Int J CARS (2011) 6 (Suppl 1):S17–S22 DOI 10.1007/s11548-011-0569-1
COMPUTED TOMOGRAPHY
Automatic determination of whole body composition from limited whole body scan for improving quantification of metabolic activity on PET/CT T. Chan1 1 The University of Hong Kong, Department of Diagnostic Radiology, Hong Kong, PR China Keywords Body composition PET/CT Quantification Purpose Positron Emission Tomography/Computed Tomography (PET/CT) has become a standard imaging examination for management of oncology patients. The accurate determination of metabolic activity is of prime importance for monitoring of treatment response, which is done by calculation of standard uptake value (SUV). SUV is commonly normalized against lean body mass (LBM), which is in turn derived from predictive equations based on parameters like body height and weight. However, such predictive equations were known to be of limited accuracy, which can in turn limit accuracy of calculating LBM normalized SUV. Whilst CT has become a standard method for measuring body composition that can also be used readily to obtain LBM, the challenge is that routine PET/CT examination for oncology applications normally covers the range from skull base to upper thighs, also known as limited whole body scan (LWBS), rather than from vertex to toes. Therefore a mechanism to approximate LBM of the whole body from CT data of a limited coverage is required. This paper aims to present a computerized method that automatically calculates LBM from CT data of a limited whole body PET/CT examination, and to compare with the results obtained using conventional predictive equations. Methods Whole body PET/CT examinations of 18 adult patients covering skull vertex to toes were retrospectively retrieved. Their body height and weight were also recorded. Different body composition, mainly fat, soft tissues, and bones at each section were segmented based on their characteristic range of CT attenuation. From this, volumes of different tissues can be calculated for any section of the body. Mass of individual tissue type in any part of the body is the product of its volume and published density of each tissue type. The sum of all tissue types from the whole body image data produces total body weight (BW). LBM was simply BW less fat mass, which was compared against the estimated LBM from predictive equations. Since the exact anatomical coverage of a LWBS can be somewhat variable, it needs to be defined anatomically for individual patients. In this system, the coverage is defined by relative distances above the top section of thoracic cage and below the bottom section of pelvic girdle. Then body composition of different sections from the training dataset was recorded to give a reference of proportional contribution to the whole body LBM at the different relative anatomical levels. The sum of proportional contribution of LBM of all the sections in a given coverage gives the ratio of LBM from that coverage to LBM of the whole body.
A computerized scheme was developed that automatically determines the levels of top of thoracic cavity and bottom of pelvis based on characteristic trend of changing composition and specific bony configurations. Building on the automatically determined information of anatomical coverage for the test data and the above reference ratio of LBM at that particular coverage allows determination of LBM from LWBS of variable coverage. Results The patients range in age from 30 to 85, with 10 males and 8 females, who had body mass index (BMI) between 16.3 and 29.1. LBM derived from whole body CT data, the gold standard, differed significantly from those predicted by regression formulae, with average error ranging from 2.22 to 12.72 kg (5.19–29.74%), confirming that predictive equations are of limited accuracy in estimation of LBM. LBM derived from limited whole body examinations provided accurate estimate of the actual LBM of the whole body, being more accurate with longer than shorter coverage, with an average error ranging from 0.44 kg (1.03%) (skull vertex to 20 cm below pelvis) to 1.18 kg (2.76%) (skull base to 5 cm below pelvis), which were still much lower than errors from results of predictive equations. Conclusion LBM of the whole body can be automatically estimated from CT data of limited coverage typically acquired in routine PET/CT examinations. The results are more accurate than that derived from conventional predictive equations. This will allow more accurate quantification of metabolic activity of tumors, using LBM normalized SUV, important for management of oncology patients.
