Medical image processing methods, devices, medical imaging equipment and storage media

Abstract

This invention discloses a medical image processing method, apparatus, medical imaging device, and storage medium. First, a target image sequence including the medical image to be processed is acquired. Then, feature extraction is performed on multiple consecutive frames of medical images within the target image sequence to obtain sequence association features characterizing the spatial and / or temporal correlations between medical images within the target image sequence. Finally, based on the sequence association features, a target medical image corresponding to the medical image to be processed is generated. This achieves image quality optimization of the medical image to be processed through correlation information (such as spatial and / or temporal correlation information) between target image sequences, resulting in high-quality medical images and reducing dependence on high-dose radiation during imaging.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a medical image processing method, apparatus, medical imaging device, and storage medium. Background Technology

[0002] Computed tomography (CT) technology can visualize information about a patient's internal tissue structure through images.

[0003] In related technologies, the patient's tissue of interest is scanned repeatedly to obtain medical images of the corresponding tissue of interest, and the abnormalities in the tissue of interest are determined by the obtained medical images.

[0004] However, the methods used in these technologies to obtain medical images rely on high doses of radiation, which require improvement. Summary of the Invention

[0005] The embodiments described in this specification aim to at least partially solve one of the technical problems in the related art. To this end, the embodiments of this specification propose a medical image processing method, apparatus, medical imaging device, and storage medium.

[0006] This specification provides a medical image processing method, the method comprising:

[0007] Acquire the target image sequence, which includes the medical images to be processed;

[0008] Feature extraction is performed on multiple consecutive medical images within the target image sequence to obtain the sequence association features of the target image sequence; wherein, the sequence association features are used to characterize the spatial correlation and / or temporal correlation between medical images within the target image sequence;

[0009] Based on the sequence association features, a target medical image corresponding to the medical image to be processed is generated.

[0010] In one implementation, the target image sequence includes a spatially continuous image sequence corresponding to a region of interest; the step of extracting features from multiple consecutive medical images within the target image sequence to obtain sequence association features of the target image sequence includes:

[0011] Feature extraction is performed on multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial association features of the region of interest in the spatial dimension; wherein, the sequence association features include the spatial association features.

[0012] In one implementation, the medical image to be processed corresponds to a specified scan position; the target image sequence includes a temporally continuous image sequence corresponding to the specified scan position; the step of extracting features based on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence includes:

[0013] Feature extraction is performed on multiple time-continuous medical images within the time-continuous image sequence to obtain the temporal correlation features of the specified scan location in the time dimension; wherein, the sequence correlation features include the temporal correlation features.

[0014] In one implementation, the medical image to be processed corresponds to a specified scanning location; the target image sequence includes a spatially continuous image sequence corresponding to a region of interest and a temporally continuous image sequence corresponding to the specified scanning location; the step of extracting features based on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence includes:

[0015] Feature extraction is performed on multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial association features of the tissue of interest in the spatial dimension.

[0016] Feature extraction is performed on multiple time-continuous medical images within the time-continuous image sequence to obtain the temporal correlation features of the specified scan location in the time dimension.

[0017] The spatial correlation features and the temporal correlation features are fused to obtain the sequence correlation features.

[0018] In one implementation, the sequence association features are obtained by any of the following methods: extracting optical flow information from multiple consecutive medical images within the target image sequence to obtain optical flow map data, and determining the sequence association features based on the optical flow map data; or inputting multiple consecutive medical images within the target image sequence into a target feature extraction network for feature extraction to obtain the sequence association features.

[0019] In one implementation, generating the target medical image corresponding to the medical image to be processed based on the sequence association features includes:

[0020] The sequence association features are input into a target symmetric network for encoding and decoding to obtain the target medical image; wherein the imaging dose corresponding to the medical image to be processed is less than the imaging dose corresponding to the target medical image.

[0021] In one implementation, the target image sequence is extracted from an initial image sequence obtained by multiple scans of a region of interest, wherein the medical image to be processed corresponds to a specified scan location; the target image sequence is obtained through any of the following methods:

[0022] Using the medical image to be processed as a reference, multiple frames of medical images that correspond to the spatial structure of the tissue region of interest and are spatially continuous are extracted from the initial image sequence in the spatial dimension to construct the target image sequence;

[0023] Using the medical image to be processed as a reference, multiple frames of medical images corresponding to any scanning position and being temporally continuous are extracted from the initial image sequence in the time dimension to construct the target image sequence;

[0024] If the medical image to be processed is located at the edge of the initial image sequence, the medical image to be processed is copied, and the copied medical image is used as the edge image of the target image sequence.

[0025] In one implementation, the step of extracting features from multiple consecutive medical images within the target image sequence to obtain sequence association features of the target image sequence includes:

[0026] Motion detection is performed on any two adjacent frames of medical images within the target image sequence to obtain first motion evaluation data;

[0027] If the first motion evaluation data indicates that motion occurs between two adjacent medical images, the two adjacent medical images are registered to obtain a registered image sequence.

[0028] Feature extraction is performed on multiple consecutive frames of registered medical images within the registered image sequence to obtain the sequence association features.

[0029] In one implementation, the two adjacent medical images include a first medical image and a second medical image; the motion detection performed on any two adjacent medical images within the target image sequence to obtain first motion evaluation data includes:

[0030] The first medical image is binarized to segment out the bone tissue, thus obtaining the first bone tissue image;

[0031] The second medical image is binarized to segment out the bone tissue, thus obtaining a second bone tissue image;

[0032] The first motion assessment data is obtained by performing cross-union ratio calculation based on the first bone tissue image and the second bone tissue image.

[0033] In one embodiment, the target image sequence is extracted from an initial image sequence obtained by scanning a region of interest (ROI) multiple times. The ROI includes multiple scanning locations, and the image sequences obtained from two adjacent scans are denoted as the first image sequence and the second image sequence. The method further includes:

[0034] Determine the sequence motion evaluation data between the first image sequence and the second image sequence;

[0035] If the sequence motion evaluation data indicates that motion occurs between the first image sequence and the second image sequence, the first image sequence and the second image sequence are registered to obtain a registered image sequence.

[0036] Feature extraction is performed on multiple consecutive frames of registered medical images within the registered image sequence to obtain the sequence association features.

[0037] In one implementation, determining the sequence motion evaluation data between the first image sequence and the second image sequence includes:

[0038] The medical image group corresponding to each scanning position is constructed based on the medical images corresponding to each scanning position in the first image sequence and the second image sequence, respectively.

[0039] Motion detection is performed on the medical image group corresponding to each scanning position to obtain the second motion evaluation data corresponding to each scanning position;

[0040] The sequence motion evaluation data is obtained by averaging the second motion evaluation data corresponding to each scanning position.

[0041] In one implementation, the target medical image is obtained by encoding and decoding the sequence association features through a target image generation network; the target image generation network is trained in the following manner:

[0042] First biodata obtained by scanning with a first imaging dose; wherein, medical images in a reference image sequence are reconstructed from the first biodata as labels;

[0043] The first imaging data is simulated and processed to obtain the second imaging data corresponding to the second imaging dose; wherein the first imaging dose is greater than the second imaging dose;

[0044] Image sequence samples are constructed based on the image sequences reconstructed from the second biological data.

[0045] The initial image generation network is trained using the image sequence samples and the corresponding labels to obtain the target image generation network; wherein the labels are medical images from the baseline image sequence reconstructed based on the first image data.

[0046] This specification provides a medical image processing apparatus, the apparatus comprising:

[0047] The target image sequence acquisition module is used to acquire a target image sequence including the medical image to be processed;

[0048] The sequence association feature extraction module is used to extract features based on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence; wherein, the sequence association features are used to characterize the spatial correlation and / or temporal correlation between medical images within the target image sequence;

[0049] The target medical image generation module is used to generate a target medical image corresponding to the medical image to be processed based on the sequence association features.

[0050] This specification provides a medical imaging device, which includes: a memory, and one or more processors communicatively connected to the memory; the memory stores instructions executable by the one or more processors, which, when executed by the one or more processors, cause the one or more processors to perform the steps of the method described in any of the above embodiments.

