinearly expanded, they are then processed by an unsupervised target detection algorithm, called automatic target detection algorithm (ATGP), to detect potential unknown targets without prior knowledge. The procedure is briefly described as follows.
Let X be a data matrix formed by data sample vectors,, (i.e., .)The norm of the data matrix X is defined by the equation, as follows: where is the length of the vector as defined by and assuming that . The norm of data matrix X in (6) can further be expressed by
which is exactly the brightest pixel and its norm has the maximum vector length. The maximum l_{2} norm defined by (7) is indeed the maximum pixel vector length corresponding to the brightest pixel vector in the data set.
Using (7), the ATGP produces a sequence of orthogonal subspace projections (OSPs) as follows:Thus, if X is the original hyperspectral image cube, the ATGP first selects an initial target pixel that yields the norm of the space X, denoted by via (6). It then projects the space X into a subspace orthogonal to via with . The resulting subspace is denoted by . The ATGP selects a first target pixel that yields the norm of space X_{1}, denoted by via (7), and then projects space X into a subspace orthogonal to via with . The resulting subspace is denoted by . The same procedure is repeated until the stopping rule is satisfied, (i.e., the number of target pixels for the ATGP required to extract). The details of its algorithmic implementation are as follows.
(1) Initial condition: the variable is the prescribed error threshold, and is a pixel with the brightest intensity value (i.e., the maximal gray level value.) Set k=0.
(2) For , is applied via (8) with U=[t_{0}] to all image pixels; r in the image. The target t_{k} is generated at the stage, which has the maximum orthogonal projection, as follows:If in which can be any target discrimination measure, then Step 2 is repeated. The algorithm is otherwise terminated. At this point, all generated target pixels are the desired targets.
2.5. Spectral Angle Mapper
The targets generated by ATGP are single targets. To find other targets that have spectral signatures similar to those of the ATGPdetected targets, we used the SAM [42] as a spectral measure to calculate the similarity between two vectors by finding their angles [33]. In particular, s and t were two vectors. The SAM is defined byIn general, it is used to compare a spectral vector against a reference vector in terms of their between angle. Therefore, the larger the angle is in (10), the less is the similarity.
2.6. KernelBased Constrained Energy Minimization
Once targets were determined and identified by using SAM, a target detector was further designed for target detection. In this section, a finite impulse response (FIR) linear filter, called constrained energy minimization (CEM) developed in several studies [42–44], was used for this purpose.
More specifically, CEM is derived from the linearly constrained minimum variance originally proposed by Frost for adaptive beam forming [45]. If a hyperspectral image is represented by a collection of image pixel vectors, denoted by in which for is an Ldimensional pixel vector, then N is the total number of pixels in the image and L is the total number of spectral channels. Furthermore, is specified by a desired signature of interest to be used for target detection. The goal is to find a target detector that can detect data samples specified by the desired target signal d via a FIR filter with L filter coefficients, , denoted by an Ldimensional vector , which minimizes the filter output energy, subject to the constraint . More specifically, if y_{i} denotes the output of the designed FIR filter resulting from the input , then y_{i} can be expressed by the following equation:
The average energy of the filter output is given by (12), in which (13) provides the autocorrelation sample matrix of the image:
The goal was to solve the following linearly constrained optimization problem:in which is the variance resulting from signals not passing through the filter. The optimal solution to (14), as described in [17–19], is presented in (15) and (16), as follows:which is the CEM error considered as the least energy resulting from unwanted signal sources impinging on an array of sensors. With the optimal weight, , specified by (15), a filter, called CEM and denoted by , was derived from a previous study [43] and can be specified byConstrained energy minimization is a linear filter and IVIM images are nonlinearly generated multispectral images; therefore, nonlinear separability may present an issue for CEM. To mitigate this problem, CEM was further expanded to a kernel version of CEM, called kernel CEM (KCEM), which was derived as follows [39]: . The R in (13) can be expanded in feature space F as presented in The CEM derived from (17) can be extended to a kernel version of CEM (KCEM), given by
By virtue of (19), the KCEM can be calculated in feature space F via the kernel trick without mapping the original data sample vectors into the feature space. Several commonly used kernel functions can be used for (19). In this paper, the Gaussianbased radial basis function (RBF) kernel is given bywhich was used for experiments to implement KCEM. There are two reasons for choosing Gaussianbased RBF kernel: (1) it is a translation invariant kernel, and (2) its associated nonlinear map is smooth. Even when the spectral signatures of a given hyperspectral data set are subject to irregular illumination, the translationally invariant kernels typically provide robust detection. This finding is because it depends only on the difference between x and y, not the absolute positions of a single spectral vector. The smooth nonlinear mapping associated with the Gaussian RBF kernel means that, after nonlinear mapping, the topographic ordering of the data in the input space is preserved in the feature space. The mapped data sample vectors in the feature space also occupy the small subspace of the feature space, in which data belonging to different classes can be separated to a larger extent than in the input space. As a result, the Gaussian RBF kernelmapped feature space in the target detection task is more effective.
