Method for in-system calibration of the position and orientation of gamma radiation detectors of a positron emission tomograph
The method addresses the inefficiencies of existing calibration methods by using a parameterized model and statistical optimization to align gamma radiation detectors in positron emission tomographs, enhancing image quality and reducing operational complexity.
Patent Information
- Application Number
- PCT/EP2026/063156
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-07
- Filing Date
- 2026-05-06
- Publication Date
- 2026-11-12
AI Technical Summary
Existing methods for calibrating gamma radiation detectors in positron emission tomographs are time-consuming, costly, and inflexible due to the need for precise positioning and complex acquisition protocols, leading to misalignment and reduced image quality.
A method utilizing a parameterized model and statistical optimization to adjust the position and orientation of radiation detectors based on a phantom geometry, eliminating the need for precise point sources and motor stages, allowing for flexible and precise calibration.
Enables efficient, accurate, and cost-effective calibration of gamma radiation detectors without moving parts, improving image quality by aligning detectors to their actual positions and orientations.
Smart Images

Figure EP2026063156_12112026_PF_FP_ABST
Abstract
Description
[0001] Bremen, May 6, 2026
[0002] Our reference: RA 2123-02 WO MSC / MSC / cze
[0003] Applicant / Owner: RWTH Aachen University,
[0004] abbreviated RWTH Aachen.
[0005] Official file number: New registration
[0006] RWTH Aachen University, a public corporation
[0007] Templergraben 55, 52062 Aachen
[0008] Method for in-system calibration of the position and orientation of gamma radiation detectors of a positron emission tomograph
[0009] The invention relates to a method for in-system calibration of the position and orientation of gamma radiation detectors of a positron emission tomograph, wherein the tomograph has several radiation detectors and is configured to reconstruct images, in particular cross-sectional images, of an object to be measured, wherein the reconstruction is based on a data set containing a coincidence distribution C measured by means of the radiation detectors and on the basis of predefined reference positions Pd and orientations Xd; Yd, Zd of the radiation detectors.
[0010] The precise spatial alignment of the detector blocks in PET systems is crucial for obtaining high-quality images through reconstruction. However, mechanical tolerances, temperature fluctuations, and system integration into MRI scanners complicate accurate assembly and lead to deviations of the actual positions and orientations of the radiation detectors from their reference positions and orientations, which are derived from the tomograph's design specifications.
[0011] The deviations from the reference positions and orientations resulting from the misalignment of the detector blocks lead to incorrect positioning of the lines of response (LORs) during reconstruction, where the recorded gamma interactions are consequently assigned to incorrect spatial coordinates. These misassignments impair image quality: artifacts, blurring, and reduced spatial resolution arise because the reconstructed activity distribution no longer corresponds to the actual radioactivity distribution.
[0012] Known methods for calibrating tomographs rely on single-point source measurements combined with specialized acquisition protocols (Pierce et al., 2009, 2012; Abella et al., 2012). These methods utilize known reference values for the annihilation positions, allowing the use of simple regression algorithms to calculate the alignment parameters. However, this approach requires a 3D motor stage setup for precise positioning and rotation of the point source, significantly increasing system complexity, costs, and operational effort. Furthermore, these methods are limited to a restricted number of predefined point source locations and rigid acquisition protocols, limiting flexibility and introducing potential sources of error due to mechanical tolerances or positioning inaccuracies.In known methods, the reference lines (LORs) used to connect coincident annihilation positions must be assigned to the point sources. The point sources must not be too close together. While a motorized position control (motor stage) allows measurements to be taken first at one position and then at another, enabling the assignment of LORs to point source positions over time, known methods primarily require precise point source positions. This means that the positioning of a motor stage within the scanner must first be accurately calibrated. This is time-consuming and difficult.
[0013] US patent 6727502 B1 discloses a method for calibrating a coincidence imaging system comprising multiple radiation detectors. The method involves measuring coincidence radiation events, assigning initial values to fitting parameters, and applying a minimization algorithm to calculate response lines (LORs). A key feature is the optimization of these parameters to extract a correction factor that corrects for the position of the detectors.
