AI-driven PET reconstruction from tissue images

Abstract

Systems and methods include acquiring a computed tomography image of a subject, determining a linear attenuation coefficient map based on the computed tomography image, acquiring positron emission tomography (PET) data of the subject, determining a tissue image of the subject based on the PET data, determining a scattered tissue image based on the tissue image and the linear attenuation coefficient map, and outputting the scattered tissue image. A scatter-corrected tissue image is determined based on the tissue image and the scatter tissue image, and the computed tomography image and the scatter-corrected tissue image are input to a trained neural network to generate a PET image.

Description

Background Art

[0001] Positron emission tomography (PET) generates functional images that represent biological processes occurring in a patient's body (e.g., glucose metabolism, receptor affinity). In PET technology, a radioactive tracer is administered to the patient via intravenous injection, inhalation, oral administration, or direct organ injection. As the tracer travels through the patient's body, it radioactively decays to produce positrons, which eventually encounter electrons and are annihilated.

[0002] The annihilation produces two 511 keV photons that travel in approximately opposite directions. A coincidence is identified when two detector crystals detect the arrival of the two photons within a short time window, indicating that they were produced by the same positron annihilation. Because the two "coincident" photons travel in approximately opposite directions, the positions of the two detector crystals determine the line of response (LoR) along which the annihilation may have occurred. PET data represents each detected annihilation as a LoR between two detector crystals.

[0003] Images are reconstructed from PET data using known algorithms such as filtered back projection (FBP) and ordered subset expectation maximization (OSEM). PET image reconstruction is complex, resource-intensive, and time-consuming. For example, reconstruction typically includes determining and subtracting random coincidence and scatter coincidence from the PET data, attenuation correction, motion correction, system sensitivity correction, and other appropriate steps.

[0004] Advances in PET scanner technology have exacerbated the challenges of PET image reconstruction. Longer axial field-of-view scanners and smaller detector crystals (i.e., finer sampling) have independently increased the number of LoRs and the complexity of correction. Furthermore, dynamic workflow technologies have placed even higher demands on the speed and quality of image reconstruction.

[0005] These challenges are currently managed by reducing (i.e., mashing) the PET data, modifying the reconstruction algorithms to compromise image quality characteristics (such as spatial resolution and noise), and / or using more powerful processing hardware with an associated increase in memory requirements. This hardware is expensive, consumes a lot of power, and generates a lot of heat. Furthermore, these remedial approaches are not infinitely scalable.

[0006] It is desirable for the system to reconstruct appropriate PET images in a resource-efficient manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a block diagram of a system for generating corrected PET images according to some embodiments.

[0008] Figure 2is a flow chart of a process for generating a corrected PET image according to some embodiments.

[0009] Figure 3 The training of a neural network to generate a scattered tissue image based on a normalized tissue image (histoimage) is shown in accordance with some embodiments.

[0010] Figure 4 The generation of neural network training data according to some embodiments is shown.

[0011] Figure 5 Training a neural network to generate a corrected PET image based on a CT image and a corrected tissue image is shown in accordance with some embodiments.

[0012] Figure 6 The generation of neural network training data according to some embodiments is shown.

[0013] Figure 7 is a block diagram of an imaging system according to some embodiments. DETAILED DESCRIPTION

[0014] The following description is provided to enable any person skilled in the art to make and use the described embodiments. However, various modifications will be apparent to those skilled in the art.

[0015] Embodiments utilize a trained neural network to derive PET images from normalized PET tissue images and corresponding anatomical images. Embodiments can thereby generate fully rectified, low-noise, high-spatial-resolution PET images within a practical timeframe using commonly available computing resources.

[0016] The PET tissue image is a simple back-projection of each detected annihilation along the LoR using time-of-flight (ToF) localization. Briefly, list mode data is acquired and normalized based on scanner sensitivity. The list mode data can also be attenuation corrected based on previously acquired anatomical information. A normalized PET tissue image is generated from the normalized list mode data. A scatter tissue image is generated based on the PET tissue image and the corresponding mu map, and the scatter tissue image is used to scatter correct the PET tissue image. The scatter-corrected PET tissue image and an anatomical image (e.g., a computed tomography (CT) image) are then input to a trained neural network to generate a corrected PET image.

