Transforming UTE-mDixon MR Abdomen-Pelvis Images into CT by Jointly Leveraging Prior Knowledge and Partial Supervision

Abstract

Computed tomography (CT) provides information for diagnosis, PET attenuation correction (AC), and radiation treatment planning (RTP). Disadvantages of CT include poor soft tissue contrast and exposure to ionizing radiation. While MRI can overcome these disadvantages, it lacks the photon absorption information needed for PET AC and RTP. Thus, an intelligent transformation from MR to CT, i.e., the MR-based synthetic CT generation, is of great interest as it would support PET/MR AC and MR-only RTP. Using an MR pulse sequence that combines ultra-short echo time (UTE) and modified Dixon (mDixon), we propose a novel method for synthetic CT generation jointly leveraging prior knowledge as well as partial supervision (SCT-PK-PS for short) on large-field-of-view images that span abdomen and pelvis. Two key machine learning techniques, i.e., the knowledge-leveraged transfer fuzzy c-means (KL-TFCM) and the Laplacian support vector machine (LapSVM), are used in SCT-PK-PS. The significance of our effort is threefold: 1) Using the prior knowledge-referenced KL-TFCM clustering, SCT-PK-PS is able to group the feature data of MR images into five initial clusters of fat, soft tissue, air, bone, and bone marrow. Via these initial partitions, clusters needing to be refined are observed and for each of them a few additionally labeled examples are given as the partial supervision for the subsequent semi-supervised classification using LapSVM; 2) Partial supervision is usually insufficient for conventional algorithms to learn the insightful classifier. Instead, exploiting not only the given supervision but also the manifold structure embedded primarily in numerous unlabeled data, LapSVM is capable of training multiple desired tissue-recognizers; 3) Benefiting from the joint use of KL-TFCM and LapSVM, and assisted by the edge detector filter based feature extraction, the proposed SCT-PK-PS method features good recognition accuracy of tissue types, which ultimately facilitates the good transformation from MR images to CT images of the abdomen-pelvis. Applying the method on twenty subjects’ feature data of UTE-mDixon MR images, the average score of the mean absolute prediction deviation (MAPD) of all subjects is 140.72±30.60 HU which is statistically significantly better than the 241.36±21.79 HU obtained using the all-water method, the 262.77±42.22 HU obtained using the four-cluster-partitioning (FCP, i.e. external-air, internal-air, fat, and soft tissue) method, and the 197.05±76.53 HU obtained via the conventional SVM method. These results demonstrate the effectiveness of our method for the intelligent transformation from MR to CT on the body section of abdomen-pelvis.

1. Introduction

Deriving synthetic Computed Tomography (sCT) images from Magnetic Resonance (MR) data remains a highly-active area of research. While significant progress has been made for sCT generation in the brain[1][2], head and neck[3]-[5], and pelvis sections[6]-[10], which can easily be imaged by MR due to relatively few tissue types and little physiologic motion [10]-[16], sCT generation for the abdomen represents an unmet need. One application where sCT is valuable is in supporting the MR-only radiation treatment planning (RTP) workflow wherein MR provides soft-tissue contrast superior to that of CT for delineating tumors and organs at risk [14],[15]. A sCT derived from MR could support radiation dose calculations and thus obviate CT scanning the patient [17]. Another application of sCT is for accurate attenuation correction (AC) for PET/MR [10]-[13],[16],[18]-[19]. In fact, Kuhn et al.[20] indicate that Dixon MR pulse sequences provide anatomic lesion correlation that is complementary to abdominal PET so that CT could be obviated if the technical issue of MR-based AC is solved.

From an acquisition perspective, CT and MR technologies are based on different physical phenomena and properties. CT relies on X-ray photons interaction with tissue. X-ray beams are directed through a body section at different angles and beams transmittances are measured. Measurements are then digitally processed to generate three-dimensional images [30]. As for MR imaging, it uses magnetic fields and radiofrequency excitation to quantify the magnetic properties of the hydrogen nuclei (protons) at a certain body section [30]. Different excitation schemes, usually referred to as pulse sequences, are used to improve tissue contrast and spatial localization. Among these, Dixon pulse sequences [78] are commonly available, relatively fast, provide soft tissue contrast, and are used alone or in combination with other pulse sequences. Still, problems remain. Occasionally there is soft tissue vs. fat misclassification and this can lead to significant impact on standard uptake value (SUV) quantification when used for PET AC [21]. Furthermore, lungs and bone both have very low signal when using a Dixon pulse sequence, which makes it difficult to recognize bone and thus leads to particularly large errors in RTP and AC, as bone has the highest attenuation of all human tissues [22]-[27]. As such, an ultra-short echo time (UTE) or zero echo time (ZTE) MR pulse sequence is sometimes used as it is able to capture bone signal before the fast T2* relaxation [28]-[34].

As for algorithms in use for sCT generation, methods include segmentation-based, atlas-based, and voxel-based techniques. In segmentation-based algorithms, images are first roughly segmented into several groups and each group is assigned a tissue bulk density. Although, segmentation-based methods are simple, they require manual segmentation of the organs which is time consuming and user dependent and thus challenging for clinical use [14],[15],[35],[73],[74]. In contrast, some atlas-based techniques have the advantage of being fully automatic. They entail the matching between the current patient’s MR and a previously-collected MR-CT pair, and deformable interpatient image registration. This, however, is prone to errors especially in regions where patients have large tumor volume or surgical void [14],[15],[36],[37]. Sometimes making an atlas customized for a particular patient group, e.g. pediatric, is viable approach [38]. Nevertheless, patients being heterogeneous or things such as being thin compared to the mean atlas or patients having muscular atrophy make the approach implausible for routine clinical use [39]. Voxel-based analyses overcome these limitations. They work based on the composition of individual voxels. If a voxel contains tumor or surgical resection cavity, the MR signals are altered and the resultant sCT voxel value reflects this altered composition [2],[9],[15],[40]-[45].

