Apparatus for generating photon counting spectral image data

Abstract

This invention relates to an apparatus (10) for generating photon-counted spectral image data, comprising: an input unit (20); a processing unit (30); and an output unit (40). The input unit is configured to receive non-photon-counted X-ray spectral energy data. The processing unit is configured to implement a deep learning regression algorithm to generate photon-counted X-ray spectral data, and the generation includes utilizing the non-photon-counted X-ray spectral energy data. The output unit is configured to output the photon-counted X-ray spectral data.

Description

Technical Field

[0001] The present invention relates to an apparatus for generating photon counting spectral image data, and an imaging system. Background Technology

[0002] Traditional computed tomography (CT) scanners typically consist of an X-ray tube mounted on a rotating gantry, opposite a detector array comprising one or more rows of detector pixels. The X-ray tube rotates around an examination area located between the X-ray tube and the detector array, emitting polychromatic radiation that passes through the examination area and onto an object or subject positioned within it. The detector array detects the radiation passing through the examination area and generates projection data indicating the examination area and onto the object or subject positioned there. A reconstructor processes the projection data and generates volumetric image data indicating the examination area and onto the object or subject positioned there. The volumetric image data can be processed to generate one or more images that include the scanned portion of the object or subject. The resulting images include pixels, typically represented by grayscale values ​​corresponding to relative radioactivity. This information reflects the attenuation characteristics of the scanned subject and / or object and often reveals structures, such as anatomical structures within a patient, the physical structure of inanimate objects, and so on. The detected radiation also includes spectral information, as the absorption of radiation by the subject and / or object depends on the energy of the X-rays. This spectral information can provide additional information, such as information indicating the elemental or material composition (e.g., the number of atoms) of the tissue and / or substance of the subject and / or object. However, using such a scanner, the projected data does not reflect the spectral characteristics because the signal output by the detector array is proportional to the energy flux integrated on the energy spectrum. Therefore, the resulting data is actually monochromatic.

[0003] Therefore, one development in this traditional CT method is the acquisition of spectral data. Computed tomography scanners (spectral scanners) configured for spectral imaging utilize this spectral information to provide further information indicating the elemental or material composition. One approach involves using two X-ray tubes, each emitting X-ray beams with different energy spectra. Another approach involves rapid kVp switching, where the voltage on the tubes is switched between two different voltages, allowing measurements to be taken at both energies. Yet another approach involves multi-layer indirect conversion detectors, having an uppermost layer for detecting low-energy X-rays and a lowermost layer for detecting high-energy X-rays. Such spectral data is referred to herein as non-photon-counting X-ray spectral data.

[0004] The output of such a spectroscopic scanner (which can be a dual-energy scanner that acquires data at two different X-ray energies) can be two images, one at a high X-ray energy and one at a low X-ray energy.

[0005] However, the mass attenuation coefficients of matter used to reconstruct computed tomography (CT) images actually integrate several physical phenomena that affect the attenuation of matter for X-ray photons. These phenomena include the photoelectric effect, Compton scattering, and the K-edge effect. See, for example, J. Hsieh, *Computed Tomography: Principles, Design, Artifacts, and Recent Advances*, SPIE, 2015. Therefore, non-photon count spectral data can also be processed using two attenuation values ​​acquired simultaneously at two photon energies to resolve the fundamental components of the photoelectric effect and Compton scattering. Since any two fundamental functions are linearly independent and span the entire attenuation coefficient space, any matter can be represented by a linear combination of the two fundamental components. The fundamental material components can be used to generate both Compton scattering and photoelectric images.

[0006] Therefore, dual-energy CT systems utilize two attenuation values ​​obtained at two different photon energies to address the photoelectric effect and Compton scattering, while neglecting the contributions of other components of the mass attenuation coefficient (such as one or more K-edges). See, for example, EALvarez and A. MacOvski, “Energy-selective reconstructions in X-ray computerized tomography,” Phys. Med. Biol., 1976. The dual-energy approach allows for the reconstruction of high-energy and low-energy X-ray images, or Compton scattering and photoelectric images. However, the fundamental function can also be used to determine virtual monochromatic images, iodine concentration maps, virtual non-contrast images, etc., as well as conventional CT images. This allows for the reconstruction of virtual monochromatic images, iodine concentration maps, virtual non-contrast images, etc., as well as conventional CT images. Some studies have demonstrated the potential of spectral CT in better representing human anatomical structures and functions. For example, see TRC Johnson's "Dual-Energy CT: General Principles", Am. J. Roentgenol., vol. 199, no. 5, supplement, pp. S3-S8, Nov. 2012, and CHMcCollough, S. Leng, L. Yu, and JG. G. Letcher's "Dual-and Multi-Energy CT: Principles, Technical Approaches, and Clinical Applications", Radiology, vol. 276, no. 3, pp. 637-653, 2015.