2.7. Thresholding
The KCEMgenerated detection map is real valued; therefore, it requires a thresholding technique to produce a binary image to show detected lesion tissues. In the literature, many thresholding techniques have been proposed such as Otsu’s method [46], two entropic thresholding methods, local entropy and joint entropy methods [45], and three relative entropy methods, local relative entropy, joint relative entropy, and global relative entropy methods in [47, 48]. Based on our extensive experiments, the local entropy method [45] was the best among entropybased methods and outperformed the most widely used Otsu’s method. The local entropy method was consequently selected as the thresholding technique to segment lesions from the KCEMproduced detection maps.
2.8. Decision Tree
A decision tree is a nonparametric supervised learning method used for classification and regression. It builds classification or regression models in the form of a tree structure, in which each internal node denotes a test on an attribute, each branch represents an outcome of the test, and leaf nodes represent classes or class distributions. It breaks a dataset into smaller and smaller subsets while incrementally developing an associated decision tree at the same time.
Decision trees can be used for categorical and numerical data. The goal is to create a model that can be used to predict the value of a target variable by learning simple decision rules inferred from the data features.
The lesions detected by the local entropythresholder KCEM were further used as the inputs of a decision tree. In order to perform breast lesion classification, the decision tree is specifically developed by a nonparametric supervised learning method, which is generally used for classification and regression. It builds classification or regression models in the form of a tree structure, in which each internal node denotes a test based on an attribute. Each branch represents an outcome of the test, and leaf nodes represent classes or class distributions. Categorical and numerical data can be used in such decision trees. The goal was to create a model that could be used to predict the value of a target variable by learning simple decision rules inferred from the data features.
3. Materials and Methods
3.1. Patient Selection
After obtaining institutional review board approval to conduct this prospective study, 25 patients with breast cancer (20 malignant lesions and 5 cysts) underwent MR examination in our institution from May 2014 to June 2015. Written informed consent was obtained from all patients. The patient inclusion criteria were as follows: (1) newly diagnosed with breast cancer confirmed by needle biopsy; (2) the patient was not receiving neoadjuvant chemotherapy or hormonal treatment at the time of imaging; (3) tumor size (solid portion, referring to the T2weighted and contrastenhanced T1weighted images)> 1 cm in DW imaging; (4) unilateral breast cancer; (5) no previous breast surgery history, and (6) no motion artifact. The definitive diagnosis of cyst was obtained by means of ultrasonographic (US) and breast MR imaging findings. The simple cyst shows thin wall, circumscribed margin and anechoic at US, low T1 signal intensity, high T2 signal intensity at fatsuppressed images, and no enhancement at postcontrast fatsuppressed images. All MRI studies were reviewed by an experienced radiologist (S.W.C) who had access to all patient information.
3.2. Experimental Materials
MRI was conducted using a 1.5T system (Aera; Siemens, Erlangen, Germany) using a body coil as the transmitter and a dedicated 16channel receiver coil. The patient was prone and head first. Conventional T1 and T2weighted images were acquired with and without fat suppression. The axial IVIM images were obtained by using singleshot spinecho echoplanar imaging (EPI). The axial IVIM images with bilateral breast coverage were acquired (TR/TE: 5800/68 ms; FOV: 320 mm^{2} x 264 mm^{2}; matrix: 132 x 160 x 30; reconstructed voxel size: 2 mm^{3}×2 mm^{3}×3 mm^{3}; with spectral presaturation inversion recovery and diffusion sensitization in the anteriorposterior direction applied with weighting factors of b = 0, 15, 30, 45, 60, 100, 250, 400, 550, 700, 850, 1000 sec/mm^{2}. Total scan time required by an IVIM imaging scan was 6.3 minutes. Axial T1weighted DCE_MRI (TR/TE, 4.5/1.8 ms; FA, 12; FOV, 320 mm^{2} x 320 mm^{2}; matrix size, 512 x 512; slice thickness, 1.5 mm) was acquired using the gradient echo sequence. The contrast media were captured after precontrast and four consecutive time points (i.e., 60second interval time) after administrating of gadolinium (Gadovist, 1.0 mmol/mL) by using a power injector, at a flow rate of 2.0 mL/s, followed by 20 cc normal saline flush.