In this paper, to address the unmet need for sCT in the abdomen-pelvis area, we present a personalized, voxelwise method. We use a novel acquisition method: multi-echo UTE combined with modified Dixon (mDixon) reconstruction in one pulse sequence using a 3D stack-of-stars sampling [33], [43]. Furthermore, the commercial Philips PET/MR acquisition software was modified to support dual torso coil receive and tests were performed to validate patient safety. This efficiently captures soft tissue and bone signal of the abdomen-pelvis anatomy without repositioning the coils on the patient. We create a new systematic method for synthetic CT generation jointly leveraging prior knowledge and partial supervision (SCT-PK-PS for short) for the body section of abdomen and pelvis. SCT refers to Synthetic CT. This is a CT synthesized (predicted) from MR data. PK refers to Prior Knowledge. It comes from the historical cluster prototypes for the different tissue types of the abdomen and pelvis: air, bone, bone marrow, fat, and soft tissue. The prior knowledge is utilized in the knowledge-leveraged transfer fuzzy c-means (KL-TFCM) [40] to reliably initialize the MR voxels into the groups of different tissue types. PS refers to Partial Supervision, i.e., a small quantity of labeled examples used in the Laplacian support vector machine (LapSVM) [42]. LapSVM exploits not only a few labeled examples but also the potential manifold information embedded in numerous unlabeled data. As such, we can figure out insightful tissue type classifiers with only labeling a few examples and thereby saving time and labor-costs. It is highly advantageous for clinical trials that in SCT-PK-PS only a fraction of the training data has to have the tissue types labeled as human labeling of all voxels is really labor-intensive and seems impractical. The input data of SCT-PK-PS are images obtained using multi-echo UTE-mDixon acquisition and reconstruction. Also, an effective feature extraction strategy for MR images, i.e., the edge detector convolution, is enlisted. Key contributions of our method are:

1. Guided by our previous work on sCT generation for abdomen and pelvis [80], we incorporate in this study the multi-echo UTE-mDixon images as input to SCT-PK-PS. By applying a short echo time and a fast radiofrequency pulse excitation [79], UTE-mDixon provides supplementary input features that enables more precise clustering and tissue labeling in the further steps of SCT-PK-PS.

2. Using prior knowledge, KL-TFCM is able to preferably group the MR data into five initial clusters of fat, soft tissue, air, (dense) bone, and bone marrow. These initial partitions are evaluated based on a few labeled examples (partial supervision) for the subsequent semi-supervised classification by means of the Laplacian support vector machine (LapSVM).

3. To further precisely identify tissue classes, e.g., bone, bone marrow, air, and soft tissue, LapSVM is enlisted that is capable of training tissue-recognizers by exploiting not only the limited supervision but also the potential manifold information embedded in numerous unlabeled data. This means that we needn’t to provide a sufficient number of manually labeled examples for specific tissue types, which greatly saves us time and labor costs.

4. With jointly using KL-TFCM, LapSVM, and the edge feature extraction, the proposed SCT-PK-PS method features good recognition accuracy of tissue types, which ultimately facilitates excellent performance of synthetic CT generation. Specifically, we identify the requisite tissue types needed to achieve PET AC having errors less than 5% [27], [47], [48] and thus are well within the National Cancer Institute/American College of Radiology Imaging Network (NCI/ACRIN) 10% specification for accuracy [49]. Moreover, as radiation therapy uses photon energies that are greater than those of 0.511 MeV energy of PET, the accuracy for RTP is expected to even surpass that of PET AC.

5. Experimental results indicate that SCT-PK-PS generally obtains the best quality of synthetic CTs on the body section of abdomen-pelvis against existing methods.

The remainder of this manuscript is organized as follows. Related work, e.g., KL-TFCM and LapSVM, is introduced briefly in section 2. The proposed SCT-PK-PS method is stated phase by phase in section 3. The experimental studies are presented in section 4. Conclusions are given in Section 5.

2. Related work

2.1. Knowledge-Leveraged Transfer Fuzzy C-Means (KL-TFCM)

Bezdek J C et.al [47] proposed the fuzzy c-means (FCM) algorithm to divide data instances into C disjoint clusters so that the deviation within clusters is minimum and the separation among clusters is maximum. Recently, FCM has been applied in many fields, such as bioinformatics [50],[51] and image analysis [40],[52],[53]. However, when applied to data that are quite sensitive to interference, e.g., noise and outliers [40], the classic FCM algorithm usually possesses unstable clustering performance and becomes ineffective. Inspired by transfer learning [40],[41],[54],[55],[75]-[77], we proposed the KL-TFCM algorithm [56] to improve the fuzzy clustering performance on the target data set (called the target domain) by referring to the prior knowledge obtained from the historical data set (called the source domain).

Let $X_{\text{T}}=\left\{x_{j,\text{T}}\mid x_{j,\text{T}}\in R^d,j=1,\dots ,N_{\text{T}}\right\}$ denote the target data set (i.e., the target domain) in which $x_{j,\text{T}},j=1,\dots ,N_{\text{T}}$, present the data instances, N_T is the total data capacity in the target domain, and d signifies the data dimension. Suppose that there exist C_T (1 <C_T <N_T) potential clusters in the target data set. The framework of KL-TFCM can be reformulated as:

$$\begin{gathered} \underset{U,V}{\min }\left(J_{\text{KL}-\mathrm{TFCM}}\left(U_{\text{T}},V_{\text{T}}\right)=\sum _{i=1}^{C_{\text{T}}}\sum _{j=1}^{N_{\text{T}}}{u}_{ij,\text{T}}^m{\Vert x_{j,\text{T}}-v_{i,\text{T}}\Vert }^2+\lambda \sum _{i=1}^{C_{\text{T}}}{\Vert {\widehat{v}}_{i,\text{S}}-v_{i,\text{T}}\Vert }^2\right) \\ \text{s}.\text{t}.0\le u_{ij}\le 1and\sum _{i=1}^{C_{\text{T}}}{u}_{ij}=1 \end{gathered}$$ (1)

where $V_{\text{T}}=\left\{v_{i,\text{T}}\mid v_{i,\text{T}}\in R^d,i=1,\dots ,C_{\text{T}}\right\}$ denotes the cluster centroid matrix in the target domain composed of all cluster centroids (namely, cluster centers or cluster prototypes); $U_{\text{T}}\in R^{C\times d}$ signifies the membership matrix and each entry u_ij,T denotes the fuzzy membership of data instance x_j,T to cluster centroid $v_{i,\text{T}};{\widehat{v}}_{i,\text{S}}$ denotes the ith referable cluster centroid in the source domain; m> 0 is the fuzzy index; and λ ≥ 0 is the regularization coefficient for transfer learning.

In (1), the first term, ${\sum }_{i=1}^{C_{\text{T}}}{\sum }_{j=1}^{N_{\text{T}}}{u}_{ij}^m{\Vert x_{j,\text{T}}-v_{i,\text{T}}\Vert }^2$, aims to partition the N_T data instances in the target domain into C_T disjoint clusters and the second term, $\lambda {\sum }_{i=1}^{C_{\text{T}}}{\Vert {\widehat{v}}_{i,\text{S}}-v_{i,\text{T}}\Vert }^2$, is the transfer learning term measuring the total approximation degree between the estimated cluster centroids in the target domain and the referable cluster centroids in the source domain. The parameter λ controls the overall learning extent between the source and target domains. Large values of λ mean that the target domain should learn much from the source domain, i.e., V_T should be close to ${\widehat{V}}_{\text{S}}$; conversely, small values of λ relax such constraint.

