A T1 image guided ASL perfusion image reconstruction method

Abstract

The application discloses a T1 image guided ASL perfusion image reconstruction method. In the method, a computing device acquires an ASL difference image sequence corresponding to a target part of a patient, aggregates the ASL difference image sequence to generate an ASL aggregate image. Then, a first encoding module is used to encode the ASL aggregate image to generate a first feature map, a second encoding module is used to encode a T1 image corresponding to the target part to generate a second feature map, a bottleneck layer between the first encoding module and a corresponding first decoding module is used to fuse the first feature map and the second feature map to generate a third feature map, and finally, the first decoding module is used to generate a reconstructed ASL perfusion image according to the third feature map. In the method, the anatomical prior information of the T1 image is used as a guide in the reconstruction process of the ASL perfusion image, so that the reconstruction quality of the ASL perfusion image can be improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a T1 image-guided ASL perfusion image reconstruction method. Background Technology

[0002] Arterial spin labeling (ASL) obtains perfusion information of a patient's brain tissue by magnetically labeling endogenous arterial blood molecules. It does not require the injection of contrast agents and is completely non-invasive to the patient, making it an important means of obtaining cerebral blood flow (CBF) information in clinical practice and research.

[0003] In existing technologies, ASL (Absorbed Surge) technology obtains ASL perfusion images by acquiring pairs of control and label images, and then determining the difference image between the control and label images. Clinically, to obtain high-quality ASL perfusion images, control and corresponding label images are typically acquired multiple times consecutively (e.g., 12 times), and a difference image is generated between them. These difference images are then averaged to produce the final ASL perfusion image.

[0004] However, this method requires acquiring multiple control images and corresponding labeled images, thus necessitating prolonged MRI scans of the subject. Furthermore, when the signal-to-noise ratio of the differential images is low, the simple averaging method used in existing technologies is insufficient to generate high-quality ASL perfusion images.

[0005] There is currently no effective solution to the technical problems of excessive image acquisition and difficulty in generating high-quality ASL perfusion images in existing ASL image reconstruction techniques. Summary of the Invention

[0006] The embodiments of this disclosure provide a method and apparatus for ASL perfusion image reconstruction based on T1 image guidance, so as to at least solve the technical problem of difficulty in generating high-quality ASL perfusion images in the prior art.

[0007] According to one aspect of the present disclosure, a method for ASL perfusion image reconstruction based on T1 image guidance is provided, comprising: acquiring an ASL differential image sequence corresponding to a target site of a patient; aggregating the ASL differential image sequence to generate a corresponding ASL aggregated image; encoding the ASL aggregated image through a first encoding module to generate a first feature map; encoding the T1 image corresponding to the target site through a second encoding module to generate a second feature map; fusing the first feature map and the second feature map through a bottleneck layer between the first encoding module and a corresponding first decoding module to generate a third feature map; and generating a reconstructed ASL perfusion image based on the third feature map through a first decoding module.

[0008] According to another aspect of the present disclosure, a storage medium is also provided, the storage medium including a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0009] According to another aspect of the present disclosure, an ASL perfusion image reconstruction device based on T1 image guidance is also provided, comprising: an acquisition module for acquiring an ASL differential image sequence corresponding to a target site of a patient; an aggregation module for aggregating the ASL differential image sequence to generate a corresponding ASL aggregated image; an ASL encoding module for encoding the ASL aggregated image through a first encoding module to generate a first feature map; a T1 encoding module for encoding the T1 image corresponding to the target site through a second encoding module to generate a second feature map; a fusion module for fusing the first feature map and the second feature map through a bottleneck layer between the first encoding module and a corresponding first decoding module to generate a third feature map; and a perfusion image generation module for generating a reconstructed ASL perfusion image based on the third feature map through the first decoding module.

