Methods, non-transitory computer-readable media, and systems for imaging

Abstract

A non-transitory computer-readable medium stores instructions that can be read and executed by at least one electronic processor (181, 182, 20) to perform an imaging method (100). The method includes: reconstructing emission imaging data to generate an emission image of a lesion; converting intensity values ​​of the emission image into at least one standardized uptake value (SUV value) for the lesion; using a recurrent neural network (NN) (28) to process input data to output a SUV correction factor for the lesion, wherein the input data includes at least two of the following: (i) image data including the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in the reconstruction; and controlling a display device (24) to display at least one of the following: (I) the SUV correction factor and (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value.

Description

Technical Field

[0001] The following generally relates to the fields of medical imaging, positron emission tomography (PET) imaging, medical image interpretation, image reconstruction, and related fields. Background Art

[0002] Positron emission tomography (PET) (and more generally, nuclear imaging) is an important tool for early tumor diagnosis. Image reconstruction in PET is performed by converting pairs of coincident 511 keV photons registered in the PET detector (e.g., annular) gantry into estimates of the underlying radiotracer concentration located inside the patient. In clinical applications - using a combination of a radiotracer with a specific selection of, for example, metabolic or receptor affine ligands - the resulting visual two-dimensional or three-dimensional (2D or 3D) representation of the radiotracer distribution is used to visualize / analyze (in vivo) biological processes linked to specific diseases. In oncology-related diagnostic applications, deviations from "normal" (i.e., physiological) tracer accumulation (resulting in cold spots or hot spots) in the assessed images are used to reveal, for example, locally increased glucose uptake as present in hypermitotic cancerous lesions.

[0003] Therefore, diagnostic PET imaging is useful for identifying and quantifying suspected lesions down to the PET resolution limit, which depends on detector geometry, crystal size, positron range, sensitivity, etc. The underlying physics in PET and the wide range of effects that affect image quality have been well understood and modeled in a variety of reconstruction algorithms, including compensation techniques for inelastic scattering, random coincidences, and even motion, which increases the chances of early tumor detection.

[0004] Quantification of PET imaging, such as lesion standard uptake value (SUV) measurements, facilitates accurate diagnosis and treatment planning. However, in actual clinical studies, lesions in PET images often suffer from quantitative degradation due to partial volume effects. Resolution recovery methods based on system point spread functions can be used in image reconstruction or image reconstruction post-processing to compensate for the loss of lesion quantification due to partial volume effects. However, these resolution recovery methods often complicate the reconstruction algorithm, increase noise propagation in image reconstruction, and potentially produce image overshoot artifacts. In addition, these methods usually cannot fully recover the loss of lesion quantification accuracy and have variable performance from case to case and from lesion to lesion, so that clinical users cannot confidently rely on lesion quantitative measurements in images even when resolution recovery is achieved.

[0005] Depending on the lesion size, the local lesion / background activity ratio of the SUV determined clinically during quantitative tumor assessment and other influencing factors (e.g., noise regularization applied during / after image reconstruction) are often significantly underestimated. In general, these effects tend to enhance the so-called partial volume effects caused by the limited spatial resolution of the scanner and / or the spatial binning of the reconstructed images (i.e., the voxel / spot size used). This leads to increased variability in the outcome analysis not only between patients but also in individual follow-up studies (e.g., during treatment monitoring).

[0006] Known methods to compensate for SUV underestimation typically focus on only a single influencing factor (such as tumor size) and typically use lookup tables generated from dedicated phantom experiments in order to provide individual correction factors. However, PET imaging devices are complex and highly configurable, typically offering features such as multiple image reconstruction modes using different image resolutions, a range of different noise reduction options, etc., which, individually or in combination, may affect the subsequent SUV assessment. The resulting (multidimensional) parameter space is complex and difficult to accommodate using existing methods. Therefore, relevant corrections recommended for reducing inter- and intra-patient variations in tumor staging and treatment monitoring may not be applied or may be incorrectly applied, resulting in incorrect quantitative SUV adjustments.

[0007] New and improved systems and methods are disclosed below. Summary of the invention

[0008] In one disclosed aspect, a non-transitory computer-readable medium stores instructions that can be read and executed by at least one electronic processor to perform an imaging method. The method includes: reconstructing emission imaging data to generate an emission image of a lesion; converting intensity values ​​of the emission image into at least one standardized uptake value (SUV value) for the lesion; using a recurrent neural network (NN) to process input data to output a SUV correction factor for the lesion, wherein the input data includes at least two of the following: (i) image data including the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in the reconstruction; and controlling a display device (24) to display at least one of the following: (I) the SUV correction factor and (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value.