Using the Lagrange optimization, the updating equations of membership u_ij,T and cluster centroids v_j,T in (1) can be separately derived as

$${\mu }_{ij,\text{T}}=\frac{1}{\sum _{l=1}^{C_{\text{T}}}{\left(\frac{{\Vert x_{i,\text{T}}-v_{j,\text{T}}\Vert }^2}{{\Vert x_{i,\text{T}}-v_{l,\text{T}}\Vert }^2}\right)}^{\frac{1}{m-1}}}$$ (2)

$${\mathbf{v}}_{j,\text{T}}=\frac{\sum _{i=1}^{N_{\text{T}}}{\mu }_{ij,\text{T}}^m{x}_{i,\text{T}}+\lambda {\widehat{v}}_{j,\text{S}}}{\sum _{i=1}^{N_{\text{T}}}{\mu }_{ij,\text{T}}^m+\lambda }$$ (3)

2.2. Laplacian Support Vector Machine (LapSVM)

Support vector machine (SVM) is a well-known classification technique in machine learning. Instead of seeking to minimize empirical risk, SVM achieves overall risk minimization by minimizing the upper bound of the generalization error. The learning performance of traditional SVM is greatly dependent on the quality and quantity of training examples. To improve the classification accuracy when facing the condition in which the labeled examples are insufficient but when numerous unlabeled data instances are available, Mikhail Belkin et.al proposed the manifold regularization mechanism with organically incorporating the theories of manifold learning and spectral graph into the framework of conventional SVM [42].

Let $S=\left\{x_i\in R^d,i=1,\dots ,l+u\right\}$ denote the training set consisting of l labeled examples and u unlabeled data instances and d be the data dimension. Let y_i ∈ {1, −1}(i = 1, …, l) signify the labels of the corresponding labeled examples in S. Suppose that HK denotes the reproducing kernel Hilbert space (RKHS) associated with one Mercer kernel K and f(·) represents the decision function. Then the framework of manifold regularization can be reformulated as:

$$\underset{f\in H_K}{\min }\left(\frac{1}{l}\sum _{i=1}^{l}V\left(x_i,y_i,f\left(x_i\right)\right)+{\gamma }_A\Vert f{\Vert }_K^2+{\gamma }_I\Vert f{\Vert }_I^2\right)$$ (4)

where V(·) is the loss function and γ_A and γ_I, are two regularization parameters associated with two regularization terms.

There are three terms in (4). The first term utilizes the loss function to control the empirical risk, and the second term avoids the overfitting issue by imposing the smoothness condition on possible solutions in RKHS. The third term, based on the manifold learning, attempts to exploit the intrinsic geometric distribution regarding all data instances. To depict the intrinsic manifold of the data distribution, the data adjacency graph, G = (W, f), is used, i.e.,

$$\Vert f{\Vert }_I^2=\frac{1}{{(u+l)}^2}\sum _{i,j=1}^{l+u}{\left(f\left(x_i\right)-f\left(x_j\right)\right)}^2{W}_{ij}=\frac{1}{{(u+l)}^2}{f}^{\text{T}}Lf,$$ (5)

in which $f={\left[f\left(x_1\right),\dots ,f\left(x_{l+u}\right)\right]}^{\text{T}}$, $W_{ij}\in W,i,j=1,\dots ,u+l$, are the edge weights in the data adjacency graph, L = D ‒ W is the graph Laplacian, and D is the diagonal degree matrix of which the entries $D_{ii}\text{=}{\sum }_{j=1}^{l+u}{W}_{ij}$ and the others are 0.

Let V(·) be the hinge loss function, based on (4), Mikhail Belkin et al. proposed the framework of LapSVM as

$$\underset{{}_{f\in H_k}}{\min }\left(\frac{1}{l}\sum _{i=1}^l{\left(1-y_{i}f\left(x_i\right)\right)}_{+}+{\gamma }_A\Vert f{\Vert }_K^2+\frac{{\gamma }_I}{{(u+l)}^2}{f}^{\text{T}}Lf\right)$$ (6)

According to the Representer Theorem [42], [57], the solution of (6) can be expressed as

$$f^{*}(x)=\sum _{i=1}^{l+u}{\alpha }_{i}K\left(x,x_i\right)$$ (7)

where α_i ∈ R. Then (6) can be rewritten as [42]

$$\begin{gathered} \underset{a\in R^{l+u},{\xi }_i\in R}{\min }\left(\frac{1}{l}\sum _{i=1}^l{\xi }_i+{\gamma }_A{\alpha }^{\text{T}}K\alpha +\frac{{\gamma }_I}{{(u+l)}^2}{\alpha }^{\text{T}}KLK\alpha \right), \\ \mathrm{s}.\mathrm{t}.y_i\left(\sum _{j=1}^{l+u}{\alpha }_{j}K\left(x_i,x_j\right)+b\right)\ge 1-{\xi }_i,i=1,\dots ,l, \\ {\xi }_i\ge 0,i=1,\dots ,l. \end{gathered}$$ (8)

where ξ_i, i = 1, …, l, are the introduced slack variables.

Based on the Karush-Kuhn-Tucker (KKT) condition, the dual form of (8) can be derived as

$$\begin{gathered} \underset{\beta \in {\text{R}}^l}{\max }\left(\frac{1}{l}\sum _{i=1}^l{\beta }_i-\frac{1}{2}{\beta }^{\text{T}}Q\beta \right), \\ \mathrm{\text{s}}.\mathrm{\text{t}}\text{.}\sum _{i=1}^l{\beta }_i{y}_i=0, \\ 0\le {\beta }_i\le \frac{1}{l},i=1,\dots ,l, \end{gathered}$$ (9)

where $Q=YJK{\left(2{\gamma }_{A}I+\left(2{\gamma }_I/{(u+l)}^2\right)LK\right)}^{-1}{J}^{\text{T}}Y$, J = [I 0] is a l × (l + u) matrix with I being the l × l identity matrix, $\mathbf{Y}=\mathrm{diag}\left(y_1,y_2,\dots ,y_l\right)$, and K is the (l + u) × (l + u) kernel matrix.

In terms of the solution of (9), the solution of (8) can be obtained using ${\alpha }^{*}={\left(2{\gamma }_{A}I+\left(2{\gamma }_I/{(u+l)}^2\right)LK\right)}^{-1}{J}^{\text{T}}Y{\beta }^{*}$ [41].