Visual tracking of treatment response in PET-CT image sequences J. Kim1, A. Kumar1, L. Wen1,2, S. Eberl1,2, M. Fulham1,2,3, D.D. Feng1,4 1 University of Sydney, Biomedical and Multimedia Information Technology (BMIT) Research Group, School of Information Technologies, Sydney, Australia 2 Royal Prince Alfred Hospital, Department of PET and Nuclear Medicine, Sydney, Australia 3 University of Sydney, Sydney Medical School, Sydney, Australia 4 Hong Kong Polytechnic University, Centre for Signal Processing, Department of EIE, Hong Kong, PR China Keywords Treatment response Visual tracking PET-CT Purpose Multi-modal PET-CT (positron emission tomography—computed tomography) is the preferred imaging technique for determining the extent (stage) of many common cancers, and is increasingly being used to best assess treatment response because it can provide functional (from PET) as well as anatomical (from CT) information. To assess treatment response scans are acquired at various time intervals during and after the completion of therapy. The conventional
123
S18 approach to evaluate response from these images is to view the PET-CT scans side-by-side and manually navigate through each scan individually to find corresponding points. However, the image volume is massive and it is not unusual to have to compare many individual PET-CT studies over the course of a patient’s study. Other approaches include the automated registration of the scans, which warps individual images to reach alignment. Such methods, however, transform the images and thus distort the anatomy depicted in the original scan and one cannot interactively navigate and visualise large scan sequences to track progressive changes in the scans over time. In this study, we propose a framework for visual tracking of treatment response in PET-CT sequences. Rather than warping the PET-CT images, we extract mapping parameters from the registration transformation and use them to calculate the reverse transform of the PET-CT images on a pixel-by-pixel basis in real-time. We demonstrate our framework with a prototype viewer for interactive visualisation of PET-CT sequences. Methods Whole-body FDG PET-CT scans were obtained on a Siemens Biograph TruePoint scanner with PET (168 9 168 pixels at 4.07 9 4.07 mm) and CT (512 9 512 pixels at 0.98 9 0.98 mm) images aligned (hardware co-registration). The slice thickness was 3.0 mm for both modalities. The PET and CT images were re-sampled to equal dimensions of 256 9 256 pixels at 1.98 9 1.98 mm. The test data consisted of 10 PET-CT studies of lymphoma patients (8 patients each with 4 separate scans performed at different times; 1 patient with 5 scans, and 1 with 6 scans). The proposed framework is illustrated in Fig. 1. For each sequence of N PET-CT scans I1,…, IN (Ii [ R3; Ii = {IiPET, IiCT}, the IiCT for i = 2,…,N, images were registered to I1CT, and the resulting parameters were then used in the construction of the transformation indexes T1/i, discarding the transformed CT images. The registration consisted of computing the mappings T1/i to transform Ii to I1 using the publicly available Elastix registration package. Affine followed by non-rigid B-spline registration algorithms were used to automatically transform the CT pairs. The CT was used in the registration due to its better anatomical definition and higher resolution compared to its PET counterpart. The indexes T1/i provided the reverse transform of a user-selected point p, which could be in either a PET or CT image,
Fig. 1 Proposed PET-CT visual tracking framework. Initially, IiCT for i = 2,…,N, were registered to I1CT resulting in the creation of transformation indexes. Using the indexes, a point p may be selected which triggers the corresponding point to be highlighted in all other images. Here, p is selected from I1PET resulting in the realtime calculation of the reverse transform on I2 (based on T1/2) to IN (T1/N)
123
Int J CARS (2011) 6 (Suppl 1):S17–S22
Fig. 2 A PET-CT sequence of a patient diagnosed with high-grade lymphoma. The top row is the PET (inverted grey scale LUT) and the bottom is the CT (grey scale LUT, soft tissue window) with each column representing a portion of a coronal PET-CT scan to calculate a set of corresponding points P, which contains one corresponding point for all other images in the sequence. Results A prototype viewer was developed based on our proposed framework by extending the ImageJ software produced by the NIH. User input (a point selection via a mouse click or drag) on any of the PET or the CT image results in the calculation of corresponding points (via a reverse transformation lookup) on all of the other images in the sequence. Standard visualisation tools such as lookup table (LUT) adjustment, brightness and contrast manipulation, and scaling were also included. The method was tested on clinical cases. Using our viewer, we traced the disease from its onset through its treatment response or remission in follow-up studies, visually tracking the response across each sequence. New sites of disease in follow-up scans were also tracked to check for abnormalities in prior studies. The ability to track the effects of treatment across all of the scans with our viewer was measured by comparing our visual tracking to the documented changes in the physician’s reports. In all of the clinical cases, our viewer was able to identify the changes to within a few pixels of the regions of interest. Real-time performance ([25 frames per second) was achieved in image navigation running on consumer hardware (Intel Core2 Q6600 @ 2.4 Ghz; 4 G RAM; 64 bit Windows7). The 4 scans of Fig. 2 were taken approximately 3 months apart. Selection of the abnormal mass (disease) in the abdomen of the 1st PET image (indicated by a cross cursor) automatically aligned all of the other images to relative anatomical localisation showed an excellent initial response to treatment and no evidence of disease recurrence or new sites of disease on subsequent scans. Conclusion We presented a framework for the visual tracking of treatment response in large volume PET-CT studies. In our tests, the viewer was able to track tumour changes across the sequences in real-time visualisation. We will further develop our viewer to include additional functionality and computational optimisations. Acknowledgement This work was supported in part by ARC grants.
