An image data processing method and apparatus

Abstract

This application provides an image data processing method and apparatus. Specifically, it involves acquiring an image to be processed; the image to be processed includes multiple projected images, each of which is obtained by scanning a target object using SPECT within a sparse angular range at a first interval angle; the first interval angle is greater than a second interval angle for scanning the target object across the entire angular range; denoising the image to be processed is performed to obtain a denoised image; tomographic reconstruction is performed on the denoised image to obtain image domain volume data; denoising is performed on the image domain volume data to obtain denoised image domain volume data; super-resolution processing is performed on the denoised image domain volume data to obtain a target image; the resolution of the target image is higher than the resolution of the image to be processed. In other words, the image data processing method provided by this application can improve scanning speed and reduce scanning time while ensuring image quality.

Description

Technical Field

[0001] This application relates to the field of computed tomography imaging technology, and in particular to an image data processing method and apparatus. Background Technology

[0002] Single-photon emission computed tomography (SPECT) is a commonly used clinical imaging technique in nuclear medicine. SPECT works by injecting a tracer containing a specific radionuclide into the patient's body. These tracers release gamma rays during metabolism. The SPECT scanner rotates around the patient, using a gamma camera with a collimator to capture these rays from different angles, thus acquiring multi-angle projection data. Subsequently, tomographic reconstruction of this projection data generates three-dimensional images reflecting physiological and pathological changes within the body.

[0003] Currently, to ensure image quality, SPECT is often used to perform full-angle scanning of the scanned object. That is, for multiple tomographic beds, SPECT rotates 360 degrees at each tomographic bed and acquires a frame of projection data at angular intervals (e.g., 3 degrees), resulting in multiple frames of projection data. However, acquiring complete 360-degree projection data for multiple tomographic beds often requires a long acquisition time. Summary of the Invention

[0004] This application provides an image data processing method and apparatus for reducing the scanning time of SPECT full-body scanning of a scanned object.

[0005] In a first aspect, embodiments of this application provide an image data processing method applied to SPECT imaging, the method comprising:

[0006] Obtain the image to be processed;

[0007] The image to be processed includes multiple frames of projection images. Each frame of projection image is obtained by scanning the target object using SPECT within a sparse angular range at a first interval angle. The first interval angle is greater than the second interval angle when scanning the target object within the full angular range.

[0008] The image to be processed is subjected to noise reduction processing to obtain a noise-reduced image;

[0009] The denoised image is then subjected to tomographic reconstruction to obtain image domain volume data.

[0010] The image domain volume data is denoised to obtain denoised image domain volume data.

[0011] The denoised image volume data is subjected to super-resolution processing to obtain a target image; the resolution of the target image is higher than the resolution of the image to be processed.

[0012] Optionally, the step of performing noise reduction processing on the image to be processed to obtain a noise-reduced image includes:

[0013] The image to be processed is denoised using a projection domain network model to obtain a denoised image; the projection domain network model includes a residual U-shaped network model, which is a network model that introduces residual blocks on a Unet structure.

[0014] Optionally, the projection domain network model is a network model trained using the first training samples. The first training samples include label data and noise data of the simulated projection image. The method for obtaining the first training samples includes:

[0015] Obtain digital human body models with different human body characteristics; the digital human body models are used to simulate the distribution of radioactive tracers in the human body;

[0016] The digital human body model is input into the target simulation platform to obtain a simulated projection image. The target simulation platform is equipped with multi-bed tomographic data scanning data. After receiving the digital human body model, it rotates 360 degrees at each bed position and scans the digital human body model according to the second interval angle to obtain the simulated projection image.

[0017] A first Poisson noise is added to the simulated projection image to obtain the label data of the simulated projection image; a second Poisson noise is added to the simulated projection image to obtain the noise data of the simulated projection image; the first Poisson noise is less than the second Poisson noise.

[0018] Optionally, the step of denoising the image domain volume data to obtain denoised image domain volume data includes:

[0019] The image domain volume data is denoised using an image domain denoising network model to obtain denoised image domain volume data. The image domain denoising network model includes a ResUnet network model, which refers to a network model that introduces residual blocks on the Unet structure.

[0020] Optionally, the image domain denoising network model is a network model trained based on the second training samples. The label data of the second training samples is full-angle image domain volume data within the full-angle scanning range, and the noise data of the second training samples is sparse image domain volume data within the sparse angle scanning range. The method for obtaining the second training samples includes:

[0021] Tomographic reconstruction is performed using the full-angle projection data obtained by scanning the full-angle range of the SPECT to obtain the corresponding full-angle image domain data.

[0022] The corresponding full-angle image domain volume data is stitched together according to the position information of multiple beds in the target direction to obtain the stitched full-angle image domain volume data, and the full-angle image domain volume data is used as the label data of the second training sample;

[0023] The sparse angle projection data obtained by scanning within the sparse angle range of the SPECT is used to perform tomographic reconstruction, and the corresponding sparse image domain volume data is obtained.

[0024] The corresponding sparse image domain data is stitched together according to the position information of multiple beds in the target direction to obtain stitched sparse image domain data. The stitched sparse image domain data is used as the noise data of the second training sample. The target direction indicates the direction of the target object from head to toe.