[0009] In another disclosed aspect, an imaging system includes an image acquisition device configured to obtain emission imaging data of a patient. At least one electronic processor is programmed to: reconstruct the emission imaging data to generate an emission image of a lesion; convert intensity values ​​of the emission image into at least one SUV value for the lesion; process input data using a recurrent neural network (NN) to output a SUV correction factor for the lesion, wherein the input data includes at least three of the following: (i) image data including the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in the reconstruction; generate a confidence value for the SUV correction factor by repeating the processing when one or more values ​​of the input data are changed to perform a sensitivity analysis of the SUV correction factor to the one or more values; and control a display device to display at least one of the following: (I) the SUV correction factor; (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value; and (III) the generated confidence value.

[0010] In another disclosed aspect, an imaging method includes: training a regression NN using training emission images of lesions labeled with true data SUV values; using the NN to process input data to output a SUV correction factor for the lesion, wherein the input data includes at least three of the following: (i) image data including the emission image or a feature vector representing the emission image, (ii) at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in the reconstruction; and controlling a display device to display at least one of the following: (I) the SUV correction factor and (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value.

[0011] One advantage resides in providing more accurate lesion SUV.

[0012] Another advantage resides in obtaining more accurate lesion SUV without requiring complex and uncertain resolution recovery processing.

[0013] Another advantage resides in providing a statistical confidence measure for the SUV of a lesion.

[0014] Another advantage resides in allowing the user to set the reconstruction for optimal visual image quality for detecting small lesions, while also obtaining accurate quantitative values ​​for any particular lesion of interest in the image.

[0015] Another advantage resides in providing improved SUV accuracy using a neural network that is efficiently trained using a large pool of clinical datasets of synthetically inserted lesions with known specifications and mathematical models of lesions.

[0016] Another advantage resides in providing improved SUV accuracy using a neural network that utilizes normal reconstructed PET images to make fast, accurate and adaptive predictions about true lesion SUV to a user at runtime without the need to perform resolution restoration during or after reconstruction.

[0017] Another advantage resides in deriving an estimate of the expected numerical mismatch in measurements between acquired data and training data to directly correct the SUVmax and SUVmean values ​​displayed to the clinician (this improves quantitative SUV accuracy, especially in the case of small lesions), leading to more precise and reproducible results in PET-based tumor staging and treatment response assessment.

[0018] Another advantage resides in providing variability estimates (ie, statistical confidence values) for displayed SUVmax and SUVmean values ​​that are related, for example, to specific lesion shape / size, tracer activity contrast, and spatial image noise properties.

[0019] A given embodiment may provide none of the aforementioned advantages, provide one, two, more, or all of the aforementioned advantages, and / or may provide other advantages as will become apparent to one of ordinary skill in the art upon reading and understanding the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present disclosure may take the form of various components and arrangements of components, and may take the form of various steps and arrangements of steps. The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the present disclosure.

[0021] Figure 1 Diagrammatically illustrating an emission imaging system according to one aspect;

[0022] Figure 2 Shows a depiction Figure 1 An exemplary flow chart of a method of operating a system of

[0023] Figure 3 shows the method for generating Figure 1 An exemplary flow chart of a suitable method for training data for a neural network of a system;

[0024] Figure 4 shows the method for generating Figure 1 Another exemplary flow chart of a suitable method for training data for a neural network of a system; and

[0025] Figure 5 Shows a depiction Figure 1 An exemplary flow chart of a method for applying a neural network of a system. DETAILED DESCRIPTION

[0026] In quantitative PET, intensity values ​​are converted to SUV values. However, errors in the SUV values ​​of small lesions can be introduced by factors such as partial volume effects, systematic blurring, or noise. The first two factors tend to spread the enhanced uptake at the lesion over an artificially larger area, typically (but not necessarily) leading to an underestimation of the mean and / or maximum SUV values ​​of the lesion. The magnitude of the error can depend on many systematic factors, such as voxel size, reconstruction parameters (including post-reconstruction filtering that can affect blurriness), etc., and can also depend on the lesion size (smaller lesions are more strongly affected by partial volume effects and blurring). The mean and / or maximum SUV values ​​of lesions are sometimes used in clinical decision making, making SUV errors clinically significant.

[0027] Methods for SUV correction, such as resolution restoration based on the system point spread function, are known to remedy these error sources. However, resolution restoration can introduce noise, and the resulting correction may result in SUV overshoot, which is also problematic for clinical tasks.

[0028] In embodiments disclosed herein, a recurrent neural network is employed to generate an SUV correction factor for each lesion. The input to the network typically includes a portion of the image containing the lesion and values ​​of reconstruction parameters known to affect the SUV error. Another optional input is a delineated boundary of the lesion. Such input may alternatively be represented by salient features, for example an image may be represented by a suitable feature vector, and a lesion boundary (if used as input) may be represented by a lesion volume and optionally an aspect ratio. Inputting actual image data and boundaries increases the number of inputs, and therefore increases the complexity of the neural network; inputting corresponding salient features reduces the required neural network complexity, at the expense of some possible reduction in accuracy due to information loss.

