PET/CT-guided biopsy with respiratory motion correction

Abstract

Purpose

Given the ability of positron emission tomography (PET) imaging to localize malignancies in heterogeneous tumors and tumors that lack an X-ray computed tomography (CT) correlate, combined PET/CT-guided biopsy may improve the diagnostic yield of biopsies. However, PET and CT images are naturally susceptible to problems due to respiratory motion, leading to imprecise tumor localization and shape distortion. To facilitate PET/CT-guided needle biopsy, we developed and investigated the feasibility of a workflow that allows to bring PET image guidance into interventional CT suite while accounting for respiratory motion.

Methods

The performance of PET/CT respiratory motion correction using registered and summed phases method was evaluated through computer simulations using the mathematical 4D extended cardiac-torso phantom, with motion simulated from real respiratory traces. The performance of PET/CT-guided biopsy procedure was evaluated through operation on a physical anthropomorphic phantom. Vials containing radiolabeled 18F-fluorodeoxyglucose were placed within the physical phantom thorax as biopsy targets. We measured the average distance between target center and the simulated biopsy location among multiple trials to evaluate the biopsy localization accuracy.

Results

The computer simulation results showed that the RASP method generated PET images with a significantly reduced noise of 0.10 ± 0.01 standardized uptake value (SUV) as compared to an end-of-expiration image noise of 0.34 ± 0.04 SUV. The respiratory motion increased the apparent liver lesion size from 5.4 ± 1.1 to 35.3 ± 3.0 cc. The RASP algorithm reduced this to 15.7 ± 3.7 cc. The distances between the centroids for the static image lesion and two moving lesions in the liver and lung, when reconstructed with the RASP algorithm, were 0.83 ± 0.72 mm and 0.42 ± 0.72 mm. For the ungated imaging, these values increased to 3.48 ± 1.45 mm and 2.5 ± 0.12 mm, respectively. For the ungated imaging, this increased to 1.99 ± 1.72 mm. In addition, the lesion activity estimation (e.g., SUV) was accurate and constant for images reconstructed using the RASP algorithm, whereas large activity bias and variations (± 50%) were observed for lesions in the ungated images. The physical phantom studies demonstrated a biopsy needle localization error of 2.9 ± 0.9 mm from CT. Combined with the localization errors due to respiration for the PET images from simulations, the overall estimated lesion localization error would be 3.08 mm for PET-guided biopsies images using RASP and 3.64 mm when using ungated PET images. In other words, RASP reduced the localization error by approximately 0.6 mm. The combined error analysis showed that replacing the standard end-of-expiration images with the proposed RASP method in PET/CT-guided biopsy workflow yields comparable lesion localization accuracy and reduced image noise.

Conclusion

The RASP method can produce PET images with reduced noise, attenuation artifacts and respiratory motion, resulting in more accurate lesion localization. Testing the PET/CT-guided biopsy workflow using computer simulation and physical phantoms with respiratory motion, we demonstrated that guided biopsy procedure with the RASP method can benefit from improved PET image quality due to noise reduction, without compromising the accuracy of lesion localization.

Introduction

Over the last decade, PET/CT imaging, which combines the strengths of anatomical and functional imaging, has become a mainstay in the management of many types of cancer [1–3]. PET/CT has also been shown to have better overall sensitivity and specificity in cancer staging and disease assessment compared to PET alone [4–6]. However, a number of benign conditions such as inflammatory changes and fibrosis may result in high 18F-fluorodeoxyglucose (FDG) uptake leading to tumor over-staging or under-staging [7, 8]. Therefore, percutaneous biopsy is often required to extract tissue for analysis of malignancy.

Percutaneous biopsy is an effective procedure to distinguish benign from malignant tissue and is a commonly performed procedure. The biopsy needle is placed using a percutaneous (through the skin) approach using CT images for guidance. Percutaneous biopsy has a much lower risk to patients compared to open surgical procedures. Even with the advances in percutaneous devices and biopsy strategies, a definitive diagnosis can be elusive due to several reasons [9]:

1. Tumors are often heterogeneous in composition, requiring multiple sampling attempts at different locations to obtain sufficient tissue for pathological evaluation.