[0051] This specification provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the above embodiments.

[0052] This specification provides a computer program product that includes instructions that, when executed by a processor of a computer device, enable the computer device to perform the steps of the method described in any of the above embodiments.

[0053] In the above-described embodiment, firstly, a target image sequence including the medical image to be processed is acquired. Then, feature extraction is performed on multiple consecutive frames of medical images within the target image sequence to obtain sequence association features characterizing the spatial and / or temporal correlations between medical images within the target image sequence. Finally, based on the sequence association features, a target medical image corresponding to the medical image to be processed is generated. This achieves image quality optimization of the medical image to be processed using spatial and / or temporal correlation information in the target image sequence, resulting in high-quality medical images and reducing dependence on high-dose radiation during imaging. Attached Figure Description

[0054] Figure 1a A schematic diagram illustrating the medical image processing method provided in the embodiments of this specification;

[0055] Figure 1b A schematic flowchart illustrating the medical image processing method provided in the embodiments of this specification;

[0056] Figure 2a A schematic diagram illustrating the implementation of the spatial feature extraction network provided for the embodiments of this specification;

[0057] Figure 2b A schematic diagram illustrating the implementation of the time feature extraction network provided in the embodiments of this specification;

[0058] Figure 2c A flowchart illustrating the process of obtaining sequence association features provided for embodiments of this specification;

[0059] Figure 2d A schematic diagram of a time and space feature extraction network provided for the implementation of this specification;

[0060] Figure 2e A schematic diagram of the target symmetric network provided for the embodiments of this specification;

[0061] Figure 3a A schematic diagram of a medical image sequence provided for an embodiment of this specification;

[0062] Figure 3b A schematic diagram of the target image sequence provided for the embodiments of this specification;

[0063] Figure 3c A schematic diagram of the target image sequence provided for the embodiments of this specification;

[0064] Figure 4 A flowchart illustrating the process of obtaining sequence association features provided for embodiments of this specification;

[0065] Figure 5a A flowchart illustrating the process of obtaining the first motion evaluation data provided for the embodiments of this specification;

[0066] Figure 5b A schematic diagram of a first bone tissue image provided for an embodiment of this specification;

[0067] Figure 5c A schematic diagram of a second bone tissue image provided for an embodiment of this specification;

[0068] Figure 5d A schematic diagram of the intersection image provided for embodiments of this specification;

[0069] Figure 5e A schematic diagram of a union image provided for embodiments of this specification;

[0070] Figure 5f A schematic diagram of a first medical image provided for an embodiment of this specification;

[0071] Figure 5g A schematic diagram of a second medical image provided for embodiments of this specification;

[0072] Figure 5h A schematic diagram of a first bone tissue image provided for an embodiment of this specification;

[0073] Figure 5i A schematic diagram of a second bone tissue image provided for an embodiment of this specification;

[0074] Figure 6 A flowchart illustrating the process of obtaining sequence association features provided for embodiments of this specification;

[0075] Figure 7 A flowchart illustrating the process of obtaining sequence motion evaluation data provided in this specification for embodiments;

[0076] Figure 8 A schematic diagram of the process for obtaining the target image generation network provided in the embodiments of this specification;

[0077] Figure 9 A schematic flowchart illustrating the medical image processing method provided in the embodiments of this specification;

[0078] Figure 10 A schematic diagram of a medical image processing apparatus provided for embodiments of this specification. Detailed Implementation

[0079] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0080] Computed tomography (CT) is an important technique for visualizing the internal structures of a patient's body. Taking X-rays as an example, during a CT scan, a high-dose X-ray emitting device surrounds the patient's body. Human tissue absorbs the X-rays, causing them to attenuate. A detector on the other side of the body receives the attenuated X-rays, and then a reconstruction algorithm is used to obtain a reconstructed image of the scanned area.

[0081] Generally, the higher the X-ray dose delivered to a patient, the higher the quality of the reconstructed image. However, X-rays are radiation, and excessive exposure can cause irreversible damage to the human body. Therefore, it is necessary to reconstruct high-quality images based on low-dose scanning procedures.

[0082] Contrast-enhanced perfusion scanning involves injecting a contrast agent into the patient and then repeatedly scanning a specific area. By observing the flow of the contrast agent within the blood vessels, the presence of abnormally supplied tissue in the scanned area can be determined. Because this requires repeated scans over a long period, patients face a significant risk of exposure to high doses of X-rays.

[0083] In related technologies, plain scan images can be used as prior images, and a prior image-constrained diffusion tensor can be constructed. A brain perfusion CT reconstruction model can then be solved using an alternating optimization algorithm to obtain brain perfusion CT images. Reducing radiation dose through sparse sampling requires pre-acquiring a non-sparse baseline image sequence. This is used to manually extract contrast-enhanced image sequences and a prior matrix from the sparsely sampled perfusion image sequence. The contrast-enhanced images are then reconstructed using an iterative reconstruction method, and finally, the optimized image is added to the baseline image. However, these technologies require high-dose medical images as reference images, and the image optimization requires iterative reconstruction methods, meaning the time spent on image reconstruction needs to be reduced.

[0084] Low-dose perfusion imaging is a technique that can reconstruct high-quality images based on low-level radiation dose scans. During perfusion scanning, because multiple scans of the same location are required, the data obtained from each scan are highly correlated. To address this characteristic, a medical image processing method is proposed that can obtain high-quality perfusion images under low-dose scanning conditions.

[0085] Based on this, firstly, a target image sequence including the medical image to be processed is acquired. Then, feature extraction is performed on multiple consecutive frames of medical images within the target image sequence to obtain sequence association features characterizing the spatial and / or temporal correlations between medical images within the target image sequence. Finally, based on the sequence association features, the target medical image corresponding to the medical image to be processed is generated. This achieves image quality optimization of the medical image to be processed using correlation information (such as spatial and / or temporal correlation information) between target image sequences to obtain high-quality medical images, thereby reducing dependence on high-dose radiation during imaging.

[0086] This specification provides an example of a medical image processing scenario. Perfusion scanning requires multiple repeated scans of a fixed area. The scanning method can be 3D scanning or 4D reciprocating scanning. The interval between each scan is generally about 1 second, and a total of about 20 to 40 scans are performed. After scanning, 20 to 40 sets of human three-dimensional data arranged in chronological order can be obtained. For blood flow information, it is constantly changing in these sequences. For other tissues that are not in the blood flow range, their morphology can be approximated as remaining basically unchanged throughout the scanning process, and they have a very strong similarity to each other.

[0087] Please see Figure 1a Low-dose perfusion scan data is constructed using FBP (Filtered Backprojection Algorithm) to maintain its spatial and temporal order, resulting in an initial image sequence. Medical images within this initial image sequence exhibit strong correlations in both spatial and temporal dimensions. Centering on a specific frame from the initial image sequence, consecutive medical images of a certain length are selected from both the spatial and temporal sequences contained within the initial sequence, and these selected images are extracted to obtain spatial and temporal image sequences. Motion detection is performed on both the spatial and temporal image sequences. If motion exists between medical images in the spatial image sequence, image registration is performed to ensure the continuity of the spatial image sequence input into the spatiotemporal dual-domain network, preventing abrupt changes that could affect network processing, thus obtaining a spatially registered image sequence. Similarly, if motion exists between medical images in the temporal image sequence, image registration is performed to obtain a temporally registered image sequence. The temporally and spatially registered image sequences are then input into the spatiotemporal dual-domain network for information fusion and image quality optimization, outputting a high-quality image corresponding to the central image.

[0088] This specification provides a medical image processing method. Please refer to [link / reference]. Figure 1b The medical image processing method may include the following steps:

[0089] S110. Obtain the target image sequence, which includes the medical images to be processed.

[0090] S120. Based on multiple consecutive frames of medical images within the target image sequence, feature extraction is performed to obtain the sequence association features of the target image sequence.

