Full-dose PET image estimation from low-dose PET images using deep learning

Abstract

Reconstruct emission imaging data to generate a low-dose reconstructed image. Apply a standardized uptake value (SUV) transform (30) to convert the low-dose reconstructed image to a low-dose SUV image. Apply a neural network (46, 48) to the low-dose SUV image to generate an estimated full-dose SUV image. Before applying the neural network, filter the low-dose reconstructed image or the low-dose SUV image using a low-pass filter (32). Train the neural network on a set of training low-dose SUV images and corresponding training full-dose SUV images using a loss function having a mean square error loss component (34) and a loss component (36) that penalizes image texture and / or a loss component (38) that promotes edge preservation to transform the training low-dose SUV images to match the corresponding training full-dose SUV images.

Description

Technical Field

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

[0002] In oncology, PET imaging is used to complete tasks such as evaluating the malignancy of lesions and the stage of the disease. In a typical workflow, a radioactive drug is administered to a patient (e.g., intravenously). In some PET imaging workflows, the patient is loaded into a PET imaging scanner and PET imaging data is collected, usually after a waiting time of tens of minutes to an hour. During the waiting period, it is expected that the radioactive drug is preferentially collected in tissues of interest (e.g., malignant lesions with high radiotracer uptake). Therefore, the PET imaging data represents the distribution of the radioactive drug in the patient's body, and thus presents an image of the tissue or organ containing the radioactive drug.

[0003] The use of radiotracers for lesion detection in PET is of concern due to the high levels of radiation to which patients and technicians are exposed during PET scans. Exposure to high levels of radiation increases the risk of cancer. Therefore, it is desirable to reduce the dose of radiotracer used to inject patients to minimize radiation exposure. However, within a given PET imaging data acquisition period, lower doses of radiopharmaceuticals translate into lower total counts; lower total counts in turn translate into higher relative noise in the reconstructed PET image, and finer details may also be lost. The resulting lower overall quality of "low-dose" PET images may lead to misdiagnosis, such as missing smaller lesions and / or misinterpreting the status of detected lesions. Efforts have been made to address the image quality degradation of low-dose PET through post-acquisition image processing techniques. For example, edge-preserving image regularization can reduce noise.

[0004] Efforts have also been made to mitigate the effects of low doses by denoising via deep learning. The relationship between low-dose images and full-dose images is learned by a model. Some examples of this method are described in the following literature: "Deep auto-context convolutional neural networks for standard-dose PET image estimation from low-dose PET / MRI" by Xiang et al. (Neural Computing, Vol. 267, No. 1, pp. 406-416, June 2017); "CT Image Denoising with Perceptive Deep Neural Networks" by Yang et al. (14th International Conference on Full Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine, Xi'an, China, 2017, pp. 858-863); and "Generative Adversarial Networks for Noise Reduction in Low-Dose CT" by Wolterink et al. (IEEE Transactions on Medical Imaging, Vol. 36, No. 12, December 2017).

[0005] "200x Low-dose PET Reconstruction using deepLearning" by Xu Junschen et al. (arXiv, December 11, 2017) describes that positron emission tomography (PET) is widely used in various clinical applications, including cancer diagnosis, heart disease, and neurological diseases. It is described that the use of radioactive tracers in PET imaging has raised concerns due to the risk of radiation exposure. It is described that efforts have been made to reduce the amount of radioactive tracers used in order to minimize this potential risk in PET imaging. However, it is described that reducing the dose leads to a decrease in signal-to-noise ratio (SNR) and information loss, both of which can seriously affect clinical diagnosis. It is described that in addition to this, the pathological regulation of low-dose PET image reconstruction makes iterative reconstruction algorithms difficult. It is described that previously proposed methods are generally complex and slow, but still fail to produce satisfactory results at very low doses. This paper proposes a deep learning method to solve this problem, which adopts an encoder-decoder residual deep network with cascaded skip connections. It is described that experiments show that the proposed method is able to reconstruct low-dose PET images to the quality of standard doses at a dose of only one-hundredth of the dose. It is described that different cost functions for training the model are investigated. It is described that a multi-slice input strategy is introduced to provide more structural information to the network and make it more robust to noise. It is described that an evaluation on ultra-low-dose clinical data shows that the proposed method is able to achieve better results than the state-of-the-art methods and reconstruct images of comparable quality using only 0.5% of the original conventional dose.

[0006] Certain improvements are disclosed below. Summary of the invention

[0007] In some embodiments disclosed herein, an emission imaging data reconstruction device includes an electronic processor and a non-transitory storage medium storing instructions that can be read and executed by the electronic processor to perform an image reconstruction and enhancement process, the image reconstruction and enhancement process including: reconstructing emission imaging data to generate a low-dose reconstructed image (e.g., a gray value image); applying a standardized uptake value (SUV) transformation to convert the low-dose reconstructed image into a low-dose SUV scaled image; and applying a neural network to the low-dose SUV image to generate an estimated full-dose SUV image. In some embodiments, the SUV image may include a lean body SUV (SUL) image.

[0008] In some embodiments disclosed herein, an emission imaging data reconstruction device includes an electronic processor and a non-volatile storage medium, wherein the non-volatile storage medium stores instructions that can be read and executed by the electronic processor to perform an image reconstruction and enhancement process, wherein the image reconstruction and enhancement process includes: reconstructing emission imaging data to generate a low-dose reconstructed image; filtering the low-dose reconstructed image using a low-pass filter; and after the filtering, applying a neural network to the low-dose image to generate an estimated full-dose image.

[0009] In some embodiments disclosed herein, an emission imaging data reconstruction device includes an electronic processor and a non-volatile storage medium storing instructions that can be read and executed by the electronic processor to perform an image reconstruction and enhancement process, the image reconstruction and enhancement process including: training a neural network on a set of training low-dose images and corresponding training full-dose images to transform the training low-dose images to match the corresponding training full-dose images, wherein the training uses a loss function having a mean square error loss component (or other smoothing loss component) and at least one additional loss component; reconstructing emission imaging data to generate a low-dose reconstructed image; and applying the trained neural network to the low-dose image to generate an estimated full-dose image.

[0010] In some embodiments disclosed herein, an emission imaging system includes: an emission imaging device configured to acquire emission imaging data; and an emission imaging data reconstruction device as described in any of the first three paragraphs. A corresponding emission imaging data reconstruction method is also disclosed herein.