[0029] The neural network is a regression network that outputs a SUV correction factor (as opposed to a classification network that outputs a classification). Many neural network architectures can be used, preferably with 3 or more layers. The regression neural network may or may not include a feedback path.

[0030] Supervised training of the neural network receives as training data images of lesions labeled with appropriate ground truth data regarding SUV values. In an illustrative training method, experimental PET imaging data from an actual patient is modified by adding "synthetic" lesions of known SUV distribution. This can be done using Monte Carlo simulations of 511 keV gamma ray emissions from synthetic lesions at desired locations in the patient's PET images, preferably adjusted for attenuation using an attenuation map (typically from CT) for the patient's PET images. (Advantageously, it is common in clinical practice to acquire CT images of the patient along with the PET images in order to perform attenuation correction during PET image reconstruction; PET / CT scanners are commercially available and facilitate the acquisition of CT images that are spatially registered with the PET images). The Monte Carlo simulated 511 keV coincident events are added to the original PET data for the patient, and the combined data set (patient PET data with added lesion simulation data) is reconstructed using the desired image reconstruction parameters (and again using the CT images for attenuation correction) to generate an attenuation-corrected reconstructed image converted to SUV units. Thus, the truth data SUV is known for the simulated lesion, and the "measured" SUV data for that lesion is the data of the reconstructed image converted to SUV units. A neural network is trained (e.g., by back-propagation) on the reconstructed image converted to SUV units to optimize the network to output correction factors adjusted to the known truth data SUV of the synthetic lesion. In another approach, PET imaging of a phantom of a "lesion" with a priori known activity levels can be used as training samples with a priori known truth data SUV values.

[0031] In some embodiments disclosed herein, the lesion boundary is delineated manually or by an automated segmentation process or by a semi-automatic method combining these, and the lesion boundary (or a volume or other feature derived from the boundary) is used as input to a neural network that outputs a SUV correction factor. In a variant approach, the lesion size is roughly approximated by fitting a 3D Gaussian probability to the lesion, and this approximated lesion size is used as input to the neural network.

[0032] In other embodiments disclosed herein, lesion boundaries or estimated sizes are not inputs, and in such cases, the neural network can optionally be trained to output both the SUV correction factor and the lesion volume (and / or other lesion size information).

[0033] Another optional output of the neural network is a confidence value characterizing the SUV correction factor. Two factors that influence this confidence are: (1) the confidence of the neural network itself; and (2) the uncertainty in the lesion boundary or estimated size (for embodiments in which this is an input to the network). In one approach, the confidence value can be generated by re-running the neural network for a specified percentage change in the input (e.g., for lesion size, 1.05x lesion size and 0.95x lesion size) and calculating the confidence metric as the + / - 5% change in the SUV correction factor produced for these two neural network runs.

[0034] The neural network can be trained offline (eg, at the factory or in the field during PET system installation or maintenance), and then the trained neural network is applied directly to clinical images to estimate the SUV correction factor.

[0035] Neural networks can be trained for different anatomical regions and lesion types (e.g., lung lesions, prostate lesions), or a single neural network can be trained where the anatomical structure is a parameter input to the neural network. Since the point spread function depends on the type of radiopharmaceutical, this can be another input to the neural network (or alternatively, different neural networks can be trained for different radiopharmaceuticals).

[0036] The output may suitably be a SUV correction factor output by the recurrent neural network, and / or a corrected SUV value calculated by applying the SUV correction to the maximum and / or mean SUV values. In some embodiments, the corrected SUV value(s) are displayed along with the uncorrected SUV value(s) for consideration by the clinician. A confidence value (if calculated) is also displayed.

[0037] The disclosed training method produces a recurrent neural network that takes into account the complex SUV dependencies on various (user-dependent and independent) influencing factors captured by a set of appropriate training data. Thereafter, when provided with matching input from a new PET study performed for clinical diagnosis or assessment of a patient, the trained recurrent neural network automatically derives an estimate of the expected numerical mismatch between the SUV value(s) of a lesion generated by the PET imaging study and the actual lesion activity (i.e., the "true" SUV). This mismatch estimate can be applied to (a) directly correct the SUVmax and SUVmean displayed to the clinician (which improves quantitative SUV accuracy, especially in the case of small lesions, whose SUV values ​​are most adversely affected by confounding factors such as partial volume effects, systematic blurring, and noise). This leads to more accurate and reproducible quantitative results in PET-based tumor staging and treatment response assessment), and (b) optionally also provide variability estimates (i.e., confidence values) for the displayed SUVmax and SUVmean, which are, for example, related to specific lesion shape / size, tracer activity contrast, and spatial image noise properties. With smaller lesion sizes, in particular when approaching the limits of the spatial resolution of the target scanner, SUV errors increase significantly. Therefore, the highest benefit can be expected in this application area (i.e. for "small" lesions). However, the impact of errors in the applied lesion segmentation also increases. Therefore, providing a combination of correction factors together with an indication of how much the proposed / applied correction depends on (one or more) of this (or other) parameters appears to be most attractive or beneficial. Thereby, the clinician can more easily judge the reliability of observed SUV numbers derived from, for example, different PET (e.g., follow-up) scans of the same patient.