[0007] This approach is well-suited for substances such as iodine, whose k-edge energy is close to the average of the diagnostic energy range. Therefore, a linear combination of two measurements at the lower and higher end of the energy range can represent the substance. In more complex cases, where more accurate substance quantification is required, or where multiple substances with different k-edge energies exist, the two-energy method may lead to suboptimal results because it cannot distinguish between substances with different k-edge energies.

[0008] Photon-counting CT systems have been developed that use direct conversion detector technology to acquire data at multiple energy levels. This makes it possible to quantify different X-ray interactions, including the K-edge energy component, in greater detail, rather than being limited to approximate models consisting only of the photoelectric effect and Compton scattering. See, for example, E. Roessl and R. Proksa, “K-edge imaging in x-ray computed tomography using multi-bin photoncount detectors,” Phys. Med. Biol., 2007, and E. Roessl, B. Brendel, K. J. Engel, J. J. Schlomka, A. Thran, and R. Proksa, “Sensitivity of photon-counting based K-Edge Imaging in X-ray computed tomography,” IEEE Trans. Med. Imaging, 2011.

[0009] This photon counting system has the following advantages:

[0010] Improved SNR compared to integral detectors

[0011] Reduced dose (reduced electronic noise)

[0012] Improved tissue differentiation / material markers

[0013] Quantitative Imaging Improved by CT

[0014] Enables new imaging techniques, such as K-edge imaging.

[0015] Reduce beam hardening artifacts

[0016] Furthermore, more advanced CT techniques, such as phase-contrast imaging, can achieve improved sensitivity at reduced doses. See, for example, H. Hetterich et al., “Phase-Contrast CT: Qualitative and Quantitative Evaluation of Atherosclerotic Carotid Artery Plaque,” ​​Radiology, 2014.

[0017] Therefore, advanced CT systems incorporating photon-counting detectors are considered a technology that can better characterize human anatomical structures and functions, with a variety of clinical applications. Photon-counting CT systems utilize direct conversion detector technology and sophisticated reconstruction algorithms to address the contribution of various components to the overall mass decay coefficient of a substance by using multiple attenuation values ​​acquired at multiple different photon energies. Further details can be found in, for example, the following literature: MJ Willemink et al., Photon-counting CT: Technical Principles and Clinical Prospects, Radiology 2018, 00, 1-20 and S. Leng et al., Photon-counting Detector CT: System Design and Clinical Applications of an emerging Technology, Radio Graphics, 2019, 39, 729-743. The data acquired from such photon-counting systems is referred to herein as photon-counted X-ray spectral data and can refer to the raw data before it is processed into a relevant image, or to the processed data involved in generating the image data.

[0018] However, the complexity of these systems, the need for specialized imaging protocols, and their very high price have hindered their adoption by clinical users.

[0019] These problems need to be addressed.

[0020] US 2019 / 251713 A1 discloses a system and method for reconstructing images of a subject acquired using a tomographic imaging system.

[0021] Lisha Yao et al.’s paper, “Direct Energy-resolving CT imaging via Energy-intergrating CT images using a Unified Generative Adversarial Network (ARVIV.ORG, Cornell University Library, October 14, 2019),” describes a mechanism for generating energy-resolved computed tomography (ErCT) images from energy-integrating computed tomography images. Summary of the Invention

[0022] Having improved means for providing X-ray spectral data for photon counting would be advantageous. The object of the invention is achieved using the subject matter of the independent claims, wherein further embodiments are incorporated in the dependent claims. It should be noted that the aspects and examples of the invention described below are also suitable for apparatuses and imaging systems for generating photon counting spectral data.

[0023] In a first aspect, an apparatus for generating photon counting spectral image data is provided, comprising:

[0024] Input unit;

[0025] Processing unit; and

[0026] Output unit.