Int J CARS (2011) 6 (Suppl 1):S17–S22 Detecting percolation transition in trabecular bone CT images E. Rokita1, G. Taton1 1 Jagiellonian University Medical College, Biophysics, Krako´w, Poland Keywords Trabecular bone CT image Percolation theory Fracture risk Purpose In clinical practice bone mineral density (BMD) itself is assumed to be a principal determinant of bone strength and fracture risk. A quantification of trabecular structure in addition to BMD might improve predictions of bone strength and fracture risk. In spite of decades of research there is still no consensus, how to adequately describe trabecular architecture, basing on data obtainable in vivo. There are two groups of methods currently used to characterize quantitatively bone architecture. The first group requires a previous binarization of the graylevel images for the calculations of the measures of bone structure. Since, the resolution achievable in vivo (*0.2 mm) is comparable with a typical trabecular thickness, the binarization remains a critical step of the trabecular bone characterization. The second group comprises calculations of the parameters that are based directly on gray-level images. In the present study a new method originating from the percolation theory is applied to perform quantitative description of the trabecular bone gray-level images. Methods The experiment comprised 12 autopsy cases (males, age 60, 81 years). From each individual the third lumbar vertebral body was used for investigations. Two-D images of the samples were obtained using a SOMATOM Sensation 16 CT system (Siemens, Germany). For each vertebral body two CT scans were performed and ten central sections were selected for further investigation. In each section a 160 9 160 pixels (pixel size = 0.146 9 0.146 mm2) ROI was selected manually in the center of the section and transferred to 8-bit bitmap image. For each sample mean of the results was calculated from 10 sections. Comparison of the results obtained for two independent CT scans of each vertebral body gave an estimate of the reproducibility. To obtain biomechanical characterization of each vertebral body the load-displacement curve was recorded using a material testing machine. Additionally, BMD values were measured using quantitative CT method. Percolation theory enables a detection of paths (percolating clusters) within the structure, along which loads applied at one of its sides are transferred to the opposite side. The concept of the percolating cluster and the problem of its detection have been generalized to the gray-level images via the concept of fuzzy connectedness. Namely, let the path L(p, q) between any two pixels p, q of the image was defined as any 4-connected sequence of pixels starting from p, and terminated at q. The conventional fuzzy degree of connectedness from p to q is defined as, C(p, q) = maxL[minzIˆLG(z)], where maxL is applied to all paths L(p,q), minzIˆL is applied to all pixels z along selected path L and G(z) denotes gray-level value at z. Then the degree of connectivity of opposite sides of the ROI is, D = maxIO[C(p,q)], where I and O denote the set of pixels being the opposite edges of the ROI. The computations of C(p,q) were based on Dijkstra’s algorithm. Finally, having evaluated D, it is checked if this value is consistent with non-bone background intensity (Gaussian distributed random variable). For this purpose the cumulative probability P(D) of the distribution of background intensities is evaluated for D. Results The BMD values of the vertebral bodies ranged from 25 to 45 (mg of hydroxyapatite)/cm3 while the ultimate compression stresses (UCS) were in the range (2.7–5.6) MPa. Because of low BMD and UCS values high risk of osteoporotic fracture could be attributed to these samples. The analysis based on percolation theory confirmed that two groups of cases may be distinguished. The first group (A) contained samples with