[0014] US 10820881 B2 describes a procedure in which calibration is to be based on the comparison of two scans of the same patient tissue.
[0015] US 7655913 B2 proposes a method for correcting detector misalignments in reconstructed images obtained from medical imaging devices with rotating detectors. This method involves calculating detector misalignments with respect to yaw, pitch, and roll errors, as well as gravity-induced deflections. These calculations are intended to be integrated into image reconstruction algorithms to significantly improve image quality.
[0016] Against this background, the invention was based on the objective of proposing a method of the type described above that overcomes the disadvantages described above as far as possible. In particular, the invention was based on the objective of providing a method with which the calibration of radiation detectors can be carried out in a time- and cost-efficient manner, especially with high accuracy.
[0017] The invention solves this problem by proposing a method which comprises the following steps:
[0018] - Providing a parameterized model p«,r, which represents an optimization position Pd° and an optimization orientation Xd°, Yd°, Zd° as a function of the reference position Pd and orientation Xd; Yd, Zd and a deviation from the reference position Pd and orientation Xd; Yd, Zd, - Introducing a phantom geometry as the object to be measured into the tomograph,
[0019] - Determining an expected spatial coincidence distribution L for the phantom geometry using the parameterized model p«,r,
[0020] - Detecting coincidences between one or more pairs of radiation detectors, - Determining gamma interaction positions in the radiation detectors corresponding to the detected coincidences, and
[0021] - Providing a data set wherein the data set represents an actual spatial distribution C of the coincidences for the pair of radiation detectors in their actual position and orientation, and preferably includes time-of-flight data of the coincidences, and
[0022] - Adjusting the optimization position Pd° and optimization orientation Xd°, Yd°, Zd° of the radiation detectors to their actual position Pd* and orientation Xd*, Yd*, Zd* using a metric differentiable according to one or more parameters T^T of the parameterized model / metric.
[0023] In a first advantageous embodiment of the method, the phantom geometry is known, and the method comprises determining an expected spatial coincidence distribution L for the introduced phantom geometry 5 using the parameterized model PR, T, wherein the fitting step—and therein the differentiable metric—comprises an iterative approximation of the expected coincidence distribution L to the actual coincidence distribution ^ by means of a maximum likelihood estimation. In this embodiment, the invention is based on the approach of dispensing with establishing an explicit mapping between measured events and known annihilation positions, but instead applying an optimization that only requires the actual activity distribution of the phantom geometry to be known. This eliminates the need for specific point sources or complex acquisition protocols with 3D motor stages.According to the invention, calibration is carried out by statistical optimization, in which the distribution of the actual coincidences, which are preferably recorded as flight-time-bound count values, and the expected coincidence distribution are adjusted.
[0024] According to the invention, the optimization position and optimization orientation are understood to be the quantity which is to be approximated to the actual position Pd* and orientation Xd*, Yd*, Zd* of the radiation detectors, which are still unknown at the beginning of the process.
[0025] The method according to the invention can be operated with a wide variety of phantom geometries, each offering distinct advantages. While spatially extended phantom geometries with complex shapes enable efficient, motor-free calibration, even with a single measurement, the method can also be performed multiple times, with point source measurements being carried out in each run, for example. Moving the phantom geometry between measurements can be accomplished using motor stages. While the position of the phantom does not need to be precisely known when performing only a single measurement, it is preferable for all measurements to be taken in the same coordinate system when performing multiple measurements. For this purpose, it is preferable if the relative positioning of the phantom geometries to each other is known from measurement to measurement. A motor stage can, for example, be placed anywhere within the scanner and move the phantom in a controlled manner for several measurements.Multiple measurements can enable even higher precision and improved scalability, as can the simultaneous use of multiple sources or distributed source positions. Furthermore, the need for explicit separation of individual sources or a known initial starting position is eliminated, since the method uses a coordinate transformation to optimize detector alignment.