[0027] The input unit is configured to receive non-photon-counted X-ray spectral energy data. The processing unit is configured to implement a deep learning regression algorithm to generate photon-counted X-ray spectral data. The generation of photon-counted X-ray spectral data includes utilizing non-photon-counted X-ray spectral energy data. The output unit is configured to output photon-counted X-ray spectral data.

[0028] In one example, non-photon-counted X-ray spectral energy data includes non-photon-counted image data, or the non-photon-counted image data is generated from non-photon-counted X-ray spectral image data. The non-photon-counted image data may include a first spectral image at a first X-ray energy and a second spectral image at a second X-ray energy. Photon-counted X-ray spectral data may include at least one photon-counted spectral image.

[0029] In one example, the non-photon-counted X-ray energy data includes non-photon-counted image data, or the non-photon-counted image data is generated from non-photon-counted X-ray spectral data image data. The non-photon-counted image data may include Compton scattering images and photoelectric images. The photon-counted X-ray spectral data may include at least one photon-counted spectral image.

[0030] In one example, the processing unit is configured to implement a reconstructor to process non-photon count X-ray spectral data to generate non-photon count image data.

[0031] In one example, at least one photon counting spectral image includes one or more photon counting spectral images from the group consisting of: a photon counting image at a first X-ray energy, a photon counting image at a second X-ray energy, a photon counting Compton image, a photon counting photoelectric image, a photon counting virtual monochromatic image, a photon counting contrast agent quantitative image, a photon counting non-contrast image, a photon counting cancellation image, a photon counting iodine image, and a photon counting k-edge image.

[0032] In one example, the input unit is configured to receive reconstruction parameters adopted by the reconstructor to generate non-photon count image data. The generation of photon count X-ray spectral data may include utilizing the reconstruction parameters.

[0033] In one example, the input unit is configured to receive acquisition parameters adopted by the image acquisition unit to obtain non-photon-counted X-ray spectral energy data. The generation of photon-counted X-ray spectral data may include utilizing the acquisition parameters.

[0034] In one example, the input unit is configured to receive patient parameters, and the non-photon-counted X-ray spectral energy data is acquired from the patient by the image acquisition unit. The generation of the photon-counted X-ray spectral data may include utilizing the patient parameters.

[0035] In one example, the input unit is configured to receive reference non-photon count X-ray spectral data and reference photon count X-ray spectral data. The processing unit is configured to train a deep learning regression algorithm that utilizes the reference non-photon count X-ray spectral data and the reference photon count X-ray spectral data.

[0036] In one example, the reference non-photon counting X-ray spectral data includes reference non-photon counting image data. The input unit is configured to receive reconstruction parameters used to generate the reference non-photon counting image data. Training the deep learning regression algorithm may include utilizing the reconstruction parameters.

[0037] In one example, the reference photon count X-ray spectral data may include image data.

[0038] In one example, the input unit is configured to receive acquisition parameters adopted by one or more image acquisition units to obtain reference non-photon count X-ray spectral energy data. Training of the deep learning regression algorithm may include utilizing the acquisition parameters.

[0039] In one example, the input unit is configured to receive patient parameters from at least one patient, with reference non-photon count X-ray spectral energy data acquired from the patient by one or more image acquisition units. Training of the deep learning regression algorithm may include utilizing the patient parameters.

[0040] In a second aspect, an imaging system is provided, comprising:

[0041] Image acquisition unit; and

[0042] The device as described in the first aspect.

[0043] The image acquisition unit is configured to acquire non-photon count X-ray spectral data and provide the non-photon count X-ray spectral data to the input unit of the device.

[0044] In one example, the processing unit of the device is configured to implement a reconstructor to process non-photon count X-ray spectral data to generate non-photon count image data.

[0045] Advantageously, the benefits provided by any of the above aspects also apply to all other aspects, and vice versa.

[0046] The above aspects and examples will become apparent and illustrated with reference to the embodiments described below. Attached Figure Description

[0047] Exemplary embodiments will now be described with reference to the following figures:

[0048] Figure 1 An example of a device for generating photon counting spectral image data is shown;

[0049] Figure 2 An example of an imaging system for generating photon counting spectral image data is shown;

[0050] Figure 3 A detailed example of an imaging system for generating photon counting spectral image data is shown;

[0051] Figure 4 The functional elements of an exemplary imaging system for generating photon counting spectral image data are shown in detail.