[0091] S130. Generate the target medical image corresponding to the medical image to be processed based on the sequence association features.

[0092] Sequence association features are used to characterize the spatial and / or temporal correlations between medical images within the target image sequence. The medical image to be processed can be any frame from the initial image sequence obtained by scanning the region of interest. The target medical image can be an optimized, high-quality image. The target medical image can be an image that improves the perfusion parameters (such as the TDC curve) of the medical image to be processed, or an image that reduces noise in the medical image to be processed. The target image sequence can be a sequence of multiple frames of medical images extracted from the initial image sequence according to pre-set extraction requirements for feature extraction.

[0093] Specifically, the region of interest is scanned to obtain an initial image sequence. The medical image to be processed is determined from the initial image sequence. Based on the medical image to be processed, multiple frames of medical images including the medical image to be processed can be selected from the initial image sequence to obtain a target image sequence including the medical image to be processed. Alternatively, multiple frames of medical images including the medical image to be processed can be pre-extracted from the initial image sequence and stored locally on the computer. When feature extraction is required, the required multiple frames of medical images can be read from the local memory of the computer and recorded as the target image sequence. Feature extraction is performed on consecutive multiple frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence. Based on the sequence association features, the target medical image corresponding to the medical image to be processed is generated.

[0094] In some implementations, a high-precision image sequence corresponding to the target image sequence is generated based on sequence association features. Based on the position of the medical image to be processed within the target image sequence, medical images at the same position in the high-precision image sequence can be used as the target medical image. In other implementations, the medical image to be processed can be placed at a fixed position within the target image sequence, and the target image generation network is trained based on the target image sequence including the medical image to be processed at the fixed position. The target image sequence is used as input to the target image generation network to output the target medical image corresponding to the medical image to be processed.

[0095] In the above implementation, firstly, a target image sequence including the medical image to be processed is acquired. Then, feature extraction is performed on multiple consecutive frames of medical images within the target image sequence to obtain sequence association features characterizing the spatial and / or temporal correlations between medical images within the target image sequence. Finally, based on the sequence association features, a target medical image corresponding to the medical image to be processed is generated. This achieves high-quality medical image optimization of the medical image to be processed by utilizing correlation information (such as spatial and / or temporal correlation information) between target image sequences, thereby reducing dependence on high-dose radiation during imaging.

[0096] In some implementations, the target image sequence includes a spatially continuous image sequence corresponding to the region of interest. Feature extraction is performed based on multiple consecutive medical images within the target image sequence to obtain sequence association features of the target image sequence, including: feature extraction based on multiple spatially continuous medical images within the spatially continuous image sequence to obtain spatial association features of the region of interest in the spatial dimension.

[0097] Among these, sequence association features include spatial association features. Spatial association features can be features extracted from image data considering the spatial relationships between adjacent data points. Spatial association features mainly include the spatial structural information of tissues, such as the three-dimensional shape of an organ or lesion in space. Taking a target image sequence containing an aneurysm as an example, the aneurysm can be a saccular aneurysm, fusiform aneurysm, tortuous aneurysm, or scaphoid aneurysm. The combination of the contour information of the aneurysms contained in each image of the target image sequence should conform to the shape characteristics of the aneurysm. This combination of contour information of the aneurysms contained in each image conforming to the shape characteristics of the aneurysm can be understood as spatial association features.

[0098] Specifically, when the target image sequence includes a spatially continuous image sequence corresponding to the region of interest, feature extraction is performed on multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial association features of the region of interest in the spatial dimension.

[0099] For example, feature extraction based on multiple consecutive frames of medical images within a target image sequence can be achieved using a spatial feature extraction network. See also... Figure 2a The spatially continuous image sequence 202 is used as the input of the spatial feature extraction network 204. The spatial feature extraction network 204 extracts features from multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial association features 206 of the tissue region of interest in the spatial dimension.

[0100] In some implementations, before inputting the spatially continuous image sequence into the spatial feature extraction network, special processing can be performed on the spatially continuous multiple frames of medical images within the spatially continuous image sequence, such as sequence truncation, sequence edge image duplication, etc.

[0101] In the above embodiments, feature extraction is performed on multiple spatially continuous medical images within a spatially continuous image sequence to obtain the spatial correlation features of the tissue region of interest in the spatial dimension. This can provide input data for the target symmetric network in order to obtain high-quality medical images.

[0102] In some implementations, the medical image to be processed corresponds to a specified scan location. The target image sequence includes a temporally continuous image sequence corresponding to the specified scan location. Feature extraction is performed based on multiple consecutive frames of medical images within the target image sequence to obtain sequence association features of the target image sequence, including: feature extraction based on multiple temporally continuous frames of medical images within the temporally continuous image sequence to obtain temporal association features of the specified scan location in the time dimension.

[0103] Sequence correlation features include temporal correlation features. Temporal correlation features can be the correlations and patterns of change between different time points in a medical image sequence. By extracting and analyzing these temporal correlation features, the temporal information and dynamic evolution characteristics in medical images can be revealed. For example, the change of contrast agent in a blood vessel over time can be represented on the image as an increase in pixel value (or CT value) as time increases within a certain time range.

[0104] Specifically, when the target image sequence includes a time-continuous image sequence corresponding to a specified scanning position, feature extraction is performed on multiple time-continuous medical images within the time-continuous image sequence to obtain the time-related features of the specified scanning position in the time dimension.

[0105] For example, feature extraction based on multiple consecutive frames of medical images within a target image sequence can be achieved using a temporal feature extraction network. See also... Figure 2b The temporally continuous image sequence 208 is used as the input of the temporal feature extraction network 210. The temporally continuous image sequence is processed by the temporally continuous image extraction network 210 to extract features from multiple frames of medical images within the temporally continuous image sequence, thereby obtaining the temporal correlation features 212 of the specified scan position in the temporal dimension.

[0106] In some implementations, before inputting the temporally continuous image sequence into the temporal feature extraction network, special processing can be performed on multiple temporally continuous medical images within the temporally continuous image sequence, such as sequence truncation, sequence edge image duplication, etc.

[0107] In the above embodiments, feature extraction is performed on multiple time-continuous medical images within a time-continuous image sequence to obtain the temporal correlation features of the specified scanning position in the time dimension. This can provide input data for the target symmetric network in order to obtain high-quality medical images.

[0108] In some implementations, please refer to Figure 2c The medical images to be processed correspond to a specified scanning location. The target image sequence includes a spatially continuous image sequence corresponding to the region of interest and a temporally continuous image sequence corresponding to the specified scanning location. Feature extraction is performed on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence, which may include the following steps:

[0109] S210. Based on multiple spatially continuous medical images within a spatially continuous image sequence, feature extraction is performed to obtain the spatial correlation features of the tissue region of interest in the spatial dimension.

[0110] S220. Based on multiple time-continuous medical images within a time-continuous image sequence, feature extraction is performed to obtain the temporal correlation features of the specified scan location in the time dimension.

[0111] S230. The spatial correlation features and temporal correlation features are fused to obtain the sequence correlation features.

[0112] Specifically, when the target image sequence includes a spatially continuous image sequence corresponding to the region of interest (ROI) and a temporally continuous image sequence corresponding to a specified scan position, feature extraction is performed on multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial correlation features of the ROI in the spatial dimension. Feature extraction is also performed on multiple temporally continuous medical images within the temporally continuous image sequence to obtain the temporal correlation features of the specified scan position in the temporal dimension. To fully explore the correlation between spatial and temporal data in the image sequence, the spatial and temporal correlation features can be fused to obtain sequence correlation features, thereby improving image quality.