[0011] In some embodiments disclosed herein, an emission imaging data processing device includes an electronic processor and a non-volatile storage medium storing instructions that can be read and executed by the electronic processor to perform a neural network training process, the neural network training process including: (i) generating a training full-dose image by reconstructing a training emission imaging data set; (ii) generating multiple training low-dose emission imaging data sets by sampling the training emission imaging data set, including generating the training low-dose emission imaging data sets representing two or more different low doses by sampling different numbers of samples for the training low-dose emission imaging data sets representing the two or more different low doses; (iii) generating multiple training low-dose images by reconstructing each of the multiple training low-dose emission imaging data sets; and (iv) training a neural network on the multiple training low-dose images and the training full-dose image to transform the training low-dose image to match the training full-dose image. The neural network training process may include repeating the generating operations (i), (ii) and (iii) for a plurality of training emission imaging data sets, and the training operation (iv) trains the neural network on the training low-dose images and corresponding training full-dose images generated by the repeating to transform the training low-dose images to match the corresponding training full-dose images. The non-transitory storage medium may also store instructions that can be read and executed by the electronic processor to perform an image reconstruction and enhancement process, the image reconstruction and enhancement process comprising reconstructing emission imaging data to generate a low-dose reconstructed image; and applying the trained neural network to the low-dose image to generate an estimated full-dose image.

[0012] In some embodiments disclosed herein, an emission imaging data reconstruction device includes an electronic processor and a non-transitory storage medium storing instructions that can be read and executed by the electronic processor to perform an image reconstruction and enhancement process, the image reconstruction and enhancement process including: segmenting each image in a training low-dose image set and a corresponding training full-dose image into different anatomical regions; training a neural network for each anatomical region on the resulting segmented portion of the training low-dose image set and the resulting segmented portion of the corresponding training full-dose image to transform the resulting segmented portion of the training low-dose image to match the resulting segmented portion of the corresponding training full-dose image; reconstructing emission imaging data to generate a low-dose reconstructed image; segmenting the low-dose reconstructed image into different anatomical regions; and applying the trained neural network for each anatomical region to the resulting segmented portion of the low-dose reconstructed image to generate an estimated full-dose image. In some embodiments, the different anatomical regions include: an anatomical region including the brain, an anatomical region including the heart, an anatomical region including the liver, and an anatomical region including the pelvis.

[0013] One advantage resides in providing an improved estimate of a full-dose emission (eg, PET or SPECT) image from a low-dose emission image.

[0014] Another advantage resides in providing an estimated full-dose emission image with reduced image texturing degradation from a low-dose emission image.

[0015] Another advantage resides in providing an estimated full-dose emission image with reduced image feature degradation from a low-dose emission image.

[0016] Another advantage resides in providing a full dose emission image from the low dose emission image with an estimate having improved robustness to differences in radiopharmaceutical dose of the low dose image compared to the radiopharmaceutical dose of the training image.

[0017] Another advantage resides in providing improved training data for training a low-dose image enhancement process.

[0018] A given embodiment may provide none, one, two, more, or all of the aforementioned advantages, and / or may provide other advantages, as will be readily appreciated by one of ordinary skill in the art upon reading and understanding the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The invention may take form in various components and arrangements of components, and in 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 invention.

[0020] Figure 1 A positron emission tomography (PET) imaging system with low-dose image enhancement and a training system for training low-dose image enhancement according to embodiments disclosed herein are schematically illustrated.

[0021] Figure 2 Schematically illustrates Figure 1 A suitable embodiment of low-dose image enhancement.

[0022] Figure 3 A method for foreground image and patch extraction used in an illustrative low-dose image enhancement training process is schematically illustrated.

[0023] Figure 4 and Figure 5 Schematically illustrates a deep learning model used in an illustrative low-dose image enhancement training process, which includes an estimator network ( Figure 4 ) and the adversarial discriminator network ( Figure 5 ).

[0024] Figure 6-9 PET images of various parts of the test subject's body are presented. Figure 6-9 In each image, the left column is a low-dose image acquired from a patient who received 1 / 10 of the full radiopharmaceutical dose; the middle column is the full-dose image estimated based on the deep learning model; and the right column is the ground-truth full-dose image of the same subject. Figure 6 An image of the brain is shown; Figure 7 An image of a heart is shown; Figure 8 an image of a liver is shown; and Fig. 9 An image showing the pelvis.

[0025] Fig.10 Presented for Figure 6-9 Distribution of root mean square error (RMSE), mean structural similarity index (MSSIM), and peak signal-to-noise ratio (PSNR) of the low-dose images and estimated full-dose images.

[0026] Fig.11 Schematic diagram of a method for training a Figure 1 Training workflow for low-dose image enhancement.

[0027] Fig.12 Schematic illustration of a specific patient training for a fractionated radiation therapy regimen Figure 1 Training workflow for low-dose image enhancement. DETAILED DESCRIPTION

[0028] Existing methods for applying deep learning to improve low-dose PET images have certain disadvantages. Some such techniques introduce excessive smoothing, resulting in a loss of image texture. This can lead to errors in cancer staging, because metastases in PET images sometimes show up in the image texture. Similarly, edge and feature preservation can also deteriorate. Robustness is another concern: for example, a convolutional neural network (CNN) trained on low-dose images acquired using, for example, 20% of the full dose may work well for patients imaged after receiving that dose, but may work poorly on improved low-dose PET images acquired using other doses (e.g., only 10% of the full dose). In practice, the "ground truth" dose at the time of PET imaging data acquisition may vary due to a variety of factors (e.g., different patient metabolic rates, precise time differences between administration of the radiopharmaceutical and the start of PET imaging data acquisition, errors in the administered dose or the radioactivity of the administered radiopharmaceutical (particularly problematic in cases where the radiopharmaceutical has a short lifetime), etc.).

[0029] More generally, attention needs to be paid to the availability of high-quality training data for training deep learning. Ideally, the training set should include low-dose training images acquired and reconstructed for a population of patients that is very similar to the patients currently undergoing clinical imaging, and corresponding "ground truth" images in the form of full-dose training images acquired and reconstructed for the same set of patients using the same PET imaging scanner and image reconstruction process. It is generally unacceptable to generate such training data by performing low-dose PET imaging and full-dose PET imaging on the same set of patients, as this would result in the patient receiving excessive radiation exposure compared to performing only a single full-dose PET imaging session. An alternative approach is to perform training imaging on imaging phantoms, however, these are only approximations of human anatomy.