[0026] Overall, the method represents a flexible and precise calibration process that requires no moving parts but can be used with both point sources and extended phantoms. This enables broader applicability in various PET systems and lays the foundation for generalized, efficient, and highly precise in-system calibration. The method is advantageously further developed with the following steps:
[0027] - End the adjustment once a predetermined number of iterations have been performed, and
[0028] - Adopting the optimization position Pd° and optimization alignment Xd°, Yd°, Zd° as a new reference position and reference alignment.
[0029] In preferred embodiments, the number of iterations is in the range of 100 to 500, preferably in the range of 150 to 300.
[0030] Alternatively or additionally, the procedure may include the following steps:
[0031] - Identify, after completion of adjusting the optimization position (P) < / °) und Optimierungsausrichtung Xd°, Yd°, Zd°), derjenigen Iterationsstufe, in welcher die Likelihood-Schätzung maximal war, und
[0032] - Taking over the optimization position (P < / °) und Optimierungsausrichtung Xd°, Yd°, Zd°) jener Iterationsstufe als neue Referenzposition und Referenzausrichtung. Hierdurch wird berücksichtigt, dass nicht jede Iterationsstufe zwangsläufig eine Minimierung der Optimierungsfunktion liefert, so dass in der Auswertung, beispielsweise im Nachgang, diejenigen Iterationsstufen herausgefiltert werden können, in denen besonders gute Annäherungen gelungen sind.
[0033] In a preferred embodiment, the step of adjusting the optimization position Pd° and optimization orientation Xd°, Yd°, Zd° is performed assuming a Poisson error distribution, in particular by applying the Poisson log likelihood model. This model is well-known from conventional image reconstruction and is easily manageable. The logarithmic formulation allows for efficient optimization.
[0034] In a further preferred embodiment, the optimization step is performed using a gradient-based optimizer, preferably a stochastic gradient-based optimizer, and most preferably an algorithm for optimizing first-order stochastic objective functions based on adaptive estimates of lower-order moments (see Diederik P. Kingma and Jimmy Ba. “Adam: A method for stochastic optimization”. In: arXiv preprint arXiv: 1412.6980 (2014)). Stochastic gradient-based optimization does not optimize the entire dataset at once, but rather uses only a subset for estimation. This increases computational and memory efficiency and can promote faster convergence, especially compared to deterministic methods.Preferably, ordered subsets are used in stochastic gradient-based optimization, with which a selection of coincident radiation detectors, in particular crystal pairs, is made in each iteration, representing an angular coverage of the entire perimeter of the recording space in the tomograph.
[0035] In a further preferred embodiment, each radiation detector, and in particular each crystal volume Äi, Ä2 of each radiation detector, is assigned a detection efficiency EM, EM, wherein the expected coincidence distribution L is normalized using the detection efficiencies. This approach is based on the fact that each crystal volume A has an associated detection efficiency EA, which is influenced by factors such as bonding, crystal defects, and positioning deviations. This efficiency represents the probability of detecting a gamma photon when an interaction has occurred within the crystal.
[0036] The detection efficiencies (EM, EM) are preferably determined using a component-based approach (see Theodorakis et al., “A review of PET normalization: striving for count rate uniformity”. In: Nuclear medicine communications 34.11 (2013), pp. 1033–1045) and / or using a fan-sum method. The latter method assumes a symmetrical, cylindrical arrangement of the radiation detectors, thus representing a design-related approximation. Preferably, the raw count fraction per crystal within the range is used directly as an efficiency estimate without additional corrections, as this eliminates the need to consider geometric factors and prior knowledge of the scanner's orientation. In other preferred embodiments, different normalization components can be used that do not require alignment geometry.For future image reconstruction after performing the adaptation of the optimization position Pd° and optimization orientation Xd°, Yd°, Zd° of the radiation detectors to their actual position Pd* and orientation Xd*, Yd*, Zd* according to the invention, the estimated efficiencies can be recalculated and successively used.
[0037] In a further preferred embodiment, to determine the actual coincidence distribution C, several interaction positions within both radiation detectors, in particular within both crystal volumes, are sampled, and an averaged line integral is formed from the sample. This sampling allows the use of larger crystal volumes, and conversely, the larger the crystals and crystal volumes, the greater the advantage becomes. It takes into account that the actual gamma interactions can occur anywhere within the crystal and do not necessarily have to occur in the center of the crystal. The use of sampling is also advantageous if the tomograph has no information on the interaction depth (DOI) for the radiation detectors. Clinical scanners are often currently unable to consider DOI and only provide two-dimensional positional data of the gamma interactions.