[0113] For example, feature extraction based on multiple consecutive frames of medical images within a target image sequence can be achieved using a temporal feature extraction network, while feature extraction based on multiple consecutive frames of medical images within a target image sequence can be achieved using a spatial feature extraction network. Both the spatial and temporal feature extraction networks can be Siamese network structures, which have two identical inputs that receive spatially and temporally consecutive image sequences, respectively. Please refer to [link / reference]. Figure 2dThe spatially continuous image sequence 202 is used as input to the spatial feature extraction network 204. The spatial feature extraction network 204 extracts features from multiple spatially continuous medical images within the spatially continuous image sequence, obtaining spatial correlation features 206 of the region of interest in the spatial dimension. The temporally continuous image sequence 208 is used as input to the temporally continuous feature extraction network 210. The temporally continuous image extraction network 210 extracts features from multiple temporally continuous medical images within the temporally continuous image sequence, obtaining temporally correlated features 212 of the specified scan location in the temporal dimension. The spatial correlation features 206 and the temporally correlated features 212 are then fused to obtain the sequence correlation features 214.

[0114] In the above embodiments, feature extraction is performed on multiple spatially continuous medical images within a spatially continuous image sequence to obtain spatial correlation features of the region of interest in the spatial dimension. Feature extraction is also performed on multiple temporally continuous medical images within a temporally continuous image sequence to obtain temporal correlation features of a specified scan location in the temporal dimension. The spatial and temporal correlation features are then fused to obtain sequence correlation features. By fusing spatial and temporal correlation features, more spatiotemporal contextual information can be captured, resulting in more comprehensive and accurate information, and providing stronger support for complex decision-making and analysis tasks. This approach allows for better mining and utilization of spatial and temporal correlation information in image sequences, and also enables the acquisition of high image quality even at low doses.

[0115] In some implementations, sequence association features are obtained using any of the following methods: optical flow information is extracted from multiple consecutive frames of medical images within the target image sequence to obtain optical flow map data, and sequence association features are determined based on the optical flow map data. Alternatively, multiple consecutive frames of medical images within the target image sequence are input into a target feature extraction network for feature extraction to obtain sequence association features.

[0116] Optical flow information refers to the motion information of each pixel position in an image sequence, that is, the displacement of each pixel over time. Optical flow information can reflect the pixel motion patterns and velocities between adjacent frames. Optical flow map data can be image data representing motion patterns and velocities obtained by calculating optical flow information. It visually displays the direction and velocity of motion at each pixel position between adjacent frames.

[0117] Specifically, optical flow algorithms can be used to calculate optical flow in multiple consecutive frames of medical images within a target image sequence, obtaining optical flow information representing the direction of motion of the same objects in the target image sequence. This optical flow information can then be processed to obtain optical flow map data. For example, the direction and magnitude of the optical flow information can be visualized to obtain the optical flow map data. Feature extraction is then performed on the optical flow map data to obtain the sequence association features of the target image sequence. For example, the optical flow algorithm can be the Lucas-Kanade algorithm or the Horn-Schunck algorithm.

[0118] Training a network suitable for extracting sequence association features yields a target feature extraction network. Multiple consecutive frames of medical images within the target image sequence are input into the target feature extraction network for feature extraction, resulting in sequence association features of the target image sequence. For example, the target feature extraction network can be a symmetric network.

[0119] In some implementations, sequence association features include spatial association features. The target feature extraction network can be a spatial feature extraction network. See also... Figure 2a The spatially continuous image sequence 202 corresponding to the region of interest included in the target image sequence is used as the input of the spatial feature extraction network 204. The spatial feature extraction network 204 extracts features from multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial association features 206 of the region of interest in the spatial dimension.

[0120] In other implementations, sequence association features include temporal association features. The target feature extraction network can be a temporal feature extraction network. See also... Figure 2b The temporally continuous image sequence 208 corresponding to the region of interest included in the target image sequence is used as the input of the temporal feature extraction network 210. The temporally continuous image sequence is processed by the temporally continuous image extraction network 210 to extract features from multiple frames of medical images within the temporally continuous image sequence, thereby obtaining the temporal correlation features 212 of the specified scan position in the temporal dimension.

[0121] In the above embodiments, obtaining sequence association features in different representations through various methods can increase the information content and diversity of association features.

[0122] In some implementations, generating a target medical image corresponding to the medical image to be processed based on sequence association features includes: inputting the sequence association features into a target symmetric network for encoding and decoding to obtain the target medical image.

[0123] In this context, the imaging dose corresponding to the medical image to be processed is less than the imaging dose corresponding to the target medical image. A symmetric network is a neural network structure characterized by symmetrical encoding and decoding parts. It is commonly used for tasks such as image segmentation, image reconstruction, and generation to extract and restore features from the target image.

[0124] Specifically, the sequence association features are input into the encoder part of the target symmetric network for encoding processing, progressively extracting features and reducing their dimensionality. The features processed by the encoder are then input into the decoder part of the symmetric network for decoding processing, gradually restoring the shape and content of the target image sequence from the features extracted by the encoder, thus obtaining the target medical image.

[0125] For example, please refer to Figure 2e The target symmetric network can be a UNet network. Sequence association features 216 are used as input to the UNet network. After encoding and decoding by the UNet network, the target medical image 218 is obtained.

[0126] It should be noted that the target symmetric network is not limited to UNet networks, but can also be VNet networks, DenseNet networks, and other symmetric networks.

[0127] In the above embodiments, sequence association features are input into a target symmetric network for encoding and decoding to obtain the target medical image. The target symmetric network can perform deep fusion encoding of sequence association features and decode to output the final high-quality image.

[0128] In some implementations, the target image sequence is extracted from an initial image sequence obtained by multiple scans of the region of interest, with the medical image to be processed corresponding to a specified scan location. The target image sequence is obtained through any of the following methods:

[0129] Using the medical image to be processed as a reference, multiple spatially continuous medical images corresponding to the spatial structure of the tissue region of interest are extracted from the initial image sequence in the spatial dimension to construct the target image sequence.

[0130] Using the medical image to be processed as a reference, multiple frames of medical images corresponding to any scanning position and being temporally continuous are extracted from the initial image sequence in the time dimension to construct the target image sequence.

[0131] If the medical image to be processed is located at the edge of the initial image sequence, the medical image to be processed is copied, and the copied medical image is used as the edge image of the target image sequence.

[0132] Spatial dimension refers to the different directions representing the location and structure of the region of interest (ROI) in an image. Each dimension represents a coordinate axis that can be used to determine the position of each pixel or voxel in the image. Spatial structure refers to the position and orientation of the scanning device relative to the patient or the object being scanned during image acquisition. This spatial structure is crucial for understanding the relationship between structures and anatomical information in the image. Spatial continuity refers to the continuity and consistency in the position and orientation between adjacent slices in an image. Temporal dimension refers to the fact that in a medical image sequence, each image represents the observation at a different point in time. This means that each image in a medical image sequence corresponds to the changes or dynamic processes of pathological or anatomical structures at different points in time. Temporal continuity refers to the temporal order and continuity between image sequences acquired at different points in time for the same patient or object. The region of interest (ROI) can be set according to the actual examination site; for example, the ROI could be liver tissue, heart tissue, or brain tissue.

[0133] Specifically, the region of interest (ROI) is identified, and data is acquired from this region using a scanner or other medical imaging equipment to obtain initial raw data. Backprojection reconstruction is performed on the initial raw data, preserving its spatial and temporal order to obtain an initial image sequence. This initial image sequence contains continuous images of the ROI. Spatially, using the medical image to be processed as a reference, multiple frames of medical images are continuously extracted from the initial image sequence according to the actual situation and requirements, to preserve the spatial continuity and spatial structure of the ROI. Based on the extracted multiple frames of medical images that correspond to the spatial structure of the ROI and are spatially continuous, a target image sequence is constructed. For example, backprojection reconstruction can be implemented using an FBP module.

[0134] In the temporal dimension, using the medical image to be processed as a reference, multiple temporally continuous images corresponding to the scan position of interest are extracted from the initial image sequence according to the actual situation and requirements, so as to preserve the temporal continuity of the region of interest. Based on the extracted multiple temporally continuous images corresponding to the scan position of interest, a target image sequence is constructed.