[0030] In the method disclosed herein, low-dose PET images (e.g., 1 / 10 dose PET images) are enhanced while preserving edge and structural details by specifically considering edge and structural details in a loss function during training, and maintaining image texture by features specified in the loss function, and introducing an adversarial discriminator network through training. Some aspects of the disclosed method include: (1) applying a Gaussian filter (or more generally a low-pass filter) to the low-dose PET images before inputting them into the neural network, which helps to train by removing certain noise without compromising key structural details; (2) employing a loss function that combines certain features (e.g., gradient and total variation) with a mean squared error (MSE) component and adding an adversarial network to ensure that the estimated full-dose image preserves edge, structural, and texture details; (3) in the case of imaging the whole body (or other large anatomical regions), segmenting the body into different regions and training a low-dose image enhancement neural network for each region to account for the distinct structures and textures that occur between regions; and (4) applying a standardized uptake value (SUV) transformation (e.g., conventional SUV or lean body SUV (i.e., SUL)) to the images to reduce inter-patient variability in training images and enhanced low-dose images. It will be appreciated that these improvements can be applied in various combinations, and that particular embodiments can omit one or more of these improvements while still achieving advantageous benefits.

[0031] In the illustrative embodiment, the enhanced low-dose emission image is a positron emission tomography (PET) image; however, it will be appreciated that the disclosed improvements are generally applicable to other types of emission images as well, such as single photon emission computed tomography (SPECT) images (e.g., acquired using a gamma camera).

[0032] refer to Figure 1 , the illustrative imaging device 2 includes a PET / CT imaging device having a computed tomography (CT) gantry 4 including an X-ray tube and an X-ray detector array (internal components not shown) on a rotating gantry, and a PET gantry 6 including one or more PET detector rings (internal components not shown) for detecting 511keV gamma rays. The CT gantry 4 and the PET gantry 6 have coaxial bores for receiving an imaging subject (e.g., a medical patient), and a patient table or examination table 8 is provided for loading the imaging subject into the appropriate CT gantry or PET gantry. The CT gantry 4 is optional, but it is advantageous to provide it because it can be used to acquire CT images of the subject for evaluating the patient's anatomy, locating organs or other internal features for imaging, generating attenuation maps to correct the PET image for 511keV gamma ray absorption, etc.

[0033] Prior to PET imaging data acquisition, a radiopharmaceutical is administered to an imaging subject (e.g., a medical patient) that includes a positron-emitting radiotracer and has a chemical substance designed to preferentially accumulate in an organ or tissue of interest. After the radiopharmaceutical is administered, there is typically a waiting period during which the radiopharmaceutical is collected in the organ or tissue of interest. During the waiting period, a CT gantry 4 may optionally be used to acquire a scout CT image to axially align the patient's organ or region of interest, and / or to acquire a CT image from which an attenuation map of the patient is generated. Depending on the imaging purpose and possible other clinical factors, PET imaging may be intended to acquire a full-dose PET image or a low-dose PET image. The term "full-dose" PET imaging or similar terms refers to a dose of a radiopharmaceutical that is designed to provide an image with a certain desired image quality, as measured by a metric such as a signal-to-noise ratio (SNR), a total count within a given acquisition time interval, or the like. The term "low-dose" PET imaging or similar terms refers to a dose of a radiopharmaceutical that is less than the full dose for full-dose PET imaging and can be usefully quantified as a fraction or percentage of the full dose. Thus, for example, low-dose PET imaging may use 1 / 10 (i.e., 10%) of the full dose of the radiopharmaceutical, or may use 15% of the full dose, etc. In some embodiments, low-dose PET is defined as using 50% or less of the full dose, but other definitions (e.g., 75% or less of the full dose) may also be used. Performing low-dose PET imaging advantageously reduces radiation exposure to patients who have been injected with or otherwise administered a radiopharmaceutical having radioactivity; low-dose PET imaging also reduces radiation exposure to imaging technicians, nurses, or other medical personnel who are near the patient after administration of the radiopharmaceutical. (Although medical personnel are exposed to much less radiation than the imaged patient during a single PET imaging session, medical personnel participate in many such PET imaging sessions in a given work shift and throughout their careers, making it particularly important to limit radiation exposure to medical personnel during PET imaging). On the other hand, low-dose PET imaging has the disadvantage of a reduction in the total counts collected in a given imaging time, for example, the counts in a given imaging time at 10% of the full dose can statistically be expected to be 1 / 10 of the counts obtained in that given imaging time if the full dose of the radiopharmaceutical is administered to the patient.

[0034] Compared to equivalent full-dose PET imaging, the lower total counts acquired in low-dose PET imaging can be expected to result in reduced image quality. In principle, this can be addressed by increasing the acquisition time, for example, increasing the acquisition time tenfold for 10% low-dose imaging may provide the same total counts as full-dose imaging. However, such a large increase in acquisition time (or indeed any acquisition time increase) is often impractical because PET imaging laboratories are expected to maintain a certain workload, such as imaging a certain number of patients in each work shift. In addition, extended imaging times increase the likelihood of patient movement during imaging data acquisition, which can render the collected imaging data compromised or even unusable. In addition, the radioactivity of the administered radiopharmaceutical decreases over time, wherein the radioactivity of the radiotracer decreases by a factor of two for each half-life time interval of the radiotracer. In the case of a short radiotracer lifetime, this may be another limitation on the feasible imaging data acquisition time. Therefore, it is generally preferred to perform low-dose PET imaging with the same (or at least similar) imaging data acquisition time compared to equivalent full-dose PET imaging.

[0035] After the waiting period, the patient is loaded into the PET gantry 6, thereby starting PET imaging data acquisition. In the imaging data acquisition, the 511keV gamma rays detected by the PET detector of the PET gantry 6 are processed by the following operations: using an energy window to filter out stray radiation (except for 511keV gamma rays), and using a coincidence time window to detect coincident gamma rays, each of which is attributed to a common electron-positron annihilation event. The resulting coincidence counts are collected in a list mode PET imaging data storage device 10. Each coincidence count is defined by a coincidence pair of 511keV gamma rays attributed to a common electron-positron annihilation event and has a defined response line (LOR) that connects two coincident 511keV detection events (so the expected response line is along the electron-positron annihilation event). In time-of-flight (TOF) PET, the time difference between the timestamps of two 511keV gamma rays is also used to locate the event along the LOR. As previously described, the total counts (total coincident 511 keV gamma-ray pairs) in a given time interval will be reduced by an amount commensurate with the dose reduction compared to full-dose PET, e.g., for a 10% low-dose PET, the total counts will be approximately 10% of the total counts collected in a full-dose PET.