[0038] In another preferred embodiment, the expected coincidence distribution L is determined using LOR modeling. This involves calculating line integrals for each crystal and time unit in an attenuation map of the phantom geometry along the entire response line, while weighting the integrals in a tracer distribution of the phantom geometry with a Gaussian TOF kernel (KTOF). This approach relies on the assumption that coincidence events occur in two crystal containers Äi, Ä2 and are assigned to a TOF bin t derived from the time difference of the photon arrivals. The expected count values are proportional to the tracer line integrals, corrected for detection efficiencies and attenuation.The tracer distribution is a spatial representation of radiation emanating from the phantom geometry, known in advance based on the phantom geometry itself—essentially a 3D map of the radioactivity emitted by the phantom geometry. The attenuation map is a spatial representation of the attenuation coefficients of the phantom material, also known in advance based on the phantom geometry, analogous to the phantom geometry.
[0039] In a further preferred embodiment, using a radiation distribution F corresponding to the phantom geometry, the attenuation map A, and the detector efficiencies EM, EM, the expected coincidence distribution L is determined according to the function
[0040]
[0041] obtained, wherein the integration is linear along a path (^) where y( / ) denotes a spatial position along the path ( / ) between two coincident interaction positions p(Xi) and p(Xz) in the radiation detectors, in particular crystal volumes.
[0042] In another preferred embodiment, the actual coincidence distribution C is available as a coincidence histogram for the coincident pair of radiation detectors, in particular crystal volumes, and a count of the detected coincidences per unit of time is carried out according to the following function:
[0043]
[0044] where δ denotes the Kronecker delta. In a further preferred embodiment, the position and orientation of each radiation detector d is defined by a reference coordinate Pd and a rotation matrix Rxd, Yd,zd(Rd), which describes the rotation of the radiation detector around the reference coordinate. In other words, each radiation detector has its own local coordinate system with the origin of the reference coordinate, which is a vector or three-dimensional quantity, and the rotation matrix comprises the three axes of this local coordinate system, so that each detector has six degrees of freedom in which its position and orientation can deviate from the theoretical reference position and orientation.
[0045] In preferred embodiments, the parameterized model is represented with two parameters, where Q is used for the optimization position (Pd°).
[0046] Pd + Td = Pd°,
[0047] where 7^ is a three-dimensional first parameter that specifies a translational deviation from the reference coordinate Pd.
[0048] For the optimization orientation Xd°, Yd°, ^ the following preferably applies:
[0049] Rxd, Yd, Zd (Rd) Xd — Xd°,
[0050] Rxd, Yd,zd (Rd) Yd = Yd°, and nd
[0051] Rxd, Yd, Zd (Rd) Zd = Zd°,
[0052] where Rxd, Yd,zd (Rd) < X\e denotes rotation matrix, Xd, Yd, / the respective local rotation axes of the radiation detectors which intersect at the reference coordinate / ^, and Ä / is a three-dimensional second parameter that specifies an angular deviation between the optimization alignment and the reference alignment.
[0053] To determine the interaction positions in the local detector coordinate system, a lookup table L: N can be used, for example. cr - IR 3, where Ncr denotes the number of crystals per detector block: These positions are transferred to the global coordinate system using alignment parameterization. In a preferred embodiment, the parameterized model PR, T is obtained for a crystal volume X of a crystal er in a radiation detector d with A = (d, er) e [0, Nd - 1] x [0, Ncr - 1] ZU
[0054] p s , r a| PXY Z) = P d ° + L(cr x X d ° + L(cr) y y°+L(cr) z Z d °, where L is a lookup table containing the interaction positions within the radiation detector.
[0055] In another preferred embodiment, the following applies to the optimization:
[0056]
[0057] In another preferred embodiment, the insertion position of the phantom geometry is chosen arbitrarily. This is made possible by the fact that the phantom geometry is known in advance and the expected spatial coincidence distribution is determined as a function of the inserted phantom geometry. This enables calibration with, but also without, precise motion control and supervision of the positioning of the phantom geometry in the imaging space of the tomograph.