[0017] Figure 1is a block diagram of a system 100 for generating corrected PET images, according to some embodiments. The illustrated components of system 100 can be implemented in computer hardware, program code, and / or one or more computing systems that execute such program code as is known in the art. Such a computing system may include one or more processing units that execute program code stored in a memory system. In some embodiments, more than one functional component may be implemented by a single computing system. One or more computing systems may include virtual machines, and one or more computing systems may include cloud-based computing resources that provide on-demand scalability and failure recovery.

[0018] The scanner 110 may include a PET / CT scanner capable of generating PET data 115 using any suitable PET imaging protocol. For each detected annihilation, the PET data 115 provides the positions of the two detector crystals that detected the annihilated photon (i.e., and defines the LoR along which the annihilation occurred), the time at which each photon of the annihilation arrived at each detector crystal, and the time difference between the arrivals of the two photons at the detector crystals (i.e., ToF data). The ToF data is used to estimate the position along the LoR where the annihilation occurred.

[0019] Tissue image 120 is generated from PET data 115. Tissue image 120 includes the three-dimensional image coordinates of each annihilation represented in the PET data. The coordinates are located along the LoR of the annihilation, at a point determined by the ToF data of the annihilation. For example, if the annihilated photons were detected simultaneously, the point would be located midway along the LoR, but if one detector detected its photon before the other, the point would be closer to one detector. In this regard, the tissue image can be considered an imprecise estimate of the reconstructed tomographic image.

[0020] The scanner 110 also generates a three-dimensional CT image 125 of the patient using any suitable CT imaging protocol. The scanner 110 can generate the CT image 125 substantially simultaneously with the acquisition of the PET data 115. For example, while the patient lies in a given position on the bed of the scanner 110, the PET imaging system of the scanner 110 can be operated to acquire the PET data 115, and while the patient remains in the given position on the bed, the CT imaging system of the scanner 110 can be operated shortly thereafter to acquire the CT data 125. Because the geometric transformation (if any) between the coordinates of the PET imaging system and the CT imaging system is known, the resulting images can be easily registered with each other.

[0021] As is known in the art, a linear attenuation coefficient (LAC) map ("mu map") 130 is derived from the CT image 125. The mu map provides the attenuation coefficient of the subject's tissue and is commonly used for attenuation correction of emission data, such as PET data and single photon emission computed tomography (SPECT) data.

[0022] The normalization component 135 generates a normalized tissue image 145 based on the tissue image 120 and the scanner sensitivity data 140. The scanner sensitivity data 140 can specify the sensitivity of the scanner 110 at various locations within the field of view of the scanner 110. In some embodiments, the most sensitive location of the scanner 110 is designated as exhibiting a sensitivity of 1.0, while other locations exhibit lower sensitivities. As is known in the art, the normalization component 135 can account for this sensitivity by dividing the tissue image 120 by the location-specific sensitivity data 140.

[0023] The normalization component 135 may also perform attenuation correction on the tissue image 120 based on the mu map 130. The normalized tissue image 145 may thus be a sensitivity and attenuation corrected version of the tissue image 120. Some embodiments may omit sensitivity correction and / or attenuation correction of the tissue image 120.

[0024] The normalized tissue image 145 and the mu-map 130 are input to a trained network 150 to generate a scatter tissue image 155. At this point, the trained network 150 has been trained to estimate the scatter coincidence of the normalized tissue image 145. The network 150 may include hardware and software for performing the mapping defined by the specified neural network architecture and trained network parameters. According to some embodiments, the scatter tissue image 155 is generated from the normalized tissue image 145 and the mu-map 130 using an algorithm rather than a trained neural network.

[0025] The scatter tissue image 155 appears to be the same size as the normalized tissue image 145. According to some embodiments, the normalized tissue image 145 and the mu-map 130 are downscaled before being input to the network 150, thereby producing a similarly downscaled scatter tissue image 155. The downscaling of the normalized tissue image 145 and the mu-map 130 may include spatial downsampling to a resolution lower than the resolution of the tissue image 120.