[0025] Optionally, the ResUnet network model includes:

[0026] The system comprises an n-layer encoder, an m-layer bottleneck layer, and a p-layer decoder; the encoder includes a first residual block and a pooling layer, the bottleneck layer includes a second residual block, and the decoder includes an upsampling module and a third residual block, wherein the upsampling module and the third residual block communicate via a skip connection; n is a positive integer, m is a positive integer, and m is a positive integer.

[0027] Optionally, the step of performing tomographic reconstruction on the denoised image to obtain image domain volume data includes:

[0028] The ordered subset expectation-maximization algorithm (OSEM) is used to perform tomographic reconstruction on the denoised image to obtain image domain volume data.

[0029] Optionally, the step of performing super-resolution processing on the denoised image domain volume data to obtain the target image includes:

[0030] The target image is obtained by super-resolution processing of the denoised image domain volume data using an image domain enhancement network model.

[0031] The image domain enhancement network model is an enhancement network model trained on the third training sample. The third training sample includes positive samples and negative samples. The positive samples include the label image and the image domain volume data corresponding to the label image. The negative samples include the low-resolution image and the image domain volume data corresponding to the low-resolution image. The label image is an image obtained by adding the first Poisson noise to the full-angle simulated projection image. The low-resolution image is an image obtained by adding the first Poisson noise to the sparse-angle simulated projection image.

[0032] Optionally, the image domain enhancement network model includes a feature extraction layer, a residual learning layer, and an upsampling layer; the feature extraction layer includes four 3x3 convolutional layers, the residual learning layer includes 64 residual blocks, and the upsampling layer includes two upsampling layers.

[0033] Secondly, embodiments of this application provide an image data processing apparatus for SPECT imaging, the apparatus comprising:

[0034] An acquisition unit is used to acquire an image to be processed; the image to be processed includes multiple frames of projection images, each frame of projection image is a projection image obtained by scanning the target object at a first interval angle within a sparse angle range using SPECT; the first interval angle is greater than a second interval angle for scanning the target object within the full angle range.

[0035] The noise reduction processing unit is used to perform noise reduction processing on the image to be processed to obtain a noise-reduced image; to perform tomographic reconstruction on the noise-reduced image to obtain image domain volume data; and to perform noise reduction processing on the image domain volume data to obtain noise-reduced image domain volume data.

[0036] A super-resolution unit is used to perform super-resolution processing on the denoised image domain volume data to obtain a target image; the resolution of the target image is higher than the resolution of the image to be processed.

[0037] Thirdly, embodiments of this application provide a computer program product, the computer program product comprising: a computer program (also referred to as code or instructions), which, when the computer program is run, causes a computer to perform a method in any of the possible implementations of any of the above aspects.

[0038] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program (also referred to as code or instructions) that, when run on a computer, causes the computer to perform the method in any of the possible implementations of any of the above aspects.

[0039] Fifthly, embodiments of this application provide a chip system including one or more processors for calling and executing instructions stored in memory, causing the methods in any of the above aspects or possible implementations to be executed. The chip system may be composed of chips or may include chips and other discrete devices.

[0040] This application provides an image data processing method and apparatus. When executing the method, firstly, an image to be processed is acquired. The image to be processed includes multiple projected images, each of which is obtained by scanning a target object using SPECT within a sparse angular range at a first interval angle. The first interval angle is greater than a second interval angle when scanning the target object across the entire angular range. The image to be processed is then denoised to obtain a denoised image. The denoised image is then subjected to tomographic reconstruction to obtain image domain volume data. The image domain volume data is then denoised to obtain denoised image domain volume data. Finally, the denoised image domain volume data is subjected to super-resolution processing to obtain a target image. The resolution of the target image is higher than the resolution of the image to be processed. This application utilizes SPECT and sparse angular scanning to reduce acquisition scanning time and improve scanning efficiency. However, sparse angular scanning inevitably leads to a decrease in image quality. Therefore, this application further improves the image quality of the scanned image by performing projection domain preprocessing, image domain denoising, and image domain super-resolution processing. Therefore, the image data processing method provided in this application can improve scanning speed and reduce scanning time while ensuring image quality. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart of an image data processing method provided in an embodiment of this application;

[0043] Figure 2 A flowchart illustrating a method for obtaining a first training sample provided in an embodiment of this application;

[0044] Figure 3 A flowchart illustrating a method for obtaining a second training sample provided in an embodiment of this application;

[0045] Figure 4 This is a schematic diagram of the structure of an image data processing device provided in an embodiment of this application. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present application, the technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] As mentioned earlier, performing a full-angle scan using SPECT often requires a long acquisition time. Currently, a full-angle SPECT scan takes approximately 15 to 20 minutes, and depending on the height of the person, about 4 to 5 beds are needed. The entire SPECT scan takes 60 to 100 minutes. Therefore, the acquisition time is quite long.

[0048] In view of this, embodiments of this application provide an image data processing method. Using SPECT, sparse angular range scanning is employed to reduce scanning time and improve scanning efficiency. However, sparse angular scanning inevitably leads to a decrease in image quality. Therefore, this application further improves the image quality of the scanned image by performing projection domain preprocessing, image domain noise reduction, and image domain super-resolution processing. Thus, the image data processing method provided by this application can improve scanning speed and reduce scanning time while ensuring image quality.