[0052] Figure 5 A detailed representation of an exemplary deep multiscale neural network for photon counting reconstruction using atrous convolutions is shown; and

[0053] Figure 6 A detailed representation of the training workflow for an exemplary deep multiscale neural network used for photon count reconstruction is shown. Detailed Implementation

[0054] Figure 1 An example of an apparatus 10 for generating photon-counted X-ray spectral image data is shown. The apparatus includes an input unit 20, a processing unit 30, and an output unit 40. The input unit is configured to receive non-photon-counted X-ray spectral energy data. The processing unit is configured to implement a deep learning regression algorithm to generate photon-counted X-ray spectral data. The generation of the photon-counted X-ray spectral data includes utilizing the non-photon-counted X-ray spectral energy data. The output unit is configured to output the photon-counted X-ray spectral data.

[0055] According to one example, non-photon-counted X-ray spectral energy data includes non-photon-counted image data, or the non-photon-counted image data is generated from non-photon-counted X-ray spectral data image data. The non-photon-counted image data, whether included within or generated from the non-photon-counted X-ray spectral energy data, can include a first spectral image at a first X-ray energy and a second spectral image at a second X-ray energy. Photon-counted X-ray spectral data can include at least one photon-counted spectral image.

[0056] According to one example, non-photon-counted X-ray energy data includes non-photon-counted image data, or the non-photon-counted image data is generated from non-photon-counted X-ray spectral data image data. Non-photon-counted image data, whether included within or generated from non-photon-counted X-ray spectral energy data, can include Compton scattering images and photoelectric images. Photon-counted X-ray spectral data can include at least one photon-counted spectral image.

[0057] According to one example, the processing unit is configured to implement a reconstructor to process non-photon count X-ray spectral data, thereby generating non-photon count image data.

[0058] According to one example, at least one photon counting spectral image includes one or more photon counting spectral images from the group consisting of: a photon counting image at a first X-ray energy, a photon counting image at a second X-ray energy, a photon counting Compton image, a photon counting photoelectric image, a photon counting virtual monochromatic image, a photon counting contrast agent quantitative image, a photon counting non-contrast image, a photon counting cancellation image, a photon counting iodine image, and a photon counting K-edge image.

[0059] According to one example, the input unit is configured to receive reconstruction parameters adopted by the reconstructor to generate non-photon count image data. The generation of photon count X-ray spectral data may include utilizing the reconstruction parameters.

[0060] According to one example, the input unit is configured to receive acquisition parameters adopted by the image acquisition unit to obtain non-photon-counted X-ray spectral energy data. The generation of photon-counted X-ray spectral data may include utilizing the acquisition parameters.

[0061] According to one example, the input unit is configured to receive patient parameters, and the non-photon-counted X-ray spectral energy data is acquired from the patient by the image acquisition unit. The generation of the photon-counted X-ray spectral data may include utilizing the patient parameters.

[0062] According to one example, the input unit is configured to receive reference non-photon count X-ray spectral data and reference photon count X-ray spectral data. The processing unit is configured to train a deep learning regression algorithm that utilizes the reference non-photon count X-ray spectral data and the reference photon count X-ray spectral data.

[0063] The reference non-photon count X-ray spectral data can be the same type of non-photon count X-ray spectral data used by the processing unit to generate photon count X-ray spectral data.

[0064] The reference photon count X-ray spectral data can be the same type of photon count X-ray spectral data generated by the processing unit.

[0065] In one example, the reference non-photon count X-ray spectral data includes at least one reference non-photon count spectral image.

[0066] In one example, the reference photon count X-ray spectral data includes at least one reference photon count spectral image.

[0067] In one example, at least one reference photon counting spectral image includes one or more photon counting spectral images from the group consisting of: a photon counting image at a first X-ray energy, a photon counting image at a second X-ray energy, a photon counting Compton image, a photon counting photoelectric image, a photon counting virtual monochromatic image, a photon counting contrast agent quantitative image, a photon counting non-contrast image, a photon counting cancellation image, a photon counting iodine image, and a photon counting K-edge image.

[0068] As an example, the reference non-photon counting X-ray spectral data includes reference non-photon counting image data. The input unit is configured to receive reconstruction parameters used to generate the reference non-photon counting image data. Training the deep learning regression algorithm may include utilizing the reconstruction parameters.

[0069] As an example, reference photon count X-ray spectral data includes image data.

[0070] According to one example, the input unit is configured to receive acquisition parameters adopted by one or more image acquisition units to obtain reference non-photon count X-ray spectral energy data. Training of the deep learning regression algorithm may include utilizing the acquisition parameters.

