Johnny H. Reed

SPMD cluster-based parallel 3D OSEM

This study empirically compares two approaches to parallel 3D OSEM that differ as to whether calculations are assigned to nodes by projection number or by transaxial plane number. For projection space decomposition (PSD), the forward projection is completely parallel, but backprojection requires a slow image synchronization. For image space decomposition (ISD), the communication associated with forward projection can be overlapped with calculation, and the communication associated with backprojection is more efficient. To compare these methods, an implementation of 3D OSEM for three PET scanners is developed that runs on an experimental, 9-node, 18-processor cluster computer. For selected benchmarks, both methods exhibit speedups in excess of 8 for 9 nodes, and comparable performance for the tested range of cluster sizes.

Continuous bed motion acquisition on a whole body combined PET/CT system

The continuous motion of the patient bed during the acquisition of PET data represents an alternative to the standard step and shoot protocol where a sequence of discrete, overlapping steps are performed. This PET acquisition mode, similar to the whole body CT scan in which the patient bed moves continuously through the scanner has significant advantages, including uniform axial signal-to-noise ratio (SNR), elimination of resolution artifacts by sampling continuously in the axial direction and a reduction in noise from detector normalization. We have implemented and evaluated continuous bed motion acquisition on the combined PET/CT scanner (CPS Innovations, Knoxville, TN). Methods: Emission data were acquired in list mode format with the bed moving at a constant velocity. The position of the bed was recorded in a separate file together with the current event count. The list mode data were then rebinned into sinograms and reconstructed using the same routines applied to reconstruct the standard images. Each sinogram corresponded to an axial plane thickness of 2.4 mm. Phantom data were acquired with both the standard discrete, overlapping steps and with continuous movement of the bed. Data were collected for a whole body phantom containing lesions carefully positioned in the overlap region between two bed positions to specifically emphasize the effect of the improved SNR. Results: The continuous movement of the patient bed resulted in a uniform axial SNR, improving lesion detectability in the bed overlap region. The continuous axial sampling allowed the 10mm lesion in the whole body phantom to have 45% improvement in contrast compared to the step and shoot acquisition. Patient studies demonstrated improved detectability due to finer axial sampling and uniform SNR.

Lossless data compression for short duration 3D frames in positron emission tomography

H/sub 2/O/sup 15/ 3D bolus studies and other dynamic 3D protocols in positron emission tomography (PET) require frame durations of five seconds or less as the protocol begins, with required frame durations increasing as the study progresses. A goal in PET acquisition is sustained 3D frame duration of ten seconds, with shorter frame durations allowed for a limited time. Transfer of projection arrays to an acquisition hard disk is a major limit of frame duration. Data compression can be used to increase the effective disk throughput and therefore decrease the sustained frame duration. The authors discuss the use of lossless compression hardware in a modern PET 3D acquisition system. The hardware uses an implementation of Lempel-Ziv compression with an estimated sustained throughput of 20 megabytes per second and compression ratios of 3 to 10 for short duration 3D projections arrays. Compression use can decrease the minimum sustainable frame duration to less than 10 seconds for an ECAT EXACT HR.<<ETX>>

Continuous bed motion data processing for a resolution LSO PET/CT scanner

Continuous whole-body PET scanning, continuous bed motion (CBM) acquisition has a number of advantages over the traditional step-and-shoot (SS) mode. Strengths of CBM include a uniform axial signal-to-noise ratio, continuous sampling in the axial direction that reduces resolution artifacts, reduction of noise from detector normalization, and reduction of sensitivity to small patient movements. This work highlights the acquisition and data handling methodology that was implemented for a series of phantoms and over 40 patient studies acquired on a high resolution, 16-slice LSO combined PET/CT scanner (CPS Innovations, Knoxville, TN). CBM data were acquired in 32-bit listmode with the bed moving at a constant speed of typically 0.6 mm/s to match the acquisition time per plane of the SS mode. CBM data were processed using the novel, virtual scanner concept that can be applied to a scanner of any axial length. For the high resolution scanner, the LSO PET detectors are arranged in a truncated spherical geometry and therefore normalization and geometrical corrections are applied on an event-by-event basis during histogramming of the 32-bit listmode data. Scatter correction is calculated on the entire image volume, in contrast to the SS mode where scatter is estimated for each bed position. The final 3D data set was reconstructed using ordinary Poisson OSEM3D. This paper will present results from phantom studies and compare clinical patient scans acquired in both SS and CBM modes