[0049] The image processing method provided in the embodiments of this application will be described below with reference to the accompanying drawings. It should be noted that the executing entity of the image processing method provided in the embodiments of this application can be a computing device, specifically a server, etc., and this application embodiment is not specifically limited to this. For ease of explanation, the following description uses a computing device as the executing entity as an example.

[0050] Appendix Figure 1 A flowchart of an image data processing method provided in this application embodiment, the method being applied to SECT, the method comprising:

[0051] S10, acquire the image to be processed.

[0052] First, we will introduce full-angle range scanning and sparse-angle range scanning.

[0053] Full-angle range scanning is a multi-bed tomographic data scanning method where the SPECT scanner rotates 360 degrees at each bed position, and a projection image is acquired every second interval angle. For example, if the second interval angle is 3 degrees, then 120 frames of projection data can be acquired at each bed position.

[0054] Sparse angular range scanning is a multi-bed tomographic data scanning technique where each bed is rotated 360 degrees within the SPECT scanner, and one frame of projected image is acquired at every first angular interval. The first angular interval is longer than the second angular interval. For example, if the second angular interval is 3 degrees and the first angular interval is 12 frames, then only 30 frames of projected images need to be acquired per bed. This significantly reduces scanning time and improves scanning efficiency.

[0055] The image to be processed is described below.

[0056] The image to be processed refers to multiple frames of projection data obtained by scanning the target object using SPECT within a sparse angular range at a first interval angle. The target object refers to the object to be scanned by SPECT, which can be a person, an animal, etc., and is not specifically limited in this embodiment. For example, if the first interval angle is 12 frames and the target object corresponds to bed 4, then the image to be processed includes 120 frames of projection images.

[0057] It should be noted that, in the embodiments of this application, the projected image is a two-dimensional image.

[0058] S20, the image to be processed is subjected to noise reduction processing to obtain the noise-reduced image.

[0059] The image to be processed was obtained by scanning within a sparse angular range. Compared to images obtained by scanning the target object across the entire angular range, this image suffers from lower resolution and poorer quality, which affects subsequent processing results. Therefore, the computing device performs noise reduction processing on the image to be processed, thereby improving image resolution by reducing noise in the projected image.

[0060] In one specific implementation, the computing device can employ a projection domain network model to perform noise reduction on the image to be processed. Specifically, the projection domain network model can be a Residual U-Net (ResUnet) network model. The ResUnet network model is a network model that introduces residual blocks into the Unet structure. This network model can reduce image noise while preserving the original image texture.

[0061] Furthermore, this application embodiment also provides a ResUnet network model structure. This ResUnet network model includes an n-layer encoder, an m-layer bottleneck layer, and a p-layer decoder. The encoder includes a first residual block and a pooling layer; the bottleneck layer includes a second residual block; and the decoder includes an upsampling module and a third residual block. The upsampling module and the third residual block communicate via skip connections; n is a positive integer, m is a positive integer, and p is a positive integer. All residual blocks in this network model have the same structure; for example, they are all composed of two 3×3 convolutional layers and include BatchNorm and ReLU residual blocks. The residual blocks directly add the input data to the output via skip connections.

[0062] For example, the following explanation uses n=4, m=2, and p=4. First, the encoder extracts features from the image to be processed layer by layer, ensuring that the acquired first features have low noise and high-level semantic features. In this embodiment, each encoder first processes the features output by the previous encoder (referred to as input features) through a first residual block to extract the basic texture features of the input features and initially suppress noise. Then, a pooling layer compresses the dimensionality of the extracted features to obtain low-dimensional features with high-level semantic features (referred to as output features). It should be noted that the input features acquired by the first encoder layer are the image to be processed.

[0063] The bottleneck layer processes the first feature output by the encoder using a two-layer structure containing only residual blocks, preserving the most critical semantic and structural information within the first feature. For example, if all residual blocks have the same structure, such as using two 3×3 convolutional layers and including BatchNorm and ReLU residual blocks, then the first residual block in the bottleneck layer first performs feature transformation on the first feature through the first 3×3 convolutional layer. After BatchNorm normalization and ReLU activation, non-linear features are introduced. Then, the second 3×3 convolutional layer further extracts features, which are also processed by BatchNorm. The input features are then directly superimposed onto the current output as the output feature through skip connections, and output to the second residual block. The second residual block undergoes the same processing as the first residual block, ensuring that the features output to the decoder retain the key structural information extracted by the encoder while enhancing noise suppression through non-linear transformation. In this embodiment, the bottleneck layer processes the first feature, and the output feature is the second feature.

[0064] The decoder receives the second feature output from the bottleneck layer and processes it through four decoder layers to finally output a projected image with the same dimensions as the image to be processed and lower noise. Specifically, each decoder layer upsamples the features input to that layer and, through skip connections, superimposes the features into the corresponding third residual block for processing and outputs the corresponding feature, which can suppress artifacts.

[0065] It should be noted that n, m, and P can be the same or all different, and this application does not limit this.