[0036] Although described using PET imaging as an example, in SPECT imaging, a radioactive drug is also administered to the patient, but each radioactive decay event emits one or more individual gamma rays that are not collected within the timing coincidence window. Gamma cameras are used for SPECT imaging, where events (single events in the case of SPECT, not coincidence pairs) are spatially confined to response lines defined along the viewing direction of the cell bore using, for example, a cell collimator.

[0037] Continue to refer Figure 1 , a PET imaging data set including coincident counts acquired from a patient is reconstructed by a PET image reconstruction process 12 to generate a corresponding reconstructed PET image. The PET image reconstruction process 12 adopts a suitable reconstruction algorithm, such as an iterative maximum likelihood expectation maximization (ML-EM) algorithm, an ordered subset expectation maximization (OSEM) algorithm, a block sequence regularized expectation maximization (BSREM), a non-iterative reconstruction (e.g., filtered back projection), etc. In the case of low-dose PET imaging, the low-dose image enhancement process 14 disclosed herein is applied to improve image quality in order to partially compensate for the image quality degradation caused by the reduced total counts compared to full-dose PET imaging. Since the low-dose image enhancement process 14 is designed to transform the low-dose reconstructed PET image into an image that more closely mimics the image obtained by full-dose PET imaging, the low-dose image enhancement process 14 generates an enhanced low-dose PET image 16 (also referred to herein as an estimated full-dose PET image). The enhanced low-dose PET image 16 is displayed on (or operatively connected to) a display 20 of a computer or other electronic data processing device 22.

[0038] The low-dose image enhancement process 14 is generated by deep learning performed by a low-dose image enhancement trainer 24. The illustrative deep learning operates on training data generated as follows. The PET imaging data acquisition gantry 6 is used to acquire a list-mode full-dose PET imaging dataset (i.e., a full dose of radiopharmaceutical is administered to a patient and the patient is imaged). The dataset is sampled by a sampler 26 using a random (or pseudo-random) sampling process to generate (i.e., synthesize) a low-dose PET imaging dataset. For example, to generate a 10% low-dose PET imaging dataset, the sampler 26 will extract random conforming samples (without replacement) from the full-dose dataset until the 10% low-dose PET imaging dataset includes 10% of the conforming samples of the (pseudo) randomly selected full-dose PET imaging dataset. In a similar manner, a 15% low-dose PET imaging dataset can be generated by (pseudo) randomly extracting 15% of the conforming samples of the full-dose PET imaging dataset (starting from the original full-dose PET imaging dataset, i.e., replacing those conforming samples extracted for the 10% dataset). In a similar manner, a 20% low-dose PET imaging dataset can be generated by (pseudo)randomly sampling 20% ​​of the conforming samples of the list-mode full-dose PET imaging dataset. If the full-dose dataset is compressed into a sinusoidal format, as is done in some commercial PET scanner configurations, such a sampling method for generating a low-dose training set is not easy to implement, although it is envisioned that a bootstrap method (e.g., Poisson method) for modeling appropriate acquisition statistics is applied to extract the low-dose samples.

[0039] It should also be noted that the sampler 26 can generate more than two or more low-dose PET imaging data sets while simulating the same low dose. For example, two 10% low-dose PET imaging data sets can be generated by the following operations: (1) randomly extracting a sample (without replacement) from the full-dose data set until the first 10% low-dose PET imaging data set is obtained; and then (2) starting again from the original full-dose PET imaging data set, randomly extracting a sample (without replacement) from the full-dose data set again until the second 10% low-dose PET imaging data set is obtained. Since the sampling is random (or pseudo-random), the two synthesized 10% low-dose PET imaging data sets are usually different, that is, contain different parts of the full-dose PET imaging data set.

[0040] It will be appreciated that this method for generating a training low-dose PET imaging data set has great advantages. One, two or more low-dose PET imaging data sets in which the same and / or different low doses are simulated can be generated from a single full-dose PET imaging data set. Each of these low-dose PET imaging data sets is reconstructed by the PET image reconstruction processor 12 to produce a training low-dose PET image; and the original full-dose PET imaging data set is reconstructed by the PET image reconstruction processor 12 to produce a training full-dose PET image, which is used as a "ground truth" image for the training low-dose image enhancement 14. That is, the low-dose image enhancement 14 is trained to transform the training low-dose PET image obtained by the sampler 26 to match the corresponding training full-dose reconstructed PET image. Since the training low-dose imaging data set is extracted from the full-dose imaging data set, the corresponding training full-dose PET image is inherently acquired and reconstructed for the same patient using the same PET imaging scanner 6 and image reconstruction process 12, and can therefore be used as an ideal "ground truth" image for the training low-dose image enhancement 14. These massive, high-quality training PET imaging datasets are synthesized from a single full-dose imaging dataset so that the underlying patients are not exposed to unnecessary radiopharmaceutical tracers.

[0041] Although described with respect to a single patient, it will be appreciated that the training can utilize multiple full-dose PET imaging data sets acquired for multiple (training) patients (wherein each full-dose PET imaging data set is sampled by sampler 26 to produce one or more low-dose PET imaging data sets reconstructed by reconstruction process 12) as well as reconstruction of the original full-dose PET imaging data sets (to provide one or more training low-dose PET image sets and corresponding training full-dose PET images as ground truth).