A digital architecture for routinely storing and buffering the entire 64-bit event stream at maximum bandwidth for every acquisition in clinical real-time 3-D PET: Embedding a 400 Mbyte/sec SATA RAID 0 using a set of four solid-state drives

A new “Stream Buffer” data acquisition architecture is proposed for positron emission tomography (PET). This new architecture significantly improves performance in high-count-rate (e.g. Rubidium-82) clinical 3-D PET. This Stream Buffer concept improves PET by removing several long-running limitations found in current data acquisition architectures. Stream buffering ensures the non-volatile storage of the entire, raw 64-bit PET coincidence event stream. In addition, this buffering benefits on-line, downstream processing - e.g. LOR-to-bin rebinning and histogramming, processing which remains critical to an effective clinical environment. Primarily this new architecture makes use of multiple high-performance solid-state drives (SSD) to form a single, very-high-speed (400 Mbyte/sec) buffer for coincidence event (list-mode) data. For reference, SSD make use of NAND flash chips for storage instead of rotating media. Today, an SSD may readily exceed 100 Mbyte/sec for read/write throughput. Here a set of 4 SATA SSD are configured as an embedded 64-Gbyte RAID 0. A single FPGA implements the striping RAID controller. The resulting 4-channel RAID is expected to have a sustainable, aggregate bandwidth of at least 400 Mbyte/sec. The Stream Buffer concept requires high-speed, time-shared write/read access into/from this RAID. With both read and write accesses each available for sustainable 200 Mbyte/sec throughput, stream buffering improves PET data acquisition in several ways. Once the PET event stream is delivered to the FPGA which controls the embedded RAID - e.g. via 2 Gbps Fibre Channel, none of the event stream data need be lost because of insufficient bandwidth. A non-volatile copy of the raw 64-bit PET event stream data can always be preserved for optional post-acquisition processing whether on-line downstream processing is selected or not. The RAID (read) output proceeds only at the available downstream throughput rate - i.e. fully eliminating the criticality of higher downstream throughput needed to prevent event loss. On-line cardiac and respiratory gating should also benefit.

Beyond list mode: On-line rebinning and histogramming for continuous bed motion in clinical whole-body TOF PET/CT

In this article, methods are described which help move the concept of continuous bed motion (CBM) for TOF PET (i.e. for PET/CT applications) out of the realm of list-mode-only research and into everyday clinical usage. Long proposed for PET, CBM moves the patient horizontally (at a more or less steady rate) through the PET FOV during the acquisition and offers several points of advantage over the more common patient-held-stationary-to-PET-FOV approach. One primary obstacle to frequent clinical application of CBM in PET has been the inability to provide on-line, real-time detector-pair-to-projection-space-bin-address rebinning calculations along with the associated on-line histogramming. Here are presented details for one approach to overcome just such obstacles — significantly allowing the on-line generation of single TOF projection data sets which may each represent an entire whole-body scan. A set of 5 real list-mode data files were collected from a TOF PET/CT — a system which has been outfitted with a patient handling system (PHS or “bed”) which is shown to be largely sufficient for CBM in PET. For an F-18 point source — placed for first a 1cm and then a 10cm transaxial offset, list-mode data was collected for both stationary and horizontal bed motion cases. In addition, a CBM list-mode data set was collected for a custom, F-18, 100cm-long, cylindrical phantom. These data sets were processed in a manner computationally equivalent to that proposed for the realtime, on-line processing case. By comparing stationary to CBM image quality, the resulting analysis strongly suggests these proposed methods will provide real-time CBM processing which is compatible with the needs of clinical-grade TOF PET. As perhaps expected from CBM, the Z-axis FWHM is shown to generally improve over the stationary case via finer axial sampling in the rebinning step. As may be atypical, one specific data point showed a Z-axis improvement of 6.5%. Other examples — which may be more typical — showed little or no Z-axis FWHM improvement when applying CBM.