[0066] To further improve the noise reduction effect, the projection domain network model can be a network model trained using the first training samples. The first training samples include the label data and noise data of the simulated projection image. The simulated projection image refers to the projection image obtained by simulating the scanning of the object using SPECT.

[0067] In one example, the first training sample can be attached Figure 2 Obtain it as shown. (Attached) Figure 2 A flowchart of a method for obtaining a first training sample provided in an embodiment of this application is shown. The method includes:

[0068] S210, acquire digital human body models with different human body features.

[0069] In this embodiment of the application, a digital human body model refers to a human body model that simulates the distribution of radioactive tracers within the human body. In this embodiment of the application, the computing device can create digital human body models with different human characteristics based on at least one of different heights and weights.

[0070] For example, the computing device can use XCAT simulation to design and manufacture [devices] for different heights and weights. 99 Digital human model of mTc-MDP tracer uptake.

[0071] S220: Input the digital human body model into the target simulation platform to obtain the simulation projection image.

[0072] The computing device can input digital human body models with different human features into the target simulation platform to obtain simulated projection images. In this embodiment, the target simulation platform is equipped with multi-bed tomographic data scanning. After receiving the digital human body model, it rotates 360 degrees at each bed position and scans the digital human body model at preset interval angles to obtain simulated projection images. In this embodiment, the preset angle is the second interval angle. For example, the target simulation platform is the GATE simulation platform.

[0073] It is understood that the simulated projection image is a projection image obtained by scanning the entire angular range, and the simulated projection image has a high resolution.

[0074] S230: Add a first Poisson noise to the simulated projected image to obtain the label data of the simulated projected image. Add a second Poisson noise to the simulated projected image to obtain the noise data of the simulated projected image; wherein, the first Poisson noise is less than the second Poisson noise.

[0075] In actual SPECT scanning, the acquired image data is often noisy and not ideally clean due to various factors such as the physical characteristics of the device itself, environmental interference, and the complex physiological structure of the human body. In this embodiment, the computing device can add different levels of Poisson noise to the simulated projection image, making the image closer to the actual imaging situation. This allows the model trained based on these images to better cope with noise interference in real-world applications, improving the model's practicality and robustness.

[0076] In this embodiment, the computing device can add a first Poisson noise to the simulated projected image to obtain label data. This label data is a reference data that approximates a real noise-free state. For example, the first Poisson noise is 20%, meaning that 20% Poisson noise is added to the simulated projected image as label data.

[0077] Noise data indicates a high level of noise interference and can be used to test a model's ability to handle complex noisy environments. In this embodiment, the computing device trains the model using noise data and label modulus. The trained model can use low-noise label data as reference data and learn from and process high-noise data to attempt to extract effective image information to approximate the high-quality image effect represented by the label data.

[0078] In this embodiment, the computing device can add a second Poisson noise to the simulated projected image to obtain noise data for the simulated projected image; the first Poisson noise is less than the second Poisson noise. For example, the first Poisson noise is 20%, and the second Poisson noise is 50%.

[0079] Furthermore, the loss function involved in the training process of this projection domain network model is the first loss function. The first loss function refers to the loss function that is correlated with mean squared error and structural similarity. In a specific implementation, the first loss function = mean squared error (MES) + structural similarity index measure (SSIM).

[0080] MES (Mean Estimation and Target Estimation) is a metric that measures the difference between a predicted value and the true value. For images, in this projection domain network model, if there exists an output image (the predicted image) processed by the model and a corresponding real, noise-free, or high-quality reference image, the MES can be obtained as follows: First, calculate the difference between the predicted and true values ​​for each corresponding pixel. Then, square these differences and calculate the average of the squared differences for all pixels. During training, the projection domain network model continuously adjusts its parameters (such as the weights of convolutional layers in the encoder and decoder) to continuously decrease the MES value, thus making the output image increasingly closer to the real, high-quality image.

[0081] SSIM measures the similarity between two images from the perspective of how the human visual system perceives image quality, comprehensively considering brightness, contrast, and structural information. SSIM places greater emphasis on structural consistency within the images. During noise reduction, relying solely on MES may result in reduced pixel differences but compromised image structure; for example, previously sharp organ outlines may become blurred, or some texture details may be lost. SSIM, by constraining structural similarity, guides the network to preserve the structural features of the image while reducing noise.

[0082] Therefore, when training this projection domain network model, using loss functions related to MES and SSIM can ensure that the projection domain network model can acquire high-quality images and maintain the structural integrity of the images while reducing noise.

[0083] It should be noted that the embodiments of this application can reduce noise in images in other ways, and the embodiments of this application are not specifically limited.

[0084] S30 performs tomographic reconstruction on the denoised image to obtain image domain volume data.

[0085] It should be noted that the denoised image is a two-dimensional image.

[0086] In this embodiment, the computing device can perform tomographic reconstruction on the denoised image to obtain image volume data. This image volume data is three-dimensional image data, capable of reflecting the full-body data of the target object.

[0087] In one example, the computing device can use the Ordered Subsets Expectation Maximization (OSEM) algorithm to perform tomographic reconstruction on the denoised image to obtain image domain volume data corresponding to each bed.