[0038] Deep learning neural networks (NNs) can help link between observed lesion quantification (e.g., measured SUV and true lesion quantification, i.e., true SUV) by learning from a large pool of training data on partial volume effects of the imaging chain. For example, due to the partial volume effect, the observed lesion has a reduced SUV and an enlarged size. However, this approach may have feasibility issues, namely when the true lesion size is unknown, and especially when the lesion size is comparable to the partial volume kernel size of the system, the mapping may not be one-to-one in reality. Lesions with different SUVs and different sizes can result in similar observed lesions (e.g., a narrow but tall lesion may end up resembling a wide and short lesion) when the same partial volume effect is applied. When the lesion is large compared to the partial volume kernel, then the SUV can be accurately measured, which makes the NN prediction less useful.

[0039] Although described herein with respect to a PET imaging system, the disclosed methods can be disclosed in other emission imaging modalities in which radiopharmaceuticals are administered to a patient, such as a single photon emission computed tomography (SPECT) imaging system, a hybrid PET / CT or SPECT / CT imaging system, and the like.

[0040] refer to Figure 1 , an illustrative medical imaging system or apparatus 10 is shown. Figure 1 , the system 10 includes an image acquisition device 12. In the illustrative example, the image acquisition device 12 includes a PET gantry of a PET / CT imaging system, which also includes a computed tomography (CT) gantry 13. A patient table 14 is arranged to load a patient into an examination region 16 of the PET and / or CT gantry 12, 13. The PET gantry 12 includes an array of (diagrammatically indicated) radiation detectors 17. As previously mentioned, it is common to acquire both PET imaging data of a patient using the PET gantry 12 and CT images of the patient using the CT gantry 13 in clinical PET imaging, and the attenuation map is derived from the CT image to perform PET image reconstruction.

[0041] The system 10 also includes a computer or workstation or other electronic data processing device 18 having typical components, such as at least one electronic processor 20, at least one user input device (e.g., a mouse, keyboard, trackball, a dictation microphone for dictating radiology reports, etc.) 22, and a display device 24. In some embodiments, the display device 24 can be a separate component from the computer 18. In a common clinical implementation, the at least one electronic data processing device 18 includes a first electronic data processing device 181 used as an imaging device controller (e.g., a PET scanner controller) and a second electronic data processing device 182 used as a radiology workstation. In a typical workflow, a radiologic technician or other medical professional uses the PET controller 181 to operate the PET scanner 12 to acquire PET images and convert them into SUV units (e.g., SUV images and / or average or maximum SUV values ​​for lesions). These imaging study results are stored in a picture archiving and communication system (PACS) 26. PACS may use another term, such as a radiology information system, RIS, etc.

[0042] Image processing also employs a neural network (NN) 28 to correct SUV values, as disclosed herein. NN 28 can be a regression NN that is trained to determine SUV correction factors for lesions in images acquired by the PET gantry 12. Data can be input to NN 28, including (i) image data including an emission image or a feature vector representing an emission image, (ii) at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in reconstruction; (v) the identity of the radiopharmaceutical used in acquiring the emission imaging data; and others. The input data can be processed by NN 28 to determine the SUV correction factor.

[0043] In an illustrative embodiment, a radiologist may operate the radiology workstation 182 to perform reading of PET images, including retrieving and comparing PET images from a current PET study and a previous PET study (from the PACS 26). For example, the previous PET study may have been performed prior to the start of chemotherapy, radiation therapy, or other tumor treatment, while the current PET study may have been performed after such treatment. As another example, during fractionated chemotherapy or radiation therapy, the previous and current PET studies may have been performed at different times during the ongoing fractionated treatment. For example, during a fractionated chemotherapy or radiation therapy, the previous and current PET studies may have been performed at different times during the ongoing fractionated treatment. Figure 1 As shown in , each of the PET controller 181 and the radiology workstation 182 includes one or more display devices 24; the illustrative radiology workstation 182 includes illustratively two displays 24, for example, one for displaying images and the other for displaying radiology reports under drafts or other text information; display tasks may be distributed among the various displays 24 in other ways. As part of the radiology reading process, the radiologist may employ a neural network 28 to call on the SUV correction estimate to correct the SUV value, or provide a proposed SUV correction. Although in this illustrative example, the neural network 28 is applied at the radiology workstation 182, in other embodiments, it may be applied at the PET controller 181 as an adjunct to PET image reconstruction, so that the corrected SUV value or proposed SUV correction factor is uploaded to the PACS 26 along with the imaging study for consideration by the radiologist viewing the study at the radiology workstation 182.