[0026] The reduction in size of the tissue image 145 and the mu-map 130 reduces the amount of data input to the trained network 150, thereby improving the operating efficiency of the network 150. Furthermore, the present inventors have discovered that because scattering is a low-frequency and source distribution-dependent phenomenon, an accurate scattering tissue image can be generated from a significantly reduced tissue image.

[0027] Scatter correction component 160 generates a scatter-corrected tissue image 165 based on normalized tissue image 145 and scattered tissue image 155. According to some embodiments, scatter correction component 160 subtracts scattered tissue image 155 from normalized tissue image 145 to generate scatter-corrected tissue image 165.

[0028] The scatter-corrected tissue image 165 and the CT image 125 are input to the trained network 170 to generate a PET image 175. Compared to existing methods, the PET image 175 can exhibit improved signal-to-noise ratio (SNR), spatial resolution, and scatter correction while requiring fewer computational resources. If the normalized tissue image 145 and the scattered tissue image 155 have reduced resolution relative to the tissue image 120, the scatter-corrected tissue image 165 is expanded to the size of the tissue image 120 before being input to the network 170. The parameters of the network 170 can be trained as described below, and the network 170 can include any hardware and software for performing the mapping defined by the specified neural network architecture and the trained network parameters.

[0029] The examples provided herein are directed to PET imaging, but the embodiments are not limited thereto. For example, anatomical image data is described herein as being acquired using CT imaging, but may also be acquired using MR imaging, where any desired mu-map is generated from the MR data. Embodiments may be implemented using other MI modalities, including SPECT imaging.

[0030] Figure 2 is a flow chart of process 200 for generating a corrected PET image, according to some embodiments. Process 200 can be performed by any combination of hardware and software that is or becomes known. Program code embodying the processes described herein can be stored on any non-transitory tangible medium, including fixed disks, volatile or non-volatile random access memory, DVDs, flash drives, and tapes, and executed by any suitable processing unit, including but not limited to one or more microprocessors, microcontrollers, processor cores, and processor threads. Embodiments are not limited to the examples described below.

[0031] A CT image of the subject is acquired at S210 using any suitable CT imaging protocol. Next, at S220, a tissue image of the subject is acquired. Acquiring the tissue image may include operating a PET scanner to acquire PET data of the subject and generating the tissue image from the PET data.

[0032] At S230, the tissue image is normalized. The tissue image may be normalized to account for the spatial sensitivity of the PET scanner. The normalization at S230 may also include attenuation correction based on a mu map derived from the acquired CT image. The normalized tissue image may also be compressed at S230 to reduce the amount of data used to represent the normalized tissue image.

[0033] At S240, a scatter tissue image is determined based on the normalized tissue image and the mu-map. According to some embodiments, at S240, the normalized tissue image and the mu-map are input to a trained neural network to generate the scatter tissue image. In other embodiments, a known algorithm is used to generate the scatter tissue image from the normalized tissue image and the mu-map. At S250, scatter correction is performed on the normalized tissue image based on the scatter tissue image. At S260, the CT image and the scatter-corrected tissue image are input to the trained network to generate a PET image.

[0034] Figure 3 A neural network 310 is shown trained to generate a scattered tissue image based on a normalized tissue image and a mu-map, according to some embodiments. The network 310 is shown as a supervised learning compatible network. The network 310 can include any type of supervised or unsupervised learning compatible network, algorithm, decision tree, etc., that receives image data and outputs known or transformed image data. For example, the network 310 can include a generator of a generative adversarial network (GAN) with trainable parameters, as is known in the art.

[0035] The network 310 may include multiple layers of neurons that receive inputs, change internal states based on the inputs, and generate outputs based on the inputs and internal states. The outputs of certain neurons are connected to the inputs of other neurons to form a directed and weighted graph. The weights and functions for calculating the internal states are iteratively modified during training. The network 310 trained in this manner may be implemented by a set of linear equations, executable program code, a set of hyperparameters defining the model structure and a set of corresponding weights, or any other representation of the mapping of inputs to outputs learned as a result of training.