[0071] According to one example, the input unit is configured to receive patient parameters for at least one patient, with reference non-photon count X-ray spectral energy data acquired from the patient by one or more image acquisition units. Training of the deep learning regression algorithm may include utilizing the patient parameters.

[0072] Figure 2 An example of an imaging system 100 is shown. This imaging system includes an image acquisition unit 104 and, as described above... Figure 1 The device 10 for generating photon-counted spectral image data. The image acquisition unit is configured to acquire non-photon-counted X-ray spectral data and provide the non-photon-counted X-ray spectral data to the input unit of the device.

[0073] According to one example, the processing unit of the device is configured to implement a reconstructor to process non-photon count X-ray spectral data, thereby generating non-photon count image data.

[0074] Therefore, a photon-counting CT system is provided that can provide photon-counting results from dual-energy CT data acquisition hardware and protocols that do not involve photon-counting data acquisition. In this way, complex and expensive hardware-based photon-counting CT systems are not required to generate photon-counting CT results. This system uses a deep learning regression method to provide photon-counting CT results based on dual-energy CT data, acquisition protocols, and reconstruction parameters. The deep learning regression model predicts photon-counting results from dual-energy CT data using inter-voxel local statistics within the input data.

[0075] Now for reference Figures 3 to 6 A more detailed description is given of the device and the imaging system used to generate photon count spectral image data.

[0076] Figure 3A detailed example of an imaging system 100 for generating photon-count spectral image data is shown. An image acquisition system 104, such as a spectral CT scanner, is shown that does not acquire photon-count data but acquires the aforementioned spectral data by utilizing two X-ray tubes, rapidly switching X-ray tubes between different voltages, or a detector that acquires data related to low-energy X-ray photons in one layer and data related to high-energy X-ray photons in a second layer. The image acquisition unit 104 includes a generally fixed gantry 106 and a rotating gantry 108, the rotating gantry 108 being rotatably supported by the fixed gantry 106 and rotating about a z-axis around an examination area 110. A radiation source 112 is rotatably supported by the rotating gantry 108 and rotates with the rotating gantry 108, emitting radiation through the examination area 110.

[0077] A radiation-sensitive detector array 114, an example of which may be a two-layer detector array discussed above, is positioned across the angular arc opposite the radiation source 112 in the inspection region 110. The radiation-sensitive detector array 114 shown includes one or more rows of layer indirect conversion detector elements (e.g., scintillators / photosensors). The array 114 detects radiation passing through the inspection region 110 and generates projection data (line integral) indicating it.

[0078] Therefore, the non-photon counting spectral data has been acquired by the image acquisition unit 104 at this time.

[0079] Such non-photon counting spectral data can be directly provided to the processing unit 30. However, the non-photon counting spectral data can first be passed to a reconstructor that generates, for example, high-energy X-ray photon images and low-energy X-ray photon images, or base images as discussed above, such as Compton scattering images and photoelectric images, and then these non-photon counting spectral images can be provided to the processing unit 30.

[0080] Then, processing unit 30 receives data related to the acquisition parameters of image acquisition unit 104 when acquiring non-photon counting spectral data, as well as patient parameters, via input unit 20. Processing unit 30 can then reconstruct the non-photon counting spectral data itself into a non-photon counting spectral image, and use the parameters forming part of the reconstruction along with the acquisition parameters and patient parameters to determine photon counting spectral data, such as image data, based on the non-photon counting spectral image. Alternatively, if processing unit is provided with a reconstructed non-photon counting spectral image, then the input unit is provided with the reconstruction parameters used in the reconstruction, and the reconstruction parameters are again used along with the acquisition parameters and patient parameters to determine photon counting spectral data, such as an image, based on the non-photon counting spectral image. The resulting photon counting spectral image can then be presented, for example, on output unit 40, such as a display monitor, and / or provided to a storage medium to be saved as digital data.

[0081] Therefore, it is clear that the processing unit 30 can operate separately from the image acquisition unit and is actually an offline device that takes the acquired non-photon counting spectral data or image and generates photon counting spectral data or image based on the data, or the processing unit 30 can be intrinsically linked to the non-photon counting spectral image acquisition unit, which can generate photon counting spectral data or image in real time in a cost-effective manner.