[0044] At least one electronic processor 20 is operatively connected to one or more non-transitory storage media (not shown; such as a disk, RAID or other magnetic storage media; a solid-state drive, a flash drive, an electrically erasable read-only memory (EEROM) or other electronic memory; an optical disk or other optical storage device; various combinations thereof; etc.) that stores instructions readable and executable by the at least one electronic processor 20 to perform the disclosed operations, including performing the imaging method or process 100. In some examples, various portions of the imaging method or process 100 are performed by a technician using a PET controller 181 and / or by a radiologist operating a radiology workstation 182, and may be performed at least in part by cloud processing.

[0045] refer to Figure 2 , an illustrative embodiment of the imaging method 100 is diagrammatically shown as a flow chart. At 102, the image acquisition device 12 (e.g., a PET imaging device) is processed by at least one electronic processor 20 (particularly Figure 1 The PET controller 181 in the illustrative example of is configured or controlled to acquire emission (eg, PET) imaging data. At 104, the PET controller 181 is programmed to reconstruct the PET imaging data into a PET emission image of the lesion.

[0046] At 106, the PET controller 181 is programmed to convert the intensity values ​​of the PET emission image to at least one SUV value for the lesion. (Alternatively, this can be done at the radiology workstation 182). The SUV value can be, for example, the maximum SUV value anywhere in the lesion; the average SUV value over the lesion, or some other representative SUV value that represents the activity of the lesion as a whole. The statistical value (e.g., maximum and / or average SUV) can then be corrected using a SUV correction factor. The same is true in the SUV image, where each voxel (both inside and outside the lesion) can have an SUV value. In other words, the original intensity image is converted to an SUV image by converting the value of each voxel to an equivalent SUV value. Therefore, in an alternative correction sequence, the SUV correction factor can be applied on a voxel-by-voxel basis to correct the SUV value of each individual voxel. Then, if the maximum SUV or average SUV is calculated for the lesion using these corrected per-voxel SUV values, the same (corrected) result should be obtained.

[0047] Converting the intensity value of the emission image into at least one SUV value of the lesion includes converting the intensity value I into a corresponding SUV value according to equation (1):

[0048]

[0049] where i is the index of a voxel of the PET image, vi is the value of voxel i in the image converted to an SUV value (expressed as the radiotracer activity concentration in the tissue at voxel i, e.g., in MBq / mL or equivalent units, calculated from the raw pixel values ​​based on source phantom calibration and pixel volume), D is the radiopharmaceutical dose, M is the patient's body mass (or weight), t is the waiting time between administration of the radiopharmaceutical and PET imaging data acquisition, and t 1 / 2 is the half-life of the radiopharmaceutical. In other contemplated embodiments, at least one SUV value can be determined by, for example, the operations described in the following documents: (1) Boellaard's Standards for PET Image Acquisition and Quantitative Data Analysis (The Journal of Nuclear Mediccine, Vol. 50, p. 5 (Suppl), p. 11S-20S (May 2009)); and (2). Kinahan et al. PET / CT Standardized Uptake Values ​​(SUVs) in Clinical Practice and Assessing Response to Therapy (Semin Ultrasound CT MR 31 (6): 496-505 (2010, p. 12)).

[0050] At 108, the PET controller 181 (or alternatively, the radiology workstation 182, or alternatively a cloud computing resource or (one or more) other remote computers operatively connected to one of these systems 181, 182) is programmed to process input data using the regression NN28 to output an SUV correction factor for the lesion. The input data can include at least two of the following: (i) image data including the emission image or a feature vector representing the emission image, (ii) at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in the reconstruction. In some examples, the input data may include at least three of these options. In other examples, the input data may also include (v) an identification of a radiopharmaceutical used in acquiring the emission imaging data.

[0051] In some embodiments, when the input data includes the size of the lesion, the size of the lesion can be represented as a delineated boundary of the lesion in the emission image. In one example, the radiology workstation 182 is programmed to receive a manual delineation of the lesion boundary via one or more user input devices 22, for example, using a manual feature contour drawing graphical user interface (GUI) for delineating types of targets and organs at risk in radiation therapy planning and similar applications. In another example, the radiology workstation 182 is programmed to delineate the lesion boundary by automatic segmentation of the emission image or by fitting a Gaussian volume to the lesion, etc.