[0010] According to another aspect of the present disclosure, an ASL perfusion image reconstruction apparatus based on T1 image guidance is also provided, comprising: a processor; and a memory connected to the processor, configured to provide the processor with instructions for processing the following steps: acquiring an ASL differential image sequence corresponding to a target site of a patient; aggregating the ASL differential image sequence to generate a corresponding ASL aggregated image; encoding the ASL aggregated image through a first encoding module to generate a first feature map; encoding the T1 image corresponding to the target site through a second encoding module to generate a second feature map; fusing the first feature map and the second feature map through a bottleneck layer between the first encoding module and a corresponding first decoding module to generate a third feature map; and generating a reconstructed ASL perfusion image based on the third feature map through the first decoding module.

[0011] In this embodiment, a computing device acquires an ASL differential image sequence corresponding to the target site of the patient, aggregates the ASL differential image sequences to generate an ASL aggregated image. Then, a first encoding module encodes the ASL aggregated image to generate a first feature map, and a second encoding module encodes the T1 image corresponding to the target site to generate a second feature map. A bottleneck layer between the first encoding module and the corresponding first decoding module fuses the first and second feature maps to generate a third feature map. Finally, a first decoding module generates a reconstructed ASL perfusion image based on the third feature map. In this method, the anatomical prior information of the T1 image is used as a guide during the reconstruction of the ASL perfusion image, thereby improving the reconstruction quality of the ASL perfusion image. Furthermore, because this method can improve the image quality of the reconstructed ASL perfusion image, it can reduce the number of ASL differential images used in the ASL differential image sequence. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure. In the drawings: Figure 1 This is a hardware structure block diagram of a computing device for implementing the method described in Embodiment 1 of this disclosure; Figure 2 This is a schematic flowchart of an ASL perfusion image reconstruction method based on T1 image guidance according to the first aspect of Embodiment 1 of this disclosure; Figure 3 This is a schematic flowchart of ASL perfusion image reconstruction according to the first aspect of Embodiment 1 of this disclosure; Figure 4A This is a schematic diagram of the first encoding module and the first decoding module according to the first aspect of Embodiment 1 of this disclosure; Figure 4B This is a schematic diagram of the first encoding module and the first decoding module according to the first aspect of Embodiment 1 of this disclosure; Figure 5A This is a flowchart illustrating the first-stage training process according to the first aspect of Embodiment 1 of this disclosure; Figure 5B This is a flowchart illustrating the second-stage training process according to the first aspect of Embodiment 1 of this disclosure; Figure 6 This is a schematic diagram of an ASL perfusion image reconstruction apparatus based on T1 image guidance, according to the first aspect of Embodiment 2 of this disclosure; Figure 7This is a schematic diagram of an ASL perfusion image reconstruction apparatus based on T1 image guidance, according to the first aspect of Embodiment 3 of this disclosure. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0014] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] Example 1

[0016] According to this embodiment, an embodiment of an ASL perfusion image reconstruction method based on T1 image guidance is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0017] Furthermore, since existing techniques typically require 5-10 minutes of MRI scanning of the subject, motion artifacts are easily generated due to the subject's activity. The simple averaging method is also prone to producing poor-quality difference images (artifact images) due to motion artifacts, thereby reducing the quality of the final ASL perfusion image generated.

[0018] The method embodiments provided in this example can be executed on mobile terminals, computer terminals, servers, or similar computing devices. Figure 1 A hardware block diagram of a computing device for implementing a T1 image-guided ASL perfusion image reconstruction method is shown. Figure 1As shown, a computing device may include one or more processors (processors may include, but are not limited to, microprocessors such as MCUs or programmable logic devices such as FPGAs), a memory for storing data, a transmission device for communication functions, and an input / output interface. The memory, transmission device, and input / output interface are connected to the processor via a bus. In addition, it may also include a display, keyboard, and cursor control device connected to the input / output interface. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, a computing device may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0019] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element in a computing device. As involved in the embodiments of this disclosure, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0020] The memory can be used to store software programs and modules of application software, such as the program instruction / data storage device corresponding to the T1 image-guided ASL perfusion image reconstruction method in this embodiment of the present disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the aforementioned application program and the T1 image-guided ASL perfusion image reconstruction method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the computing device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0021] The transmission device is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the computing device's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0022] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows users to interact with the user interface of the computing device.

