Patent Grant
6915004
Not Applicable
Not Applicable
1. Field of Invention
This invention pertains to apparatus and processes for three-dimensional image reconstruction from data acquired in a positron emission tomograph (PET). More particularly, this invention pertains to apparatus and methods based on a parallel/pipelined architecture for processing data acquired as the bed moves through the tomograph.
2. Description of the Related Art
In a positron emission tomograph (PET) imaging system, a patient is injected with a radioactively tagged substance that the body normally metabolizes in some fashion. The radioactive tag used is a positron-emitting isotope of either an element found in the substance or an element that is substituted for another element in the substance. For example, a widely used isotope is the positron-emitting isotope of fluorine, 18F. This isotope is substituted, through a chemical synthesis process, for hydrogen in complex compounds such as glucose-forming fluro-deoxyglucose (FDG). When FDG is injected into a patient, the body will attempt to use it in the same fashion as it would normal glucose. Thus, there will be higher concentrations of positron emitters in areas where glucose is metabolized at higher levels, such as the brain, muscle tissue (the heart), and tumors.
As the FDG or other radiopharmaceutical isotopes decay in the body, they discharge positively charged particles called positrons. Upon discharge, the positrons encounter electrons, and both are annihilated. As a result of each annihilation event, gamma rays are generated in the form of a pair of diametrically opposed photons approximately 180 degrees (angular) apart. By detecting these annihilation “event pairs” for a period of time, the isotope distribution in a cross section of the body can be reconstructed. These events are mapped within the patient's body, thus allowing for the quantitative measurement of metabolic, biochemical, and functional activity in living tissue. More specifically, PET images (often in conjunction with an assumed physiologic model) are used to evaluate a variety of physiologic parameters such as glucose metabolic rate, cerebral blood flow, tissue viability, oxygen metabolism, and in vivo brain neuron activity.
Mechanically, a PET scanner consists of a bed or gurney and a gantry, which is typically mounted inside an enclosure with a tunnel through the center, through which the bed traverses. The patient, who has been treated with a radiopharmaceutical, lies on the bed, which is then inserted into the tunnel formed by the gantry. Traditionally, PET scanners are comprised of one or more fixed rings of detectors, surrounding the patient on all sides. Some newer scanners use a partial ring of detectors and the ring revolves around the tunnel. The gantry contains the detectors and a portion of the processing equipment. Signals from the gantry are fed into a computer system where the data is then processed to produce images.
Detectors on the detector rings encircling the patient detect the gamma rays, one on either side of the patient, and the time at which they were detected. Therefore, when two detectors on opposite sides of the patient have detected gamma rays that occurred within some time window of each other, it is safe to assume that the positron-electron interaction occurred somewhere along the line connecting the two detectors. If the detectors that detected the pair of gamma rays are located on the same ring, the coincidence plane, which is a transaxial plane, is called a direct plane. If the detectors are located on different rings, the coincidence plane, which is an oblique plane, is called a cross plane.
By histogramming the detected occurrences based on these lines of response (LOR), a pattern that uniquely describes the distribution of radioactivity is formed. The array in which the histogram is formed is typically called a sinogram. An image of the isotope distribution can be formed from these sinograms using any number of techniques that have been described in the prior art. However, the image that is produced is inaccurate due to several factors. As the gamma rays pass through the patient's body (and other objects, such as the patient bed), they are attenuated and scattered. Additionally, each gamma ray detector has a different response. All of these factors produce either noise or artifacts. Methods for correcting these effects are described in the prior art.
Positron emission tomography is one of the medical imaging modalities for which the transition from two-dimensional to three-dimensional acquisition has been most successful. Following pioneering work in the 1980s, the development after 1989 of multi-ring scanners equipped with retractable septa has led to the present widespread utilization of volume PET-scanners. These scanners have an open, collimator-less cylindrical geometry, which allows the measurement of coincidences between all pairs of detectors on the cylindrical surface.