S19 P(D) [ 0.97 and the second group (B) was composed of samples with P(D) \ 0.96. P(D) and SD(P(D)) values were significantly different in both groups (p = 0.003 and 0.013, respectively), while BMD, UCS and age were not different (p = 0.82, 0.43 and 0.2 for BMD, UCS and age, respectively). SD(P(D)) was estimated from the measurements of 10 sections analyzed for each sample. Keeping in mind the definition of P(D) it could be concluded that group A of P(D) represents decrease of BMD which is not followed by the significant degradation of the trabecular bone architecture. Therapy increasing bone mineral content would probably lead to recovery of stiffness of trabecular bone for this group. The group B represents trabecular structures of low mineral mass, which are in addition highly perforated. Therapy increasing bone mineral content would be probably much less effective in these cases. Another feature that should be noted is the increase of standard deviation of the P(D) measurement with decreasing P(D). The correlation coefficient between P(D) and SD(P(D)) was equal to -0.93. This increase is a manifestation of increasing inter-section inhomogeneity of the trabecular structure. Conclusion In this study a new architectural parameter P(D), derived from the percolation theory has been defined and applied to the analysis of gray-level CT images of trabecular bone, acquired in vitro. P(D) is intended to measure a global ability of trabecular bone to transfer applied loads. The reproducibility of the P(D) measurement is quite good. The variations of P(D) value obtained in two measurements were of the order of 10% of SD(P(D)). Adjusting specimens with small trabecular bone density for age of the donors, BMD and UCS, we have shown that P(D) significantly differentiates different types of trabecular architecture. It seems that the proposed technique is potentially interesting for in vivo applications. The next important issue is the fact that the analysis is based directly on gray-level images. This allows us avoiding all controversies related to the still unresolved problem of segmentation bone from marrow in CT images of trabecular bone.
A cone-beam CT system for musculoskeletal extremities with advanced multi-mode imaging capabilities W. Zbijewski1, P. De Jean1, P. Prakash1, Y. Ding1, J.W. Stayman1, N. Packard2, R. Senn2, D. Yang2, J. Yorkston2, A. Machado3, J. Carrino3, J. Siewerdsen1 1 Johns Hopkins University, Biomedical Engineering, Baltimore, MD, USA 2 Carestream Health, Inc., Rochester NY, USA 3 Johns Hopkins University, Radiology, Baltimore, MD, USA Keywords Musculoskeletal radiology Cone-beam CT Flat-panel detector Dual-energy CT Extremities imaging Purpose A dedicated cone-beam CT (CBCT) scanner has been developed to complement conventional musculoskeletal (MSK) imaging modalities by addressing the following clinical requirements: (1) imaging of weight-bearing extremities; (2) low imaging dose; (3) compact design with better workflow compared to whole-body scanners; and (4) combined radiography, fluoroscopy and tomographic imaging in the same system. We report the scanner design and summarize advanced imaging modes under development, such as ultra-high resolution imaging of bone and joint morphology and dual-energy imaging. Methods As illustrated in Fig. 1, two gantry configurations include: a standing configuration for weight-bearing lower extremities in a natural stance (Fig. 1a) and a sitting configuration for unloaded lower and upper extremities (with the ability to apply tension, Fig. 1c). The gantry is self-shielded, simplifying radiation safety site requirements. The outer diameter is *110 cm, and the footprint is *110 9 180 cm2, with
123
S20 weight *200 kg. The inner bore is *20 cm and incorporates accessories for patient immobilization (e.g., inflatable air bladder). Inside the gantry, a flat-panel detector (FPD) and fixed-anode X-ray source are mounted on sickle-shaped arm allowing the patient to enter easily from the side of the bore. The source-detector distance is 550 mm, and source-isocenter distance is 420 mm. The scanning arc is *240, exceeding the short-scan range for this geometry. The field of view is *20 9 20 9 20 cm3. The FPD (Varian 3,030+, 1,536 9 1,536 pixels, 0.194 mm pixel size) provides radiographic and fluoroscopic acquisition and optionally binned read-out. A custom 10:1 antiscatter grid (transmission * 70%) is used. The source (Source-Ray XRS-125-7K-P) delivers X-ray techniques up to 125 kVp, 7 mA (*0.8 kW); the focal spot is 0.5 mm. The system can be powered from a standard 120 V, 25 A electrical power outlet. The prototype is being deployed in patient pilot studies. Various imaging modes were studied on an experimental testbench emulating the scanner geometry. CBCT acquisition involved 480 projections collected over 240 at 90 kVp (+0.3 mm Cu and 4 mm Al filtration). The Feldkamp algorithm with extended Parker weights and Hann-apodized ramp filter with a cutoff at the Nyquist frequency was used for reconstruction. Two imaging protocols were identified: a ‘‘standard’’ protocol at 0.1 mAs/projection, with 2 9 2 detector binning (0.388 mm pixels) and 0.5 mm voxels; and a ‘‘sharp’’ protocol at 0.25 mAs/projection, with full-resolution detector readout and 0.15 mm voxels. The dose for each protocol was measured with an ionization chamber at the center of a 16 cm CTDI phantom. Dual-energy CBCT imaging capability was assessed using fresh cadaveric specimens after intra-articular iodine injection. The lowenergy technique was 60 kVp (+2 mm Al filtration), 0.4 mAs/projection; the high-energy technique was 120 kVp (+0.5 mm Ag +0.2 mm Cu +2 mm Al), 0.5 mAs/projection. A single-energy scan was collected at 100 kVp (+0.2 mm Cu +2 mm Al), 0.1 mAs/projection. Calcium and iodine decomposition images were obtained as weighted sums of low- and high-energy reconstructions, with weights
Fig. 1 CAD rendering and photographs of the prototype scanner, showing (a, b) standing configuration and (c, d) sitting configuration
123
Int J CARS (2011) 6 (Suppl 1):S17–S22 selected to extinguish iodine and bone, respectively. A composite image with equal weights for low- and high-energy components was also computed. Other imaging modes under development include 4D cine studies of articular motion using retrospective gating of projections under periodic motion. Quantitative studies of dose, image noise, resolution, and contrast-to-noise ratio were complemented by qualitative assessment of image quality in cadavers by two musculoskeletal radiologists. Results Figure 2a, b presents a slice through a cadaveric hand imaged using the ‘‘standard’’ protocol. An analogous pair of images for a cadaveric knee is in Fig. 2c, d. Dose was *6.4 mGy/scan (0.064 mSv), *1/3 of the dose reported for a conventional CT knee exam. The spatial resolution apparent in visualization of bony details exceeds that of conventional CT. Soft-tissue discrimination is comparable to conventional CT (10–20 HU contrast). Figure 2e, f illustrate a cadaveric hand imaged using the ‘‘sharp’’ protocol (*0.15 mSv/scan). Trabecular bone and joint space morphology are reconstructed with exquisite detail. Despite slightly increased noise, the high-resolution protocol was considered valuable beyond conventional CT in the diagnosis and assessment of treatment response in diseases such as arthritis.
Fig. 2 a, b Coronal slice and volume rendering of a cadaveric hand imaged using the ‘‘standard’’ protocol. c, d Sagittal slice and volume rendering of a knee imaged using the ‘‘standard’’ protocol. e, f Coronal slice and volume rendering of a hand imaged using the ‘‘sharp’’ protocol. g Axial slice through a single-energy reconstruction of a hand following intra-articular iodine injection. h, i Axial and coronal slices through a dual-energy CBCT image, iodine (red) and calcium (blue)
Int J CARS (2011) 6 (Suppl 1):S17–S22 Figure 2g shows a reconstruction of an iodine-injected wrist obtained from a single-energy scan at 100 kVp. Delineation of iodine and bone is challenging due to similar reconstructed value of both materials. In Fig. 2h, i, calcium (blue) and iodine (red) images obtained using a dual-energy technique display clear separation of the two materials. Abnormal distribution of iodine (yellow arrows) from communication defects between joint spaces is easily visualized using the dual-energy technique. Conclusions The adaptable design and imaging modes of the dedicated CBCT scanner for MSK extremities allow a variety of new capabilities for MSK imaging tasks and pathologies. Standing and sitting configurations allow visualization of pathology (e.g., tissue impingement) visible only under load. Digital radiography and real-time fluoroscopy available on the same system as volumetric imaging provides multimode acquisition within a single session. Image quality similar to that of conventional CT can be achieved at *1/3 the dose. A variety of task-specific tomographic protocols were developed, including ultrasharp bone imaging with significantly improved spatial resolution over conventional CT. Promising results were obtained with dualenergy imaging, indicating improved material differentiation valuable for soft-tissue visualization, analysis of inflammatory disease, and imaging pathologies with distinct exogenous and endogenous material characteristics (e.g., gout). Other capabilities under development include imaging joints in motion, reconstruction in the presence of metallic implants, and automatic assessment of joint space morphology and bone/tissue changes in longitudinal studies.