[0042] Continue to refer Figure 1, the illustrative low-dose image enhancement trainer 24 processes the training low-dose images and the full-dose images as follows. A standardized uptake value (SUV) transform 30 is applied to transform each training low-dose reconstructed image into a training low-dose SUV image, and likewise, a standardized uptake value (SUV) transform 30 is applied to transform each training full-dose reconstructed image into a training full-dose SUV image. A low-pass filter 32 is applied to filter each training low-dose SUV image. In an alternative embodiment, the pre-processing is reversed, i.e., the low-pass filter 32 can alternatively be applied first to filter each training low-dose reconstructed image, and then the SUV transform 30 is applied to the images so filtered. The pre-processed images (pre-processed by the SUV transform 30 and the low-pass filter 32, or alternatively only one of them, or in another illustrative embodiment without any pre-processing) are input to a neural network training process using back-propagation or (one or more) other neural network training techniques to train a neural network on the training low-dose SUV images and the corresponding training full-dose SUV images, thereby transforming the training low-dose SUV images to match the corresponding training full-dose SUV images. The illustrative neural network training employs a multi-component loss function to quantify the match between the transformed low-dose SUV image and the corresponding full-dose SUV image. The illustrative multi-component loss function includes: a smoothing loss function component 34, such as a mean square error (MSE) loss; an image texture preservation loss function component 36, such as a total variation (TV) loss component; an edge preservation loss function component 38, such as a gradient loss component. In some embodiments, the neural network training also employs a second stage in which an additional (i.e., a second) adversarial network loss function 40 is added. These loss functions and some suitable embodiments of neural network training employing these loss functions are also described herein.

[0043] Reference now Figure 2 , describes Figure 1 A suitable embodiment of the low-dose image enhancement 14 is Figure 1 The SUV transform 30 is applied to the reconstructed low-dose image 42 to generate a low-dose SUV image filtered by the low-pass filter 32, and the SUV transform 30 is applied by training and using a multi-component loss function (see Figure 1 ) to generate the enhanced low-dose PET image 16. It will be appreciated that while the SUV transform 30 and the low-pass filter 32 can employ a range of embodiments, the trainer 24 ( Figure 1 )and Figure 1 and Figure 2The low-dose image enhancement 14 should all apply the same SUV transformation 30 and the same low-pass filter 32.

[0044] Return to reference Figure 1 , Figure 1 The various processing components 12, 14, 24, 26 of the PET imaging system of the present invention may be implemented by a computer or other electronic data processing device 22 programmed to perform the disclosed processing operations. For example, the electronic data processing device 22 may include a computer including a microprocessor and auxiliary components (e.g., RAM memory, support circuits / electronic devices, etc.) that are connected to run instructions that can be used to cause the computer 22 to perform these processing operations. The instructions are suitably read from a non-transitory storage medium storing such instructions. The non-transitory storage medium may, for example, include a hard drive or other magnetic storage medium; an optical disk or other optical storage medium; a solid state drive (SSD) or other electronic storage medium; various combinations thereof, etc.

[0045] Already referenced Figure 1 and Figure 2 An overview of a PET imaging system with low-dose image enhancement is provided, and illustrative embodiments of the various components are described in further detail below.

[0046] In more detail, the SUV conversion 30 operates to scale the voxel values ​​of the PET image to SUV values ​​using a scaling factor including at least a body shape metric and a dose metric. For example, the following non-limiting illustrative SUV formula may be employed:

[0047]

[0048] Where i is the index of the voxel in the PET image, v i is the value of voxel i in the image (that is, the activity concentration of the radiotracer in the tissue at voxel i), which is converted to an SUV value, D is the radiopharmaceutical dose, M is the patient's 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. i Calibrate so that v iis the unit of radioactivity per unit volume, for example, MBq / ml or mCi / ml. If body weight in grams is used as the measurement result M of body size, the injected activity D divided by the body weight M (that is, the ratio D / M) gives units of MBq / g or equivalent. This will then give units of SUV in g / ml or equivalent. Since soft tissue is approximately considered to be 1 g / ml, SUV values ​​are usually expressed as unitless parameters. In addition, before acquiring PET imaging data, there is usually a waiting time t indicated for dose uptake after injection of the radiopharmaceutical. This results in a dose decay in the time interval t between the time of the injected dose measurement and the time of the activity concentration measurement in the image (scan time). Therefore, The term is a dose decay scaling factor that is used to account for dose decay over the waiting time t. The SUV formula of equation (1) is merely an example, and other formulas are also known. For example, in equation (1), body shape is captured by body weight M, while in some other formulas, body shape is captured by a body surface area term or lean body mass SUV (SUL) to account for lower radiopharmaceutical uptake by adipose tissue.

[0049] refer to Figure 3 , various image preprocessing can be performed. As just noted, because the pixel values ​​of PET images (quantified as counts or activity concentrations, for example) typically have a large range, a SUV transform 30 is applied to convert the PET images to their SUV scale to facilitate CNN training. (Before or after the SUV transform) the low-dose image is passed through a low-pass filter 32, which in the illustrative embodiment herein is a Gaussian filter with σ=1.5 voxels. The low-pass filter 32 operates to reduce certain noise without losing too much structural detail. Typically, there is a large portion of the background of a PET image that does not contain any information relevant to the diagnosis. Additional optional preprocessing can be performed to isolate the foreground portion so that only the foreground image containing relevant information is used to estimate the denoised estimated full-dose image. As Figure 3 As shown, the foreground image is determined by locating the top left pixel and the bottom right pixel of the bounding box BB that contains the relevant information. In a suitable method, these pixels are defined as the minimum and maximum coordinates such that the SUV of the pixel is greater than 0.2 compared to the SUV of the pixel two coordinates away. In order to reduce computational cost and increase the data for training, the cropped portion of the image is then divided into 16×16 pixel patches that overlap by N pixels, as shown in Figure 3 As shown in the rightmost image of . In the illustrative embodiment of this article, n=2. Patches are extracted from the low-dose image and the full-dose image at the same position, and are finally fed back through the deep learning model.

[0050] refer to Figure 4 and Figure 5 , describes an illustrative embodiment of a neural network and its training. A deep learning model architecture is used to estimate a full-dose PET image from a low-dose PET image. In the illustrative architecture, an estimator network 46 ( Figure 4 ) attempts to estimate the true full-dose image and deceive the adversarial discriminator network 48 ( Figure 5 ), the adversarial discriminator network 48 attempts to distinguish the ground truth full-dose image from the estimated full-dose image. Figure 4 The illustrative estimator network 46 has 4 hidden convolutional layers (labeled as "convolution+elu" layers), each of which computes a two-dimensional (2D) convolution of the previous layer using a learned kernel to extract features from the input. Figure 4 In , the first layer operates on the low-dose SUV image after filtering by a low-pass filter 32. The four "convolution+elu" layers are followed by 4 hidden deconvolution layers (labeled as "deconvolution+elu" layers), which use the learned kernels to compute the 2D transposed convolution of the previous layer. Layer 1 uses a 3×3×1×128 kernel, layers 2-7 use a 3×3×128×128 kernel, and layer 8 uses a 3×3×128×1 kernel. All kernels have a stride of 2 and all hidden layers are followed by elu activations, which can be implemented as:

[0051]

[0052] where x is a feature vector whose elements store the features for a given layer (element-wise comparison with 0). Figure 4 Σ) is used between layers of the same size, where features from the previous layer are added to the features of the next layer. In the last layer, a skip connection is made between the residual image patch R and the input image patch X, which can be defined as:

[0053]

[0054] in, is the estimated “full dose” image patch.