Data collected in the transaxial or direct plane and in the oblique planes is three-dimensional (3D) data. These 3D data approximate line integrals of the radioactive tracer distribution along LORs which are not restricted to lie within transaxial planes. This is in contrast with the two-dimensional (2D) data acquired when the scanner is operated in 2D mode, in which the data collected is limited to LORs in the transaxial planes. The transition from 2D acquisition to 3D acquisition leads to a significant improvement of the scanner sensitivity, due to the increased number of measured LORs and to the elimination of the detector's shadowing by the septa.
Usually, 3D PET data are reconstructed using a reprojection algorithm (3DRP), which is a 3D filtered-backprojection (FBP) method obtained by discretizing an analytical reconstruction formula. Owing to the considerable number of LORs measured in 3D mode, it is not surprising that the 3DRP algorithm is much more time consuming than the 2D slice-by-slice FBP used to reconstruct data acquired in 2D mode. A further reason for this increased complexity is that the reconstruction of the 3D image is not decomposed into the reconstruction of a set of independent slices. Other algorithms relying on exact analytical formulae have so far been unable to reduce reconstruction time by factors larger than 2 compared to the 3DRP algorithm. In contrast, significant improvements in the reconstruction speed have been achieved using various combinations of the three following approaches. The first one is the introduction of faster, but often expensive, hardware. The second approach uses a reduced sampling of the 3D data to decrease the number of LORs which must be backprojected. Reduced sampling is achieved by adding groups of adjacent LORs in such a way that the resulting loss of spatial resolution remains acceptable for a given type of study. Finally, the third approach to faster 3D reconstruction is the use of approximate algorithms based on axial rebinning. The Fourier rebinning (FORE) process is one such approximate algorithm. The FORE algorithm is described in “Exact and Approximate Rebinning Algorithms for 3D PET data,” M. Defrise, P. Kinahan, D. Townsend, C. Michel, M. Sibomana, and D. Newport, IEEE Transactions on Medical Imaging, pp. 145-58, 1997.
The advantages of using a continuous axial scanning motion are described in “Implementation of True Continuous Whole Body PET Scanning,” M. Dahlbom, J. Reed, and J. Young, IEEE 2000 Medical Imaging Conference. This paper describes performing a scan by moving the patient bed in small, discrete steps. True continuous movement of the patient bed is described in “Methods for Improving Image Quality in Whole Body PET Scanning,” M. Dahlbom, DC Yu, S. Cherry, A. Chatziioannou, and E. Hoffman, IEEE Transactions on Nucl. Sci., Vol. 39, No. 4, pp. 1079-83, 1992. This second paper describes scanning a continuously moving subject and storing the data in list mode, which is later sorted into sinograms for reconstruction.
Apparatus and methods for processing continuous bed motion, three-dimensional (3D) positron emission tomography (PET) acquisitions based on a parallel/pipelined architecture are provided. As the patient bed crosses predetermined positions, specific portions of the acquired data are inserted into the processing pipeline. At each stage of the pipeline, a different processing step is performed on the data in parallel to the others. One of these stages is the conversion of the 3D data set to a two-dimensional (2D) data set. The final result of the pipeline is a single reconstructed image plane corresponding to the acquired data inserted in the pipeline at an earlier time. As the patient bed moves, new image planes are continually produced in a periodic fashion. At the completion of the acquisition, only the portions of the data not in the pipeline and those remaining in the pipeline have to be processed through the pipeline.
During acquisition, the emission and/or transmission events are received from an acquisition processor, along with information on the current position of the patient bed. These events are histogrammed into a 3D sinogram space based on the current patient bed position. When the patient bed has moved a predetermined amount, the histogramming is shifted based on this amount. With this shift, a portion of the 3D sinogram space is no longer within the histogramming region, which corresponds to the portion of the patient and patient bed that has traversed, and is no longer within, the axial field-of-view of the tomograph. This portion of the 3D sinogram space is transferred to either an attenuation processing process (for transmission data) or a normalization process (for emission data). When normalization has been completed, the normalized emission data is transferred to an attenuation correction process. After attenuation correction has been completed, the corrected data is transferred to the Fourier Rebinning (FORE) process. The FORE process is a conversion of the data from a 3D data set to a 2D data set.