2. With advanced imaging and increased screening practices, tumors are being detected at earlier stages and biopsies are therefore being performed on smaller targets. The interventional radiologist must make a mental map from the CT images to the patient anatomy to precisely place the needle at the desired location. Even with the best techniques, smaller lesions will have higher false negative biopsy outcomes.

3. Lesions on CT images can be inconspicuous and impossible to target with CT alone.

4. Finally, even when a diagnostic PET/CT image is available for pre-procedural biopsy planning, it is typically only available on a workstation suitable for diagnostic interpretation, rather than a workstation that displays image data as is needed during a biopsy procedure. This forces the interventional radiologist to mentally integrate and register pre-acquired images within the real-time biopsy environment. Although many interventional radiologists have become quite skilled at this process, it is challenging and potentially subject to significant errors.

Having PET/CT fused information readily available as a roadmap for tissue sampling is needed to address these issues. PET images can be used to reduce sampling error by guiding biopsy in places where the tumor has a mix of necrotic and malignant regions, and direct integration of PET/CT images into the biopsy environment eliminates the need of mentally registering images from different workstations. Several investigators have described clinical techniques that are used to perform PET/CT-guided biopsies, underscoring the clinical need for these procedures [10–12].

However, the diagnostic and operational benefits of using PET/CT images are gravely affected by respiratory motion artifacts. Respiratory motion causes a discrepancy in spatial correspondence between the CT and PET data, leading to errors in the attenuation corrections required for accurate PET imaging. CT data are acquired in a short time and represent an instantaneous snapshot during the breathing cycle. In contrast, PET data acquisition takes an average of 1–10 min per gantry table position. Depending on the region to be scanned, multiple gantry table positions may be required, as each position covers 10–15 cm along the craniocaudal axis of the body. For typical whole body image acquisitions, 5–7 gantry table positions are required. During this time, voluntary patient motion and involuntary motion due to respiratory and cardiac motion are unavoidable. As a result, the PET image generated is a motion-averaged PET image. This blurred PET image may lead to inaccurate tumor localization or biopsy guidance resulting in incorrect staging of the tumor [13–16]. Recent data [17] show that PET uptake values are reduced by approximately 28% on average and that apparent size is more than doubled for a 1-cm-diameter lesion, and in general, the lower thorax and upper abdomen have the largest errors. These errors will also affect the accuracy and/or success rate of biopsy that is based on the PET image.

To facilitate PET/CT-guided needle biopsy, in this study we developed a workflow that allows us to bring PET image guidance into the interventional CT suite while accounting for respiratory motion. In the simulation studies described below, we focus on the lower thorax and upper abdomen. We adopted the registered and summed phases (RASP) method [18] for PET/CT respiratory motion correction. This method relies on respiratory gated CT images aligned using image registration that is then applied to the phase-matched attenuation-corrected PET images. We demonstrate the feasibility of the proposed PET/CT-guided biopsy workflow via computer simulations and physical phantom studies. We note that for terminology, ‘PET’ refers to only the PET acquisition or image, ‘PET/CT’ refers to the dual-mode PET and CT acquisitions or images from a PET/CT scanner, and ‘PET/CT’ refers to combined operations on PET and CT images, such as the RASP algorithm described below.

Methods

Overview of proposed clinical workflow

Our proposed clinical workflow for PET/CT-guided biopsy is shown in Fig. 1. In this workflow, we first obtain the preoperative images in the PET/CT suite. A respiratory motion correction algorithm is then applied to these preoperative PET/CT images. During the biopsy procedure, the intra-operative CT images are registered to the preoperative CT images and the resulting deformation maps are then applied to the motion-corrected PET images. These deformed PET images are then fused with the intra-operative CT images to provide visualization guide during biopsy.

Fig. 1.

System diagram of a PET/CT-guided biopsy workflow

To plan the path for needle placement, we first register the interventional CT coordinate system with the patient coordinate system, where fiducials are placed on the patient’s body before the interventional CT acquisition for point-to-point registration. During the needle path planning, the clinicians first select the area of interest on the overlaid PET/CT image on the user interface as the target point. Then, the skin entry point is selected and the proposed needle path is drawn between those two points over the CT image. The clinicians may adjust the skin entry point to form a projected path avoiding vessels and other anatomy. During the biopsy procedure, the needle will be tracked in real time and updated on the virtual image using an electromagnetic external tracking device. The clinicians will then use this display as guidance for the biopsy.