[0055] Now go to Figure 5 , the adversarial discriminator network 48 has 1 hidden convolutional layer (labeled as "Convolution + Tanh Layer"), followed by 1 fully connected layer (labeled as "Fully Connected + Tanh Layer"). In the illustrative example, layer 1 uses a 3×3×1×64 kernel with a stride of 1, and layer 2 uses 16384 hidden units. Both layers are followed by a hyperbolic tangent (tanh) activation, which can be implemented as:

[0056]

[0057] The fully connected layer outputs the logarithm of the patch, which is then passed through the final sigmoid activation, where sigmoid can be implemented as:

[0058]

[0059] This final activation yields the probability that the patch is from the ground truth image.

[0060] In the following, it is described Figure 4 and Figure 5 Illustrative training of neural networks 46 and 48. First, train Figure 4 The estimated full-dose image is relatively close to the corresponding ground-truth full-dose image. Similarly, the ground-truth full-dose image is obtained by Figure 1 The PET image reconstruction process 12 reconstructs the training full-dose PET images based on the full-dose PET imaging data set; since the training low-dose PET images are reconstructed based on a randomly selected subset of the full-dose PET imaging data set extracted by the sampler 26, the training full-dose PET images extracted therefrom to form each training low-dose PET image are used as the ground truth images, that is, the images ideally reconstructed based on the low-dose PET imaging data set. This completes the Figure 4 The estimator network 46 is trained separately, so that when it is subsequently introduced Figure 5 When adversarially adversarially adversarially trained discriminator network 48 is used, it learns features (e.g., texture) in addition to structure and pixel values ​​to distinguish between the ground truth full-dose images and the generated full-dose images.

[0061] The loss function to be minimized before introducing the adversarial network 48 is a weighted sum of the MSE between the estimated "full dose" image patch and the true full dose image patch and various image features expected in the final estimate. This can be implemented as a mean square error (MSE) loss component (e.g., Figure 1 ), a total variation (TV) loss component (e.g., Figure 1 The image texture preservation loss function component 36 of the embodiment) and the loss component of the MSE of the gradient (for example, Figure 1 ), which are weighted by weights w1, w2 and w3 respectively:

[0062]

[0063] Where N is the number of tiles (see Figure 3 ), θ represents the learned parameters (i.e., kernel and bias values), represents the estimated “full dose” tablet, Y i represents the true full-dose patch, j represents the pixel for a given patch, and and Represent the estimated gradients of the patch in the horizontal and vertical directions respectively.

[0064] The TV loss function component 36 quantifies the total variation of the estimated patch. This term is maximized (subtracted) in the multi-component loss function of equation (6) to reduce the smoothing effect caused by averaging in the MSE loss function component 34. This ensures that the estimated image maintains the texture and edge details found in the low-dose image. The gradient MSE loss function component 38 quantifies the MSE of the gradient between the estimated full-dose image patch and the true full-dose image patch. This term is minimized so that the structural component of the estimated image is as similar as possible to the true full-dose image.

[0065] In an illustrative embodiment, the ADAM optimization algorithm for deep learning is adapted for training with a learning rate equal to 0.001 and L1 regularization applied to the kernel Figure 4 The estimator network 46 of . The weight values ​​are optimized by changing one weight at a time while keeping the other two weights unchanged. In the illustrative actual implementation of the optimization described herein, the weights that achieve the lowest loss are empirically determined to be w1=1, w2=0.00005, and w3=0.075. The scale of each item is different.

[0066] After 100 generations of training, the estimator network 46 converged and introduced Figure 5 The adversarial network 48 is trained. At this time, the adversarial loss caused by the estimated image is incorporated (e.g., Figure 1 ), and the loss becomes:

[0067]

[0068] Wherein, L(θ) is the multi-component loss function given in equation (6), and the second (adversarial) loss function 40 is added to the fourth weight w4 to obtain the loss function L * (θ). In equation (7), Item indicates Figure 5 The probability that the patch predicted by the adversarial discriminator network 48 comes from the real image, z i represents the true label of the patch (1 = real image, 0 = generated image), and w4 = 0.1. This term is the cross entropy loss due to the estimated image. The adversarial (second) loss function term 40 is maximized in the loss function of equation (7) so that Figure 4 The estimator network 46 learns how to deceive Figure 5 The discriminator network 48 increases (that is, the error rate of correctly distinguishing real images from generated images increases). When the adversarial (second) loss function is included to form the loss L * (θ), the learning rate is reduced by one-tenth, allowing the estimator network 46 to learn finer details (e.g., texture) without changing the already learned structure and edge details.

[0069] In the actual test performed, the low-dose image enhancement method disclosed in this article was tested on PET image slices (from brain to legs) with 1 / 10 of the given full dose, and compared with the corresponding real full-dose image slices for a single patient. Due to the limited availability of low-dose PET medical data and full-dose PET medical data, only one patient was used. The patient had 482 slices, and all images were 288×288 pixels, where the size of the isotropic voxel was 2mm. The dataset was divided into 4 regions, each with 100 slices, and the low-dose image enhancement process was trained for each region14. Since different regions of the body have distinct textures and structures, the entire body image was divided into four parts to help with training. The dataset for each region was randomly divided into training and test sets, where approximately 70% of the slices were used for training and 30% of the slices were used for testing. A total of 272 slices (24495 patches) were used for training. The root mean square error (RMSE), mean structural similarity index (MSSIM) and peak signal-to-noise ratio (PSNR) between the estimated full-dose image (output by the low-dose image enhancement 14) and the ground truth full-dose image are used as measures of image quality.

[0070] Figure 6-9 Representative images are presented for four regions (brain images ( Figure 6 ), Heart Image( Figure 7 ), liver images( Figure 8 ) and pelvic images ( Fig. 9 )) to visually compare each area in the Figure 6-9 The image quality is clearly improved between the left column (1 / 10 dose images) and the middle column (estimated full dose images output by low dose image enhancement 14). The close visual similarity between the right column (ground truth full dose images) and the middle column (estimated full dose images) is also evident.