[0135] When the medical image to be processed is the first frame of the initial image sequence, it is copied to obtain a copied image. This copied image is then placed in front of the original medical image and used as the edge image of the target image sequence. When the medical image to be processed is the last frame of the initial image sequence, it is copied again to obtain a copied image. This copied image is then placed behind the original medical image and used as the edge image of the target image sequence.

[0136] For example, please refer to Figure 3a A first scan of the region of interest yields five medical images, forming the first image sequence. These five images are numbered 1, 2, 3, 4, and 5. A second scan of the region of interest yields five more images, forming the second image sequence. These five images are numbered 6, 7, 8, 9, and 10. A third scan of the region of interest yields five more images, forming the third image sequence. These five images are numbered 11, 12, 13, 14, and 15. These image numbers can also be called image indices.

[0137] The number of image frames extracted can be set to 3 frames, depending on the requirements. Spatially, using the medical image to be processed as a reference, multiple spatially continuous medical images corresponding to the spatial structure of the region of interest can be extracted. For example: Please refer to... Figure 3b The image index of the medical image to be processed can be 2, and medical images with image indices 1, 2, and 3 can be extracted to construct the target image sequence. The image index of the medical image to be processed can be 8, and medical images with image indices 7, 8, and 9 can be extracted to construct the target image sequence.

[0138] The number of image frames extracted can be set to 3 frames, depending on the requirements. In terms of time, using the medical image to be processed as a reference, multiple consecutive medical images corresponding to any scan position are extracted. For example: Please refer to [link / reference]. Figure 3c The image index of the medical image to be processed can be 7. Medical images with image indices of 2, 7, and 12 can be extracted to construct the target image sequence.

[0139] In the spatial dimension, taking the medical image to be processed as a reference, when the medical image to be processed is located at the edge of the initial image sequence, for example, when the index of the medical image to be processed is 1, the medical image to be processed with index 1 can be copied to obtain the copied medical image 1'. The copied medical image 1' is placed in front of the processed medical image 1, and the medical images with image indices 1', 1, and 2 are extracted to construct the target image sequence. For example, when the index of the medical image to be processed is 5, the medical image to be processed with index 5 can be copied to obtain the copied medical image 5'. The copied medical image 5' is placed after the processed medical image 5, and the medical images with image indices 4, 5, and 5' are extracted to construct the target image sequence.

[0140] In terms of time dimension, taking the medical image to be processed as a reference, when the medical image to be processed is located at the edge of the initial image sequence, for example, when the index of the medical image to be processed is 2, the medical image to be processed at index 2 can be copied to obtain the copied medical image 2'. The copied medical image 2' is placed before the processed medical image 2, and the medical images with image indices 2', 2, and 7 are extracted to construct the target image sequence. For example, when the index of the medical image to be processed is 14, the medical image to be processed at index 14 can be copied to obtain the copied medical image 14'. The copied medical image 14' is placed after the processed medical image 14, and the medical images with image indices 9, 14, and 14' are extracted to construct the target image sequence.

[0141] It should be noted that each frame of the medical image in the initial image sequence contains location stamp information. The location stamp information is the same, and the scan location is the same.

[0142] In the above embodiments, a target image sequence is constructed to provide input data for the target image generation network in order to obtain high-quality medical images.

[0143] In some implementations, please refer to Figure 4 Feature extraction based on multiple consecutive frames of medical images within a target image sequence to obtain sequence association features of the target image sequence may include the following steps:

[0144] S410. Perform motion detection on any two adjacent frames of medical images within the target image sequence to obtain the first motion evaluation data.

[0145] S420. If the first motion evaluation data indicates that motion occurs between two adjacent medical images, the two adjacent medical images are registered to obtain a registered image sequence.

[0146] S430 extracts features from multiple consecutive registered medical images within a registered image sequence to obtain sequence-related features.

[0147] Motion detection refers to the detection and tracking of patient or organ movement within a continuous sequence of medical images. This is crucial for many medical applications, such as radiotherapy, cardiac function assessment, and surgical navigation.

[0148] In some cases, patient movement during an examination can cause shifts in the scan area of ​​interest. This shift can blur medical images, reducing image clarity and detail, making accurate diagnosis difficult. It can also lead to image loss or incomplete data acquisition, leaving doctors with insufficient information for accurate diagnosis. Significant patient movement during the scan can disrupt spatiotemporal continuity, further degrading the final image quality.

[0149] Specifically, two adjacent medical images are arbitrarily selected from the target image sequence, and motion detection is performed on these two adjacent medical images to obtain the first motion evaluation data. A motion detection threshold t is set. s For example, motion detection threshold t s It can be set to 0.8 or 0.9. When the first motion evaluation data is less than the motion detection threshold t... s When motion occurs between two adjacent medical images corresponding to the first motion evaluation data, it indicates that motion has occurred between them. When motion occurs between two adjacent medical images, registration processing is performed on the two adjacent medical images to obtain a registered image sequence. When the first motion evaluation data is greater than or equal to the motion detection threshold t... s When this is the case, it indicates that there is no significant motion between the two adjacent medical images corresponding to the first motion assessment data, and registration processing of the two adjacent medical images is unnecessary. Here, the motion detection threshold t... s It can be configured according to actual needs.

[0150] When the registered image sequence is a spatially continuous image sequence, the registered image sequence is used as the input of the spatial feature extraction network. The spatial feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain sequence association features.

[0151] When the registered image sequence is a temporally continuous image sequence, the registered image sequence is used as the input of the temporal feature extraction network. The temporal feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain sequence association features.

[0152] When the registered image sequence includes a temporally continuous image sequence and a spatially continuous image sequence, the registered image sequence is used as input to a spatial feature extraction network. The spatial feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain spatial correlation features. The registered image sequence is also used as input to a temporal feature extraction network. The temporal feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain temporal correlation features. The spatial correlation features and temporal correlation features are then fused to obtain sequence correlation features.

[0153] In some implementations, when motion occurs between two adjacent medical images, image processing can be performed on the two adjacent medical images to obtain the tilt angle of the region of interest. The region of interest in that frame of medical image is then corrected to achieve registration processing of the two adjacent medical images, resulting in a registered image sequence.

[0154] It should be noted that registration processing includes, but is not limited to, 2D registration, 3D registration, traditional method registration, and deep learning registration.

[0155] In the above embodiments, motion detection is performed on any two adjacent frames of medical images within the target image sequence to obtain first motion evaluation data. If the first motion evaluation data indicates that motion occurs between the two adjacent frames, registration processing is performed on the two adjacent frames to obtain a registered image sequence. Feature extraction is then performed on multiple consecutive registered medical images within the registered image sequence to obtain sequence association features. By obtaining the sequence association features, input data can be provided to the target symmetric network to obtain high-quality medical images.

[0156] In some implementations, please refer to Figure 5a Two adjacent medical images include a first medical image and a second medical image. Motion detection is performed on any two adjacent medical images within the target image sequence to obtain first motion evaluation data, which may include the following steps:

[0157] S510. Perform binarization processing on the first medical image to segment out the bone tissue and obtain the first bone tissue image.

[0158] S520. Perform binarization processing on the second medical image to segment out the bone tissue and obtain the second bone tissue image.

[0159] S530. The first motion assessment data is obtained by performing crossover ratio calculation based on the first bone tissue image and the second bone tissue image.

[0160] Binarization, in this context, involves converting grayscale levels in a medical image into two discrete numerical values, typically 0 and 1. Intersection over Union (IoU) calculation is a commonly used metric for evaluating image segmentation and object detection, measuring the degree of overlap between two regions. It can be used to compare the similarity between predicted results and ground truth annotations.

[0161] In some cases, when a patient moves, bone tissue is more stable and less prone to deformation compared to soft tissue containing contrast agents. Changes in bone position are more easily observed than in soft tissue containing contrast agents.