Clinical time OSEM3D: infrastructure issues

In previous work we compared two parallel algorithms for calculating 3D forward and backprojection on a distributed-memory cluster computer. These two methods were used to develop an implementation of fully three-dimensional ordered subset expectation maximization iterative reconstruction for emission tomography (OSEM3D). It is, however, necessary to embed these computational kernels in an environment that supports efficient data movement and other infrastructural operations, such as process management. Here we briefly describe two particular components of the infrastructure: the I/O subsystem and the service demon. For the I/O subsystem, a fortuitous relationship between the traditional representation of the fully three-dimensional sinogram S/sub 4DO/(/spl theta/,z,/spl phi/,r) and the distributed representation for image space decomposition permits an efficient solution involving minimal data movement. The service demon is similar to others, but contains additional recovery-oriented mechanisms for minimizing mean time to repair. We conclude with performance benchmarks for fully 3D reconstruction for a scanner that produces a significant volume of data.

Tracking coincidence events in pet even when count rates are extremely high: The Lost-Event Tally packet concept

We describe techniques that extend the usefulness of real-time data-handling architectures designed for clinical positron emission tomography (PET)-especially for instances of extremely high (>; 10 M events/s) count rate. As is widely known, Rubidium-82 (82Rb) with a 1.3-min half-life is often used in clinical PET. When used, 82Rb is more often applied for dynamic and/or gated studies-typically with little or no delay between tracer injection and start of acquisition. The use of 82Rb for short-duration-framed studies in clinical PET has its own set of challenges. For example, a “too large” dose may temporarily exceed the acquisition throughput. When such saturation occurs, coincidence events are lost. In the case of 82Rb, such loss is typically short-lived and limited to the count-rate peak. Such saturation also leads to a truncation of the count-rate profile-a truncation that often limits curve fitting essential for an accurate compartmental model across the entire study. The techniques offered here help to resolve this issue. By electronically keeping track of those events that are subject to saturation loss in the acquisition channel, the complete count-rate profile itself is preserved. Such tracking is done in real time with an accurate log of loss quantities added to the list-mode stream. This approach enables the clinician to raise the 82Rb dose without a specific concern over count-rate profile truncation. We describe a special 64-bit nonevent (i.e., “tag”) packet that is automatically inserted into the list-mode stream. The “Lost-Event Tally” tag packet stream enables a full recovery of the rate profile even when event packets are lost. During acquisition, this tag packet is repeatedly inserted into the stream to report the counted event-packet loss (up to 1 048 575) since the previous such tag packet was inserted.

PET gantry simulation: Concepts and methods for inexpensively reproducing the PET data acquisition environment using a single PC with the PDT card

A custom PCI-application hardware card installed in a server-class PC forms a system to aid with PET data acquisition testing. This system: 1) is called the PET Gantry Simulator PC, 2) mimics most random processes as output in a coincidence-event stream from a PET detector array via fiber optics, 3) will primarily aid the development, testing and validation of PET data acquisition systems, and 4) uses a conventional server-class PC running Windows XP. This Gantry Simulator PC contains a conventional RAID disk system for fast read access into large files. In addition, this PC contains a custom-designed PETLINKTM DMA Transceiver (PDT) card with fiber-optic transceivers and PCI interface. The fiber-optic output from this PDT-in-Gantry-Simulator effectively reproduces - in several important ways - the coincidence- detection data-stream output of many types of real PET detector arrays. Significantly, this PETLINK-compatible fiber-optic retransmission of PET data (real or synthetic) can be repeated exactly and endlessly. Typically, real 64-bit PET coincidence-event and tag packet data - i.e. data collected previously in list mode form - are initially loaded onto this RAID. Under control of Windows-based custom application code, this data can be read from the RAID, input to the PDT via the PCI bus, and retransmitted from the PDT via fiber optic cables and at controlled rates. The rates of packet re-transmission can either be set to mimic that of the original PET acquisition or can be arbitrarily modified to conform to other desired scenarios - i.e. constant rate, bolus injections, half-life decay, etc. Note that the various rates of packet re-transmission are typically only updated by the PDT hardware with each passing millisecond. In one operational mode, the PDT card can be set to mimic even the Poisson distribution of the individual time delays between consecutive retransmitted coincidence-event packets. The effective range of packet transmission rates for the Gantry Simulator approaches a low of 1 kHz and a high of 1 to 8 MHz or more.