3. The Proposed SCT-PK-PS Method

The proposed SCT-PK-PS method consists of six phases, as shown in Fig. 1. Phase I extracts fifteen features from given abdomen-pelvis MR images acquired using the UTE-mDixon pulse sequence. Phase II obtains the prior knowledge, i.e. the historic cluster centroids regarding the five tissue types of air, bone, bone marrow, fat, and soft tissue from all available MR feature data. Phase III initializes the MR feature data to preliminarily identify fat, using the transfer learning based KL-TFCM algorithm. Phase IV prepares the partial supervision for LapSVM on all given subjects and then obtains multiple tissue-recognizers. Phase V optimizes the selection of the multiple candidate tissue-recognizers for new subjects’ prediction using the genetic algorithm. Lastly, Phase VI predicts the tissue types of voxels in the new subject’ MR images using the selected tissue-recognizers and generates the synthetic CT.

Fig 1.

The overall workflow of SCT-PK-PS

3.1. Phase I: Extract Features from Given Abdomen- Pelvis 3D MR Images

Effective methods of feature extraction are indispensable for medical image processing. At each voxel of the abdomen-pelvis images we consider 9 features. Three correspond to the basic UTE-mDixon signals: free induction decay (FID), first echo (echo 1), and second echo (echo 2). Three features are from reconstructed images: Dixon-water, Dixon-fat, and R2* [31] [33][44][58]. Three features are the 3D spatial coordinates. Acquisition details are given in section 4.1.

In addition to the contrast differences concerning different tissue types in MR images, texture features, particularly at interfaces between different tissue types, belongs to significant information conducive to reliable recognition. Inspired by the edge detector convolution¹ [59], the specific texture filter matrix M_3×3 defined in Fig. 2(b) is used. We extract the texture features in image planes perpendicular to the long-axis of human body (i.e., the Z axis). Specifically, for pixel (i, j) in plane z in one of the acquired MR images, of which the intensity value is denoted as v(i, j), its texture feature is calculated using

$$v^{\prime }(i,j)=\sum _{p\in \{-1,0,1\}}\sum _{q\in [-1,0,1\}}[v(i+p,j+q)\times M(2+p,2+q)]$$ (10)

Fig 2.

Illustration of texture feature extraction for pixel (i, j) in one slice of MR image

Eq. (10) is able to delineate whether the texture structures of voxels are adjacent to tissue boundaries or not. The value of v′(i, j), Eq. (10), would be equal or close to zero if pixel (i, j) is located fully within one homogeneous tissue wherein all of the eight pixels surrounding pixel (i, j), as illustrated in Fig. 2(a), have the same or similar intensity value as pixel (i, j). Conversely, the value of v′(i, j) would be much different from zero when pixel (i, j) is at or near the boundary of different tissue types wherein the intensity values of some of the eight neighbors could differ from that of pixel (i, j). This particularly benefits recognizing tissue boundaries and thereby distinguishing heterogeneous tissues.

As such, we obtain the fifteen-dimensional feature expression for each voxel in the MR images, i.e., $\left[\begin{array}{l}{v}_{\text{FID}},v_{\text{echo}1},v_{\text{echo}2},v_{\text{Dixon-water}},v_{\text{Dixon-Fat}},v_{\text{R}{2}^{*}},v_{\text{FID}}^{\prime },\hfill \\ v_{\text{echo}1}^{\prime },v_{\text{echo}2}^{\prime },v_{\text{Dixon-water}}^{\prime },v_{\text{Dixon-fat}}^{\prime },v_{\text{R}{2}^{*}}^{\prime },x,y,z\hfill \end{array}\right]$, in which [x, y, z] are the spatial coordinate values.

3.2. Phase II: Generate Prior Knowledge for Transfer Fuzzy Clustering

The historical cluster centroids of the five tissue types — air, bone, bone marrow, fat, and soft tissue — are regarded as the prior knowledge and play the vital role in our proposed SCT-PK-PS method. The transfer learning based KL-TFCM needs such prior knowledge to assist the clustering on target MR feature data.

To obtain such prior knowledge, in terms of a few existing patients’ abdomen-pelvis MR images, e.g., denoted as Subject 1 (Sub 1) to Subject n (Sub n), which were previously acquired using both the UTE and mDixon pulse sequences, we first generate the fifteen-dimensional feature data set for a subject, following Phase I. Afterwards, the classic FCM algorithm is implemented on each of the feature data sets to partition the data instances into five clusters, and the estimated cluster centroids are associated with the tissue types of air, bone, bone marrow, fat, and soft tissue, respectively. Here given empirical knowledge or the Hounsfield Unit (HU) values with regard to specific tissue types defined in existing references (e.g., [44]) are dependent to determine the appropriate match between the tissue types and the estimated cluster centroids. To strengthen the generalizability of our method, all of the n subjects’ averages of the cluster centroids are eventually treated as the historical cluster centroids, namely, the prior knowledge, for the transfer clustering using KL-TFCM.

3.3. Phase III: Initialize MR Feature Data Sets Using KL-TFCM

From Phases III to IV, with multiple given subjects (e.g., m subjects), we attempt to learn several candidate tissue-recognizers for the tissue identification in new subjects’ abdomen-pelvis MR images, and thus the synthetic CT generation.

Due to the individual diversity and because there are subject-induced B0 distortions and gradient non-linearities, there is some inconsistency in MR data across subjects. In this context, traditional clustering techniques (e.g., the classic FCM), which partition the target data according merely to the similarities among data instances, would be prone to getting fooled by noise and outliers existing in imperfect MR images. In contrast, the transfer learning based KL-TFCM algorithm proves better clustering stability by leveraging the given prior knowledge [35], thus it is enlisted throughout our study to initial each subject’s generated, fifteen-dimensional MR feature data of MR images into the five tissue types.

3.4. Phase IV: Prepare Partial Supervision for LapSVM and Obtain Multiple Tissue-Recognizers

After using KL-TFCM in Phase III, however, it was observed that only the segmentation regarding the fat tissue is reliably satisfactory. This indicates that other effective methodologies for distinguishing the remaining tissue types are needed. For this purpose, LapSVM, the well-established semi-supervised classification technique, is adopted. Leveraging the manifold structure embedded in given data instances both labeled and unlabeled, LapSVM achieves attractive classification performance under the premise of using a small quantity of labeled examples, i.e., the so-called partial supervision while training classifiers. As the identification ability of classification is usually better than that of clustering, the leftover tissue types (other than fat) in abdomen and pelvis are further divided into five subclasses: air, bone, bone marrow, soft tissue, and tissue interfaces. It was observed that particularly distinguishing the areas at the junction of different tissue types improves the overall classification accuracy. Therefore, via the openly accessible COMKAT software [61],[62] that we developed previously, a few detailed image voxels matching each of the five subclasses are manually labeled on each subject’s MR images. Thus, some human experience is leveraged. Despite the fact that manually labeling is time-consuming and labor-intensive, it is acceptable in our method as only a quite small number of labeled examples in regard to each subclass are required. As such, by means of a limited number of labeled examples (denoted as n_l) and numerous unlabeled data instances (denoted as n_u) regarding every subclass, we are able to obtain a five-subclass-classifier on each subject’s MR feature data using the LapSVM.