Sixteen sets of MR images were used for the experiments, which were acquired by spinlattice relaxation time (T1), spinspin relaxation time (T2), proton density (PD), and T1DCE along with IVIM based DW images, for which T1DCE was involved using contrast agents at different times. These IVIMDWI images were considered as the multispectral images.
Image analysis was conducted on a personal computer using inhouse software written in MATLAB (Math Works, Natick, MA). The IVIM parameters were calculated from the DW images at all bvalues by using the biexponential model, described by Le Bihan et al. [25, 26]. At first, the assumption parameters derived from the IVIM image using all bvalues were compared with DCE_MRI to obtain reliable results. A radiologist with 13 years of experience read all MRI breast images to determine the tumor region of interest (ROI) and exclude areas of necrosis, bleeding, and edema. After, parameters were mapped again onto the DCE_MRI as the ground truth. There is no need of ROI in our study because entire images were processed, pixel by pixel. Therefore, interobserver variability was limited.
3.3. Preprocessing
Before processing the data, a technique, called nonparametric nonuniform intensity normalization (N3), was used to adjust the nonuniformity of the MR images. Multiple bvalues images of IVIM displacement caused by respiration must be corrected; therefore, an alignment. was used to register all pixels. Once all images were aligned, an automatic method to extract the breast region was developed, as follows. Automatic process to extract the breast region: Step 1. A high pass filter was used to enhance the boundary of the breast image. Step 2. A Sobel edge detector was used to extract the contour of the breast image. Step 3. The hitormiss transformation function was applied to the flat area between two breasts to extract the sternum. Step 4. Otsu’s method was implemented to produce a binary image. Step 5. A morphological opening using a structural element with a window (size, 2x2 pixels) was used to detach the connected components in the binary image. Step 6. A morphological opening was used to isolate the largest connected component. Step 7. A Sobel edge detector was applied again to obtain the silhouette of the breast image.
3.4. Breast Lesion Tissue Detection
The BEP was used to create additional IVIM images for the IVIMMRI classification. After BEP, the original images were expanded to 90 band images. Since KCEM requires the knowledge of the desired signature d, therefore, we first used ATGP to determine the initial target pixels. These pixels were further used to find target pixels with similar spectral signatures measured by SAM as the training samples that could be used as the desired d for KCEM. The detected region was further verified by an experienced breast radiologist and by tumor biopsy. The detected part of each breast lesion tissue on an ADC map was calculated by all b values by using (3). For all lesion tissues, the D, D, PF, and slope values were further calculated on a pixelbypixel basis by using (4). The lesion tissues of 25 individual patients were tested and the statistical values of all parameters were used as the standard for the decision tree.
3.5. Quantitative Analysis
This study was conducted on IVIMDWI using 12 b values: b=0, 15, 30, 45, 60, 100, 250, 400, 550, 700, 850, 1000 s/mm^{2}. The biexponential model was proposed by Le Bihan et al. [25]; microcapillary perfusion of the blood has no specific orientation, which depends on the velocity of flowing blood and the vascular architecture. The effect of pseudodiffusion on signal attenuation in an IVIMDWI voxel is also b value dependent. The biexponential analysis to describe DWI comprises a mixture of perfusion and diffusion.
The analytic model of IVIM is a biexponential and monoexponential model, demonstrated in (4). The monoexponential model of IVIM is expressed in (1). The ADC was obtained by using all bvalues (01000 s/mm^{2}) and then fitted to the (4). Four parameters, D (pure molecular diffusion), D (perfusionrelated diffusion), PF (perfusion fraction), and ADC (apparent diffusion coefficient), were determined after undergoing processing by image preprocessing, detection, and analysis of the lesion. Through these parameters, the lesion area could be further compared.
3.6. Histogram Analysis
In this section, a quantitative analysis was conducted on the IVIMDWI images obtained by using 12 diffusionweighted b values 0, 15, 30, 45, 60, 100, 250, 400, 550, 700, 850, 1000 s/mm^{2}. The faster a molecule diffuses, the greater is the attenuation and the weaker is the corresponding pixel signal intensity on a DWI image. The biexponential analysis used in. (4) to describe DWI comprises a mixture of perfusion and diffusion, whereas a simple monoexponential analysis only uses the bvalues of 0 s/mm^{2} and 1000 s/mm^{2} via (3).