MSCT follow up in patients with malignant lymphoma: does semi-automated volumetry improve therapy response classification compared to manual linear measurements? B. Buerke1, M. Puesken1, A. Knauer1, C. Schuelke1, W. Heindel1, J. Wessling1 1 University of Muenster, Department of Clinical Radiology, Muenster, Germany Keywords Segmentation Lymph node Follow up Therapy response evaluation Purpose In CT manual acquisition of linear measurements in the axial plane suffers from high inter- and intraobserver variability. This may potentially lead to misinterpretation in tumour response assessment. Semi-automated 3D measurements and volumetry were found to be more reproducible and accurate than manual assessment. The aim of this study was to assess the impact of semi-automated volumetry compared to manual linear measurements on therapy response classification in CT follow-up of malignant lymphoma. Methods Contrast enhanced MSCT scans of 65 patients (mean age 56 ± 13 years) with histologically confirmed malignant lymphoma prior to therapy (baseline) and after 2 cycles of chemotherapy (follow-up) were included. All examinations were performed using a 64-multislice CT scanner (Somatom Definition; Siemens Medical Solutions, Forchheim, Germany). Images were obtained at 120 kV using CARE dose with a 32 9 0.6 mm collimation. CT data sets were reconstructed at a slice thickness of 1.5 mm with a reconstruction increment of 1.0 mm. A total of 313 target lymph nodes (56 cervical, 131 thoracic and 126 abdominal, each in baseline and follow up) were evaluated by two radiologists independently. Prior to evaluation, lymph node selection was performed by an unblinded radiologist using IWC-guidelines for target lesions. Each selected lymph node was digitally tagged with numbers in baseline and follow-up examinations in order to avoid correlation and mapping errors. Long axis diameter (LAD), short axis diameter (SAD), WHO-
S21 square and volume were determined manually and using semi-automated segmentation software (OncoTreat, MeVis, Bremen, Germany). Dedicated correction tools could be used to modify unsatisfactory segmentation results. Relative interobserver difference (RID) and time for manual and semi-automated segmentation were evaluated. Therapy response was calculated for each parameter based on ,,IWC lymphoma-guidelines and ,,RECIST-adapted guidelines. Mean of metric and volumetric measurements served as the reference standard. Statistical analysis encompassed intraclass correlation coefficients (ICC), t test and McNemar-test. Results Over all regions mean lymph node size in baseline/follow-up was 23.8 ± 10.3/17.0 ± 9.2 mm for LAD, 445.5 ± 448.1/232.2 ± 205.1 mm2 for WHO-square and 7.2 ± 13.5/3.4 ± 9.9 ml for volume. RID was consistently low in baseline and follow up for both manual and semi-automated measurements with high ICC C 0.96. Mean evaluation time for semi-automated segmentation without need for correction was shorter (16.6 ± 11.7 s) than for manual measurements (29.0 ± 14.5 s). In 52% of cases use of correction tools was necessary and mean evaluation time increased to 39.5 ± 25.9 s. Regarding follow up and therapy response, semi-automated volumetry obtained significantly more accurate classifications than semi-automated and manual LAD and SAD (e.g. volume 87.8% vs. manual SAD 78.9%, p \ 0.05). Volume and WHO-square were equivalent parameters in therapy response evaluation. Conclusion Semi-automated lymph node volumetry is more accurate for therapy response classification in patients with malignant lymphoma as compared to established LAD. Mean evaluation time is within acceptable ranges. Thus, our data support the use of semi-automated lymph node volumetry in the follow up of patients with malignant lymphoma.