[0052] In other embodiments, the input data does not include the size of the lesion. In these embodiments, NN28 is trained to output the boundary of the lesion and / or the volume of the lesion. The volume of the lesion can be determined by fitting a three-dimensional Gaussian probability distribution to the lesion in the emission imaging data.

[0053] At 110, the PET controller 181 (or alternatively, the radiology workstation 182, or alternatively a cloud computing resource or other remote computer(s) operatively connected to one of these systems 181, 182) is programmed to optionally generate a confidence value for the SUV correction factor. To this end, the PET controller 181 is programmed to apply NN28 to repeat processing of the input data, wherein one or more values ​​of the input data are changed to perform a sensitivity analysis of the SUV correction factor on the one or more values. In some examples, when the input data includes the size of the lesion, the size of the lesion is changed to perform a sensitivity analysis of the SUV correction factor on the size of the lesion.

[0054] At 112, the PET controller 181 (or alternatively, the radiology workstation 182, or alternatively a cloud computing resource or other remote computer(s) operatively connected to one of these systems 181, 182) is programmed to control the display device 24 to display the SUV correction factor. The corrected SUV value generated by applying the SUV correction factor to at least one SUV value may also (or alternatively) be displayed on the display device 24. In addition, other parameters determined (e.g., the boundaries and volume of the lesion, confidence values, etc.) may be displayed on the display device 24.

[0055] refer to Figure 3, an illustrative embodiment of a method 200 for generating training data for training NN28 is diagrammatically shown as a flow chart. NN28 is trained using training emission images of lesions labeled with true data SUV values. At 202, a Monte Carlo simulation of synthetic gamma ray emissions is performed from synthetic lesions inserted into the training emission imaging data. At 204, a combination of synthetic gamma ray emissions and training emission imaging data is reconstructed to generate a training emission image with synthetic lesions in the training emission images.

[0056] Example 1 - Mathematical Model

[0057] NN28 is programmed to predict the true SUV value. A mathematical model (e.g., a 3D Gaussian model) is used to represent the lesion contour and the partial volume blur kernel, which has the property of a one-to-one mapping from the blur model to the peak value, resolution, etc. of the original model for a given blur kernel. NN28 is then trained using a large number of patient datasets with different patient specifications, acquisition times, durations, statistics, etc., where the simulated inserted lesions have known sizes and SUVs. The output of the network is the blur kernel at different spatial locations in the image considering different imaging specifications.

[0058] For each lesion of interest in the clinical image, a blur kernel is predicted at the location of the lesion using the given imaging specifications. A 3D Gaussian model is generated based on the outline of the lesion observed in the image. The true SUV of the lesion is predicted using the blur kernel and the measured lesion outline model. The lesion outline can be modeled based on the Gaussian and total intensity of the lesion measured on the image and / or based on the lesion size measured on the CT image.

[0059] Example 2 - Training NN and Application of NN

[0060] NN28 can be trained using training data, where the volumetric radiotracer distribution is reconstructed from Monte Carlo simulated (or measured) PET scan data, considering various different lesion configurations for a given PET scanner geometry. During (batch) image reconstruction, adjustments accessible to the clinician are simulated (changes to reconstruction, noise regularization or filtering parameters). For each reconstructed image, a known lesion mask is used to determine the regional SUV value. Given the true data and the image-based SUV measure, a matching correction factor is determined. Based on these true data values, the achieved reconstruction output and the set of adjusted parameters, an artificial neural network (ANN) algorithm is trained to learn the relationship between these provided inputs.

[0061] Figure 4An example method 300 for training a NN 28 depicted as a flow chart is shown. At 302, ground truth data is generated using a particle simulation framework (such as GEANT4 / GATE) that is capable of adequately modeling the process from localized positron decay to final coincident photon detection. The geometric model takes into account the relevant scanner properties of the target PET system (e.g., detector module settings). In the simulation, the scanner / detector model is combined with a lesion background model ("phantom") that takes into account, among other factors, different (e.g., radial) lesion locations inside the scanner field of view, different lesion extensions (i.e., different radii if only spherical regions are used; other variations of those parameters characterizing the form / orientation of the applied basic shapes in space), different lesion / background material density ratios (similar to lesions in different body tissue types), different lesion / background activity ratios (matching the expected tracer accumulation range for a given radiotracer, etc.). In some examples, it is possible to combine multiple variations (e.g., different lesion extensions or material density values) in a single phantom slice. For individual relevant radioisotopes (such as 18 F), the simulation mimics the radiation detection process during various acquisition times, including realistic count statistics (and noise patterns) in the detector data, thereby generating individual list mode data sets.

[0062] At 304, each list mode data set is then reconstructed using targeted product reconstruction. Thereby, a (typical for application) range of clinically useful parameter adjustments are applied (all to achieve individual reconstruction results), including different spatial sampling; image resolution settings; different configurations with respect to subset size and number of iterations; different parameter settings related to image pre / post filtering; different configurations affecting (e.g., inherent) noise regularization; etc.