The ADC was obtained by the monoexponential model in (3) using bvalues ranging from 0 to 1000 s/mm^{2} and then fitted to the biexponential model in (4). Four parameters result from the series of processes, preprocessing, detection, and analysis of the lesion tissue: D (pure molecular diffusion), D (perfusionrelated diffusion), PF (perfusion fraction), and slope. Figure 4 shows the flowchart for IVIM imaging processing.
Details of stepbystep implementations are as follows.
Step 1. The ADC, D, D, PF, and slope values were calculated to find the threshold values for central tumor, peripheral tumor, normal tissue, and cyst.
Step 2. The ADC, slope, D, D, and PF values of each pixel were used to run the decision tree to determine to which of the three categories a pixel belongs to: tumor, normal tissue, or cyst.
Step 3. If a pixel indicated a tumor, then histogram analysis was used to determine whether the pixel indicated a central tumor or a peripheral tumor.
Step 4. Otherwise, the pixel indicated a nontumor lesion. The decision tree was then run again to determine whether the pixel was normal tissue, cyst, or other.
Histogram analysis can provide information beyond the IVIM parameters in the spatial distribution histogram, such as the skewness and kurtosis of the parameter distributions. Skewness is a measure of symmetry, where kurtosis is a measure of whether the data are peaked or flat relative to a normal distribution.
4. Results and Discussion
Figure 5 shows a case of breast cancer on IVIM images at different b values: (a) b = 0 (b) b =15 (c) b = 30 (d) b = 45 (e) b = 60 (f) b = 100 (g) b = 250 (h) b = 400 (i) b = 550 (j) b = 700 (k) b = 850 (l) b = 1000 s/mm^{2}. (m) The dynamic T1DCE MR image. The lesion mappings are shown in image (n). The red and green pixels indicate the central tumor region and the peripheral tumor region, respectively.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
(m)
(n)
Figure 6 shows the ADC, D, D, PF, and slope values of the histogram analysis in the range of the tumor. The red and green pixels show the central tumor and peripheral tumor region, respectively.
To automatically find potential lesion pixels, ATGP was used to find the brightest point, and then the SAM algorithm was applied to locate the points with spectral signatures similar to the ATGPfound point that could be used as desired target signatures for KCEM to detect a suspected tumor range. However, the signal strength was insufficiently strong to be curbed and the tumor area could not be fully detected. Therefore, a local entropic thresholding method was applied to KCEM for suspected tumor range improvement. Figure 7(a) shows the results of tumor detection using KCEM.
(a)
(b)
(c)
The local entropy method was used to find an appropriate threshold value. Threshold results are presented in Figure 7(b) in which areas with low intensity values could still be detected by KCEM. However, using the local entropy method to adjust threshold values could only capture areas that are different from the background. In addition, in this technique, the threshold value needs to be manually adjusted slice by slice; thus, it is also subjective and time consuming. To address this issue, a histogram analysis using the five parameters of ADC, D, D, PF, and slope was further used to determine the true scale of the suspected tumor area. As a result, some dense parts of the tissues could also be detected, as shown in Figure 7(c).
The monoexponential and biexponential models [i.e., (3) and (4), respectively] were then applied to the IVIMDW images in which the relative signal intensity decays of central tumor, peripheral tumor, cyst, and normal tissue are specified by the five parameters of ADC, D, D, PF, and slope. We first obtained MRIIVIM images and then obtained contrastenhanced MR images as the ground truth. We used a patient with breast cancer to demonstrate how these five parameters could be used to classify breast tissues. In Figure 7(b), the tumor area was marked by red pixels and was detected by the KCEM method using the local entropy method. For the pixels in the red area, the calculated values of D were larger than those of other areas without tumor. The D and ADC values in the tumor area were both lower than those in the normal breast tissue area. The D, PF, and D values obtained from IVIM biexponential model fitting and the ADC value obtained from monoexponential model fitting were also used to classify breast tissues into three categories, the tumor (central and peripheral tumor), cyst, and normal tissues. These values are tabulated in Tables 1 and 2.


In particular, Table 2 tabulates the average results of ADC, D, D, PF, and slope parameters of the total cases (i.e., test results of 25 patients) based on mono and biexponential model analysis. For all images, each slice contained 12 IVIM images from b=0 mm^{2}/s to b=1000 mm^{2}/s. Each case may contain 20 IVIM slices and more than 40 T1, T2, PD, and T1delay slices.