Deformable registration is a necessary preprocessing step for perfusion CT imaging of malignant pleural mesothelioma W. Sensakovic1, Z. Labby1, C. Straus1, S. Armato1 1 The University of Chicago, Dept of Radiology, Chicago, USA Keywords Registration Perfusion Computed tomography (CT) Quantitative imaging Functional imaging Purpose Malignant pleural mesothelioma (MPM) is a bulky neoplasm that grows circumferentially about the lung. Patient-specific characterization of MPM tumors to guide treatment decisions (both surgical and medical) and stratify the patient population is an active area of current research. Our institution is currently conducting a pilot study investigating the use of functional information provided by perfusion computed tomography (CT) to augment the structural information provided by clinically indicated contrast-enhanced CT scans. Unfortunately, cardiac and respiratory motion can cause severe artifacts that corrupt the functional data obtained from MPM patients. Further, unlike other lung cancer patients, MPM patients lose a substantial portion of their respiratory volume making respiratory gating difficult. This study investigated the application of deformable registration to align the multiple time points of perfusion CT scans and evaluate their impact on the subsequent calculation of functional data. Methods Four patients underwent an experimental perfusion CT protocol on a Philips iCT scanner (fanbeam configuration with 256 detector rows). Each perfusion CT scan consisted of 25 separate scans (i.e., time points) of a 5.5–5.8 cm axial portion of the patient. The time points were aligned using a multi-resolution deformable registration method based on B-splines and a multi-resolution deformable registration method based on demons registration (Fig. 1). The alignment error
123
S22 was calculated as the mean square pixel value differences between the initial perfusion time point and all subsequent registered time points. Functional data maps were calculated from the registered images (Fig. 2) and qualitatively compared to each other and to maps derived from the non-registered scan. The perfusion time points were acquired with 120 kVp and 100 mAs. Perfusion images were reconstructed as 512 9 512-pixel matrices with image filter B, reconstruction kernel B, 3 mm slice thickness, 3 mm spacing between sections, and an average pixel spacing of 0.6441 ± 0.0317 mm/pixel for each in-plane dimension. Perfusion maps were calculated using in-house software based on standard slope-based definitions of perfusion CT metrics. Regions of interest (ROIs) used for calculating perfusion maps were manually defined for each patient in mesothelioma tumor, fat, pectoral muscle, and the aorta. ROIs in the aorta were used to normalize the contrastmedia uptake for calculation of the perfusion maps. Perfusion, mean transit time (MTT), peak enhancement, blood volume and time to peak enhancement maps were calculated for each patient. Results The error for the B-spline deformable registration method was 94.58 ± 3.75% (max: 98.35%, min: 91.80%) of the non-registered scan error. The error for demons deformable registration method was 83.96 ± 1.59% (max: 85.92%, min: 82.10%) of the non-registered scan error. Qualitatively, demons registration resulted in the best registration accuracy overall. Lung boundary, bone, pulmonary vessels, and mediastinal structures were all aligned (Figs. 1, 2). Unfortunately, the registered images seemed severely ‘‘smoothed’’ and further tests will be necessary to determine if this impacts the absolute accuracy of the functional maps. For metrics averaged over spatial regions of interest, this smoothing effect may not be overly detrimental. B-spline registration completely failed on one patient. Other B-spline registered patients demonstrated improved alignment for the lungs, bones, and pulmonary vessels, but not for mediastinal structures. Finally, non-registered data demonstrated severe motion artifacts and was found to be largely unusable for calculation of functional data maps (Figs. 1, 2). Conclusion Perfusion CT may provide functional information useful for informing therapeutic care for MPM patients. Severe cardiac and pulmonary motion artifacts will affect the perfusion CT scans making accurate
123
Int J CARS (2011) 6 (Suppl 1):S17–S22
Fig. 1 Difference between first and last perfusion CT scans. Left Non-registered perfusion CT image. Middle B-spline registered perfusion CT image, and Right Demons registered perfusion CT image. Note that non-registered images demonstrate severe artifact due to pulmonary motion (light and dark pixels). Ideal registration between image pixels would be visualized as middle gray
Fig. 2 Peak enhancement data map calculated from Left nonregistered perfusion CT data, Middle B-spline registered perfusion CT data, and Right demons registered perfusion CT data. Note that non-registered data demonstrates severe artifact due to pulmonary motion (e.g., white rim about lung and bright vessels). B-spline registered data removes artifact around the lung, but the mediastinal structures are incorrectly aligned. Demons registration provides the best overall fit, reducing artifacts around the lung, pulmonary vessels, and mediastinum calculation of functional data maps difficult, if not impossible. Deformable registration, especially the technique based on demons, substantially improved alignment of perfusion CT scans and resulted in improved calculation of functional data maps.