[0088] Because SPECT uses multi-bed computed tomography (CT) scans, to obtain image volume data reflecting the entire body of the target subject, it is necessary to stitch together the image volume data corresponding to each bed position. Since SPECT multi-bed scans are performed sequentially along the long axis of the human body (from head to toe), to ensure that the acquired image volume data reflects the entire body of the target subject, images can be stitched together according to the target direction. This target direction, also known as the Z-axis direction, refers to the direction along the target subject from head to toe.

[0089] To improve the stitching effect, the image domain volume data of each bed is first normalized to obtain a processing coefficient A (also known as the normalization baseline value). Then, the data from different beds are all normalized to coefficient A according to the proportion to achieve stitching and obtain a whole body data, i.e., image domain volume data.

[0090] S40, the image domain volume data is subjected to noise reduction processing to obtain the noise-reduced image domain volume data.

[0091] Since noise may be introduced or amplified during fault reconstruction, computing devices can further denoise the image domain data to obtain denoised image domain data.

[0092] In one example, the computing device can use an image domain denoising network model to denoise image domain volume data to obtain denoised image domain volume data; the image domain denoising network model also includes the ResUnet network model.

[0093] Furthermore, the image domain denoising network model provided in this application embodiment is a network model trained based on the second training samples. The label data of the second training samples is full-angle image domain volume data within the full-angle scanning range, and the noise data is sparse image domain volume data within the sparse-angle scanning range.

[0094] Appendix Figure 3 A flowchart of a method for obtaining a second training sample provided in this application embodiment, the method including:

[0095] S310 performs tomographic reconstruction on the full-angle projection data obtained from scanning within the full-angle range to obtain the corresponding full-angle image domain data.

[0096] In one example, the computing device can use OSEM to perform tomographic reconstruction on full-angle projection data acquired through a real-device scan. A real-device scan refers to scanning a real object using a physical device. Full-angle projection data is projection data collected over a complete angular range. After processing with the OSEM algorithm, image domain volume data reflecting the internal structure of the human body can be obtained. Because full-angle scan data is relatively more complete and comprehensive, it is used as low-noise image domain data.

[0097] S320: The full-angle image domain volume data is stitched together according to the position information of multiple beds in the target direction to obtain the stitched full-angle image domain volume data. This full-angle image domain volume data is used as the label data of the second training sample.

[0098] Since actual scans often involve multiple beds (e.g., scanning zones divided according to different parts of the body and different heights), to obtain whole-body data, the data from scans at different beds must first be integrated. Specifically, the computing device first stitches images together according to the positional information of multiple beds in the target direction to determine the relative position of each bed's data within the overall picture. Then, the volume data from each bed is normalized, that is, a processing coefficient A is obtained through a certain calculation method. This coefficient is equivalent to a unified standard scale. Afterwards, the data from different beds are all normalized and adjusted according to the corresponding proportion of the processing coefficient A. The purpose of this is to eliminate the differences that may exist between the data from different beds due to factors such as scanning conditions, so that the data from each bed can be better connected, achieving stitching, and finally obtaining a complete volume data that reflects the condition of the whole body, providing a foundation for subsequent more comprehensive and accurate analysis of the internal condition of the human body.

[0099] S330 performs tomographic reconstruction on the sparse angle projection data obtained by scanning within the sparse angle range to obtain the corresponding sparse image domain volume data.

[0100] In this embodiment of the application, the computing device can obtain the corresponding sparse image domain volume data through the same tomographic reconstruction method as the full-angle projection data.

[0101] S340, the sparse image domain volume data is stitched together according to the position information of multiple beds in the target direction to obtain the stitched sparse image domain volume data, and the stitched sparse image domain volume data is used as the noise data of the second training sample; the target direction indicates the direction of the target object from head to toe.

[0102] It should be noted that in the embodiments of this application, S310 and S330 can be executed simultaneously, or S310 can be executed first and then S320, or S320 can be executed first and then S310. The embodiments of this application are not limited.

[0103] It should be noted that other methods can also be used to denoise image volume data in the embodiments of this application, and the embodiments of this application are not limited thereto.

[0104] S50, perform super-resolution processing on the denoised image domain data to obtain the target image; the resolution of the target image is higher than the resolution of the image to be processed.

[0105] Computing devices can perform super-resolution processing (i.e., super-resolution processing) on ​​the denoised image volume data to obtain the target image. The target image obtained in this way has the same dimensional information as the image to be processed, but its resolution is higher than that of the image to be processed. For example, computing devices can utilize the Enhanced Deep Residual Networks for Single Image Super-Resolution (EDSR) algorithm to perform super-resolution processing and obtain the target image.

[0106] In one example, a computing device can utilize an image domain augmentation network model to perform super-resolution processing on the denoised image domain volume data to obtain the target image. The image domain augmentation network model is an augmentation network model trained on a third set of training samples. These third training samples include positive and negative samples. The positive samples include a label image and its corresponding image domain volume data. The negative samples include low-resolution images and their corresponding image domain volume data. The label image is obtained by adding a first Poisson noise to a full-angle simulated projection image, and the low-resolution image is obtained by adding the first Poisson noise to a sparse-angle simulated projection image.