[0063] At 306, the reconstructed activity distribution is multiplied with the corresponding sensitivity map (point by point) and weighted by the acquisition time to derive an estimate for the spatial distribution of detected decay events. The output is spatially cropped around the lesion and (if necessary) resampled to create a standardized predefined input format for NN28. Each input data set considered is also assumed to specify only one lesion.

[0064] At 308, the individual lesion definition masks are applied to the reconstructed image data in order to extract the cumulative, regionally recovered tracer dose. Together with information about the total "injected" tracer dose and phantom weight (e.g., calculated from known subvolumes and their dispensed material), the image-derived lesion SUV (SUVi) is calculated as usual. Alternatively, SUV calibration information extracted from existing scanners (or an average from multiple scanner calibrations) may also be applied.

[0065] At 310, regional ground truth SUV values ​​(SUVg) are calculated for individual lesions based on known phantom settings. Again, SUV calibration information extracted from existing scanners (or averages from multiple scanner calibrations) may also be applied.

[0066] At 312 , for each lesion sample, ie, each set of individually selected parameters (generated at 302 and 304 ) and the resulting reconstructed local tracer distribution (from 306 ), a SUV correction factor is calculated according to 308 and 310 as SUVg / SUVi.

[0067] At 314, the mapping between lesion samples and individual SUV correction factors is trained (e.g., via back-propagation) to a multi-layer NN. Modern NN architectures applied in image processing, such as (deep) convolutional neural networks (CNNs), typically consist of a stack of individual layers, including convolutional layers (convolution-based filtering / feature enhancement), max pooling layers (dimensionality reduction), fully connected layers (feature reorganization), etc. They also include additional layers / techniques (such as dropout and regularization) to maintain the ANN's ability to abstract the feature space based on the training examples seen. Advanced techniques that allow publicly available pre-trained ANN layers to be included in dedicated, specific NNs, such as transfer learning or fine-tuning, can also be considered.

[0068] The (iterative) NN weight adaptation can be performed on a single computer or a computing cluster using a CPU or a GPU. There are different sources of publicly available NN modeling toolkits (TensorFlow, Torch, Caffe / Caffe2, Keras, etc.) that allow the creation of various types of NN structures.

[0069] Once trained, NN28 can be used to output SUV correction factors. When presented with a new set of PET study parameters and pre-segmented volumetric activity distribution data, NN28 responds with SUV correction factors based on the learned parameter space. The correction factors are applied directly to compensate for the estimated SUV errors, or they are displayed along with the uncorrected lesion SUV for subsequent (manual) correction by the clinician. Additional information about the local variability of the estimated SUV correction factors is determined, for example via small-range variations in selected parameters (such as the size and / or shape of lesion segmentations). The SUV variability estimates are displayed (graphically or numerically) along with the corrected or uncorrected SUV assessment results to indicate the "confidence level" for the relevant analysis.

[0070] Figure 5An example method 400 depicted as a flow chart using a trained NN28 is shown. At 402, manual or semi-automatic lesion delineation for a given PET study is performed by a clinician. Using the lesion mask, weighted and lesion-centered region cropping and resampling are applied. At 404, the lesion subvolume is fed into the NN28 along with a set of reconstruction / acquisition parameters, which derives an estimate for an appropriate SUV correction factor via a multi-dimensional projection of the provided input into the learned parameter space. At 406, small (random) shape variations are applied to the segmentation mask used (as well as other parameters). The resulting variations in the SUV correction factor output of the NN28 allow the reliability of that parameter to be estimated.

[0071] The present disclosure has been described with reference to preferred embodiments. After reading and understanding the foregoing detailed description, others may think of modifications and changes. The present invention is intended to be interpreted as including all such modifications and changes, as long as it falls within the scope of the appended claims or their equivalents.

Claims

1. A non-transitory computer-readable medium storing instructions capable of being read and executed by at least one electronic processor (181, 182, 20) to perform an imaging method (100), the method comprising: reconstructing the emission imaging data to generate an emission image of the lesion; converting the intensity values ​​of the emission image into at least one standardized uptake value SUV value for the lesion; Processing input data using a recurrent neural network NN (28) to output a SUV correction factor for the lesion, wherein the input data includes at least two of the following: (i) image data including the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in the reconstruction; and A display device (24) is controlled to display at least one of: (I) the SUV correction factor and (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value.

2. The non-transitory computer readable medium of claim 1, wherein: The input data comprises the size of the lesion represented as a delineated boundary of the lesion in the emission image.

3. The non-transitory computer readable medium of claim 2, wherein: The method further comprises one of the following: receiving a manual delineation of a lesion boundary via one or more user input devices (22); or Lesion boundaries are delineated by automatic segmentation of the emission images.