So far, combining the fat segmentation generated by KL-TFCM with the five-subclass-classification achieved using LapSVM, we obtain the so-called tissue-recognizer for each subject enlisted in the training process. Via the m given subjects, we can generate m candidate tissue-recognizers.

3.5. Phase V: Optimize the Selection of Multiple Candidate Tissue-Recognizers for New Subjects’ Prediction

Starting from this phase, our task is to identify the tissue types in target subjects’ abdomen-pelvis MR images by optimally using the m candidate tissue-recognizers obtained at the end of Phase IV. To this end, the genetic algorithm (GA) [63], one well-known evolutionary computation [64] technique, is used in our method. Specifically, we first initialize a population consisting of l randomly generated individuals and each of which having the gene: $\left[g_1,\dots ,g_i,\dots ,g_m\right]$, where g_i ∈ {0, 1} (1 ≤ i ≤ m) and digit values 0 or 1 mean that tissue-recognizer i is not or is, respectively, used for prediction. Then, according to a few parameters involved in GA, e.g., the mutation probability p_m, crossover probability p_c, and maximum iteration number iter_num, and taking the Dice similarity coefficient of the predicted results with regard to all given labeled examples as the fitness function [65], our method iterates to improve the fitness. In each iteration, the mutation and crossover operators vary the current population with the pre-defined probabilities so that our method gradually approaches the potential optimal solution. To be specific, for the crossover, we pick out any two individuals in the current iteration in terms of the crossover probability p_c and exchange the half chromosome of each of them. For the mutation, under the probability p_m we choose the chromosome positions on each individual and then invert the values, i.e., change the value from 0 to 1 or from 1 to 0. In addition, before starting a new iteration, one individual with the worst fitness value, obtained in the current iteration, is replaced with the one having the best fitness value such that good genes can be inherited throughout the overall population evolution. Our method eventually stops at the maximum iteration number or whenever obtaining successive n_{max_acc} times of the maximum prediction accuracy (n_{max_acc} < iter_num).

After the GA optimization, in terms of the individual (chromosome) in the final population that has the best fitness value, we get the accessibly optimal selection concerning the m candidate tissue-recognizers. All of the tissue-recognizers corresponding to the genes in the chromosome of which the values equal 1 are recruited to predict the tissue types in new subjects’ MR images. Subsequently, the voting strategy is used and the tissue type of each MR image voxel is determined by means of the majority principle.

3.6. Phase VI: Predict Tissue Types of Voxels in New Subjects’ MR Images and Generate Synthetic CTs

As the orders of magnitude of subjects’ MR feature data are usually as high as millions in our study, it is infeasible to directly apply the recruited tissue-recognizers, which are determined using GA in Phase V, on one new subject’s whole MR feature data. Hence the “down-sampling + K nearest neighbors (KNN)” mechanism is enlisted to speed up the processing. Specifically, with the arrival of data from a new subject, the fat tissue in its MR images are obtained using the KL-TFCM. After that, we randomly sample the remaining, non-fat voxels according to the given sampling-capacity S_c to obtain a subset denoted as SB_vi. For each voxel in SB_vi, we obtain its tissue type through the outputs of the recruited tissue-recognizers and based on the majority principle. Then, referring to the obtained tissue types of voxels in SB_VI, all of the other unknown voxels are determined using the KNN algorithm.

Based on the obtained tissue types of all voxels in the target subject’s abdomen-pelvis MR images, with assigning a matching CT value to each tissue type, we can eventually synthesize a corresponding CT image. Referring to [66], the CT values with respect to air, bone, bone marrow, fat, soft tissue, and tissue interface are set to −466, 1524, 538, −98, 120, and 120 HU, respectively, in our proposed SCT-PK-PS method.

4. Experimental Studies

4.1. Setup

We aim to evaluate the objective performance of the proposed SCT-PK-PS method in this section. To this end, within a protocol approved by the University Hospitals Cleveland Medical Center Institutional Review Board, data were acquired from twenty subjects using the UTE-mDixon pulse sequence. Using a Philips 3T Ingenuity PET-MR system, the UTE-mDixon acquisition collected data at three TEs: TE1=0.14, TE2=1.14, and TE3=2.14ms [33] [43]. Three-point analyses were used for both Dixon water/fat reconstruction and R2* estimation. To achieve anatomic coverage over the abdomen and extending into the pelvis without having to reposition coils, the commercial Philips PET/MR acquisition software was modified to support dual torso coil receive and tests were performed to validate patient safety. Each of the abdomen-pelvis MR image is composed of 104 image planes and each plane has the 320 × 320 pixel resolution with spacing being 1.4312 × 1.4312 mm and 2.9 mm between planes. The raw MR data have been reconstructed to generate the free induction decay (FID), first echo (echo 1), second echo (echo 2), Dixon-water, Dixon-fat, and R2* images [43], [58].

Three existing methods, i.e., the four-cluster-partitioning (FCP) [27], all-water (AW) [67],[68], and SVM [42],[69],[70], were enlisted to compare with our SCT-PK-PS method. FCP straightforwardly segments the MR image into four different clusters: external air, internal air, fat, and soft tissue, using the classic FCM algorithm; whereas AW indiscriminately regards all voxels within human body as the water. To demonstrate the advantages regarding our semi-supervised-leveraged SCT-PK-PS method, the traditional, purely-supervised SVM method was also used as the competitor. The root mean square error (RMSE), mean absolute prediction deviation (MAPD), and R value [67],[80] were adopted as the validity metrics for performance comparisons.

It should be particularly clarified that no measured CTs are available in our study, thus what can be used as the benchmark to measure the performance of synthetic CT generation is a challenge in our experiments. In this regard, we suppose that for one specific subject, if the training data are available, the tissue-recognizer learned using its own training data both labeled and unlabeled is more convincing than the ones from other subjects. Under such premise, on one specific subject, the prediction outcomes of the tissue-recognizer learned using the training data from the subject itself can be regarded as the benchmark, which is the obtainable optimum in our current study, and be used to measure the performance of other methods.