PET/CT respiratory motion correction

In the proposed workflow shown in Fig. 1, we adopted the registered and summed phases (RASP) method described by Kinahan et al. [18] for PET/CT respiratory motion correction using respiratory gated CT images. In summary, the motion correction algorithm is illustrated in Fig. 2 and proceeds as follows:

Fig. 2.

Illustration of the registered and summed phases (RASP) method for respiratory motion correction

1. Each respiratory phase image (1 to N in Fig. 2) of the CT acquisition is matched to the corresponding respiratory phase of the PET acquisition and used to correct for attenuation for each respiratory phase. The N attenuation correction PET images are then each reconstructed used standard methods.

2. In parallel, a deformable registration algorithm is used to compute deformation maps for each of the N CT image volumes with respect to a reference image (e.g., the first-phase CT image volume in Fig. 2).

3. The deformation maps are then used to align each attenuation-corrected PET image to the reference phase volume (e.g., the first-phase PET image volume in Fig. 2).

4. The registered PET volumes are summed to generate a single motion-free, low-noise PET image volume.

Since the RASP algorithm uses all of the events detected in the PET acquisition, it has the same noise properties as a PET acquisition without respiratory gating, but with minimal attenuation correction artifacts and respiratory motion.

Motion-corrected PET/CT to interventional CT registration

We used the B-spline algorithm for PET/CT deformable image registration for the motion correction and CT image registration between PET/CT suite and interventional CT suite [19]. More precisely, each CT image was directly registered to the reference image to obtain the deformation map, where the non-rigid transformation was obtained through optimization using B-spline algorithm. In this study, we used the open-source BRAINSfit implementation of the B-spline registration algorithm [20]. We used the following parameters for this implementation:

• samplingPercentage: 5%

• splineGridSize: 9 × 9 × 9

• interpolationMode: Linear

• maxBSplineDisplacement: 5 cm

• costMetric: RMSE.

We have previously demonstrated the feasibility of this method in [19].

Interventional CT to patient coordinate system registration

We used landmark-/point-based registration to register the interventional CT coordinate system to the patient coordinate system in the interventional suite. We relied on the tracked biopsy needle and fiducials placed on the patient’s body before the interventional CT acquisition to register the tracker coordinate system with the interventional CT coordinate system. For this, we used electromagnetically external tracking device (Aurora, Northern Digital, Waterloo, Canada). Magnetic sensors were embedded in the needle tip to facilitate tracking. In a ferromagnetically clean environment, the average positioning error for this system was measured to be 1.15 ± 0.78 mm [21].

During registration, the needle tip was placed on a fiducial point, and then, that point was selected in the interventional CT image using our software. Once at least four points had been registered this way, the software calculated the linear transform between tracker coordinates and image coordinates. Applying the computed transformation to the tracker coordinates consequently registered the patient with the interventional CT image. All previous PET and CT images were converted to the interventional CT image coordinate system for subsequent guidance.

SlicerPET: PET/CT-guided needle biopsy software application

We implemented an image-guided biopsy software application, SlicerPET, for the workflow described above using the 3DSlicer framework [22]. 3DSlicer is a free and open-source extensible cross-platform toolkit for medical image segmentation, registration, visualization and image-guided surgery. Slicer is in active use by the biomedical imaging research community as a vehicle to translate innovative algorithms into clinical research applications. The Slicer- PET software is also open source [23, 24] and provides a workflow module consisting of the following steps:

1. Data loading: Allows a user to select the three volumes that are required for the guidance: respiratory-compensated PET, respiratory-compensated CT and interventional CT scans.

2. Registration: Register the respiratory-compensated CT with the interventional CT scan.

3. Tracking: Establish a connection between the tracking system and 3DSlicer using the image-guided surgery toolkit (IGSTK) [25].

4. Tool calibration: Determine the transformation between the tip and head of the biopsy needle if needed.

5. Patient registration: Register the interventional CT coordinate system to the patient coordinate system in the interventional suite using point-based registration.