[0023] It should be noted here that, in some optional embodiments, the above... Figure 1 The computing device shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computing devices.

[0024] Under the aforementioned operating environment, according to the first aspect of this embodiment, a method for ASL perfusion image reconstruction based on T1 image guidance is provided. This method consists of... Figure 1 The computing device shown is implemented. Figure 2 A flowchart illustrating the method is shown below. (Refer to...) Figure 2 As shown, the method includes: S202: Obtain the ASL differential image sequence corresponding to the target site of the patient; S204: Aggregate the ASL differential image sequences to generate the corresponding ASL aggregated image; S206: Encode the ASL aggregated image through the first encoding module to generate the first feature map; S208: The second encoding module encodes the T1 image corresponding to the target region to generate a second feature map; S210: By using the bottleneck layer between the first encoding module and the corresponding first decoding module, the first feature map and the second feature map are fused to generate a third feature map; and S212: The first decoding module generates a reconstructed ASL perfusion image based on the third feature map.

[0025] The method provided in this embodiment is used to generate reconstructed ASL perfusion images based on ASL differential image sequences, thereby providing higher quality ASL perfusion images for clinical or experimental studies.

[0026] First, the computing device acquires the ASL differential image sequence corresponding to the target site of the patient (S202).

[0027] Then, the computing device aggregates the ASL differential image sequences to generate the corresponding ASL aggregated image (S204). The specific aggregation method for the ASL differential image sequences in this embodiment will be mentioned later. This method can eliminate motion artifacts present in the ASL differential image sequences, and will be described in detail later.

[0028] Figure 3 This is a schematic diagram of the ASL image reconstruction model in this embodiment. Figure 3 The diagram illustrates the process of generating ASL perfusion images using a T1 image-guided method in this embodiment.

[0029] refer to Figure 3 As shown, the computing device can encode the ASL aggregated image through the first encoding module to generate a first feature map (S206). The computing device can encode the T1 image corresponding to the target region through the second encoding module to generate a second feature map (S208).

[0030] Then, the computing device can fuse the first feature map and the second feature map through the bottleneck layer between the first encoding module and the corresponding first decoding module to generate a third feature map (S210).

[0031] Finally, the computing device generates a reconstructed ASL perfusion image based on the third feature map using the first decoding module (S212).

[0032] In other words, in the above steps, the computing device incorporates T1 images (T1-weighted images) of the same patient and the same target site into the reconstruction of the ASL perfusion images. (Reference) Figure 3 As shown, the ASL image reconstruction model provided in this embodiment generates a first feature map corresponding to the ASL aggregated image through a first encoding module, and generates a second feature map corresponding to the T1 image through a second encoding module. Then, the first and second feature maps are fused through a bottleneck layer between the first encoding module and the first decoding module to generate a third feature map. Finally, the ASL perfusion image is generated by the first decoding module based on the third feature map.

[0033] Therefore, in this embodiment, the anatomical prior information of the T1 image is used as a guide during the reconstruction of the ASL perfusion image to improve the reconstruction quality of the ASL perfusion image.

[0034] The target area mentioned above could be, for example, the patient's brain. That is, in this embodiment, the ASL differential image can be generated by acquiring images of the patient's brain using ASL technology, and the T1 image can be obtained by performing a magnetic resonance imaging scan of the patient's brain.

[0035] As described in the background section, in clinical practice, to obtain high-quality ASL perfusion images, control images and corresponding labeled images are typically acquired multiple times consecutively (e.g., 12 times), and a difference image is generated between the two. These difference images are then averaged to produce the final ASL perfusion image. However, this method requires acquiring multiple control and labeled images, necessitating prolonged MRI scans for the subject. Furthermore, when the signal-to-noise ratio of the difference images is low, the simple averaging method used in existing techniques is insufficient to generate high-quality ASL perfusion images.