Several system parameters, listed in Table I, are involved in the proposed SCT-PK-PS method. System parameters facilitate the flexibility of our method. However, too many indeterminate parameters conversely weaken the practicability. Therefore, based on the extensive empirical studies, most parameters are set to the suggested values. The grid search [71] was employed to determine these suggested values. The trial ranges of the primary parameters in SCT-PK-PS as well as their recommended settings are also listed in Table I. By the way, the regularization coefficients γ_A and γ_I, in LapSVM and the number of neighbors k_lap to build the graph Laplacian were determined using the hold-out strategy [72] in the training process, as commonly used in semi-supervised classification [42].

TABLE I.

Primary parameters used in SCT-PK-PS

Parameters	Recommended settings	Trial Ranges
Fuzzy index m and regularization parameter λ in KL-TFCM in Eq.(1)	m=1.1, λ= 1e2	m ∈ {1.1, 1.2, 1.5, 1.7, 2.0, 2.5} λ ∈ {1e−3, 1e−2, 1e−1, 1e1, 1e2, 1e3, 1e4,}
SSL and SSUL to train LapSVM classifier	(SSL, SSUL) = (500,1000)	(SSL, SSUL) ∈ {(150,350), (250,750), (500,1500), (750,1750), (1000,2000)}
The regularization coefficients γI, γA and the number of neighbors in building the graph Laplacian kLap in LapSVM	Determined by hold-out strategy	γI ∈ {2−10, 2−8, 2−6, 2−4, 2−3, 23, 27,} γA ∈ {2−10, 2−8, 2−6, 2−4, 2−3, 23, 27,} klap ∈ {3,5,7,9}
Mutation probability pm, Crossover probability pc, Maximum iteration number iter_num, Number of iteration that achieves maximum prediction accuracy, nmax_acc in Phase V	pm = 0.6 pc = 0.6 iter_num = 200 nmax_acc = 10	pm ∈ {0.1, 0.2, 0.4, 0.6, 0.8} pc ∈ {0.1, 0.2, 0.4, 0.6, 0.8} iter_num ∈ {50,100, 200,300, 400,500}
Parameter sc to obtain SBVI and parameter K for KNN in testing Phase	sc = 5e4 K=1	sc ∈ {1e4,2e4, 3e4, 4e4, 5e4, 6e4,} K ∈ {1, 3, 5, 7, 9}

Our experiments were carried out using a computer with an Intel Core i7–6850K 3. 6 GHz CPU, 128 GB of RAM, Ubuntu 16.04 operating system (64 bit), and MATLAB 2016b.

4.2. Experimental Details and Results

As already mentioned in Section 4.1, for each of the twenty subjects we first learned the tissue-recognizer using its own training data, i.e., a few labeled examples together with numerous unlabeled data instances, following Phases I to IV. Using this tissue-recognizer, we then predicted the tissue types of the subjects itself to generate a synthetic CT as the benchmark for performance measurement. This type of CT is called as the reference CT in our experiments.

Afterwards, utilizing the leave-one-out strategy, we selected the data set from one subject for validation and used the remaining 19 to train the multiple candidate tissue-classifiers (see Phase IV). The process was repeated 20 times so that each of the data sets was used for validation. Via the GA optimization, we got the optimal selection regarding the nineteen candidate tissue-classifiers, and only the ones corresponding to the genes equaling 1 at the best individual in the last population were enlisted to predict the tissue types in the validation subject’s MR images, by means of the majority principle. Finally, we generated the synthetic CT for the new subject, following Phase VI. In this way, our SCT-PK-PS method is able to create a synthetic CTs for all twenty subjects.

To evaluate the traditional, purely supervised SVM method, the leave-one-out strategy was used again. All of the labeled examples from the nineteen subjects were put together to train the classifier for tissue identification.

Using each subject’s MR feature data and applying the FCP, AW, SVM, and SCT-PK-PS methods, respectively, we obtained their individual synthetic CTs. Fig. 3 compares the objective measures of performance of the four methods across all subjects and the numbers are tabulated (see Table II). Fig. 4 shows the reference CTs and the synthetic CTs achieved by the four methods for two representative subjects. In comparing methods, our proposed method achieves the best MAPD, RMSE, and R values across all subjects.

Fig 3.

Performance curve illustrate all subjects metrics concerning MAPD, RMSE, and R

TABLE II.