A boxplot is a convenient way of graphically depicting groups of numerical data through their quartiles. Boxplots can show the minimum (Q1), the medium (Q3), interquartile range (interquartile range = Q3 – Q1), and maximum values. The boxplot in Figure 8 illustrates the distribution for all 25 patients of biexponential median values for (a) ADC, (b) slope, (c) D, (d) D, and (e) PF.
(a) ADC
(b) Slope
(c) D
(d) D
(e) PF
In practice, the histological features of parameter D in IVIM are similar to those of parameter ADC in DWI. However, it is better to use parameter D than to use the parameter ADC because it can accurately demonstrate true diffusion without being affected by perfusionrelated diffusion. In this study, we also found that parameter D has significant differences between malignant tumors and cysts, so the D value in IVIM can effectively complement existing traditional DCEMRI and DWI to distinguish between malignant breast tumors and cysts. The mean D value of cysts is higher than the mean D value of malignant tumors, close to the result of Ma et al. [31]. Inconsistent parameter results may be due to differences in the number and distribution of low and high b values used in IVIM. From the experimental results, we also found that the D and f values of benign and malignant breast lesions have relatively low effects, so the D parameter is given a higher weight in the part of the detection algorithm decision tree. One of the main limitations of our study is that the number of patients is not large enough, including variety of diseases. Second, the selection and appropriate number of b values for breast IVIM are still unknown.
Currently DCEMRI is still one of the most important and mature methods in breast lesion imaging. Although DCEMRI has high sensitivity for detecting breast cancer, its disadvantage is that it requires injection of contrast agents. However, the significant association between gadoliniumbased contrast agents and the incidence of nephrogenic systemic fibrosis (NSF) in patients with advanced renal disease has been reported. DWI is a noninvasive method that uses magnetic resonance imaging to observe the diffusion of water molecules in living tissue. Current results show that IVIMDWI helps to understand tissue characteristics and distinguish between benign and malignant lesions. Therefore, DWI may have a role as an alternative diagnostic technique for detecting breast lesion without the need of contrast agent.
5. Conclusions
In this paper, hyperspectral imaging technique was developed to detect and classify breast tissue lesions, which can be implemented in two stages. The first stage is the detection of the breast tissue, followed by the classification in the second stage. This study makes several contributions to the breast imaging process. First, the most important contribution is to process IVIMDW images as multispectral images, which can be further extended to hyperspectral images by BEP. Second, an unsupervised hyperspectral target detection algorithm, ATGP, was applied to the expanded IVIMDW images to determine potential lesion pixels, which can be used as the desired target knowledge for the followup target detection algorithm, KCEM, to locate suspected areas of breast cancer lesion tissues. Third, a thresholding technique was used to extract the lesion areas. Fourth, the detected lesion pixels were used to calculate five parameters: ADC, D, D, PF, and slope via the biexponential model specified by (4). The calculated values of ADC, D, D, PF, and slope were ultimately input into a decision tree using histogram analysis to classify the detected breast lesion tissues into three categories: malignant tumor (central and peripheral tumor), cysts, and normal tissue. It is our belief that the work presented in this paper is the first to process breast IVIMDW images as hyperspectral images. Our findings show that hyperspectral imaging techniques could be used to detect breast lesion tissues. We also believe that this work is the first to take advantage of a biexponential model to classify breast lesion tissue by using the five parameters of ADC, D, D, PF, and slope. This study was limited by its small sample size and it did not employ multiple comparison correlations. Therefore, the usefulness of the IVIM parameters for lesion differentiation requires further investigation.
Data Availability
The data used to support the findings of this study are included within the article.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this paper.
References
 L. A. Torre, R. L. Siegel, E. M. Ward, and A. Jemal, “Global cancer incidence and mortality rates and trends—an update,” Cancer Epidemiology, Biomarkers & Prevention, vol. 25, no. 1, pp. 16–27, 2016. View at: Publisher Site  Google Scholar
 World Cancer Research Fund International, Breast cancer statistics, 2018, https://www.wcrf.org/int/cancerfactsfigures/dataspecificcancers/breastcancerstatistics.