[0107] For example, the simulated projection image is a simulated projection image obtained by a computing device inputting a digital human body model with different human features into a target simulation platform. In this embodiment, the target simulation platform is equipped with multi-bed tomographic data scanning data. After receiving the digital human body model, it rotates 360 degrees at each bed position and scans the digital human body model at a second interval angle to obtain a simulated projection image. The full-angle simulated projection image is a high-resolution simulated projection image, while the sparse-angle simulated projection image is a low-resolution simulated projection image.

[0108] In this embodiment, OSEM can be used to perform tomographic reconstruction of the label image to obtain the corresponding image domain volume data. OSEM is also used to perform tomographic reconstruction of the label image on the low-resolution image to obtain the corresponding image domain volume data.

[0109] In one specific implementation, the image domain enhancement network model includes a feature extraction layer, a residual learning layer, and an upsampling layer. The feature extraction layer comprises e convolutional layers, the residual learning layer comprises f residual blocks, and the upsampling layer comprises g upsampled blocks. Here, e, f, and g are all positive integers. For example, e is 4, f is 64, and g is 2. For instance, the feature extraction layer may be a 4-layer 3x3 convolutional layer, the residual learning layer may comprise 64 residual blocks, and the upsampling layer may comprise 2 upsampled blocks.

[0110] Furthermore, when training this image domain augmentation network model, the loss function used can be a second loss function. This second loss function is correlated with L1 and SSIM.

[0111] The L1 loss, also known as the Mean Absolute Error (MAE), measures the average of the absolute differences between the predicted and true values. For image data, assuming we have the image output after processing by this image domain augmentation network and the corresponding true high-resolution image, L1 can be obtained as follows: first, calculate the absolute difference between the predicted and true values ​​for each corresponding pixel, and then average these absolute differences across all pixels. L1 can preserve strong signals and reduce sensitivity to outliers. The SSIM loss strengthens the constraints on structural features such as contours, avoiding structural distortion. Therefore, considering the second loss function when training this image domain augmentation network model can enhance the processing effect, resulting in a higher resolution target image.

[0112] This application can also perform super-resolution processing in other ways, and the embodiments of this application are not limited to these methods.

[0113] This application provides an image data processing method. First, an image to be processed is acquired. The image to be processed includes multiple projected images, each obtained by scanning a target object using SPECT within a sparse angular range at a first interval angle. The first interval angle is greater than a second interval angle for scanning the target object across the entire angular range. The image to be processed is then denoised to obtain a denoised image. The denoised image is then reconstructed using tomography to obtain image domain volume data. This image domain volume data is then denoised again to obtain denoised image domain volume data. Finally, the denoised image domain volume data is subjected to super-resolution processing to obtain a target image. The resolution of the target image is higher than that of the image to be processed. This application utilizes SPECT and sparse angular scanning to reduce acquisition time and improve scanning efficiency. However, sparse angular scanning inevitably leads to a decrease in image quality. Therefore, this application improves the image quality of the scanned image by performing projection domain preprocessing, image domain denoising, and image domain super-resolution processing. Thus, the image data processing method provided by this application can improve scanning speed and reduce scanning time while ensuring image quality.

[0114] In addition, this application also provides an image data processing apparatus.

[0115] Appendix Figure 4 This is a schematic diagram of an image data processing apparatus provided in an embodiment of this application. The apparatus 400 is used for SPECT imaging and includes:

[0116] The acquisition unit 401 is used to acquire an image to be processed; the image to be processed includes multiple frames of projection images, each frame of projection image is a projection image obtained by scanning the target object according to a first interval angle within a sparse angle range using SPECT; the first interval angle is greater than a second interval angle for scanning the target object within the full angle range.

[0117] The noise reduction processing unit 402 is used to perform noise reduction processing on the image to be processed to obtain a noise-reduced image; to perform tomographic reconstruction on the noise-reduced image to obtain image domain volume data; and to perform noise reduction processing on the image domain volume data to obtain noise-reduced image domain volume data.

[0118] The super-resolution unit 403 is used to perform super-resolution processing on the denoised image domain volume data to obtain a target image; the resolution of the target image is higher than the resolution of the image to be processed.

[0119] The noise reduction processing unit 402 is specifically used to: perform noise reduction processing on the image to be processed using a projection domain network model to obtain a noise-reduced image; the projection domain network model includes a residual U-shaped network model, which is a network model that introduces residual blocks on the Unet structure.

[0120] Optionally, the projection domain network model is a network model trained using the first training samples. The first training samples include label data and noise data of the simulated projection image. The method for obtaining the first training samples includes:

[0121] Obtain digital human body models with different human body characteristics; the digital human body models are used to simulate the distribution of radioactive tracers in the human body;

[0122] The digital human body model is input into the target simulation platform to obtain a simulated projection image. The target simulation platform is equipped with multi-bed tomographic data scanning data. After receiving the digital human body model, it rotates 360 degrees at each bed position and scans the digital human body model according to the second interval angle to obtain the simulated projection image.

[0123] A first Poisson noise is added to the simulated projection image to obtain the label data of the simulated projection image; a second Poisson noise is added to the simulated projection image to obtain the noise data of the simulated projection image; the first Poisson noise is less than the second Poisson noise.