[0058] In another preferred embodiment, exactly one measurement is performed to detect the coincidences. Alternatively, several measurements are performed to detect the coincidences, with the phantom geometry being moved between measurements, and the phantom geometry being moved again to a position, the subsequent position preferably being known relative to the initial position of the first measurement. By increasing the amount of data thus obtained, the quality of the statistical measurement can be further improved.
[0059] The invention is described in more detail below with reference to the accompanying figures and the preferred embodiment. The figures show:
[0060] Fig. 1 shows a schematic representation of a positron emission tomograph according to an exemplary embodiment,
[0061] Figs. 2-4 show different virtual representations of the arrangement of radiation detectors in the tomograph according to Fig. 1, and
[0062] Fig. 5 is a schematic representation of the method according to the invention.
[0063] Figure 1 shows a positron emission tomograph 1. The tomograph 1 has a scanning chamber 3, which is configured to accommodate an object 5 to be measured, for example, a phantom geometry. The tomograph 1 has an array of gamma radiation detectors 7. The radiation detectors 7, generally also referred to by the index "d", each have a well-known structure with scintillator crystals A1 and A2 housed within the radiation detectors.
[0064] The object 5 to be measured will regularly have a (not shown) radiation source, which is usually housed in and enclosed within a container. The enclosure may impose physical limitations, but it does not constitute shielding that would restrict the radiation source to directional emission. The radiation source, and with it the phantom geometry, is designed to emit non-directional β radiation. + - To emit radiation.
[0065] Crystals Ai and A2 are configured to register gamma interactions originating from annihilation events of the radiation source of the object 5 under investigation, using generally known methods. When such coincident gamma interactions are registered on crystals Ai and A2, it can be assumed that the triggering annihilation event lies on a reference line LOR connecting the two locations where the registration occurred in crystals Ai and A2. Images can be reconstructed from a very large number of these reference lines LOR using known methods.
[0066] The device further comprises a data processing device 9, which is equipped with a processor (not shown), memory, and other data processing means, and is configured to receive, process, and provide data signals from the radiation detectors 7 of the tomograph 1. In preferred embodiments, the data processing device 9 can also be configured to control the tomograph 1 and be connected to the tomograph for this purpose via a signal conductor.
[0067] The data processing device 9 is further equipped to carry out the calibration procedure described in the general part, which will be discussed in more detail below.
[0068] Figure 2 shows a schematic spatial representation of a plurality of radiation detectors 7. The arrangement shown in Figure 2 corresponds to the arrangement of the respective radiation detectors 7 in a reference position Pd and a reference orientation Rd in, which, in other words, corresponds to the reference model, also called the blueprint, that the data processing device 9 uses as a basis for image reconstruction of the phantom geometry 5. The radiation detectors 7 are positioned equidistantly along a cylindrical envelope and radially aligned with their respective end-face normals. However, the state shown in Figure 2 does not correspond to an actual arrangement of the radiation detectors 7 within the tomograph 1, unlike in Figure 3.The arrangement of the radiation detectors 7 around the recording chamber 3 is shown there, in an exaggerated form, in an initial optimization position Pd° and orientation Rd°, where one, several, or all of the radiation detectors 7 deviate from the reference geometry Pd, Rd according to Fig. 2 by the translational parameter T and / or the uni- or multi-axis rotation R described above. The deviations would be visible if Figures 2 and 3 were superimposed as a projection. The actual positions and orientations Pd*, Rd* are not known, and the optimization parameters Pd°, Rd° are expressed via the parameterized model PR, T described above.
[0069] Starting from the state shown in Fig. 3, the inventive method described above is carried out as shown in Fig. 5 and described in more detail below.
[0070] Following the execution of the procedure, the state shown in Fig. 4 is established, in which the virtual optimization positions Pd° and optimization orientations Rd° and the actual positions and orientations are approximated to each other, these positions being chosen such that the probability with which the optimization position and optimization rotation correspond to the actual position and rotation has been maximized.