[0162] Specifically, first, a threshold t is set.b By setting a threshold t b The value of t is used to segment bone tissue in two adjacent frames of medical images. When the computed tomography (CT) value is greater than or equal to the threshold t, bone tissue is segmented. b When, it will be greater than or equal to the threshold t b The computed tomography (CT) value at that location is set to 1. This value is greater than or equal to the threshold t. b The area can be considered the location of bone tissue. When the computed tomography (CT) value is less than the threshold t... b When, it will be less than the threshold t b The computed tomography (CT) value at that location is set to 0. This is less than the threshold t. b The region can be considered the location of soft tissue. Each computed tomography (CT) value in the first medical image is binarized using the above method to segment the bone tissue, resulting in the first bone tissue image. Similarly, each computed tomography (CT) value in the second medical image is binarized using the above method to segment the bone tissue, resulting in the second bone tissue image. Then, the first and second bone tissue images are intersected pixel-by-pixel to obtain an intersection image. The first and second bone tissue images are then unioned pixel-by-pixel to obtain a union image. Next, the pixel values ​​at each location in the intersection image are summed to obtain the number of "1"s in the intersection image, denoted as vi. The pixel values ​​at each location in the union image are summed to obtain the number of "1"s in the union image, denoted as vu. The ratio of vi to vu is calculated to achieve the intersection-union ratio of the first and second bone tissue images, resulting in the first motion evaluation data, denoted as S. move S move =vi / vu. Where, S move The value range is [0, 1]. Wherein, the first motion evaluation data S... move The closer the value is to 0, the greater the amplitude of motion between two adjacent medical images. (First motion assessment data S) move The closer the threshold is to 1, the smaller the amplitude of motion between two adjacent medical image frames. Threshold t b It can be configured according to actual needs.

[0163] For example, please refer to Figure 5b and Figure 5c , Figure 5b This is recorded as the first bone tissue image. Figure 5c This is denoted as the second bone tissue image. The intersection of the first and second bone tissue images is then processed, i.e. Figure 5b In the first bone tissue image, pixel position 502 corresponds to pixel value 1 and... Figure 5cThe intersection of pixel position 504 and pixel value 1 in the second bone tissue image yields... Figure 5d The pixel at position 506 in the intersecting image has a value of 1. Figure 5b In the first bone tissue image, pixel position 508 corresponds to pixel value 1 and... Figure 5c The intersection of pixel position 510 and pixel value 0 in the second bone tissue image yields... Figure 5d The pixel at position 512 in the intersection image has a value of 0. Figure 5b In the first bone tissue image, pixel position 514 corresponds to pixel value 0 and... Figure 5c Taking the intersection of pixel position 516 and pixel value 1 in the second bone tissue image, we get... Figure 5d The pixel at position 518 in the intersection image has a value of 0. Figure 5b In the first bone tissue image, pixel position 520 corresponds to pixel value 0 and... Figure 5c Taking the intersection of pixel position 522 and pixel value 0 in the second bone tissue image, we get... Figure 5d The pixel at position 524 in the intersecting image has a value of 0.

[0164] The first bone tissue image and the second bone tissue image are subjected to union processing, i.e. Figure 5b In the first bone tissue image, pixel position 502 corresponds to pixel value 1 and... Figure 5c The pixel value 1 corresponding to pixel position 504 in the second bone tissue image is taken as the union of the two values. Figure 5e The pixel at position 526 in the union image has a value of 1. Figure 5b In the first bone tissue image, pixel position 508 corresponds to pixel value 1 and... Figure 5c In the second bone tissue image, the pixel value 0 corresponding to pixel position 510 is taken as the union of the two values. Figure 5e The pixel at position 528 in the union image has a value of 1. Figure 5b In the first bone tissue image, pixel position 515 corresponds to pixel value 0 and... Figure 5c In the second bone tissue image, the pixel value 1 corresponding to pixel position 516 is taken as the union of the two values. Figure 5e The pixel at position 530 in the union image has a value of 1. Figure 5b In the first bone tissue image, pixel position 520 corresponds to pixel value 0 and... Figure 5c In the second bone tissue image, the pixel value 0 corresponding to pixel position 522 is taken as the union of the two values. Figure 5eThe pixel at position 532 in the union image has a value of 0.

[0165] For example, please refer to Figure 5f and Figure 5g , Figure 5f and Figure 5g These are two adjacent medical images. Figure 5f Recorded as the first medical image, Figure 5g This is denoted as the second medical image. The first medical image is binarized to segment the bone tissue, resulting in... Figure 5h First image of bone tissue. The second medical image is binarized to segment the bone tissue, resulting in... Figure 5i Second bone tissue image. The white areas represent bone tissue, and the black areas represent soft tissue.

[0166] It should be noted that motion detection for any two adjacent frames of medical images within a target image sequence is not limited to detection methods based on Intersection over Union (IOU).

[0167] In the above embodiment, the first medical image is binarized to segment bone tissue, resulting in a first bone tissue image. The second medical image is also binarized to segment bone tissue, resulting in a second bone tissue image. The intersection-over-union (IoU) ratio is calculated based on the first and second bone tissue images to obtain first motion assessment data. This first motion assessment data provides a data basis for subsequently determining whether motion has occurred between two adjacent frames of medical images.

[0168] In some implementations, please refer to Figure 6 The target image sequence is extracted from an initial image sequence obtained by scanning a region of interest (ROI) multiple times. The ROI includes multiple scan locations, and the image sequences obtained from two adjacent scans are denoted as the first image sequence and the second image sequence. The sequence may include the following steps:

[0169] S610, Determine the sequence motion evaluation data between the first image sequence and the second image sequence.

[0170] S620. If the sequence motion evaluation data indicates that motion occurs between the first image sequence and the second image sequence, the first image sequence and the second image sequence are registered to obtain a registered image sequence.

[0171] S630. Based on multiple consecutive frames of registered medical images within the registered image sequence, feature extraction is performed to obtain sequence association features.

[0172] In some cases, patient movement during adjacent scans can cause a shift in the scan points of interest (SPO) between the first and second image sequences. This shift reduces image sharpness and detail, making accurate diagnosis difficult for physicians. It may also lead to image loss or incomplete data acquisition, preventing physicians from obtaining sufficient information for accurate diagnosis. Significant patient movement during the scan can disrupt spatiotemporal continuity, further degrading the final image quality.

[0173] Specifically, the image sequences obtained from two adjacent scans are denoted as the first image sequence and the second image sequence, respectively. Motion detection is performed on the first and second image sequences to obtain sequence motion evaluation data. A motion detection threshold t is set. t When the sequence motion evaluation data is less than the motion detection threshold t t When motion evaluation data is greater than or equal to a motion detection threshold t, it indicates that motion has occurred between adjacent image sequences. When motion occurs between adjacent image sequences, the first and second image sequences are registered to obtain a registered image sequence. t When this is the case, it indicates that there is no significant motion between adjacent image sequences corresponding to the motion evaluation data, and registration processing of the first and second image sequences is unnecessary. The motion detection threshold t is... t It can be configured according to actual needs.

[0174] When the registered image sequence is a spatially continuous image sequence, the registered image sequence is used as the input of the spatial feature extraction network. The spatial feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain sequence association features.

[0175] When the registered image sequence is a temporally continuous image sequence, the registered image sequence is used as the input of the temporal feature extraction network. The temporal feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain sequence association features.

[0176] When the registered image sequence includes a temporally continuous image sequence and a spatially continuous image sequence, the registered image sequence is used as input to a spatial feature extraction network. The spatial feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain spatial correlation features. The registered image sequence is also used as input to a temporal feature extraction network. The temporal feature extraction network extracts features from multiple consecutive registered medical images within the registered image sequence to obtain temporal correlation features. The spatial correlation features and temporal correlation features are then fused to obtain sequence correlation features.

[0177] In the above embodiments, sequence motion evaluation data between the first image sequence and the second image sequence is determined. If the sequence motion evaluation data indicates that motion occurs between the first image sequence and the second image sequence, registration processing is performed on the first image sequence and the second image sequence to obtain a registered image sequence. Feature extraction is performed based on multiple consecutive registered medical images within the registered image sequence to obtain sequence association features. By obtaining the sequence association features, input data can be provided to the target symmetric network to obtain high-quality medical images.