 M. A. Helvie, L. K. Joynt, R. L. Cody, L. J. Pierce, D. D. Adler, and S. D. Merajver, “Locally advanced breast carcinoma: accuracy of mammography versus clinical examination in the prediction of residual disease after chemotherapy.,” Radiology, vol. 198, no. 2, pp. 327–332, 1996. View at: Publisher Site  Google Scholar
 B. E. Adrada, L. Huo, D. L. Lane, E. M. Arribas, E. Resetkova, and W. Yang, “Histopathologic correlation of residual mammographic microcalcifications after neoadjuvant chemotherapy for locally advanced breast cancer,” Annals of Surgical Oncology, vol. 22, no. 4, pp. 1111–1117, 2015. View at: Publisher Site  Google Scholar
 W. A. Berg and P. L. Gilbreath, “Multicentric and multifocal cancer: wholebreast us in preoperative evaluation,” Radiology, vol. 214, no. 1, pp. 59–66, 2000. View at: Publisher Site  Google Scholar
 A. A. Tardivon, L. Ollivier, C. El Khoury, and F. Thibault, “Monitoring therapeutic efficacy in breast carcinomas,” European Radiology, vol. 16, no. 11, pp. 2549–2558, 2006. View at: Publisher Site  Google Scholar
 R. Croshaw, H. ShapiroWright, E. Svensson, K. Erb, and T. Julian, “Accuracy of clinical examination, digital mammogram, ultrasound, and mri in determining postneoadjuvant pathologic tumor response in operable breast cancer patients,” Annals of Surgical Oncology, vol. 18, no. 11, pp. 3160–3163, 2011. View at: Publisher Site  Google Scholar
 F. Pediconi, C. Catalano, A. Roselli et al., “The challenge of imaging dense breast parenchyma: is magnetic resonance mammography the technique of choice? A comparative study with xray mammography and wholebreast ultrasound,” Investigative Radiology, vol. 44, no. 7, pp. 412–421, 2009. View at: Publisher Site  Google Scholar
 K. Vassiou, T. Kanavou, M. Vlychou et al., “Characterization of breast lesions with CEMR multimodal morphological and kinetic analysis: Comparison with conventional mammography and highresolution ultrasound,” European Journal of Radiology, vol. 70, no. 1, pp. 69–76, 2009. View at: Publisher Site  Google Scholar
 M. E. Starver, C. E. Loo, E. J. Rutgers et al., “MRImodel to guide the surgical treatment in breast cancerpatients after neoadjuvant chemotherapy,” Annals of Surgery, vol. 251, no. 4, pp. 701–707, 2010. View at: Publisher Site  Google Scholar
 G. L. G. Menezes, F. M. Knuttel, B. L. Stehouwer, R. M. Pijnappel, and M. A. A. J. Van Den Bosch, “Magnetic resonance imaging in breast cancer: A literature review and future perspectives,” World Journal of Clinical Oncology, vol. 5, no. 2, pp. 61–70, 2014. View at: Publisher Site  Google Scholar
 W. Huang, P. R. Fisher, K. Dulaimy, L. A. Tudorica, B. O'Hea, and T. M. Button, “Detection of breast malignancy: diagnostic MR protocol for improved specificity,” Radiology, vol. 232, no. 2, pp. 585–591, 2004. View at: Publisher Site  Google Scholar
 R. M. L. Warren, L. Pointon, D. Thompson et al., “Reading protocol for dynamic contrastenhanced MR images of the breast: sensitivity and specificity analysis,” Radiology, vol. 236, no. 3, pp. 779–788, 2005. View at: Publisher Site  Google Scholar
 W. A. Berg, L. Gutierrez, M. S. NessAiver et al., “Diagnostic accuracy of mammography, clinical examination, US, and MR imaging in preoperative assessment of breast cancer,” Radiology, vol. 233, no. 3, pp. 830–849, 2004. View at: Publisher Site  Google Scholar
 N. H. G. M. Peters, I. H. M. Borel Rinkes, N. P. A. Zuithoff, W. P. T. M. Mali, K. G. M. Moons, and P. H. M. Peeters, “Metaanalysis of MR imaging in the diagnosis of breast lesions,” Radiology, vol. 246, no. 1, pp. 116–124, 2008. View at: Publisher Site  Google Scholar
 M. B. I. Lobbes, R. Prevos, M. Smidt et al., “The role of magnetic resonance imaging in assessing residual disease and pathologic complete response in breast cancer patients receiving neoadjuvant chemotherapy: A systematic review,” Insights into Imaging, vol. 4, no. 2, pp. 163–175, 2013. View at: Publisher Site  Google Scholar
 C. E. Meacham and S. J. Morrison, “Tumour heterogeneity and cancer cell plasticity,” Nature, vol. 501, no. 7467, pp. 328–337, 2013. View at: Publisher Site  Google Scholar