[0082] Therefore, a novel, advanced CT system is provided that does not require specialized hardware or new imaging protocols beyond standard dual-energy CT data. This system utilizes inter-voxel statistics from dual-energy CT data to predict advanced CT results, such as photon counting results, by employing a deep learning regression model. This will be discussed below. Figures 4 to 6 To provide a more detailed description.

[0083] The advantages of the new system include:

[0084] A cost-effective system that does not require any dedicated photon counting acquisition hardware.

[0085] Seamless integration allows clinicians to use their routine clinical access protocols.

[0086] The output image is similar to the true photon count image, but it exhibits reduced noise in the image. This is because the electronic noise of the detector is effectively eliminated, as the photon count results can discount or reject this detector noise.

[0087] The system can operate in one of two configurations. In the first configuration, the dual-energy data used as input consists of two images acquired at different energy levels (low and high energy), and the output is a photon count image from N different energy bins. Therefore, a photon count image can be generated that can be presented at a higher energy level resolution, for example, 5 instead of 2 in the dual-energy input data.

[0088] In the second configuration, the dual-energy data used as input includes two basic images, such as a photoelectric image and a Compton scattering image, while the output photon counting data includes photon counting CT results, which include, but are not limited to, photon counting photoelectric images, photon counting Compton scattering images, and photon counting k-edge energy images.

[0089] Figure 4 The functional elements of an exemplary imaging system for generating photon counting spectral image data are shown in detail. The system has the following main components:

[0090] Dual-energy CT data. Spectral data should include image / projection data with at least two energy levels that allow for spectral analysis, along with acquisition and reconstruction parameters and patient parameters. Acquisition and reconstruction parameters include scan type, body location, mA, mAs, kVp, rotation time, collimation, spacing, reconstruction filter, reconstruction algorithm, slice thickness, slice increment, matrix size, and field of view, etc. Patient parameters include, for example, weight, age, sex, clinical test results, etc.

[0091] A photon counting reconstruction module implemented within the processing unit.

[0092] A. Dual-energy CT data

[0093] The data used as input to the system should include at least two CT data levels that allow for spectral analysis. Examples include, but are not limited to, CT images of anatomical structures of interest reconstructed from CT projection data acquired using a dual-layer detector system that splits the X-ray flux into two energy levels at the detector.

[0094] B. Photon-counting CT reconstruction

[0095] The inputs to this module include, but are not limited to, scan protocols, acquisition parameters, and dual-energy CT data. Optionally, the inputs may include dual-energy CT results generated using a standard dual-energy CT tubing. Based on the input data and potential additional acquisition and patient parameters, the module reconstructs the photon counting results using a deep neural regression network trained to predict photon counting results based on the input data within a training operating range.

[0096] The photon counting reconstruction module may include additional preprocessing steps, such as applying denoising algorithms, to reduce noise in the input data and improve overall performance.

[0097] against Figures 5 to 6 A more detailed description of photon count reconstruction.

[0098] Figure 5An architecture for a deep neural regression network for virtual photon counting CT is presented. A deep multi-scale neural network using dilated convolutions is used for the photon counting reconstruction module. More information on dilated convolutions can be found, for example, in L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. Yuille, “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,” IEEE Trans. Pattern Anal. Mach. Intell., 2018. Other potential architectures may include the U-net architecture. More details on the U-net architecture can be found, for example, in O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” Miccai, 2015.

[0099] The neural network takes dual-energy CT data as input, consisting of at least a low-energy image and a high-energy image, or two base images. The network applies several convolutional layers. Each layer consists of several convolutional kernels and activation functions such as rectified linear units (RCUs) and max-pooling layers, which reduce the image size through downsampling. The output of the first part of the neural network (as shown on the left) is used as input to additional dilated convolutional layers, which apply convolutions with varying receptive fields to the input. The result of this part is then combined with the network to acquire model parameters (top part), such as acquisition parameters, reconstruction parameters, and patient parameters. At this stage, these are all combined as a linear combination, followed by RCU activation functions. Finally, the two parts are combined through a set of convolutional layers, and then the image size is increased to its original size through upsampling layers. The output is then the generated photon count image.

[0100] Continue to refer to Figure 5 Moreover, it provides specific details. The figure shows an example of the deep regression network described above. This network consists of flow layers, where different channels of the outputs of networks 1 and 2 are cascaded together:

[0101] Network 1 (Input: Dual-energy CT):

[0102] Dual-energy CT data is input, such as Figure 5 As shown on the left side.