4. The non-transitory computer readable medium of claim 1, wherein: In said processing, said input data does not include (iii) the size of said lesion, and said processing of said input data using said recurrent neural network NN (28) also outputs a boundary of said lesion and a volume of said lesion; and The method also includes controlling the display device (24) to display the determined boundary and volume of the lesion.

5. The non-transitory computer readable medium of any one of claims 1 to 4, wherein: The method may further comprise one or more of the following: generating a confidence value for the SUV correction factor; and controlling the display device (24) to display the generated confidence value; and The regression NN (28) is trained using training emission images of lesions labeled with true data SUV values.

6. The non-transitory computer readable medium of claim 5, wherein: Generating the confidence value for the SUV correction factor includes: The process is repeated with one or more values ​​of the input data being changed to perform a sensitivity analysis of the SUV correction factor to the one or more values.

7. The non-transitory computer readable medium of claim 6, wherein: In the process: The input data includes (iii) the size of the lesion, and in the iterations, the size of the lesion is varied to perform a sensitivity analysis of the SUV correction factor to the size of the lesion.

8. The non-transitory computer readable medium of any one of claims 1, in, Converting the intensity value of the emission image into at least one SUV value for the lesion includes converting the intensity value I into a corresponding SUV value according to the following formula: Where i is the index of the voxel in the PET image, v i is the value of voxel i in the image converted to SUV value, expressed as the radiotracer activity concentration in the tissue at voxel i and in MBq / mL or equivalent units, calculated from the raw pixel values ​​based on source phantom calibration and pixel volume, D is the radiopharmaceutical dose, M is the patient's body mass or weight, t is the waiting time between administration of the radiopharmaceutical and PET imaging data acquisition, and t 1 / 2 is the half-life of the radiopharmaceutical.

9. The non-transitory computer readable medium of claim 1, wherein: The emission imaging data is positron emission tomography (PET) imaging data and the emission image of the lesion is a PET image of the lesion, and / or The input data further includes an identification of a radioactive drug used in acquiring the emission imaging data.

10. An imaging system (10), comprising: an image acquisition device (12) configured to obtain emission imaging data of a patient; as well as At least one electronic processor (181, 182, 20) programmed to: reconstructing the emission imaging data to generate an emission image of the lesion; converting the intensity values ​​of the emission image into at least one standardized uptake value SUV value for the lesion; Processing input data using a recurrent neural network NN (28) to output a SUV correction factor for the lesion, wherein the input data comprises at least two of the following: (i) image data comprising the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) the size of the lesion, and (iv) reconstruction parameters used in the reconstruction; generating a confidence value for the SUV correction factor by repeating the process with one or more values ​​of the input data being changed to perform a sensitivity analysis of the SUV correction factor to the one or more values; and The display device (24) is controlled to display at least one of: (I) the SUV correction factor; (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value; and (III) the generated confidence value.

11. The imaging system (10) according to claim 10, wherein: The input data includes the size of the lesion represented as a delineated boundary of the lesion in the emission image, the at least one electronic processor (181, 182, 20) being further programmed to determine the delineated boundary by one of: receiving a manual delineation of a lesion boundary via one or more user input devices (22); or Lesion boundaries are delineated by automatic segmentation of the emission images.

12. The imaging system (10) according to any one of claims 10 and 11, wherein: In the processing, the input data does not include (iii) the size of the lesion, and the processing of the input data using the recurrent neural network NN (28) also outputs the boundary of the lesion and the volume of the lesion; and the at least one electronic processor (181, 182, 20) is further programmed to: The display device (24) is controlled to display the determined boundary and volume of the lesion.

13. The imaging system (10) according to claim 10, wherein: In the processing, the input data includes (iii) the size of the lesion, and in the iterations, the size of the lesion is varied to perform a sensitivity analysis of the SUV correction factor to the size of the lesion.

14. An imaging method (100), comprising: Training a recurrent neural network NN (28) using training emission images of lesions labeled with ground truth data normalized uptake values ​​(SUV); Processing input data using the NN to output a SUV correction factor for the lesion, wherein the input data includes at least two of the following: (i) image data including the emission image or a feature vector representing the emission image, (ii) at least one SUV value, (iii) a size of the lesion, and (iv) a reconstruction parameter used in reconstruction; and A display device (24) is controlled to display at least one of: (I) the SUV correction factor and (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value.

15. The imaging method (100) according to claim 14, wherein: The training includes: performing a Monte Carlo simulation of synthetic gamma-ray emissions from synthetic lesions inserted into the training emission imaging data; A combination of the synthetic gamma ray shots and the training shot imaging data is reconstructed to generate a training shot image having the synthetic lesion in the training shot image.

Images