[0036] In view of this, in this embodiment, the computing device acquires the ASL differential image sequence corresponding to the target site of the patient, aggregates the ASL differential image sequence, and generates an ASL aggregated image. Then, the ASL aggregated image is encoded by a first encoding module to generate a first feature map, and the T1 image corresponding to the target site is encoded by a second encoding module to generate a second feature map. The first and second feature maps are fused through a bottleneck layer between the first encoding module and the corresponding first decoding module to generate a third feature map. Finally, the reconstructed ASL perfusion image is generated based on the third feature map by the first decoding module. In this embodiment, the anatomical prior information of the T1 image is used as a guide during the reconstruction of the ASL perfusion image. Therefore, compared to the existing method of averaging differential images to generate ASL focus images, this embodiment can reconstruct a high-quality ASL perfusion image from a smaller number of ASL differential images, even when there are ASL differential images with low signal-to-noise ratios in the ASL differential image sequence, by introducing the anatomical prior information of the T1 image. This solves the technical problem in existing ASL image reconstruction techniques that involves too many image acquisitions and makes it difficult to generate high-quality ASL perfusion images.

[0037] The cross-modal fusion mechanism of the bottleneck layer will be explained below, and how this cross-modal fusion mechanism avoids the leakage of anatomical structural information of the T1 image carried in the second feature map to the first decoding module.

[0038] Optionally, the operation of fusing the first feature map and the second feature map to generate a third feature map through a bottleneck layer between the first encoding module and the corresponding first decoding module includes: generating a query matrix and a value matrix corresponding to the first feature map through the bottleneck layer; generating a key matrix corresponding to the second feature map through the bottleneck layer; and generating the third feature map through the bottleneck layer using a cross-attention mechanism based on the query matrix, the value matrix, and the key matrix.

[0039] Specifically, the computing device can generate a query matrix corresponding to the first feature map through the bottleneck layer. ) and value matrix ( ), and generate a key matrix corresponding to the second feature map through the bottleneck layer ( The computing device can first generate a query matrix corresponding to the first feature map. Then, the matrix that is the same as the query matrix is ​​determined as the value matrix. Finally, the computing device passes through the bottleneck layer, based on the query matrix. Value matrix and the key matrix A third feature map is generated by utilizing a cross-attention mechanism.

[0040] First, the computing device can generate a query matrix corresponding to the first feature map. ) and value matrix ( ), and generate the key matrix corresponding to the second feature map ( ). Among them, for example Figure 4A As shown, both the first and second feature maps are matrices of size 256×16×16, where 256 represents the number of channels and 16×16 represents the size of the feature map. The computing device uses flattening and transpose operations to generate the query matrix corresponding to the first feature map. ) and value matrix ( ), resulting in a 256×256 query matrix ( ) and value matrix ( ), where 256×256 represents a feature vector with 256 channels, each channel corresponding to a feature vector of dimension 256. The same applies to the second feature map; the computing device uses flattening and transpose operations to generate a key matrix corresponding to the second feature map. ), resulting in a 256×256 bond matrix ( ).

[0041] In other words, during the flattening operation, the computing device unfolds the feature map into a 256-dimensional feature vector according to the channels, resulting in a sequence of feature vectors.

[0042] Then, the computing device can utilize a multi-head attention layer based on the aforementioned query matrix. Value matrix and the key matrix This generates the corresponding attention information. Specifically, based on the query matrix described above... Value matrix and the key matrix Generate corresponding attention information The formula is shown below:

[0043] In the above formula Representing the query matrix Value matrix and the key matrix The dimension of the vector in the vector.

[0044] Then the computing device can use attention information Generate the third feature map.

[0045] In other words, query matrix Sum matrix The key matrix is ​​generated from the first feature map associated with the ASL differential image sequence. It is generated through the second feature map associated with the T1 image. Therefore, the query matrix Sum matrix All of these information originate from ASL differential image sequences, while the key matrix... Derived from image T1. In the attention mechanism, the query matrix... Bond matrix Used to determine attention weights, while the value matrix The "value" of the feature map generated through the attention mechanism is determined. Therefore, in this embodiment, the query matrix... Sum matrix All are derived from ASL difference image sequences, containing only the key matrix. The information originates from the T1 image, thus avoiding the leakage of the texture information of the anatomical structure of the T1 image to the first decoding module (i.e., avoiding the leakage of the pixel information of the T1 image to the first decoding module), and avoiding the "illusion" problem of the first decoding module due to the guidance of the T1 image.