[0036] Figure 3 The training data includes N normalized tissue images 320, N mu maps 325 and N corresponding scattered tissue images 330. Each of the N pairs of normalized tissue images 320 and the N mu maps 325 can be as described above. Figure 1 Each of the N normalized tissue images 320 may be attenuation corrected, sensitivity corrected, and / or compressed in the same manner as would be applied to the normalized tissue images that would be input to the network 310 during deployment of the network.

[0037] Each "ground truth" scatter tissue image 330 is generated from a corresponding pair of normalized tissue image 320 and mu map 325. The generation of the scatter tissue image 330 can utilize any one or more techniques that are known or become known, including but not limited to single scatter simulation, phantom-based techniques, and data simulation. In some embodiments, the generation of the scatter tissue image 330 is also based on the PET data underlying its corresponding normalized tissue image 320.

[0038] Figure 4 The generation of tissue images and corresponding scatter tissue images for training a network according to some embodiments is shown. Scanner 410 generates PET data 420 of a patient, which is converted into tissue images 440 and used by reconstruction component 430 to reconstruct PET image 450. Tissue image 440 can be normalized and / or attenuation corrected based on the sensitivity of scanner 410. Similarly, reconstruction component 430 can apply sensitivity correction and / or attenuation correction during reconstruction of PET image 450.

[0039] Scatter estimation component 460 generates scatter image 470 based on PET image 450 using any algorithm that is or becomes known. In some embodiments, component 460 operates directly on PET data 420 to generate scatter image 470. Conversion component 480 converts scatter image 470 into scatter tissue image 490. Conversion component 480 can convert scatter image 470 into list-mode data and thereby generate scatter tissue image 490, as described above. Tissue image 440 and scatter tissue image 490 can be used as corresponding input data and ground truth data during network training as described below.

[0040] Figure 5 1. In some embodiments, a neural network 510 is trained to generate a PET tissue image based on a scatter-corrected tissue image and an anatomical image. The network 510 is shown as a supervised learning compatible network, but includes any type of supervised or unsupervised learning compatible system, as described above with respect to the network 310.

[0041] Figure 6 The generation of scatter-corrected tissue images, anatomical images, and corresponding PET tissue images for training a network according to some embodiments is shown. A scanner 610 generates PET data 620 of a patient and simultaneously generates a CT image 640. A reconstruction component 630 reconstructs a PET image 650 from the PET data 620. The PET image 650 may be attenuation-corrected based on attenuation data derived from the CT image 640.

[0042] The scatter correction component 660 generates a scatter-corrected PET image 670 based on the PET image 650 using any scatter correction algorithm that is or becomes known. The conversion component 680 then converts the scatter-corrected PET image 670 into a scatter-corrected tissue image 690. In some embodiments, the conversion component 680 can convert the scatter-corrected PET image 670 into list mode data and thereby generate the scatter-corrected tissue image 490. Figure 5 As shown, during network training, CT images 640 and scatter-corrected tissue images 670 may be used as input data, and scatter-corrected PET images 670 may be used as corresponding ground truth data.

[0043] During training, a batch of M scatter-corrected tissue images 520 and a batch of M CT images 530 are input to the network 510. The network 510 generates a PET image 550 for each pair of scatter-corrected tissue images 520 and CT images 530 in the batch. A loss layer 560 calculates a loss based on the difference between each of the M generated PET images 550 and its corresponding ground truth PET image 540. The loss is backpropagated to the network 510, and the network 510 is modified to minimize the loss. Additional batches are input, and the network 510 is modified accordingly, until training is determined to be complete.

[0044] Figure 7 A PET / CT scanner 700 is shown that performs one or more of the processes described herein. The embodiments are not limited to the scanner 700 or the multi-modality imaging system.

[0045] Scanner 700 includes a gantry 710 defining an aperture 712. As is known in the art, gantry 710 houses a PET imaging assembly for acquiring PET image data and a CT imaging assembly for acquiring CT image data. The CT imaging assembly may include one or more x-ray tubes and one or more corresponding x-ray detectors, as is known in the art. The PET imaging assembly may include any number or type of detectors, including background radiation emitting crystals, and may be arranged in any configuration known in the art.

[0046] The couch 715 and base 716 are operable to move a patient lying on the couch 715 into and out of the aperture 712 before, during, and after imaging. In some embodiments, the couch 715 is configured to translate on the base 716, and in other embodiments, the base 716 is movable with the couch 715 or alternatively movable from the couch 715.