Just as with the histogramming process, when the patient bed has moved a predetermined amount, the FORE processing is shifted a corresponding amount. With this shift, a portion of the 3D sinogram space is no longer within the FORE processing region. This region corresponds to the portion of the patient and patient bed that has traversed, and is no longer within, the axial field-of-view of the tomograph. This portion of the now 2D sinogram space is transferred to an image reconstruction process. After the reconstruction process is completed, the image plane is stored, scatter corrected, and/or displayed. All stages of this parallel/pipelined architecture are operating on data at the same time. However, the data for a given processing stage is different from the data in the other processing stages.
The above-mentioned features of the invention will become more clearly understood from the following detailed description of the invention read together with the drawings in which:
Apparatus and methods for processing continuous bed motion, three-dimensional (3D) positron emission tomography (PET) acquisitions based on a parallel/pipelined architecture are disclosed. A PET scanner has a bed that moves continuously as the patient is being scanned. The data from the scanner is processed as it is acquired, producing an image within a short time after the scanning is completed.
In
The histogram process 104 creates a 3D sinogram space histogram of the emission and/or transmission events received from the acquisition process 102, along with information on the current position of the patient bed. Those skilled in the art will recognize that the bed position information can be either a time signal based on a fixed bed speed or a position signal based on a bed position sensor. The emission events are histogrammed into a 3D sinogram space based on the current patient bed position. When the patient bed has moved a predetermined amount, the histogramming is shifted a corresponding amount. With this shift, a portion of the 3D sinogram space is no longer within the histogramming region, which corresponds to the portion of the patient and patient bed that has traversed, and is no longer within, the axial field-of-view of the tomograph.
The histogram process 104 outputs synchronous data as two data streams 162, 156. The first data stream 162 from the histogram process 104 transfers the contents of a transmission data file created during the histogram process 104 to a transmission/attenuation process 122. The transmission data file contains two-dimension (2D) data. The transmission/attenuation process 122 uses an existing blank transmission data file to create an attenuation data file. The transmission/attenuation process 122 outputs a data stream to both an attenuation correction process 108 and a Mu image reconstruction process 124. The Mu image reconstruction process 124 creates a Mu image data file and outputs a data stream to a scatter correction process 126.
The second data stream 156 transfers the contents of a 3D emission data file created during the histogram process 104. The second data stream 156 transfers the data to a normalization process 106. The normalization process 106 uses an existing normalization file to create a second emission data file. The existing normalization file contains the normalization coefficients. The normalization process 106 outputs a data stream to the attenuation correction process 108.
The attenuation correction process 108 accepts a data stream from the transmission/attenuation process 122 and the normalization process 106. The attenuation correction process 108 creates a sinogram data file and outputs a data stream to a Fourier rebinning (FORE) process 110, which creates an image data file and outputs a 2D data stream to an image reconstruction process 112 and the scatter correction process 126. The FORE process 110 converts the data from a 3D data set to a 2D data set.
The data passing through the FORE process 110 corresponds to the bed movement. After the patient bed has moved a predetermined amount, a portion of the 3D sinogram space is no longer within the FORE processing 110 region. This portion of the 3D sinogram space corresponds to the portion of the patient and patient bed that has traversed, and is no longer within, the axial field-of-view of the tomograph. The output of the FORE process 110, which represents a 2D sinogram space, is transferred to an image reconstruction process 112. After the reconstruction process 112 is completed, the image plane is stored, scatter corrected 126, and/or displayed 114.
The scatter correction process 126 accepts data streams from the image reconstruction process 112 and the Mu image reconstruction process 124. The scatter correction process 126 creates a final image data file and outputs a data stream to the image display process 114.
All stages of the above-described parallel/pipelined architecture are operating on data at the same time. However, the data for a given processing stage is different from the data in the other processing stages. Just as in any parallel/pipelined architecture, each stage of processing must complete processing the current data before accepting new data. Therefore, the data from one stage of processing cannot be sent to the next stage of processing until the next stage has completed processing data from the previous cycle. Thus, the overall speed of processing is determined by the slowest stage of processing. Those skilled in the art will recognize that processing stages can be omitted or additional processing stages (various corrections, such as arc correction, etc.) can be added to the architecture without departing from the spirit and scope of the present invention.