[0071] Low-dose image enhancement was tested on 128 slices (11545 blocks), and the RMSE, MSSIM and PSNR between the estimated "full-dose" image foreground and the true full-dose image foreground, as well as the RMSE, MSSIM and PSNR between the low-dose image foreground and the true full-dose image were calculated. These results are presented in Table 1. It can be clearly seen from the table that the estimated full-dose image is more similar to the ground truth full-dose image than the low-dose image. In addition, the high values ​​of MSSIM and PSNR and the low RMSE of the estimated full-dose image indicate that the image quality produced by the learned low-dose image enhancement process 14 is more comparable to the true full-dose image.

[0072] Table 1

[0073] RMSE MSSIM PSNR 1 / 10 Dose Image 0.2536±0.0699 0.9594±0.0210 32.6045±4.2252 Estimated full dose image 0.1336±0.0473 0.9792±0.0152 37.1720±4.7146

[0074] refer to Fig.10 Image quality differences were further assessed by comparing the distributions of the three metrics for low-dose images and estimated full-dose images. To determine whether these distributions were indeed statistically different, paired 2-sample t-tests were performed on the distributions of RMSE, MSSIM, and PSNR for 1 / 10 dose image slices and estimated full-dose image slices. The null hypothesis was that the distribution means were the same and a significance value of α=0.05 was used. Each of these three tests resulted in p<<0.001. These small p-values ​​for each of the three metrics indicate that the means for 1 / 10 dose image quality and estimated full-dose image quality are indeed statistically different.

[0075] Reference now Fig.11 , Fig.11 An illustrative workflow for collecting training data and using it to train low-dose image enhancement 14 to enhance low-dose PET images acquired for a population is schematically shown. In operation 50, a full-dose PET list mode imaging dataset is acquired for a sample population. To avoid unnecessary radiation exposure to the imaged subjects, operation 50 is typically performed as part of routine clinical practice, for example, acquiring a full-dose imaging dataset from a group of patients representing a target population (e.g., patients with a particular type of cancer; or adult patients undergoing cardiac imaging, etc.), who are undergoing full-dose PET imaging as part of a routine course of clinical treatment. Applying PET image reconstruction 12 ( Figure 1 ) to generate a full-dose "ground truth" PET training image 52. In operation 54, the sampler 26 ( Figure 1 ) The full-dose PET imaging data set acquired in operation 50 is sampled to generate a low-dose PET training image set, each of which is reconstructed to generate a corresponding low-dose PET training image 56. In training operation 58, as described herein with reference to Figure 1-5 As described, the low-dose PET training images 56 and their corresponding full-dose "ground truth" PET training images 52 are used as training data to train the low-dose image enhancement 14. Since the source training data 50 is representative of a population, the low-dose image enhancement 14 so trained is expected to be effective for enhancing low-dose images acquired on patients adapted to the population.

[0076] As mentioned above, in the sampling operation 54, the sampler 26 ( Figure 1 ) The full-dose PET imaging data set acquired in operation 50 is sampled to generate a low-dose PET training image set, each of which is reconstructed to generate a corresponding low-dose PET training image 56. Fig.11 In the illustrative example of , this operation is repeated with different numbers of total counts in the low-dose dataset to simulate different low doses administered to the patient (in the illustrative example of Fig.11 A low-dose imaging dataset is prepared (e.g., 10% of the full dose, 15% of the full dose, and 20% of the full dose). By doing so, it can be expected that the resulting low-dose image enhancement process 14 trained on these training low-dose images at different dose levels (10%, 15%, 20%) will be more robust against dose differences than training using only a single low dose (e.g., only the 15% dose). As previously described herein, low-dose training images at different dose levels are simulated by sampling to select an appropriate total count for the low-dose dataset, for example, to generate a 20% dataset, sampling extracts 20% of the samples from the complete imaging dataset acquired at operation 50; and to generate a 10% dataset, sampling extracts 10% of the samples from the complete imaging dataset acquired at operation 50; and so on.

[0077] Fig.11 The population-level training workflow advantageously provides a trained low-dose image enhancement process 14 that is expected to be broadly applicable to any imaging subject and task (e.g., imaged anatomical region) that falls within the population represented by the training data acquired in operation 50. Alternatively, however, it is contemplated to obtain a low-dose image enhancement process 14 that is specifically trained for a particular individual in a particular situation.

[0078] Reference now Fig.12For example, consider a situation where a particular patient is undergoing an extended treatment regimen involving an initial PET imaging examination and serial follow-up PET imaging examinations. An example of such an extended treatment regimen is fractionated radiation therapy, in which a patient undergoes serial radiation therapy treatments with treatments spaced days or weeks apart. In such a workflow, a high-quality (i.e., full-dose) PET list mode dataset 60 is typically first acquired, reconstructed to generate a full-dose planning image 62, and used in radiation therapy planning 64 to develop and optimize a fractionated radiation therapy plan. Radiation therapy planning 64 may use any suitable radiation therapy planning workflow, such as employing inverse radiation therapy planning to optimize radiation source trajectories, multi-leaf collimator (MLC) settings, the number of radiation treatment fractions, and the like. The developed and optimized radiation therapy plan may be executed over days, weeks, or months as a fractionated radiation therapy plan 66. Between certain radiation therapy delivery fractions, the oncologist may have the patient undergo one or more follow-up PET imaging examinations. The purpose of the follow-up PET imaging examination(s) is to evaluate the effectiveness of the fractionated radiation therapy regimen 66 to date, such as using measures such as tumor growth or reduction, increase or decrease in metastases, and the extent of tumor necrosis (if any).

[0079] Continue to refer Fig.12 For such a follow-up assessment, the PET images may not need to be of as high quality as the full-dose PET images 62 used in the radiation therapy planning 64. Moreover, the patient undergoing radiation therapy has been exposed to a large cumulative radiation dose, and it would be advantageous to employ a low dose of the radiopharmaceutical for the follow-up PET examination. Therefore, the illustrative follow-up PET examination requires the acquisition of a low-dose PET list mode data set in operation 70. It should be understood that operation 70 is a true low-dose PET imaging data acquisition, in which a low dose of the radiopharmaceutical is actually administered to the patient (e.g., perhaps only 10% of the full dose or only 15% of the full dose, etc.). In operation 72, the low-dose imaging data set is reconstructed to generate a low-dose PET image.