[0103] The following layers are convolutional layers with batch normalization (BN) and rectified activation function (ReLU). For more information on ReLU, see, for example:

[0104] https: / / en.wikipedia.org / wiki / Activation_function.

[0105] The lower layers include max pooling, with another convolutional layer, as well as batch normalization transformation and rectified activation function.

[0106] The lower layers also include max pooling, with another convolutional layer, as well as batch normalization transformation and rectified activation function.

[0107] Then, within the dashed box, we show the cascade of outputs from the following parallel layers, where the inputs to each layer are the outputs of the operators in max pooling:

[0108] 1-Dental / Dilated Convolution

[0109] 3-Dental / Dilated Convolution

[0110] 5-Dental / Dilated Convolution

[0111] 7-Dental / Dilated Convolution

[0112] For more details on dilated / dilated convolutions, see, for example, F. Yu and V. Koulton’s Multi-Scale Context Aggregation By Dilated Convolutions, published as a conference paper at ICLR 2016.

[0113] Network 2 (Input: parameters, get parameters, reconstruct parameters, patient parameters):

[0114] Perform a cascade of all input parameters.

[0115] The next layer consists of a fully connected network and a rectified activation function (ReLU).

[0116] The next layer is still a fully connected network and a rectified activation function.

[0117] Combination of Networks 1 and 2

[0118] The outputs of networks 1 and 2 are combined with parallel channels and fed into a convolutional layer, followed by batch normalization (BN) and rectified activation functions.

[0119] The next layer is a convolutional layer with batch normalization transformation and rectified activation function.

[0120] The next layer includes upsampling, which uses, for example, NN / Bilinear / Cubic interpolation to upscale the image by a factor of 2.

[0121] The next layer is a convolutional layer with batch normalization transformation and rectified activation function.

[0122] The next layer includes upsampling.

[0123] The next layer is a convolutional layer with batch normalization transformation and rectified activation function.

[0124] Finally, convolution is performed using a 1x1 kernel with one output channel, the output of which is photon count spectral image data / image.

[0125] The inventors also evaluated how to reduce noise in the resulting photon count spectral images and determined that this operation can be performed automatically using a deep regression network—see, for example, H. Chem et al., Low-dose CT via convolutional neural network, Biomedical Optics Express, vol.8, No.2, 679-694 (2017).

[0126] Figure 6 It shows the use of in Figure 5 The training procedure for the network used in the photon counting reconstruction module. Figure 6 The training procedure for the photon counting reconstruction module and the deep neural regression network is shown. This training procedure aims to determine... Figure 5 The program describes the network parameters (i.e., convolutional kernels, etc.). It initializes the network parameters with some random values ​​and then modifies them using gradient decay or a similar algorithm based on a predefined loss function. In this case, the predefined loss function could be the root mean square error between the predicted and reference photon count images. Other loss functions can also be used.

[0127] Formally, the goal of the training procedure is to find a function: f(CT) DE )→CT PC It will input dual-energy CT data (CT) DE Mapping to photon-counted CT data PC .

[0128] Training is accomplished by minimizing a certain loss function:

[0129]

[0130] Where D represents the loss function. A potential example of D is the root mean square error, as discussed above:

[0131] D(f(CT DE ), CT PC)=‖f(CT DE ), CT PC || 2

[0132] therefore, Figure 6 The training process of a supervised deep network is illustrated. Training can be accomplished using one of the common optimization algorithms, such as DPKingma and JLBa, ADAM's "A METHOD FOR STOCHASTIC OPTIMIZATION," published as a conference paper at ICLR 2015.

[0133] Furthermore, a hybrid adversarial training procedure can be implemented, where the training objective is to both achieve a lower RMSE for samples with reference hardware-based photon counting results and increase the number of images generated by the network that are classified as hardware-based photon counting results. For samples without reference hardware-based photon counting results, this can be done using an adversarial trainer. An example of a suitable adversarial trainer can be found here: "Generative Adversarial Networks" by IJ Goodfellow et al., June 2014.

[0134] It should be noted that embodiments of the invention have been described with reference to different subjects. However, those skilled in the art will understand from the above and below description that, unless otherwise indicated, any combination of features related to different subjects is also considered to be disclosed in this application, except for any combination of features belonging to one type of subject. However, all features can be combined to provide a synergistic effect that is more than the simple sum of features.