The longitudinal view of
δ=d·Δ·δ where d=(i−j)=0, ±1, ±2, . . . , ±dmax
z=−(L−σ)/2+n·σ/2 where n=(i+j)=|d|, |d|+2, |d|+4, . . . , 2N−2−|d|
where N=number of rings and the ring indices j, i run between 0 and N−1, σ=L/N, Δ·δ=σ/2R is the axial angular sampling and −(L−σ)/2 is the axial coordinate of the center of the first ring. The parameter dmax determines the maximum value of δ in the acquired data.
To gain both memory and reconstruction speed, 3D data is acquired with a reduced axial sampling as shown with the sampling scheme of the following equations:
δ=d·Δ·δ where d=0, ±S, ±2S, ±3S, . . . , ±d′max
z=−(L−σ)/2+n·σ/2 where n=n0, n0+1, n0+2, . . . , 2N−2−n0
where S is an integer parameter called ‘span,’ and n0=max{0, |d|−(S−1)/2}. Each discrete sample (d, n) is obtained by summing the range of LORs about d defined by |d−d|≦(S−1)/2 where d is as defined above. The number of LORs summed in this manner is approximately S/2.
The reduced axial sampling scheme is illustrated in
For example, the acquisition process 102 continuously acquires raw data and outputs data to the histogram process 104. When the histogram process 104 has processed a data set 702a, it outputs that data set 702a to the transmission/attenuation process 122 and/or the normalization process 106, which processes the data and then outputs the data set 702a to the next processing stage. Once the histogram process 104 outputs the data set 702a, the histogram process 104 prepares to output the next data set 702b, which can be output only when the transmission/attenuation process 122 and/or the normalization process 106 has completed its processing of the data set 702a and has completed the transfer of the data set 702a to the next stage. The data sets 702 flow through the parallel/pipelined architecture in this stepwise manner until all the data sets 702 acquired have been processed.
As can be seen in
The oblique coincidence planes 1, 3, 5 are cross planes and the events recorded in these planes are attributed to the space midway between the direct planes 0, 2, 4, 6. Because the cross planes 1, 3, 5 are defined by detectors in adjacent rings, the recorded events are summed. The oblique coincidence planes 7, 9, 10, 12 are second-order cross planes with a plane separation of ±2, and the events recorded in these planes approximately coincide with data recorded by the direct planes 0, 2, 4, 6. The oblique coincidence planes 8, 11 are third-order cross planes with a plane separation of ±3, and the events recorded in these planes approximately coincide with data recorded by the cross plane 3.
The counting efficiency of the cross planes 1, 3, 5, 7-12 is approximately twice that of the direct planes 0, 2, 4, 6 because the cross planes 1, 3, 5, 7-12 acquire data from twice as many detector pairs. To reorient the data acquired from the cross planes 1, 3, 5, 7-12 into axial cross sections, the difference in counting efficiency must be corrected, which is done during the normalization process 106.
The example illustrated in
From the foregoing description, it will be recognized by those skilled in the art that apparatus and methods for real-time three dimensional image reconstruction from data acquired in a positron emission tomograph (PET) has been provided. As the tomograph bed moves continuously through the scanner, the acquired data flows through a processing system with a parallel/pipelined architecture.
While the present invention has been illustrated by description of several embodiments and while the illustrative embodiments have been described in considerable detail, it is not the intention of the applicant to restrict or in any way limit the scope of the appended claims to such detail. Additional advantages and modifications will readily appear to those skilled in the art. The invention in its broader aspects is therefore not limited to the specific details, representative apparatus and methods, and illustrative examples shown and described. Accordingly, departures may be made from such details without departing from the spirit or scope of applicant's general inventive concept.
| Number | Name | Date | Kind |
|---|---|---|---|
| 6410919 | Nickles | Jun 2002 | B1 |
| 6429434 | Watson et al. | Aug 2002 | B1 |
| 6468218 | Chen et al. | Oct 2002 | B1 |
| 6490476 | Townsend et al. | Dec 2002 | B1 |
| Number | Date | Country | |
|---|---|---|---|
| 20030161521 A1 | Aug 2003 | US |