[0071] The process 100 shown in Fig. 5 begins in a first process step 101, in which a parameterized model PR, T is provided which defines the optimization position Pd° and the optimization orientation Rd°.
[0072] In step 103, a phantom geometry 5 is introduced into the acquisition chamber 3. The phantom geometry 5 can, for example, be a thin-walled plastic tube filled with a radioactive fluid, or an arrangement of several point radiation sources. Only the spatial shape of the phantom geometry needs to be known and unambiguously described geometrically, as do its attenuation map and tracer distribution.
[0073] For the phantom geometry 5 in the recording space 3, which may be arbitrarily chosen, an expected spatial coincidence distribution L for the introduced phantom geometry 5 is determined in a further step 105 using the parameterized model PR, T. For this purpose, coincidences between one or more pairs of radiation detectors 7, or their crystal volumes Ai, Ä2, in which the gamma interactions are registered, are recorded, and gamma interaction positions in these radiation detectors 7 are determined for the recorded coincidences. Additionally, in a further step 107, a data set is provided from this, wherein the data set represents an actual spatial distribution C of the coincidences for the pair of radiation detectors in their actual position and orientation, and preferably TOF (time-of-flight) data containing coincidences. Steps 105 and 107 can be executed sequentially or simultaneously.
[0074] In a next step 109, the iterative adjustment of the optimization position Pd° and optimization orientation Rd° of the radiation detectors 7 to their actual position Pd* and orientation Rd* is approximated by approximating the expected coincidence distribution L to the actual coincidence distribution C using the maximum likelihood estimation described above in the general part.
[0075] The iterative adjustment continues until, in step 111, a termination condition is detected, whereby the termination condition may consist, for example, of a predetermined number of iterations having been completed, or of a minimization target of the optimization function having been reached, and / or of successive iterations differing only by amounts that reach or fall below a predetermined threshold.
[0076] After the iteration is completed in step 113, step 115 optionally selects which of the iterations achieved the best optimization. In step 117, these parameters are used to assign the optimized position and orientation obtained with these parameters as the new reference position and orientation and make them available to the data processing device 9 for further application.
[0077] The method according to the invention generally leads to a sufficiently accurate calibration of the tomograph after a single run. However, it is possible to repeat the method one or more times. Preferably, in a further step 119, the existence of a termination condition for the method is checked. If termination is desired, the method is completed in step 121. If continuation is desired, in step 123 the phantom geometry 5 is moved to a different location within the acquisition space 3, preferably a location that is known exactly relative to the initial position of the first measurement, and the method 100 is repeated with the new position from step 105 onwards. The method can, for example, be carried out from step 105 onwards based on a data set supplemented by the further measurement, or dedicated data sets can be created.
[0078] The optimization function above
[0079]
[0080] For example, in the embodiment for 200 iterations, the optimization can be carried out using the Adam optimizer, see Diederik P. Kingma and Jimmy Ba. “Adam: A method for stochastic optimization”. In: arXiv preprint arXiv:1412.6980 (2014), preferably sampling a random sample of 5 million crystal volume combinations and TOF bins. For each combination, for example, 5 positions within the crystal boundaries of both crystals can be selected, and averaged line integrals can be used for optimization.
[0081] Learning rates of 0.1 and 1 for R and T, respectively, can be assumed. For determining the expected coincidence distribution L, a voxel size of 0.125 mm can be used, for example, for the tracer distribution and attenuation map. The optimization can be implemented in Python using the PyTorch framework; see Adam Paszke et al. “PyTorch: An imperative style, high performance deep learning library”. In: Advances in neural information processing systems 32 (2019).
[0082] The Parallelproj package can be used for fast line integral calculations on GPUs using the Joseph method (see Georg Schramm and Kris Thielemans, “PARALLELPROJ - an open-source framework for fast calculation of projections in tomography”. In: Frontiers in Nuclear Medicine 3 (2024), p. 1324562; see Peter M. Joseph, “An improved algorithm for reprojecting rays through pixel images”. In: IEEE transactions on medical imaging 1.3 (1982), pp. 192-196).