6. Needle path planning: Identify entry and target points and generate plan for the needle path.

7. Guidance: Overlay a model of the needle and provide 3D visualization guidance for needle placement.

Computer simulation of motion compensation

We simulated PET/CT scans with the 4D extended cardiac-torso (XCAT) phantom [26] to generate 3D distributions of radiotracers and linear attenuation coefficients of the patient body. To simulate lesion inhomogeneity, we inserted two 15-mm-diameter lesions with relatively low activity (cold) and three 6-mm-diameter lesion with high activity (hot). Of the three hot lesions, one is placed in the liver and close to the diaphragm and the other two are placed within the two cold lesions near their edges in order to create heterogeneous lesions. The proportionate activity concentrations of lesion (hot):lesion(cold):liver:lung:body were 13:3:3:2:1 in SUV. Figure 3 illustrates the resultant XCAT phantom in overlaid PET/CT image. In this figure, lesions 1, 2 and 3 represent the hot dynamic liver lesion, hot dynamic lung lesion and hot static lung lesion, respectively. The two dynamic lesions tend to move strongly with respiration, while the static lesion barely moves with respiration.

Fig. 3.

Illustration of the XCAT phantom in overlaid PET/CT image. Lesions 1, 2 and 3 represent the hot dynamic liver lesion, hot dynamic lung lesion and hot static lung lesion, respectively

To simulate respiratory motion, we used the method described in detail by Liu et al. [17]. In summary, we used three Varian Real-time Position Management (RPM) time-varying measurements of chest displacement previously collected during patient PET/CT scans [17]. These patterns were used to drive the rendering of XCAT images with simulated lesions with amplitudes of anterior-posterior expansion in XCAT images matching amplitudes of the RPM measurements. This approach has been shown to accurately replicate the original patient respiratory motion of the chest wall [17].

For each respiratory motion trace, we simulated phase-matched PET and CT data. We used a method similar to [27] to form noisy motion-blurred sinograms according to a given RPM trace. We also included a spatially variant PSF model, in combination with Poisson noise, scatter events and random events, in the PET simulation. We reconstructed each phase-gated PET sinogram data with a voxel size of 2.73 × 2.73 × 3.27 mm³ using the ordered-subset expectation-maximization (OS-EM) reconstruction algorithm with 3 iterations and 32 subsets. A Gaussian post-filter with 6 mm full width at half maximum (FWHM) and an axial filter of value [1/4, 1/2, 1/4] were applied to the reconstructed images.

To study the impact of respiratory phase gating and motion correction, we compare the following three reconstruction methods: the ‘ungated’ method where the final PET image is reconstructed from data without respiratory gating, with attenuation correction corresponding to the end-of-expiration phase; the ‘unregistered’ method where the final PET image is the ensemble average of the 10 reconstructed phase-gated PET images, each of which represents a distinct respiratory phase and is attenuation-corrected using phase-matched CT image; and the RASP method where the final PET image is the average of the 10 reconstructed phase-gated PET images after separate image registration for motion correction. We used ‘static’ PET image as a reference, which is acquired and reconstructed at the end of expiration without noise simulation.

For performance evaluation, we conducted active PET volume measurement, lesion centroid estimation and simulated biopsy. For active PET volume measurement, within a 1.5 cm × 1.5 cm × 1.5 cm lesion window on the PET image, we counted the number of voxels whose values were beyond 75% of the maximum value within the particular window. Then, we calculated the centroid of the volume for each lesion. The distance between the reference lesion centroids and the estimated lesion centroids from different reconstruction methods was used to evaluate the accuracy of lesion localization. The simulated biopsy was then performed by acquiring the voxel values on the XCAT activity phantom image at the end of expiration, at the locations of estimated lesion centroids obtained from the previous step. The average of the acquired activity values was compared against the reference value to evaluate the accuracy of biopsy.