Performance comparison of the proposed SCT-PK-PS to three other methods

Sub		RMSE				MAPD				R
		AW	FCP	SVM	SCT-PK-PS	AW	FCP	SVM	SCT-PK-PS	AW	FCP	SVM	SCT-PK-PS
1	mean	389.20	526.40	562.35	325.77	273.50	301.36	373.69	172.73	0.0555	0.0695	0.4025	0.6580
	std	0.00	0.35	0.35	2.91	0.00	0.42	0.42	1.92	0.0000	0.0279	0.0013	0.0121
2	mean	401.71	518.02	501.23	263.31	263.43	268.84	325.67	136.97	0.0034	0.2769	0.4722	0.7831
	std	0.00	0.63	0.63	2.30	0.00	0.45	0.45	1.46	0.0000	0.0000	0.0017	0.0098
3	mean	417.86	555.73	315.22	299.61	272.22	294.28	169.36	138.91	0.0988	0.2738	0.7070	0.7449
	std	0.00	0.81	0.81	2.59	0.00	0.41	0.41	1.48	0.0000	0.0000	0.0019	0.0076
4	mean	368.79	533.51	506.44	268.40	245.90	275.98	343.73	141.46	0.1141	0.3509	0.3741	0.8362
	std	0.00	0.33	0.33	2.13	0.00	0.32	0.32	1.72	0.0000	0.0294	0.0011	0.0055
5	mean	351.49	350.55	294.75	332.78	231.91	202.58	134.14	197.77	0.1240	0.1597	0.5724	0.3902
	std	0.00	0.82	0.82	2.02	0.00	0.33	0.33	1.20	0.0000	0.0047	0.0032	0.0101
6	mean	410.31	615.03	443.90	254.58	260.78	307.45	202.79	123.85	0.0897	0.3318	0.3506	0.8448
	std	0.00	0.81	0.81	2.27	0.00	0.47	0.47	1.35	0.0000	0.0204	0.0026	0.0112
7	mean	391.67	526.26	353.95	271.51	255.19	250.04	175.84	128.73	0.1685	0.2502	0.5512	0.8031
	std	0.00	0.62	0.62	2.52	0.00	0.33	0.33	1.67	0.0000	0.0117	0.0038	0.0105
8	mean	318.20	403.80	254.53	319.33	214.26	278.28	117.16	177.61	0.1615	0.2496	0.6392	0.2803
	std	0.00	0.74	0.74	1.77	0.00	0.40	0.40	1.13	0.0000	0.0000	0.0025	0.0111
9	mean	393.38	512.96	374.98	284.77	261.00	255.96	241.82	132.81	0.1664	0.0687	0.5477	0.7609
	std	0.00	0.61	0.61	2.97	0.00	0.47	0.47	1.83	0.0000	0.0101	0.0020	0.0146
10	mean	337.55	454.80	352.64	265.29	215.12	210.50	146.53	124.48	0.1014	0.2521	0.4422	0.7338
	std	0.00	0.26	0.26	1.79	0.00	0.14	0.14	1.10	0.0000	0.0069	0.0018	0.0103
11	mean	345.07	461.18	275.05	243.00	229.34	211.17	113.13	105.51	0.2004	0.1151	0.6657	0.7737
	std	0.00	0.70	0.70	1.36	0.00	0.39	0.39	0.79	0.0000	0.0058	0.0021	0.0069
12	mean	395.86	565.68	358.95	331.31	229.00	287.18	217.81	165.18	0.1614	0.3760	0.4945	0.7293
	std	0.00	0.69	0.69	1.79	0.00	0.36	0.36	1.30	0.0000	0.0006	0.0020	0.0093
13	mean	365.27	507.01	386.29	305.13	229.65	255.26	282.89	138.88	0.1541	0.2513	0.4866	0.7026
	std	0.00	0.63	0.63	1.13	0.00	0.44	0.44	0.68	0.0000	0.0249	0.0021	0.0053
14	mean	326.16	435.09	303.96	275.92	218.68	219.64	160.34	123.13	0.2017	0.0002	0.5209	0.7730
	std	0.00	0.76	0.76	3.16	0.00	0.44	0.44	2.57	0.0000	0.0000	0.0025	0.0155
15	mean	332.82	440.20	386.50	342.17	225.94	212.11	181.53	142.77	0.1308	0.0658	0.2500	0.6307
	std	0.00	0.85	0.85	1.79	0.00	0.53	0.53	1.17	0.0000	0.0962	0.0031	0.0086
16	mean	439.39	665.49	385.09	360.94	284.46	355.38	190.28	190.07	0.0913	0.3866	0.6042	0.7731
	std	0.00	0.34	0.34	1.57	0.00	0.20	0.20	1.95	0.0000	0.0373	0.0013	0.0046
17	mean	385.39	594.30	354.24	278.63	247.61	277.67	147.60	145.95	0.1282	0.4546	0.4836	0.8323
	std	0.00	0.84	0.84	2.98	0.00	0.31	0.31	1.76	0.0000	0.1042	0.0069	0.0091
18	mean	369.73	621.62	341.00	352.52	242.28	324.02	151.92	184.52	0.2082	0.3871	0.5929	0.7464
	std	0.00	0.54	0.54	2.28	0.00	0.29	0.29	1.81	0.0000	0.0001	0.0026	0.0042
19	mean	295.98	391.83	233.17	191.08	199.84	191.29	97.62	78.36	0.1758	0.0769	0.6450	0.8115
	std	0.00	0.74	0.74	2.35	0.00	0.39	0.39	1.09	0.0000	0.0347	0.0034	0.0126
20	mean	359.48	514.70	364.11	204.43	227.11	266.74	167.15	96.67	0.1462	0.2497	0.4223	0.8481
	std	0.00	0.45	0.45	2.58	0.00	0.36	0.36	1.64	0.0000	0.0188	0.0022	0.0121
average	mean	369.77	519.41	367.42	288.52	241.36	262.77	197.05	140.72	0.1341	0.2323	0.5112	0.7228
	std	34.94	70.56	82.11	45.10	21.79	42.22	76.53	30.60	0.0490	0.1275	0.1138	0.1420

Fig 4.

Synthetic CTs generated by the employed methods on representative subjects (Subs 2 and 3)

4.3. Discussions

As detailed in Introduction, little progress had been made in voxel-wise synthetic CT generation especially for the abdomen. We use a novel UTE-mDixon acquisition with stack-of-stars sampling and two receive coils to achieve robust, efficient anatomic coverage. The data particularly facilitate the differentiation of bone and air. Furthermore, we introduce a new analysis method that outperforms previously described methods. Specifically, we enlisted two state-of-the-art machine learning techniques, i.e., the transfer learning based KL-TFCM as well as the semi-supervised classification based LapSVM, in our proposed SCT-PK-PS method. Compared with traditional FCM, owing to referencing given prior knowledge, KL-TFCM mitigates the sensitivity to noise and outliers, and reliably initialize voxels in the MR images into the five tissue groups: air, bone, bone marrow, soft tissue, and fat. Using partial supervision (i.e., a few labeled data instances) as well as lots of unlabeled data instances, LapSVM is capable of learning insightful classifiers, and thereby overcomes the challenge of having too few labeled examples available for classifier training. Taking advantage of the synergistic efforts of both KL-TFCM and LapSVM, our proposed SCT-PK-PS method achieves superior performance in synthetic CT generation on the body section of abdomen-pelvis (see Fig. 4).

As for specific performance comparisons among the four employed methods, AW, a commonly-used uniform approximation for diagnostic imaging and radiation therapy treatment applications, straightforwardly treats all voxels within body as water, so it is only recruited as a benchmark for performance measurements in our experiments. FCP intrinsically ignores the bone tissue and partitions the image voxels into four clusters: fat, soft tissue, internal-air, and external-air using FCM. Therefore, as shown in Table II, both AW and FCP have worse metric scores, i.e., lager MAPD and RMSE values as well as smaller R values, than SVM and our proposed SCT-PK-PS. As far as SVM and SCT-PK-PS, two classification methods, are concerned, owning to the high time complexity (O(N³)), the fully-supervised SVM is merely suitable for small-scale training sets, e.g., 3000 manually-labeled samples per subject in our experiments. In this context, when the training samples are not informative enough, the trained classifiers are prone to poor performance. As indicated in Fig. 4, compared with the outcomes of SCT-PK-PS, SVM misclassified many tissues particularly in upper abdomen.

In addition, it is worth noting that a remarkable feature extraction technique is embedded in our proposed SCT-PK-PS method, i.e., the edge feature detector used in Phase I. Depending on this, SCT-PK-PS method can work on the fifteen-dimensional, advanced feature data, whereas other existing methods, e.g., SVM and FCP, work merely upon the original MR intensity features. This strategy is shown to outperform existing methods.