[0124] The noise reduction processing unit 402 is further configured to: use an image domain noise reduction network model to perform noise reduction processing on the image domain volume data to obtain noise-reduced image domain volume data; the image domain noise reduction network model includes a ResUnet network model, wherein the ResUnet network model indicates a network model that introduces residual blocks on the Unet structure.

[0125] Optionally, the image domain denoising network model is a network model trained based on the second training samples. The label data of the second training samples is full-angle image domain volume data within the full-angle scanning range, and the noise data of the second training samples is sparse image domain volume data within the sparse angle scanning range. The method for obtaining the second training samples includes:

[0126] Tomographic reconstruction is performed using the full-angle projection data obtained by scanning the full-angle range of the SPECT to obtain the corresponding full-angle image domain data.

[0127] The corresponding full-angle image domain volume data is stitched together according to the position information of multiple beds in the target direction to obtain the stitched full-angle image domain volume data, and the full-angle image domain volume data is used as the label data of the second training sample;

[0128] The sparse angle projection data obtained by scanning within the sparse angle range of the SPECT is used to perform tomographic reconstruction, and the corresponding sparse image domain volume data is obtained.

[0129] The corresponding sparse image domain data is stitched together according to the position information of multiple beds in the target direction to obtain stitched sparse image domain data. The stitched sparse image domain data is used as the noise data of the second training sample. The target direction indicates the direction of the target object from head to toe.

[0130] Optionally, the ResUnet network model includes:

[0131] The system comprises an n-layer encoder, an m-layer bottleneck layer, and a p-layer decoder; the encoder includes a first residual block and a pooling layer, the bottleneck layer includes a second residual block, and the decoder includes an upsampling module and a third residual block, wherein the upsampling module and the third residual block communicate via a skip connection; n is a positive integer, m is a positive integer, and m is a positive integer.

[0132] The step of performing tomographic reconstruction on the denoised image to obtain image domain volume data includes:

[0133] The ordered subset expectation-maximization algorithm (OSEM) is used to perform tomographic reconstruction on the denoised image to obtain image domain volume data.

[0134] Optionally, the super-resolution unit 403 is specifically used to: perform super-resolution processing on the denoised image domain volume data using an image domain enhancement network model to obtain the target image;

[0135] The image domain enhancement network model is an enhancement network model trained on the third training sample. The third training sample includes positive samples and negative samples. The positive samples include the label image and the image domain volume data corresponding to the label image. The negative samples include the low-resolution image and the image domain volume data corresponding to the low-resolution image. The label image is an image obtained by adding the first Poisson noise to the full-angle simulated projection image. The low-resolution image is an image obtained by adding the first Poisson noise to the sparse-angle simulated projection image.

[0136] Optionally, the image domain enhancement network model includes a feature extraction layer, a residual learning layer, and an upsampling layer; the feature extraction layer includes four 3x3 convolutional layers, the residual learning layer includes 64 residual blocks, and the upsampling layer includes two upsampling layers.

[0137] This application provides an image data processing apparatus that utilizes SPECT to reduce acquisition time and improve scanning efficiency through sparse angle scanning. However, sparse angle scanning inevitably leads to a decrease in image quality. Therefore, this application improves the image quality of the scanned image by performing projection domain preprocessing, image domain noise reduction, and image domain super-resolution processing. Thus, the image data processing method provided by this application can increase scanning speed and reduce scanning time while ensuring image quality.

[0138] According to the method provided in the embodiments of this application, this application also provides a chip system, which includes one or more processors for calling and executing instructions stored in memory, thereby causing the method described in the embodiments of this application to be executed. The chip system may be composed of chips or may include chips and other discrete devices.

[0139] The chip system may include input circuits or interfaces for transmitting information or data, and output circuits or interfaces for receiving information or data.

[0140] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0141] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0142] The computer-readable storage medium may be the aforementioned volatile memory or non-volatile memory, or it may include both volatile memory and non-volatile memory.

[0143] In the embodiments of this application, the terms and English abbreviations are exemplary examples given for ease of description and should not be construed as limiting the application in any way. This application does not preclude the possibility of defining other terms that can achieve the same or similar functions in existing or future agreements.

[0144] In the embodiments of this application, the terms and English abbreviations are exemplary examples given for ease of description and should not be construed as limiting the application in any way. This application does not preclude the possibility of defining other terms that can achieve the same or similar functions in existing or future agreements.

[0145] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated.

[0146] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

Claims

1. An image data processing method, characterized by, Applied to SPECT imaging, the method includes: Obtain the image to be processed; The image to be processed includes multiple frames of projection images. Each frame of projection image is obtained by scanning the target object using SPECT within a sparse angular range at a first interval angle. The first interval angle is greater than the second interval angle when scanning the target object within the full angular range. The image to be processed is subjected to noise reduction processing to obtain a noise-reduced image; The denoised image is then subjected to tomographic reconstruction to obtain image domain volume data. The image domain volume data is denoised to obtain denoised image domain volume data. An image domain enhancement network model is used to perform super-resolution processing on the denoised image domain volume data to obtain a target image; the resolution of the target image is higher than the resolution of the image to be processed. The image domain enhancement network model is an enhancement network model trained on the third training sample. The third training sample includes positive samples and negative samples. The positive samples include a label image and the corresponding image domain volume data. The negative samples include a low-resolution image and the corresponding image domain volume data. The label image is an image obtained by adding the first Poisson noise to a full-angle simulated projection image. The low-resolution image is an image obtained by adding the first Poisson noise to a sparse-angle simulated projection image.