Physical phantom study

We used the anthropomorphic phantom shown in Fig. 4a to demonstrate the feasibility of targeting a PET-visible lesion using our newly developed software application. This torso phantom has the exact shape of a human torso and is described in detail by Lin et al. [28]. In brief, it is designed to (1) simulate respiratory motion at various breathing rates, (2) be CT compatible and (3) accommodate needle punctures in a realistic manner. The torso phantom is cast of silicone rubber with properties similar to soft tissue and contains pulmonary and abdominal cavities that can be punctured repeatedly with a needle. There are an embedded rib cage and a small opening in the thorax to allow the repeated percutaneous targeting used in this study. Respiratory motion is driven by a computer-controlled air pump that inflates the thorax. By controlling the air pump, the phantom can simulate realistic respiratory motion at different breathing rates. Within the thorax, we embedded FDG filled, sealable circular vials to serve as targets. These appear as areas of increased radiopharmaceutical activity on the PET image. The dose was injected into a small hollow plastic sphere with a threaded seal. The spheres were sealed, placed inside a simulated liver made of foam and then placed in the anthropomorphic phantom. During the imaging studies described next, the phantom respiratory motion was driven by sine waveforms with a period of 5 s to simulate a human respiratory rate of 12 breaths per minute.

Fig. 4.

Phantom experiments using electromagnetic tracking and fused PET/CT in the interventional CT suite at Georgetown University. a Anthropomorphic phantom with embedded target, tracked biopsy needle, electromagnetic field generator and graphical user interface (GUI). b Close-up of the GUI displaying four quadrant view. A virtual reality view of the needle touching the PET-visible embedded target can be seen in the top right quadrant, and the extension of the needle tip through center of the target can also be visualized indicating an accurate biopsy

The phantom was imaged with the PET/CT scanner (Siemens mCT 64) in the Nuclear Medicine Department at Georgetown University Medical Center, and the images were saved in the DICOM format on CD. The PET image size was 200 × 200 × 170 with 4.0 × 4.0 × 1.5 mm³ voxels, while the CT image size was 512 × 512 × 127 with 1.5 × 1.5 × 2.0 mm³ voxels. The phantom was then moved to the Interventional Radiology Department and placed on the CT scanner (Siemens Volume Zoom) in the interventional CT suite, where the biopsy procedure would be performed. The interventional CT images were 512 × 512 × 315 with 0.98 × 0.98 × 1.0 mm³ voxel size. All images were acquired at end expiration phase. A CT scan was obtained and sent to our computer workstation, where it was fused with the PET/CT images using a landmark-based registration algorithm based on external donut fiducial markers (multimodality markers, IZI Medical, Baltimore, MD). The same fiducials were used to register the tracking system to the CT image. The GUI system then provided an image overlay for targeting the PET-visible hot spot as shown in Fig. 4b. Then using the visual guidance of the fused PET/CT image and real-time location and position information from the tracking device, the biopsy needle (MagTrax, Traxtal Technologies Inc., Canada) was inserted by a trained interventional radiologist into the phantom to target the sphere. An illustrative example of how such a needle insertion looked like is shown in Fig. 3 of the study by Yaniv et al. [29].

We conducted a total of ten needle targeting attempts, with the goal of hitting the surface of the embedded vial and for the needle trajectory to pass through the center of the sphere. After each needle placement, an additional CT scan was obtained to determine the actual needle trajectory. The perpendicular distance between the actual needle trajectory and the center of the embedded vial was then computed and was used as a metric to evaluate the tumor localization accuracy.

Results

Computer simulation study

Figure 5 shows PET images reconstructed with different strategies, overlaid with the CT image acquired at the end of expiration. Figure 5a presents the static PET image as a reference. Figure 5b presents the ‘ungated’ PET image, as reconstructed from data without respiratory gating and attenuation-corrected with the CT image at the end of expiration. Figure 5c presents the PET image reconstructed using single phase at the end-of-expiration phase, out of 10 respiratory phases generated from phase gating, with phase-matched CT attenuation correction. Figure 5d presents the ‘unregistered’ PET image, as the average of 10 reconstructed PET images, each of which corresponds to a distinct respiratory phase and is attenuation-corrected using phase-matched CT image. Figure 5e presents the RASP PET image, as the average of 10 reconstructed PET images after separate image registration for each phase for motion correction. Table 1 presents the average of liver noise measurements in PET images produced by different motion management strategies. Noise measurements were taken within same ROI in liver for different methods and were repeated for all three patients.

Fig. 5.

Overlaid PET/CT images with a static PET image as a reference; b ‘ungated’ PET image; c single phase of respiratory phase-gated PET data; d ‘unregistered’ PET image; and e RASP PET image. Note the artificially enhanced liver dome in (b), which is an artifact caused by motion-mismatched CT attenuation correction

Table 1.