[0047] Movement of the patient in and out of bore 712 may allow the patient to be scanned using the CT imaging element and the PET imaging element of gantry 710. According to some embodiments, during such scans, couch 715 and base 716 may provide continuous couch motion and / or step-and-shoot motion.

[0048] The control system 720 may include any general-purpose or special-purpose computing system. Accordingly, the control system 720 includes one or more processing units 722 and a storage device 730 for storing program code, wherein the processing unit 722 is configured to execute the program code to cause the system 720 to acquire image data and generate images therefrom. The storage device 730 may include one or more fixed disks, solid-state random access memory, and / or removable media (e.g., a thumb drive) installed in a corresponding interface (e.g., a universal serial bus port).

[0049] The storage device 730 stores program code of a control program 731. The one or more processing units 722 may execute the control program 731 to control the CT imaging elements of the scanner 700 using the CT system interface 724 and the bed interface 725 to acquire CT data and reconstruct CT images 734 therefrom.

[0050] One or more processing units 722 can execute a control program 731 to control the hardware components in conjunction with the PET system interface 723 and the bed interface 725 to inject a radiopharmaceutical into the patient, move the patient through the bore 712 past the PET detector of the gantry 710, and detect photons emitted from the patient based on pulses generated by the PET detector. The detected photons can be recorded as PET data, which can be converted into tissue images as described above.

[0051] The scatter tissue image network 732 can be executed based on the trained network parameters to generate a scatter tissue image based on the tissue image. The tissue image can be normalized and / or compressed before being input into the scatter tissue image network 732. A scatter-corrected tissue image is then determined based on the tissue image and the scatter tissue image.

[0052] The PET reconstruction network 733 is executed to generate a PET image based on the scatter-corrected tissue image and the CT image. The PET image can be saved in a PET image 735. The PET image 735 and the CT image 734 can be transmitted to a terminal 740 via a terminal interface 726. The terminal 740 may include a display device and an input device coupled to the system 720. The terminal 740 can display the received PET image 735 and CT image 733. The terminal 740 can receive user input for controlling the display of data, the operation of the scanner 700, and / or the processes described herein. In some embodiments, the terminal 740 is a separate computing device such as, but not limited to, a desktop computer, a laptop computer, a tablet computer, and a smartphone.

[0053] Each component of the scanner 700 may include other elements necessary for its operation, as well as additional elements for providing functionality other than that described herein. Each functional component described herein may be implemented in computer hardware, program code, and / or one or more computing systems that execute such program code, as is known in the art. Such a computing system may include one or more processing units that execute processor-executable program code stored in a memory system.

[0054] Those skilled in the art will appreciate that various adaptations and modifications of the above embodiments may be configured without departing from the claims. Therefore, it should be understood that the claims may be implemented in ways other than as specifically described herein.

Claims

1. A positron emission tomography (PET) scanner comprising: multiple photon detectors; as well as A processing unit for: determining an anatomical image of the subject; determining a linear attenuation coefficient map based on the anatomical image; acquiring PET data of the subject at a plurality of photon detectors; determining a tissue image of the object based on the PET data; determining a scattering tissue image based on the tissue image and the linear attenuation coefficient map; determining a scatter-corrected tissue image based on the tissue image and the scatter tissue image; as well as Inputting the anatomical image and the scatter-corrected tissue image into a trained neural network to generate a PET image; as well as A display for displaying PET images.

2. The scanner of claim 1 , wherein determining the scatter tissue image comprises: Compressed tissue images and linear attenuation coefficient maps; as well as A scattering tissue image is determined based on the compressed tissue image and the compressed linear attenuation coefficient map.

3. The scanner of claim 2, wherein determining the scatter tissue image comprises inputting the compressed tissue image and the compressed linear attenuation coefficient map into a second trained neural network to generate the scatter tissue image.

4. The scanner of claim 2, wherein determining the scatter-corrected tissue image is based on the compressed tissue image and the scattered tissue image.