[0135] Although the invention has been illustrated and described in the accompanying drawings and the foregoing description, such illustrations and descriptions are to be considered illustrative or exemplary and not restrictive. The invention is not limited to the disclosed embodiments. By studying the drawings, the disclosure, and the dependent claims, those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed invention.

[0136] In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite articles "a" or "an" do not exclude a plurality. A single processor or other unit can perform the functions of several items recited in the claims. The recitation of certain measures only in mutually different dependent claims does not imply that combinations of these measures cannot be advantageously used. Any reference numerals in the claims should not be construed as limiting the scope.

Claims

1. An apparatus (10) for generating photon counting spectral image data, comprising: Input unit (20); Processing unit (30); and Output unit (40); The input unit is configured to receive non-photon count X-ray spectral energy data, which includes non-photon count image data, and the non-photon count image data includes a first spectral image at a first X-ray energy and a second spectral image at a second X-ray energy. The processing unit is configured to implement a deep learning regression algorithm to generate photon-counted X-ray spectral data, and the generation includes utilizing the non-photon-counted X-ray spectral energy data; and The output unit is configured to output the photon counting X-ray spectral data.

2. The device according to claim 1, wherein, The photon-counted X-ray spectral data includes at least one photon-counted spectral image.

3. The device according to claim 1, wherein the non-photon counting image data includes Compton scattering images and photoelectric images, and the photon counting X-ray spectral data includes at least one photon counting spectral image.

4. The device according to any one of claims 1 to 3, wherein, The processing unit is configured to implement a reconstructor to process the non-photon count X-ray spectral data, thereby generating the non-photon count image data.

5. The device according to any one of claims 2 to 3, wherein, The at least one photon counting spectral image comprises one or more photon counting spectral images from the group consisting of: a photon counting image at the first X-ray energy, a photon counting image at the second X-ray energy, a photon counting Compton image, a photon counting photoelectric image, a photon counting virtual monochromatic image, a photon counting contrast agent quantitative image, a photon counting non-contrast image, a photon counting cancellation image, a photon counting iodine image, and a photon counting K-edge image.

6. The device according to any one of claims 1 to 3, wherein, The input unit is configured to receive reconstruction parameters adopted by the reconstructor to generate the non-photon count image data, and the generation of the photon count X-ray spectral data includes utilizing the reconstruction parameters.

7. The device according to any one of claims 1 to 3, wherein, The input unit is configured to receive acquisition parameters adopted by the image acquisition unit to obtain the non-photon count X-ray spectral energy data, and the generation of the photon count X-ray spectral data includes utilizing the acquisition parameters.

8. The device according to any one of claims 1 to 3, wherein, The input unit is configured to receive patient parameters of the patient, the non-photon count X-ray spectral energy data is acquired from the patient by the image acquisition unit, and the generation of the photon count X-ray spectral data includes using the patient parameters.

9. The device according to any one of claims 1 to 3, wherein, The input unit is configured to receive reference non-photon count X-ray spectral data and reference photon count X-ray spectral data, and the processing unit is configured to train the deep learning regression algorithm, the deep learning regression algorithm including the use of the reference non-photon count X-ray spectral data and the reference photon count X-ray spectral data.

10. The device according to claim 9, wherein, The reference non-photon counting X-ray spectral data includes reference non-photon counting image data, and the input unit is configured to receive reconstruction parameters used to generate the reference non-photon counting image data, and the training of the deep learning regression algorithm includes utilizing the reconstruction parameters.

11. The device according to claim 9, wherein, The reference photon counting X-ray spectral data includes image data.

12. The device according to claim 9, wherein, The input unit is configured to receive acquisition parameters adopted by one or more image acquisition units to obtain the reference non-photon count X-ray spectral energy data, and the training of the deep learning regression algorithm includes utilizing the acquisition parameters.

13. The device according to claim 9, wherein, The input unit is configured to receive patient parameters of at least one patient, the reference non-photon count X-ray spectral energy data is acquired from the patient by one or more image acquisition units, and the training of the deep learning regression algorithm includes utilizing the patient parameters.

14. An imaging system (100), comprising: Image acquisition unit (104); and The device (10) according to any one of claims 1 to 13; The image acquisition unit is configured to acquire non-photon count X-ray spectral data and provide the non-photon count X-ray spectral data to the input unit of the device.

15. The imaging system according to claim 14, wherein, The processing unit of the device is configured to implement a reconstructor to process the non-photon count X-ray spectral data, thereby generating non-photon count image data.

Images