Liver noise measurement in PET images produced by different motion management strategies

Motion management strategy	End of expiration	RASP	Unregistered	Ungated
Noise (SUV)	0.34 ± 0.04	0.10 ± 0.01	0.09 ± 0.01	0.09 ± 0.01

Noise measurements were taken within same ROI in liver for different methods and were repeated for all three patients. Average of the noise measurements over all patients is reported, along with the standard deviation indicating the variation in noise level among the three patients

Figure 6 plots the averaged volume for each lesion across three respiratory traces within each reconstruction method. Due to respiratory motion blurring and motion artifacts, the ‘unregistered’ and ‘ungated’ methods result in significantly increased lesion volumes. In particular, the significantly increased activity volume measured from the ‘ungated’ result is caused by reduced peak activity, resulted from respiratory motion blurring, and the artifact at the liver dome (Fig. 5b), resulted from motion-mismatched CT attenuation correction. By applying motion correction, the RASP method lessens the motion blurring and therefore produces lesions with more accurate volumes.

Fig. 6.

Averaged volume of lesions across different respiratory traces in different reconstructions. Each color bar represents the volume of a particular lesion averaged across respiratory traces within the same reconstruction method. The error bar indicates the standard deviation across three different traces

Figure 7 presents the RMSE between the ground-truth lesion centroids and the estimated lesion centroids determined from different PET images for each lesion across three respiratory traces. Table 2 presents the overall RMSE of lesion centroids between the ground truth and the estimated result for each lesion from different reconstruction methods, evaluated across all respiratory traces. The results show that both the ‘ungated’ and ‘unregistered’ results lead to imprecise lesion localization for dynamic liver and lung lesions, which are close to the diaphragm where respiratory motion has greatest influence. On the other hand, the RASP method locates all three lesions accurately across different respiratory traces due to proper respiratory motion compensation. Note that all three methods exactly localize the static lung lesion due to its lack of movement during PET/CT acquisition.

Fig. 7.

Average distance between the reference centroids and the estimated centroids determined from different PET reconstructed images for three different lesions across respiratory traces. The error bar indicates the standard deviation across three different traces

Table 2.

Average distance between the reference lesion centroids and the estimated lesion centroids for each lesion from different PET reconstructed images, averaged across all respiratory traces

Methods	RASP	Unregistered	Ungated
Dynamic liver lesion	0.83 ± 0.72	4.21 ± 1.98	3.48 ± 1.45
Dynamic lung lesion	0.42 ± 0.72	3.85 ± 2.25	2.5 ± 0.12
Static lung lesion	0	0	0

Standard deviation indicates variation among different respiratory traces

Figure 8 presents the simulated biopsy result on the XCAT phantom. The results were produced by averaging the activity measurements for each distinct lesion across all three respiratory traces. Each measurement was taken on the XCAT activity phantom image at the end-of-expiration phase, at the location of estimated lesion centroid obtained from each reconstruction method in the previous step. The results show that RASP method achieves accurate lesion activity, while the ‘ungated’ and ‘unregistered’ methods result in significantly lower activity measurement since performing biopsy with inaccurate lesion centroid estimates missed the lesions with highest activity.

Fig. 8.

Simulated biopsy result using the centroids determined from different reconstruction methods. Each bar represents an average of measured voxel values over three respiratory traces. Each voxel value measurement was taken on the XCAT activity phantom image at the end of expiration, at estimated lesion centroid for each lesion and each reconstruction method

Physical phantom study

The typical average error of CT-CT image registration is 2.52 mm, as previously measured with 10 landmarks [19]. We take this as representative of registration error between PET/CT and interventional CT. The RMSE of landmark registration between interventional CT coordinates and patient coordinates was 1.72 mm. During the ten attempts of the guided biopsy experiment, all the needles met the goal of touching the surface of the vial and the average distance computed was 2.94 mm, with a standard deviation of 0.86 mm.