 S. Kul, A. Cansu, E. Alhan, H. Dinc, G. Gunes, and A. Reis, “Contribution of diffusionweighted imaging to dynamic contrastenhanced MRI in the characterization of breast tumors,” American Journal of Roentgenology, vol. 196, no. 1, pp. 210–217, 2011. View at: Publisher Site  Google Scholar
 S. C. Partridge, H. Rahbar, R. Murthy et al., “Improved diagnostic accuracy of breast MRI through combined apparent diffusion coefficients and dynamic contrastenhanced kinetics,” Magnetic Resonance in Medicine, vol. 65, no. 6, pp. 1759–1767, 2011. View at: Publisher Site  Google Scholar
 S. Sinha, F. A. LucasQuesada, U. Sinha, N. DeBruhl, and L. W. Bassett, “In vivo diffusionweighted MRI of the breast: Potential for lesion characterization,” Journal of Magnetic Resonance Imaging, vol. 15, no. 6, pp. 693–704, 2002. View at: Publisher Site  Google Scholar
 M. Iima, K. Yano, M. Kataoka et al., “Quantitative nongaussian diffusion and intravoxel incoherent motion magnetic resonance imaging: differentiation of malignant and benign breast lesions,” Investigative Radiology, vol. 50, no. 4, pp. 205–211, 2015. View at: Publisher Site  Google Scholar
 D.M. Koh, D. J. Collins, and M. R. Orton, “Intravoxel incoherent motion in body diffusionweighted MRI: Reality and challenges,” American Journal of Roentgenology, vol. 196, no. 6, pp. 1351–1361, 2011. View at: Publisher Site  Google Scholar
 X. Chen, W. Li, Y. Zhang, Q. Wu, Y. Guo, and Z. Bai, “Metaanalysis of quantitative diffusionweighted MR imaging in the differential diagnosis of breast lesions,” BMC Cancer, vol. 10, article 693, 2010. View at: Publisher Site  Google Scholar
 Y. Mazaheri, A. Afaq, D. B. Rowe, Y. Lu, A. ShuklaDave, and J. Grover, “Diffusionweighted magnetic resonance imaging of the prostate,” Journal of Computer Assisted Tomography, vol. 36, no. 6, pp. 695–703, 2012. View at: Publisher Site  Google Scholar
 D. Le Bihan, E. Breton, D. Lallemand et al., “MR imaging of intravoxel incoherent motions: application to diffusion and perfusion in neurologic disorders,” Radiology, vol. 161, no. 2, pp. 401–407, 1986. View at: Publisher Site  Google Scholar
 D. Le Bihan, E. Breton, D. Lallemand, M.L. Aubin, J. Vignaud, and M. LavalJeantet, “Separation of diffusion and perfusion in intravoxel incoherent motion MR imaging,” Radiology, vol. 168, no. 2, pp. 497–505, 1988. View at: Publisher Site  Google Scholar
 L. Bokacheva, J. B. Kaplan, D. D. Giri et al., “Intravoxel incoherent motion diffusionweighted MRI at 3.0 T differentiates malignant breast lesions from benign lesions and breast parenchyma,” Journal of Magnetic Resonance Imaging, vol. 40, no. 4, pp. 813–823, 2014. View at: Publisher Site  Google Scholar
 C. Liu, C. Liang, Z. Liu, S. Zhang, and B. Huang, “Intravoxel incoherent motion (IVIM) in evaluation of breast lesions: comparison with conventional dwi,” European Journal of Radiology, vol. 82, no. 12, pp. e782–e789, 2013. View at: Publisher Site  Google Scholar
 E. E. Sigmund, G. Y. Cho, S. Kim et al., “Intravoxel incoherent motion imaging of tumor microenvironment in locally advanced breast cancer,” Magnetic Resonance in Medicine, vol. 65, no. 5, pp. 1437–1447, 2011. View at: Publisher Site  Google Scholar
 X. Mao, X. Zou, N. Yu, X. Jiang, and J. Du, “Quantitative evaluation of intravoxel incoherent motion diffusionweighted imaging (IVIM) for differential diagnosis and grading prediction of benign and malignant breast lesions,” Medicine, vol. 97, no. 26, p. e11109, 2018. View at: Publisher Site  Google Scholar
 D. Ma, F. Lu, X. Zou et al., “Intravoxel incoherent motion diffusionweighted imaging as an adjunct to dynamic contrastenhanced MRI to improve accuracy of the differential diagnosis of benign and malignant breast lesions,” Magnetic Resonance Imaging, vol. 36, pp. 175–179, 2017. View at: Publisher Site  Google Scholar
 M. W. Vannier, T. K. Pilgram, C. M. Speidel, L. R. Neumann, D. L. Rickman, and L. D. Schertz, “Validation of Magnetic Resonance Imaging (MRI) multispectral tissue classification,” Computerized Medical Imaging and Graphics, vol. 15, no. 4, pp. 217–223, 1991. View at: Publisher Site  Google Scholar
 C. I. Chang, Hyperspectral Imaging: Techniques for Spectral Detection and Classification, Springer Science and Business, Berlin/Heidelberg, Germany, 2003.