[0083] The optimization can be performed on multiple GPUs, utilizing Parallelproj's integrated threading function to balance line integral calculations across the GPUs. Reference symbol list:
[0084] 1 positron emission tomograph
[0085] 3 Recording room
[0086] 5. Object to be measured: Phantom geometry
[0087] 7 radiation detectors d, er
[0088] 9 Data processing device
[0089] 100 procedures
[0090] 101. Procedure step: Providing a parameterized model
[0091] 103 Procedure step: Insertion into the tomograph
[0092] 105 Procedure step: Determining an expected spatial coincidence distribution (LJ) 107 Procedure step: Providing the actual spatial coincidence distribution (C) 109 Procedure step: Iteratively adjusting the optimization position and optimization orientation
[0093] Step 111: Check if the termination condition for iterative adjustment exists. Step 113: End the iteration.
[0094] 115 Procedure step: Selection of iteration results
[0095] 117. Procedure step: Adopting the selection as the new reference geometry
[0096] 119 Check whether a termination condition for the proceedings exists
[0097] 121 Termination of the procedure
[0098] 123 Moving the phantom geometry to another position
[0099] λ1, λ2 crystal volumes
[0100] C actual coincidence distribution
[0101] L expected coincidence distribution
[0102] Pd* actual position
[0103] Rd* actual orientation
[0104] Pd Reference position
[0105] Rd Reference Alignment
[0106] Pd° Optimization Position
[0107] Rd° Optimization alignment
[0108] PR, T parameterized model
[0109] R rotational parameter
[0110] T translational parameter
[0111] LOR reference line, gamma interaction
Claims
Claims:
1. Method (100) for in-system calibration of the position and orientation of gamma radiation detectors (7) of a positron emission tomograph (1), wherein the tomograph (1) has several radiation detectors (7) and is configured to reconstruct images, in particular cross-sectional images, of an object to be measured, wherein the reconstruction is based on a data set containing a coincidence distribution (6) measured by means of the radiation detectors and on the basis of predefined reference positions (Pd) and orientations (Xd; Yd, Zd) of the radiation detectors, characterized in that the method comprises: - Providing (101) a parameterized model (pR,T) which has an optimization position (Pd°) and an optimization orientation (Xd°, Yd°, Zd°) as a function of the reference position ( / ' < / ) und -ausrichtung (Xd; Yd, Zd und einer Abweichung von der Referenzposition (Pd) und -ausrichtung Xd; Yd, ZJ) darstellt, - Introducing (103) a phantom geometry (5) as the object to be measured into the tomograph (1), - Detecting coincidences between one or more pairs of radiation detectors (7), - Determining gamma interaction positions in the radiation detectors (7) for the detected coincidences, and - Providing (107) a data set, wherein the data set represents an actual spatial distribution (C) of the coincidences for the pair of radiation detectors (7) in their actual position (Pd*) and orientation (Xd*, Yd*, Zd*), and preferably includes time-of-flight data of the coincidences, and - Adjusting (109) the optimization position (Pd°) and optimization orientation (Xd°, Yd°, Zd°) of the radiation detectors (7) to their actual position (Pd*) and orientation (Xd*, Yd*, Zd*) using a metric differentiable according to one or more parameters (P, 7) of the parameterized model (p / u).
2. Method according to claim 1, wherein the phantom geometry (5) is known beforehand, and the method comprises: - Determining (105) an expected spatial coincidence distribution (L) for the introduced phantom geometry (5) using the parameterized model (pR,T), and wherein The adjustment step (109) comprises an iterative approximation of the expected coincidence distribution (L) to the actual coincidence distribution (C) by means of a maximum likelihood estimation.
3. Method according to claim 1 or 2, comprehensively the steps: - Terminate (113) the adjustment once a predetermined number of iterations have been performed, and - Taking over (117) the optimization position ( / ' < / ') und Optimierungsausrichtung Xd°, Yd°, Zd") als neue Referenzposition und Referenzausrichtung.