Error analysis

Table 3 presents the error within the CT-CT registration step and patient registration step, as two main sources of error that affect the accuracy of biopsy location. Then one may estimate the error of localizing the biopsy point as square root of the sum quadrature combination of the above-mentioned two errors. Note that end-of-expiration PET/CT data were used in the corresponding experiment and therefore assumed free of respiratory motion error. Table 3 shows that the resultant estimated localization error (3.05 mm) is close to the actual measured localization error (2.94 mm).

Table 3.

Average error within the intermediate registration step of PET/CT-guided biopsy workflow and the final lesion localization error

Step	CT to CT registration	CT to patient registration	Estimated localization error	Measured localization error
Error (mm)	2.52	1.72	3.05	2.94

Table 4 presents the resultant estimated localization error as combining the PET/CT motion compensation strategies listed in Table 2 into the biopsy process, following the same calculation used in Table 3.

Table 4.

Estimated lesion localization error as resulted from using different motion management strategies

Motion management strategy	End of expiration	RASP	Unregistered	Ungated
Estimated localization error (mm)	3.05	3.08	4.07	3.64

Errors from CT to CT registration and CT to patient registration remain the same as those presented in Table 3

Discussion

In this study, we proposed and tested the entire PET/CT-guided biopsy workflow using computer simulation and a physical phantom with respiratory motion. In this workflow, we adopted the registered and summed phases (RASP) method to correct the respiratory motion occurred during the acquisition of PET/CT data.

Figure 5 compares the attenuation-corrected PET/CT overlaid images produced by using different motion management strategies. The ‘ungated’ PET image (b) shows severe motion artifacts at the diaphragm with heavily blurred liver lesion and lung lesion near the diaphragm, as caused by the motion-mismatched PET and CT. With respiratory phase gating, the PET image from a single phase (c) greatly reduces the blurriness due to respiratory motion as compared to the ‘ungated’ result; however, it suffers from significantly increased noise (as shown in Table 1) due to reduced detected photons within each phase bin and therefore would not be suitable for image guidance. By taking the average of PET images from all respiratory phases, (d) suppresses the noise but brings back the motion blurring (liver lesion and lung lesion near the diaphragm). By registering the phase-gated PET images first and then taking the average, the RASP method (e) results in images with a balance between motion artifact reduction and noise suppression, and is hence suitable for lesion localization and biopsy guidance.

Figures 6, 7, 8 and Tables 1, 2 quantitatively compare the impact of respiratory motion compensation methods on the accuracy of lesion detection and localization. The computer simulation results showed the RASP method generated PET images with significantly reduced noise of 0.10 ± 0.01 standardized uptake value (SUV) as compared to an end-of-expiration image with noise of 0.34 ± 0.04 SUV (Table 1). Figure 6 shows that respiratory motion increased the apparent liver lesion size from 5.4 ± 1.1 to 35.3 ± 3.0 cc. The RASP algorithm reduced this to 15.7 ± 3.7 cc.

Figure 7 and Table 2 show that the distance between the centroids for the ground-truth image lesion and those reconstructed with the RASP algorithm was 0.83 ± 0.72 mm for the dynamic liver lesion and 0.42 ± 0.72 mm for the dynamic lung lesion. For the ungated imaging, these values increased to 3.48 ± 1.45 mm and 2.5 ± 0.12 mm, respectively. Figure 8 shows that the lesion FDG activity estimation (e.g., SUV) was accurate and constant for images reconstructed using the RASP algorithm, whereas large activity bias and variations (± 50%) were observed for lesions in the ungated images. As shown in Fig. 8, RASP achieves the ground-truth activity for all three hot spot lesions, while the other two methods fail to sample the region of highest activity due to inaccurate lesion localization as shown in Fig. 7. Thus, it is demonstrated that RASP as a respiratory motion compensation method can effectively reduce the lesion distortion caused by breathing and potentially lead to more accurate lesion localization. We note that the original description of the RASP algorithm in [18] was a proof of concept and did not include any of the performance characterizations or comparisons with other approaches that are described here.

Table 3 summarizes the physical phantom studies, which demonstrated a biopsy needle localization error of 2.9 ± 0.9 mm from CT, compared to a predicted value of 3.05 mm. Table 4 combines the biopsy needle localization error with the localization error due to respiration for the PET images from the simulation studies. The overall estimated lesion localization error would be 3.08 mm for PET-guided biopsies images using RASP and 3.64 mm when using ungated PET images. In other words, RASP reduced the localization error by approximately 0.6 mm.