5. The scanner of claim 1 , wherein determining the tissue image of the object comprises normalizing the tissue image based on a sensitivity of the scanner to determine a normalized tissue image. wherein the scattered tissue image is determined based on the normalized tissue image, and The scatter-corrected tissue image is determined based on the normalized tissue image and the scatter tissue image.

6. The scanner of claim 5, wherein determining the scatter tissue image comprises: compressing normalized tissue images and compressed linear attenuation coefficient maps; as well as A scattering tissue image is determined based on the compressed normalized tissue image and the compressed linear attenuation coefficient map.

7. The scanner of claim 6, wherein determining the scatter tissue image comprises inputting the compressed normalized tissue image and the compressed linear attenuation coefficient map into a second trained neural network to generate the scatter tissue image.

8. A method comprising: acquiring a computed tomography image of the subject; determining a linear attenuation coefficient map based on the computed tomography image; Acquiring positron emission tomography (PET) data of the subject; determining a tissue image of the object based on the PET data; determining a scattering tissue image based on the tissue image and the linear attenuation coefficient map; determining a scatter-corrected tissue image based on the tissue image and the scatter tissue image; inputting the computed tomography image and the scatter-corrected tissue image into a trained neural network to generate a PET image; and Presenting PET images.

9. The method of claim 8, wherein determining the scattered tissue image comprises: Compressed tissue images and linear attenuation coefficient maps; as well as A scattered tissue image is determined based on the compressed tissue image and the compressed linear attenuation coefficient map.

10. The method of claim 9, wherein determining the scatter tissue image comprises inputting the compressed tissue image and the compressed linear attenuation coefficient map into a second trained neural network to generate the scatter tissue image.

11. The method according to claim 9, wherein The scatter-corrected tissue image is determined based on the compressed tissue image and the scattered tissue image.

12. The method according to claim 8, wherein Determining a tissue image of the subject includes normalizing the tissue image based on a sensitivity of a scanner to determine a normalized tissue image, wherein the scatter tissue image is determined based on the normalized tissue image, and The scatter-corrected tissue image is determined based on the normalized tissue image and the scatter tissue image.

13. The method of claim 12, wherein determining the scattered tissue image comprises: Compressed normalized tissue images and linear attenuation coefficient maps; as well as A scattering tissue image is determined based on the compressed normalized tissue image and the compressed linear attenuation coefficient map.

14. The method according to claim 13, wherein: Determining the scatter tissue image includes inputting the compressed normalized tissue image and the compressed linear attenuation coefficient map into a second trained neural network to generate the scatter tissue image.

15. A non-transitory medium storing program code, the program code being executable by at least one processing unit to cause a computing system to: acquiring a computed tomography image of the subject; determining a linear attenuation coefficient map based on the computed tomography image; acquiring positron emission tomography (PET) data of the subject; determining a tissue image of the object based on the PET data; determining a scattering tissue image based on the tissue image and the linear attenuation coefficient map; determining a scatter-corrected tissue image based on the tissue image and the scatter tissue image; as well as The computed tomography images and the scatter-corrected tissue images are input to a trained neural network to generate PET images.

16. The medium of claim 15, wherein determining the scatter tissue image comprises: Compressed tissue images and linear attenuation coefficient maps; as well as A scattered tissue image is determined based on the compressed tissue image and the compressed linear attenuation coefficient map.

17. The medium according to claim 16, wherein Determining the scatter tissue image includes inputting the compressed tissue image and the compressed linear attenuation coefficient map into a second trained neural network to generate the scatter tissue image.

18. The medium according to claim 16, wherein A scatter-corrected tissue image is determined based on the compressed tissue image and the scattered tissue image.

19. The medium of claim 15, wherein determining the tissue image of the subject comprises normalizing the tissue image based on a sensitivity of a scanner to determine a normalized tissue image, wherein the scatter tissue image is determined based on the normalized tissue image, and The scatter-corrected tissue image is determined based on the normalized tissue image and the scatter tissue image.

20. The medium of claim 19, wherein determining the scatter tissue image comprises: Compressed normalized tissue images and linear attenuation coefficient maps; as well as The compressed normalized tissue image and the compressed linear attenuation coefficient map are input to a second trained neural network to generate a scattering tissue image.

Images