[0080] It would be advantageous to now apply the embodiments of the low-dose image enhancement process 14 disclosed herein to improve the image quality of the reconstructed low-dose PET images. To this end, the image enhancement process 14 may be used as previously described with reference to Fig.11 The low-dose image enhancement process 14 described is trained on a population level. However, it is expected that such a trained low-dose image enhancement process 14 will not be as accurate as a low-dose image enhancement process 14 trained specifically for the current patient.

[0081] Therefore, if Fig.12As schematically shown in FIG. 7 , the full-dose planning PET image 62 and its underlying full-dose PET list mode data set 60 are used to train a patient-specific instance of the low-dose image enhancement process 14. Advantageously, these data 60, 62 are already available because they have been acquired for radiation therapy planning 64. Therefore, in operation 74, the sampler 26 ( Figure 1 ) to synthesize one or more training low-dose PET imaging datasets from the full-dose PET list mode dataset 60, and reconstruct the one or more training low-dose PET imaging datasets in operation 76 to generate a training low-dose PET image set. The full-dose planning PET image 62 is used as a "ground truth" full-dose image corresponding to these training low-dose PET images. In the training operation 78, as referred to herein, Figure 1-5 As described above, these low-dose PET training images and their corresponding full-dose "ground truth" PET training images 62 are used as training data to train the patient-specific low-dose image enhancement 14P. Since the source training data 60 is specific to that particular patient, it is expected that the low-dose image enhancement 14P so trained will be particularly effective for enhancing the follow-up low-dose images reconstructed in operation 72 and thereby generating enhanced follow-up PET images 80 for use in evaluating the efficacy of the ongoing fractionated radiation therapy regimen 66. In this way, Fig.12 The patient-specific workflow 82 schematically illustrated in FIG. 8 incorporates fractionated radiation therapy to provide improved follow-up PET images.

[0082] As previously mentioned, while the illustrative embodiments are directed to enhancing low-dose PET images, the disclosed methods are also readily applicable to enhancing other types of low-dose emission images, such as low-dose SPECT images.

[0083] The present invention has been described with reference to the preferred embodiment. Others may think of modifications and substitutions when reading and understanding the above detailed description. It is intended that this exemplary embodiment be interpreted as including all such modifications and substitutions as long as they fall within the scope of the claims and their equivalents.

Claims

1. A transmission imaging data reconstruction device, include: Electronic processor (22); as well as a non-transitory storage medium storing instructions readable and executable by the electronic processor to perform an image reconstruction and enhancement process, the image reconstruction and enhancement process comprising: reconstructing the emission imaging data to generate a low-dose reconstructed image; applying a standardized uptake value SUV conversion (30) to convert the low-dose reconstructed image into a low-dose SUV image; and applying a neural network (46, 48) to the low-dose SUV image to generate an estimated full-dose SUV image; The neural network (46, 48) is trained on a set of training low-dose SUV images and corresponding training full-dose SUV images to transform the training low-dose SUV images to match the corresponding training full-dose SUV images.

2. The emission imaging data reconstruction device according to claim 1, in, The image reconstruction and enhancement process also includes: Prior to applying the neural network (46, 48), one of the low-dose reconstructed image and the low-dose SUV image is filtered using a low-pass filter (32).

3. The emission imaging data reconstruction device according to any one of claims 1 to 2, in, The image reconstruction and enhancement process also includes: The neural network (46, 48) is trained on a set of training low-dose SUV images and corresponding training full-dose SUV images to transform the training low-dose SUV images to match the corresponding training full-dose SUV images.

4. The emission imaging data reconstruction device according to claim 1, in, Each training full-dose SUV image and the corresponding one or more training low-dose SUV images are generated by an operation including the following: reconstructing the emission imaging data set to generate a training full-dose image, and applying the SUV transformation (30) to the training full-dose image to generate the training full-dose SUV image; generating one or more training low-dose emission imaging datasets by sampling the emission imaging dataset; Each training low-dose emission imaging dataset is reconstructed to generate a training low-dose image, and the SUV transformation is applied to the training low-dose image to generate the training low-dose SUV image.

5. The emission imaging data reconstruction device according to claim 4, in, The one or more training low-dose emission imaging data sets include a plurality of training low-dose emission imaging data sets generated using different numbers of samples, whereby different training low-dose emission imaging data sets represent different low doses.

6. The emission imaging data reconstruction device according to any one of claims 1, 4 and 5, in, The neural network (46, 48) is trained using a loss function having a smoothness loss component (34) and a loss component (36) that penalizes image texture.

7. The emission imaging data reconstruction device according to claim 6, in, The loss component (36) of the loss that penalizes image texture includes a total variance loss component.

8. The emission imaging data reconstruction device according to any one of claims 1, 4, 5 and 7, in, The neural network (46, 48) is trained using a loss function having a smoothing loss component (34) and a loss component (38) that promotes edge preservation.

9. The emission imaging data reconstruction device according to claim 8, in, The loss component (38) that promotes edge preservation includes a mean square error of a gradient loss component.

10. The emission imaging data reconstruction device according to any one of claims 1, 4, 5, 7 and 9, in, The neural network (46, 48) comprises an estimator network (46) and an adversarial discriminator network (48) and is trained in the following stages: In a first stage, only the estimator network is trained using a first loss function (34, 36, 38); and In a second stage, a combination of the estimator network and the adversarial discriminator network is trained using a second loss function (34, 36, 38, 40) comprising the first loss function (34, 36, 38) and an adversarial loss function (40).

11. The emission imaging data reconstruction device according to any one of claims 1, 4, 5, 7 and 9, in, The SUV conversion (30) operates to scale voxel values ​​to SUV values ​​using a scaling factor including at least a shape metric and a dose metric.

12. The emission imaging data reconstruction device according to claim 11, in, The scaling factor also includes the dose decay calculated for the dose decay time interval.

13. A transmission imaging system, include: an emission imaging device (6) configured to collect emission imaging data; as well as An emission imaging data reconstruction device according to any one of claims 1 to 12.

14. The emission imaging system according to claim 13, in, The emission imaging device (6) comprises a positron emission tomography (PET) imaging device configured to acquire the emission imaging data including PET imaging data.

Images