To our knowledge, this is the only study that has mechanistically studied the errors in PET/CT-guided biopsies. There have been, however, prior studies with patients using PET-guided biopsies [10–12]. Tatli et al. [30] registered prior PET/CT images for 14 patients with interventional CT-guided biopsies with one diagnostic failure. Venkatesan et al. [31] registered prior PET/CT images with either ultrasound- or CT-guided biopsies for 25 patients with a total of 36 biopsies. In their study, there were 5 diagnostic failures. The differences between these studies and ours are: (1) We do not report on patient studies. (2) In the Tatli and Venkatesan studies, there was no compensation for respiratory motion, likely as the technology was not available, which may have increased the number of diagnostic errors. (3) Our study provided error estimates in targeting.

Traditionally, for a PET/CT-guided biopsy procedure, the PET data acquired at the end of expiration are used to reveal the tracer concentration, so that the region of abnormal concentration within the organ of interest can be accurately located and later biopsied. However, as shown in Fig. 5c, the end-of-expiration PET data typically contain excessive amount of noise due to lack of photons and therefore may sometimes compromise lesion detection. To reduce the noise, we need to use PET data from all respiratory phases, with a motion management strategy to mitigate the organ movement and shape distortion caused by respiratory motion. Table 4 presents the resultant estimated localization error when different motion management strategies are combined into the biopsy process. It shows that, as compared to the error when only using end-of-expiration PET data (3.05 mm), incorporating RASP method into the guided biopsy process (3.08 mm) only increases the localization error by 1%, while the other two methods increase the localization error significantly (> 20%). This result demonstrates that the guided biopsy procedure with RASP method can benefit from improved PET image quality due to noise reduction, without compromising the accuracy of lesion localization. For context and comparison of our estimated positioning error, the study by Yaniv et al. [21] evaluated the accuracy of several image-guided CT and X-ray fluoroscopy systems in different suites (interventional radiology, interventional CT and pulmonary). The estimated positioning errors ranged from less than a millimeter to almost 20 mm. For interventional CT, perhaps the best comparison with our study, the mean error ranged from 1.0 to 5.7 mm depending on the system and configuration. Thus, our results of approximately 3 mm for respiratory gated PET/CT-guided biopsy are well within the range found by Yaniv et al. [21].

Note that in this study the RASP error was obtained through computer simulation and therefore may be understated as compared to the error encountered in clinical practice. However, the computer simulation was conducted with a validated digital phantom (XCAT) and respiratory traces obtained from patients undergoing PET/CT scans. Thus, we believe the computer-simulated RASP result is still valid and representative.

One limitation of the computer simulation is that only three respiratory traces with regular breathing patterns were used in the experiment. Despite demonstrating efficacy for patients with regular respiratory patterns, the RASP method may not fully compensate the respiratory motion for patients with irregular breathing patterns, for example, patients with significant intra-phase movement where the respiratory amplitude and/or baseline changes dramatically even within the same respiratory phase. In that case, the motion blurring and shape distortion may still be present in the processed PET/CT image and therefore affect the accuracy of lesion localization for guided biopsy. An alternative strategy may be replacing the respiratory phase gating with respiratory amplitude gating [32] to track the chest location directly.

Another limitation of the study is that the physical phantom study will not completely replicate needle insertion in a breathing human. While the phantom’s materials and skeletal structure were designed to closely resemble life-like characteristics, it cannot fully simulate the mechanical action of diaphragm contraction interacting with the pleurae. However, the phantom does capture what are considered the most important effects of motion of the chest wall and CT image appearance.

Conclusion

The entire PET/CT-guided biopsy workflow has been tested using computer simulations and physical phantoms with respiratory motion. The proposed registered and summed phases (RASP) method can produce PET images with low noise, artifact and respiratory motion, resulting in accurate lesion location and reduced shape distortion, which further benefits the biopsy guidance. SlicerPET, an image-guided biopsy software application for PET/CT-guided biopsy, was also developed and tested. With computer simulation and physical phantom studies, we demonstrated that guided biopsy procedure with RASP method for respiratory motion correction can benefit from improved PET image quality due to noise reduction, without compromising the accuracy of lesion localization.