[0178] In some implementations, please refer to Figure 7 Determining sequence motion evaluation data between the first image sequence and the second image sequence may include the following steps:

[0179] S710. Construct a medical image group corresponding to each scanning position based on the medical images corresponding to each scanning position in the first image sequence and the second image sequence.

[0180] S720. Perform motion detection on the medical image group corresponding to each scanning position to obtain the second motion evaluation data corresponding to each scanning position.

[0181] S730. Calculate the mean value of the second motion evaluation data corresponding to each scanning position to obtain the sequence motion evaluation data.

[0182] Specifically, a medical image group corresponding to each scan position is constructed based on the medical images corresponding to each scan position in the first and second image sequences, respectively. Two adjacent frames of medical images in the medical image group can be denoted as the third medical image and the fourth medical image. The third medical image is binarized to segment bone tissue, resulting in the third bone tissue image. The fourth medical image is also binarized to segment bone tissue, resulting in the fourth bone tissue image. The intersection-over-union (IoU) ratio of the third and fourth bone tissue images is calculated to obtain the second motion evaluation data corresponding to each scan position. The mean of the second motion evaluation data corresponding to each scan position is calculated to obtain the sequence motion evaluation data.

[0183] For example, the indices of the five medical images in the first image sequence can be denoted as 1, 2, 3, 4, and 5. The indices of the five medical images in the second image sequence can be denoted as 6, 7, 8, 9, and 10. In the time dimension, the medical image with index 1 and the medical image with index 6 are two adjacent medical images at the same scanning position, i.e., a medical image group; the medical image with index 2 and the medical image with index 7 are two adjacent medical images at the same scanning position, i.e., a medical image group; the medical image with index 3 and the medical image with index 8 are two adjacent medical images at the same scanning position, i.e., a medical image group; the medical image with index 4 and the medical image with index 9 are two adjacent medical images at the same scanning position, i.e., a medical image group; and the medical image with index 5 and the medical image with index 10 are two adjacent medical images at the same scanning position, i.e., a medical image group. Calculate the second motion assessment data a1 between the medical image with index 1 and the medical image with index 6; calculate the second motion assessment data a2 between the medical image with index 2 and the medical image with index 7; calculate the second motion assessment data a3 between the medical image with index 3 and the medical image with index 8; calculate the second motion assessment data a4 between the medical image with index 4 and the medical image with index 9; and calculate the second motion assessment data a5 between the medical image with index 5 and the medical image with index 10. Summate the second motion assessment data a1, a2, a3, a4, and a5 to obtain a. s Then, for a... s Calculate the mean, i.e., a s / 5, to obtain sequence motion evaluation data.

[0184] In the above embodiments, a medical image group corresponding to each scanning position is constructed based on the medical images corresponding to each scanning position in the first image sequence and the second image sequence, respectively. Motion detection is performed on the medical image group corresponding to each scanning position to obtain second motion evaluation data corresponding to each scanning position. The mean of the second motion evaluation data corresponding to each scanning position is calculated to obtain sequence motion evaluation data. The sequence motion evaluation data can be used to evaluate whether motion has occurred in the first image sequence and the second image sequence, thereby correcting the image sequence that has motion and improving the accuracy of image optimization.

[0185] In some implementations, please refer to Figure 8 The target medical image is obtained by encoding and decoding sequence association features through a target image generation network. The target image generation network is trained in the following way:

[0186] S810: Acquire first biodata obtained by scanning with a first imaging dose.

[0187] S820: Simulate and process the first bio-data to obtain the second bio-data corresponding to the second imaging dose.

[0188] S830. Construct image sequence samples based on the image sequences obtained by reconstructing the second-generation data.

[0189] S840. The initial image generation network is trained using image sequence samples and the corresponding labels of the image sequence samples to obtain the target image generation network.

[0190] In this process, the first raw data is reconstructed to obtain a reference image sequence, which serves as a label for the medical images. The first imaging dose is greater than the second imaging dose. Raw data is unprocessed or unprocessed data acquired from medical imaging equipment. Raw data is typically recorded digitally and contains information on the intensity and position of X-rays or photons relative to the scanned object, measured by a detector. Raw data represents the absorption and scattering through the human body or other scanned object.

[0191] Specifically, at the first imaging dose, the region of interest is scanned using a medical imaging device to obtain first-generation data. The acquired first-generation data is then processed using simulation techniques (e.g., adding noise, using low-dose simulation tools) to obtain second-generation data corresponding to the second imaging dose. The second-generation data is reconstructed (e.g., using FBP imaging) to obtain an image sequence, and image sequence samples are constructed from this sequence. A baseline image sequence is obtained by reconstructing the first-generation data, with the medical images in the baseline image sequence serving as labels. The image sequence samples are used as input to an initial image generation network to train the network. Then, based on the output data of the initial image generation network and the labels corresponding to the image sequence samples, the loss value of the initial image generation network can be determined, and the parameters of the initial image generation network are updated based on the model loss value. This process continues, training the updated initial image generation network until a model training stopping condition is met, at which point the target image generation network is obtained. The model training stopping condition can be either the model loss value converging or the number of training epochs reaching a preset number. It should be noted that the image generation network uses a combination of various neural networks, including but not limited to commonly used networks such as convolutional neural networks, recurrent neural networks, fully connected networks, and generative adversarial networks.

[0192] For example, training data is extracted in the time and spatial dimensions and input into the initial image generation network, and the loss value is calculated based on the output of the initial image generation network and the label. The loss function is defined as follows:

[0193]

[0194] Where L is the loss value of the initial image generation network, T is the number of images output by the initial image generation network, and W and H are the width and height of the output images of the initial image generation network, respectively. t,w,h and yt,w,h These are the output images of the network that generate the initial image and their corresponding label images.

[0195] During training, the weights of the initial image generation network can be updated using the SGD algorithm. After training, the network parameters are fixed to obtain the target image generation network.

[0196] It should be noted that the calculation of loss values ​​includes, but is not limited to, the MSE loss function, the adversarial loss function, and the perceptual loss function.

[0197] In the above embodiments, by obtaining the target image generation network, the image quality of medical images obtained at low doses can be optimized to obtain high-quality medical images.

[0198] This specification also provides a medical image processing method, wherein the medical image to be processed corresponds to a specified scanning position. The target image sequence includes a spatially continuous image sequence corresponding to a region of interest and a temporally continuous image sequence corresponding to the specified scanning position. Two adjacent frames of medical images include a first medical image and a second medical image. For example, please refer to... Figure 9 The medical image processing method may include the following steps:

[0199] S902. Using the medical image to be processed as a reference, extract multiple frames of medical images that correspond to the spatial structure of the tissue region of interest and are spatially continuous from the initial image sequence in the spatial dimension. Using the medical image to be processed as a reference, extract multiple frames of medical images that correspond to any scanning position and are temporally continuous from the initial image sequence in the temporal dimension to construct the target image sequence.

[0200] Specifically, if the medical image to be processed is located at the edge of the initial image sequence, the medical image to be processed is copied, and the copied medical image is used as the edge image of the target image sequence.

[0201] S904. Based on multiple spatially continuous medical images within a spatially continuous image sequence, the first medical image is binarized to segment out bone tissue, thus obtaining the first bone tissue image.

[0202] S906. Perform binarization processing on the second medical image to segment out the bone tissue and obtain the second bone tissue image.

[0203] S908. The first motion assessment data is obtained by performing crossover ratio calculation based on the first bone tissue image and the second bone tissue image.

[0204] S910. Based on multiple time-continuous medical images within a time-continuous image sequence, the third medical image is binarized to segment out bone tissue, thus obtaining a third bone tissue image.

[0205] S912. Perform binarization processing on the fourth medical image to segment out the bone tissue and obtain the fourth bone tissue image.

[0206] S914. The intersection-union ratio is calculated based on the third and fourth bone tissue images to obtain the second motion assessment data.

[0207] S916. Construct a medical image group corresponding to each scanning position based on the medical images corresponding to each scanning position in the first image sequence and the second image sequence.