5. Conclusion

To generate synthetic CTs from UTE-mDixon MR images on the challenging body section of abdomen-pelvis, we propose the SCT-PK-PS method as effective and practical. SCT-PK-PS organically integrates multiple techniques and strategies, including the edge detector based feature extraction, transfer learning based fuzzy clustering (i.e., KL-TFCM), manifold learning based semi-supervised classification (i.e., LapSVM), and GA-optimized vote decision via multiple tissue-recognizers. Accordingly, TFC-ALC exhibits superior performance.

Last but not least, it is worth clarifying that one limitation of our study is the lack of measured CT data as the reference standard. This is difficult to do as it entails extra radiation exposure to patients. Another consideration is collecting a larger data set to test if the superior performance of our method generalizes to a larger population.

Biography

Pengjiang Qian (M’2012, SM’2019) received his Ph.D. degree from Jiangnan University in March, 2011. Now he is a Full Professor at the School of Digital Media, Jiangnan University, Wuxi, Jiangsu, China. He has authored or co-authored more than 70 papers published in international/national journals and conferences, e.g., IEEE TMI, IEEE TNNLS, IEEE TSMC-B, IEEE T. Cyber., IEEE TFS, PR, InS, and KBS. His research interests include data mining, pattern recognition, bioinformatics and their applications, such as analysis and processing for medical imaging, intelligent traffic dispatching, and advanced business intelligence in logistics

Jiamin Zheng is a M.S. candidate at the School of Digital Media, Jiangnan University, Wuxi, Jiangsu, China. Her research interests include pattern recognition and data mining.

Qiankun Zheng is a M.S. candidate at the School of Digital Media, Jiangnan University, Wuxi, Jiangsu, China. His research interests include intelligent algorithms and their applications.

Yuan Liu received his M.S. degree from Wuxi University of Light Industry in 1998. Now he is a Full Professor at the School of Digital Media, Jiangnan University, Wuxi, Jiangsu, China. He is also a member of the 863 expert panel in the information security technology domain of the Ministry of Science and Technology, a senior member of the China Computer Federation (CCF), and a member of the CyberSecurity Association of China (CSAC). His main researches focus on the software development of network information systems, network security, and digital media applications. His current research interests include network traffic measurement, social network, and digital media. He has published more than 100 academic papers in the authoritative and core journals.

Tingyu Wang is a B.S. candidate at the Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, Champaign, USA. His research interests include machine learning and data mining.

Rose Al Helo graduated with a BS in physics from the Lebanese University of Sciences in 2008 and an MS medical physics at Cleveland State University in 2011. Currently, she is a Diagnostic Medical Physicist at University Hospitals of Cleveland and a PhD student at Case Western Reserve University. Her focus is to bridge between the clinical needs and her research interest in medical imaging, especially MRI. Other scientific interests are optimizing MRI detectors for endemic diseases such as malaria, enhancing MRI coils and understanding errors in Magnetic Resonance Fingerprinting.

Atallah Baydoun received his MD degree from the Lebanese University in June 2009, and his BS in pure mathematics in June 2013. He is currently a physician attending at the Louis Stokes Cleveland VA Medical Center, Cleveland, OH, USA and a Biomedical Engineering PhD student in the Quantitative Imaging Laboratory – Case Western Reserve University, Cleveland, OH, USA. His research projects include the application of artificial intelligence in biomedical imaging and radiation therapy.

Norbert Avril is a clinical scientist and academic nuclear medicine physician exploring new applications for positron emission tomography (PET) in oncology. He is Professor of Radiology and Director of Nuclear Medicine for University Hospitals Health System in Cleveland. A particular emphasis of the clinical and preclinical research is the use of imaging procedures for optimizing cancer treatments particularly the early identification of treatment response using positron emission tomography (PET). Currently active clinical research protocols are focusing on changes in treatment induced cell proliferation using FLT-PET/MR and treatment induced changes in the expression of prostate specific membrane antigen (PSMA) in different tumor types.

Rodney J Ellis received his M.D. from NorthEast Ohio Medical University, and trained at the Ohio State University during his residency. He served as Vice-Chair at Case Western Reserve and was promoted to full professor, prior to accepting his current role a Chairman at Penn State University. His specialties include GU malignancy, brachytherapy, and MR-guided radiotherapy. He is a member of the GU Steering Committee for NRG, and serves on the Renal Task Force for the NCI. He is the national PI for NRG GU-005.

Harry Friel earned a Master’s thesis degree in Biology from The Cleveland State University in Cleveland, OH. He then worked for Picker/Marconi MR for 14 years in areas of clinical testing, customer training and MR applications and technique development. He currently works for Philips MR as a clinical scientist since 2003, including MR consultation for the development of the Philips PET-MR scanner. Areas of expertise include clinical testing of prototype coils and pre-release MR software and broad based MR clinical applications knowledge.

Melanie S Traughber earned her D.Sc. in Biomedical Engineering from Washington University in St. Louis, MO. Subsequently, she completed a post-doctoral fellowship at the National Institutes of Health (NIH) in Bethesda, MD. She previously served as Director of MR Clinical Research and Development for Philips Healthcare. Her expertise includes quantitative MRI and pulse sequence development and she specializes in cardiac MR and MR in radiation oncology.

Ajit Devaraj earned a doctoral degree in MR physics and is presently a Senior Scientist with Philips Research North America, Cambridge, MA, USA. His research is focused on developing new MR acquisition and reconstruction technologies and translating them to clinical use. Application areas include prostate and breast MRI. In the past, he has been the feature architect leading multiply disciplinary teams to bring new MR clinical features to market. He has also mentored interns and post-doctoral researchers.

Bryan Traughber received his M.D. from University of California Los Angeles (UCLA) School of Medicine. He completed a post-doctoral fellowship at the National Institutes of Health (NIH) in Bethesda, MD and residency training in radiation oncology at the Case Comprehensive Cancer Center in Cleveland, OH. He is the Director of Advanced Technologies and Brachytherapy at University Hospitals Seidman Cancer and Case Western Reserve University School of Medicine. His expertise includes GYN and GU Oncology, Brachytherapy, and MR-guided Radiotherapy, and Molecular Radiotherapy.

Raymond F. Muzic, Jr. (M’1990, SM’2000) earned a PhD in biomedical engineering and is now Professor of Radiology, Biomedical Engineering, and General Medical Sciences - Oncology in the School of Medicine at Case Western Reserve University, Cleveland, OH, USA. He is co-director of the Quantitative Imaging Laboratory and conducts research developing new technologies and translating them to human use. The application areas include medical imaging, nuclear medicine, and radiation oncology. He also teaches and mentors students and fellows.