4. Method according to one of claims 2 or 3, comprehensively the steps: - Identify (115) after completing the adjustment of the optimization position ( / ' < / ') und Optimierungsausrichtung Xd°, Yd°, Zd ), derjenigen Iterationsstufe, in welcher die Likelihood-Schätzung maximal war, und - Taking over (117) the optimization position ( / ' < / ') und Optimierungsausrichtung (Xd°, Yd°, Zd ) jener Iterationsstufe als neue Referenzposition und Referenzausrichtung.
5. Method according to any one of claims 2 to 4, where the step of adjusting the optimization position ( / ' < / ') und Optimierungsausrichtung (X d °, Yd°, Z d°) assuming a Poisson error distribution, in particular using the Poisson log likelihood model.
6. Method according to any one of claims 2 to 5, wherein the fitting step (109) is performed using a gradient-based optimizer, preferably a stochastic gradient-based optimizer, particularly preferably an algorithm for optimizing first-order stochastic objective functions based on adaptive estimates of lower-order moments.
7. Method according to any one of claims 2 to 6, wherein each radiation detector, in particular each crystal volume (λ1, λ2) of each radiation detector, is assigned a detection efficiency (Eλ1, Eλ2), and wherein the expected coincidence distribution (L) is normalized using the detection efficiencies.
8. Method according to claim 7, where the detection efficiencies (Eλ1, Eλ2) are determined using a fan sum method.
9. Method according to any of the foregoing claims, wherein to determine the actual coincidence distribution (C) several interaction positions within both radiation detectors, in particular both crystal volumes, are sampled and an average line integral is formed from the sample.
10. Method according to any one of claims 2 to 9, where an LOR modeling is carried out to determine the expected coincidence distribution (L) by calculating line integrals for each crystal and time unit in an attenuation map of the phantom geometry along the entire response line, while in a tracer distribution of the phantom geometry the integrals are weighted with a Gaussian TOF kernel (KTOF).
11. Method according to claim 10, where, using a radiation distribution (F) corresponding to the phantom geometry, the attenuation map A, and the detector efficiencies (Eλ1, Eλ2), the expected coincidence distribution (L) is determined according to the function is obtained by integrating along a straight path ( ), where X4 e i ne spatial position along the path ( ) between two coincident interaction positions p(Ai) and / >(2^in the radiation detectors, especially crystal volumes, is denoted.
12. Method according to any of the foregoing claims, wherein the actual coincidence distribution (C) is given as a coincidence histogram for the coincident pair of radiation detectors, in particular crystal volumes, and includes a count of the detected coincidences per unit of time according to the function: where δ denotes the Kronecker Delta.
13. Method according to any of the foregoing claims, wherein the position and orientation of each radiation detector is defined by a reference coordinate (Pd) and a rotation matrix (Rxd,Yd,Zd(Rd)) which describes the rotation of the radiation detector around the reference coordinate.
14. Method according to claim 13, where the following applies to the optimization position (Pd°): Pd + Td = Pd°, where (Td) is a first parameter that specifies a translational deviation, and where the optimization orientation Xd°, Yd°, Zd° holds, Rxd, Yd, Zd (Rd) Xd — Xd°, Rxd,rd,zd (Rd) Yd = Yd°, and nd Rxd, Yd, Zd (Rd) Zd = Zd°, where Rxd,Yd,Zd(Rd) is a rotation matrix, and Xd, Yd, Zd specify the respective local rotation axes of the radiation detectors that intersect at the reference coordinate Pd, and Rd is a second parameter that specifies an angular deviation between the optimization alignment and the reference alignment.
15. Method according to claim 14, where the parameterized model (pR,T) for a crystal volume (λ) of a crystal (cr) in a radiation detector (d) is given by pR,T(λ| P, X, Y, Z) = Pd° + L(cr)xXd° + L(cr)yYd°+L(cr)zZd°, where L(cr) is a lookup table with the interaction positions within the radiation detector.
16. Method according to any one of claims 4 to 15, where the following applies to the optimization:
17. Method according to any of the foregoing claims, where the insertion position of the phantom geometry is chosen arbitrarily.
18. Method according to any of the foregoing claims, where exactly one measurement is performed to record the coincidences.