[0208] S918. Perform motion detection on the medical image group corresponding to each scanning position to obtain the second motion evaluation data corresponding to each scanning position.

[0209] S920. Calculate the mean value of the second motion evaluation data corresponding to each scanning position to obtain the sequence motion evaluation data.

[0210] S922. If the sequence motion evaluation data indicates that motion occurs between the first image sequence and the second image sequence, the first image sequence and the second image sequence are registered to obtain the registered image sequence.

[0211] S924. If the first motion evaluation data indicates that motion occurs between two adjacent frames of medical images, feature extraction is performed based on multiple consecutive frames of registered medical images within the registered image sequence to obtain spatial correlation features.

[0212] S926. If the second motion evaluation data indicates that motion occurs between two adjacent frames of medical images, feature extraction is performed based on multiple consecutive frames of registered medical images within the registered image sequence to obtain time-related features.

[0213] S928. Spatial correlation features and temporal correlation features are fused to obtain sequence correlation features.

[0214] Among them, sequence association features are used to characterize the spatial and / or temporal correlations between medical images within a target image sequence.

[0215] S930. Input the sequence association features into the target symmetric network for encoding and decoding to obtain the target medical image.

[0216] The imaging dose of the medical image to be processed is less than the imaging dose of the target medical image.

[0217] This specification provides a medical image processing device 1000. Please refer to [link / reference]. Figure 10The medical image processing device 1000 includes: a target image sequence acquisition module 1010, a sequence association feature extraction module 1020, and a target medical image generation module 1030.

[0218] The target image sequence acquisition module 1010 is used to acquire a target image sequence including the medical image to be processed.

[0219] The sequence association feature extraction module 1020 is used to extract features based on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence; wherein, the sequence association features are used to characterize the spatial correlation and / or temporal correlation between medical images within the target image sequence;

[0220] The target medical image generation module 1030 is used to generate a target medical image corresponding to the medical image to be processed based on the sequence association features.

[0221] For a detailed description of the medical image processing device, please refer to the description of the medical image processing method above, which will not be repeated here.

[0222] In some embodiments, a medical imaging device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method steps described above.

[0223] This specification provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method in any of the above embodiments.

[0224] One embodiment of this specification provides a computer program product including instructions that, when executed by a processor of a computer device, enable the computer device to perform the steps of the method described in any of the above embodiments.

[0225] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

Claims

1. A medical image processing method, characterized in that, The method includes: Acquire the target image sequence, which includes the medical images to be processed; Feature extraction is performed on multiple consecutive medical images within the target image sequence to obtain the sequence association features of the target image sequence; wherein, the sequence association features are used to characterize the spatial correlation and / or temporal correlation between medical images within the target image sequence; Based on the sequence association features, a target medical image corresponding to the medical image to be processed is generated; The target medical image is obtained by encoding and decoding the sequence association features through a target image generation network; the target image generation network is trained in the following manner: First biodata obtained by scanning with a first imaging dose; wherein, medical images in a reference image sequence are reconstructed from the first biodata as labels; The first imaging data is simulated and processed to obtain the second imaging data corresponding to the second imaging dose; wherein the first imaging dose is greater than the second imaging dose; Image sequence samples are constructed based on the image sequences reconstructed from the second biological data. The initial image generation network is trained using the image sequence samples and the corresponding labels to obtain the target image generation network; wherein the labels are medical images from the baseline image sequence reconstructed based on the first image data.

2. The method according to claim 1, characterized in that, The target image sequence includes a spatially continuous image sequence corresponding to a region of interest; the step of extracting features from multiple consecutive medical images within the target image sequence to obtain sequence association features of the target image sequence includes: Feature extraction is performed on multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial association features of the region of interest in the spatial dimension; wherein, the sequence association features include the spatial association features.

3. The method according to claim 1, characterized in that, The medical image to be processed corresponds to a specified scanning position; the target image sequence includes a temporally continuous image sequence corresponding to the specified scanning position; the step of extracting features based on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence includes: Feature extraction is performed on multiple time-continuous medical images within the time-continuous image sequence to obtain the temporal correlation features of the specified scan location in the time dimension; wherein, the sequence correlation features include the temporal correlation features.

4. The method according to claim 1, characterized in that, The medical image to be processed corresponds to a specified scanning location; the target image sequence includes a spatially continuous image sequence corresponding to the region of interest and a temporally continuous image sequence corresponding to the specified scanning location; the step of extracting features based on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence includes: Feature extraction is performed on multiple spatially continuous medical images within the spatially continuous image sequence to obtain the spatial association features of the tissue of interest in the spatial dimension. Feature extraction is performed on multiple time-continuous medical images within the time-continuous image sequence to obtain the temporal correlation features of the specified scan location in the time dimension. The spatial correlation features and the temporal correlation features are fused to obtain the sequence correlation features.

5. The method according to any one of claims 1 to 4, characterized in that, The sequence association features are obtained using any of the following methods: Optical flow information is extracted from multiple consecutive frames of medical images within the target image sequence to obtain optical flow map data, and the sequence association features are determined based on the optical flow map data. Multiple consecutive frames of medical images within the target image sequence are input into a target feature extraction network for feature extraction to obtain the sequence-related features.

6. The method according to claim 1, characterized in that, The step of generating a target medical image corresponding to the medical image to be processed based on the sequence association features includes: The sequence association features are input into a target symmetric network for encoding and decoding to obtain the target medical image; wherein the imaging dose corresponding to the medical image to be processed is less than the imaging dose corresponding to the target medical image.

7. The method according to claim 1, characterized in that, The target image sequence is extracted from an initial image sequence obtained by multiple scans of the region of interest, and the medical image to be processed corresponds to a specified scan location; the target image sequence is obtained through any of the following methods: Using the medical image to be processed as a reference, multiple frames of medical images that correspond to the spatial structure of the tissue region of interest and are spatially continuous are extracted from the initial image sequence in the spatial dimension to construct the target image sequence; Using the medical image to be processed as a reference, multiple frames of medical images corresponding to any scanning position and being temporally continuous are extracted from the initial image sequence in the time dimension to construct the target image sequence; If the medical image to be processed is located at the edge of the initial image sequence, the medical image to be processed is copied, and the copied medical image is used as the edge image of the target image sequence.

8. The method according to claim 1, characterized in that, The target image sequence is extracted from an initial image sequence obtained by scanning a region of interest (ROI) multiple times. The ROI includes multiple scanning locations, and the image sequences obtained from two adjacent scans are denoted as the first image sequence and the second image sequence. The method further includes: Determine the sequence motion evaluation data between the first image sequence and the second image sequence; If the sequence motion evaluation data indicates that motion occurs between the first image sequence and the second image sequence, the first image sequence and the second image sequence are registered to obtain a registered image sequence. Feature extraction is performed on multiple consecutive frames of registered medical images within the registered image sequence to obtain the sequence association features.

9. A medical image processing device, characterized in that, The device includes: The target image sequence acquisition module is used to acquire a target image sequence including the medical image to be processed; The sequence association feature extraction module is used to extract features based on multiple consecutive frames of medical images within the target image sequence to obtain the sequence association features of the target image sequence; wherein, the sequence association features are used to characterize the spatial correlation and / or temporal correlation between medical images within the target image sequence; The target medical image generation module is used to generate a target medical image corresponding to the medical image to be processed based on the sequence association features. The target medical image is obtained by encoding and decoding the sequence association features through a target image generation network; the target image generation network is trained in the following manner: First biodata obtained by scanning with a first imaging dose; wherein, medical images in a reference image sequence are reconstructed from the first biodata as labels; The first imaging data is simulated and processed to obtain the second imaging data corresponding to the second imaging dose; wherein the first imaging dose is greater than the second imaging dose; Image sequence samples are constructed based on the image sequences reconstructed from the second biological data. The initial image generation network is trained using the image sequence samples and the corresponding labels to obtain the target image generation network; wherein the labels are medical images from the baseline image sequence reconstructed based on the first image data.

10. A medical imaging device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images