2. The processing method according to claim 1, characterized in that, The step of denoising the image to be processed to obtain a denoised image includes: The image to be processed is denoised using a projection domain network model to obtain a denoised image; the projection domain network model includes a residual U-shaped network model, which is a network model that introduces residual blocks on a Unet structure.

3. The processing method according to claim 2, characterized in that, The projection domain network model is a network model trained using the first training samples. The first training samples include label data and noise data of the simulated projection image. The method for obtaining the first training samples includes: Obtain digital human body models with different human body characteristics; the digital human body models are used to simulate the distribution of radioactive tracers in the human body; The digital human body model is input into the target simulation platform to obtain a simulated projection image. The target simulation platform is equipped with multi-bed tomographic data scanning data. After receiving the digital human body model, it rotates 360 degrees at each bed position and scans the digital human body model according to the second interval angle to obtain the simulated projection image. A first Poisson noise is added to the simulated projection image to obtain the label data of the simulated projection image; a second Poisson noise is added to the simulated projection image to obtain the noise data of the simulated projection image; the first Poisson noise is less than the second Poisson noise.

4. The processing method according to claim 1, characterized in that, The step of denoising the image domain volume data to obtain denoised image domain volume data includes: The image domain volume data is denoised using an image domain denoising network model to obtain denoised image domain volume data. The image domain denoising network model includes a ResUnet network model, which refers to a network model that introduces residual blocks on the Unet structure.

5. The processing method according to claim 4, characterized in that, The image domain denoising network model is a network model trained based on the second training sample. The label data of the second training sample is the full-angle image domain volume data within the full-angle scanning range, and the noise data of the second training sample is the sparse image domain volume data within the sparse angle scanning range. The methods for obtaining the second training samples include: Tomographic reconstruction is performed using the full-angle projection data obtained by scanning the full-angle range of the SPECT to obtain the corresponding full-angle image domain data. The corresponding full-angle image domain volume data is stitched together according to the position information of multiple beds in the target direction to obtain the stitched full-angle image domain volume data, and the full-angle image domain volume data is used as the label data of the second training sample; The sparse angle projection data obtained by scanning within the sparse angle range of the SPECT is used to perform tomographic reconstruction, and the corresponding sparse image domain volume data is obtained. The corresponding sparse image domain data is stitched together according to the position information of multiple beds in the target direction to obtain stitched sparse image domain data. The stitched sparse image domain data is used as the noise data of the second training sample. The target direction indicates the direction of the target object from head to toe.

6. The processing method according to claim 4, characterized in that, The ResUnet network model includes: The system comprises an n-layer encoder, an m-layer bottleneck layer, and a p-layer decoder; the encoder includes a first residual block and a pooling layer, the bottleneck layer includes a second residual block, and the decoder includes an upsampling module and a third residual block, wherein the upsampling module and the third residual block communicate via a skip connection; n is a positive integer, m is a positive integer, and m is a positive integer.

7. The processing method according to claim 1, characterized in that, The step of performing tomographic reconstruction on the denoised image to obtain image domain volume data includes: The ordered subset expectation-maximization algorithm (OSEM) is used to perform tomographic reconstruction on the denoised image to obtain image domain volume data.

8. The processing method according to claim 1, characterized in that, The image domain enhancement network model includes a feature extraction layer, a residual learning layer, and an upsampling layer; the feature extraction layer includes four 3x3 convolutional layers, the residual learning layer includes 64 residual blocks, and the upsampling layer includes two upsampling layers.

9. An image data processing apparatus, characterized in that, For use in SPECT imaging, the device includes: An acquisition unit is used to acquire an image to be processed; the image to be processed includes multiple frames of projection images, each frame of projection image is a projection image obtained by scanning the target object at a first interval angle within a sparse angular range using SPECT; the first interval angle is greater than a second interval angle for scanning the target object within the full angular range. The noise reduction processing unit is used to perform noise reduction processing on the image to be processed to obtain a noise-reduced image; to perform tomographic reconstruction on the noise-reduced image to obtain image domain volume data; and to perform noise reduction processing on the image domain volume data to obtain noise-reduced image domain volume data. The super-resolution unit is used to perform super-resolution processing on the denoised image domain volume data using an image domain enhancement network model to obtain a target image; the resolution of the target image is higher than the resolution of the image to be processed. The image domain enhancement network model is an enhancement network model trained on the third training sample. The third training sample includes positive samples and negative samples. The positive samples include a label image and the corresponding image domain volume data. The negative samples include a low-resolution image and the corresponding image domain volume data. The label image is an image obtained by adding the first Poisson noise to a full-angle simulated projection image. The low-resolution image is an image obtained by adding the first Poisson noise to a sparse-angle simulated projection image.

Images