 C. I. Chang, Hyperspectral Data Processing: Algorithm Design and Analysis, John Wiley & Sons, Hoboken, NJ, USA, 1st edition, 2013.
 Y. C. Ouyang, H. M. Chen, J. W. Chai, C. Chen et al., “Band ExpansionBased OverComplete Independent Component Analysis for Multispectral Processing of Magnetic Resonance Images,” IEEE Transactions on Biomedical Engineering, vol. 55, no. 6, pp. 1666–1677, 2008. View at: Publisher Site  Google Scholar
 C. M. Bishop, Pattern Recognition and Machine Learning, Springer, New York, NY, USA, 2006. View at: MathSciNet
 R. O. Duda and P. E. Hart, Pattern Classification and Scene Analysis, John Wiley & Sons, Hoboken, NJ, USA, 1973.
 R. Hsuan and C. CheinI, “Automatic spectral target recognition in hyperspectral imagery,” IEEE Transactions on Aerospace and Electronic Systems, vol. 39, no. 4, pp. 1232–1249, 2003. View at: Publisher Site  Google Scholar
 X. Jiao, S. S. Shen, P. E. Lewis, and C. Chang, “Kernelbased constrained energy minimization (KCEM),” in Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XIV, vol. 6966, p. 69661S, International Society for Optics and Photonics, Bellingham, WA, USA, 2008. View at: Publisher Site  Google Scholar
 E. M. CharlesEdwards and N. M. de Souza, “Diffusionweighted magnetic resonance imaging and its application to cancer,” Cancer Imaging, vol. 6, no. 1, pp. 135–143, 2006. View at: Publisher Site  Google Scholar
 H. Ren and C.I. Chang, “A generalized orthogonal subspace projection approach to unsupervised multispectral image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 38, no. 6, pp. 2515–2528. View at: Publisher Site  Google Scholar
 J. C. Harsanyi, Detection and Classification of Subpixel Spectral Signatures in Hyperspectral Image Seq uences, University of Maryland Baltimore County, 1993.
 W. H. Farrand and J. C. Harsanyi, “Mapping the distribution of mine tailings in the Coeur d'Alene River Valley, Idaho, through the use of a constrained energy minimization technique,” Remote Sensing of Environment, vol. 59, no. 1, pp. 64–76, 1997. View at: Publisher Site  Google Scholar
 C. CheinI, “Target signatureconstrained mixed pixel classification for hyperspectral imagery,” IEEE Transactions on Geoscience and Remote Sensing, vol. 40, no. 5, pp. 1065–1081, 2002. View at: Publisher Site  Google Scholar
 N. R. Pal and S. K. Pal, “Entropy: a new definition and its applications,” The Institute of Electrical and Electronics Engineers Systems, Man, and Cybernetics Society, vol. 21, no. 5, pp. 1260–1270, 1991. View at: Publisher Site  Google Scholar  MathSciNet
 N. Otsu, “A threshold selection method from graylevel histograms,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 9, no. 1, pp. 62–66, 1979. View at: Publisher Site  Google Scholar
 C. Chang, K. Chen, J. Wang, and M. L. Althouse, “A relative entropybased approach to image thresholding,” Pattern Recognition, vol. 27, no. 9, pp. 1275–1289, 1994. View at: Publisher Site  Google Scholar
 C. I. Chang, Y. Du, J. Wang, S. M. Guo, and P. D. Thouin, “A survey and comparative study of entropic and relative entropic thresholding techniques,” IEE Proceedings Vision, Image Signal Process, vol. 153, no. 6, pp. 837838, 2006. View at: Google Scholar
Copyright
Copyright © 2019 SiWa Chan et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.