[0046] Optionally, the operation of generating a third feature map through the bottleneck layer using a cross-attention mechanism based on the query matrix, value matrix, and key matrix includes: generating attention information corresponding to the third feature map through the bottleneck layer using a cross-attention mechanism based on the query matrix, value matrix, and key matrix; and generating the third feature map based on the attention information through the bottleneck layer using a residual gating mechanism.

[0047] In this embodiment, a residual gating mechanism can be further introduced on the basis of the cross-attention mechanism, thereby further preventing the pixel information of the T1 image from being leaked to the first decoding module.

[0048] Specifically, the computing device, through the bottleneck layer, generates attention information corresponding to the third feature map based on the query matrix, value matrix, and key matrix using a cross-attention mechanism. Then, through the bottleneck layer, the computing device generates the third feature map based on the attention information using a residual gating mechanism.

[0049] The computing device can determine the gate control parameters, utilize the residual gating mechanism, and combine the gating parameters with attention information to generate a third feature map. The expression for generating the third feature map is as follows:

[0050]

[0051] in, Represents the third feature map; This represents the feature information generated using the residual gating mechanism, which is a 256×256 matrix in this embodiment. This represents the recombination operation corresponding to the inverse of the flattening and transpose operations described above; Indicates the normalization layer; This represents the gating parameters. It can be set manually. Of course, the gating parameter can also be a learnable scalar. For example, the gating parameter can be initially set to 0.3 during the training phase, and the range of the gating parameter value can be set to [0,1]. The gating parameter changes during the training process until it reaches a fixed value after training is completed.

[0052] Due to the generation of the query matrix Value matrix and the key matrix At that time, the computing device flattens and transposes the 256×16×16 matrix shape of the first and second feature maps to obtain a 256×256 matrix shape. (Reference) Figure 4A As shown, when generating the third feature map, the computing device uses the feature information generated by the residual gating mechanism. A recombination operation is performed to obtain a matrix shape of 256×16×16. That is, the recombination operation in the above formula... This means that a feature vector of dimension 256 is converted back into a 16×16 matrix, thereby allowing the feature information generated using the residual gating mechanism to be processed. By recombining, the third feature map can be restored to a matrix shape of 256×16×16.

[0053] Therefore, in this application, by using a residual gating mechanism to control the attention information using a cross-attention mechanism, the pixel information of the T1 image can be more effectively prevented from leaking to the first decoding module.

[0054] Optionally, the operation of aggregating the ASL difference image sequence to generate a corresponding ASL aggregated image includes: calculating the mean of pixels at the same position in each ASL difference image in the ASL difference image sequence to generate an initial aggregated image; determining the pixel value deviation between each pixel in the ASL difference image and the corresponding pixel in the initial aggregated image, and determining the logarithmic variance corresponding to each pixel in the ASL difference image based on the pixel value deviation; and for each pixel position in the ASL difference image, performing a weighted summation of the pixel values ​​of the corresponding pixels in each ASL difference image based on the logarithmic variance of the corresponding pixels in each ASL difference image to generate the ASL aggregated image.

[0055] Specifically, the process by which the computing device aggregates the ASL differential image sequences using an aggregator is as follows: The average value of pixels at the same position in each ASL difference image in the ASL difference image sequence is calculated to generate an initial aggregated image.

[0056] The expression for any frame of the ASL differential image sequence is as follows: , Indicates the first ASL difference images, where, =1~ , This indicates the number of ASL difference images in the ASL difference image sequence. This indicates the size of the ASL difference image. First, the computing device can take the pixel average of each ASL difference image in the ASL difference image sequence, according to each pixel position, to generate an initial aggregated image. The expression is:

[0057] Then, the pixel value deviation between each pixel in the ASL difference image and the corresponding pixel in the initial aggregate image is determined (here, the pixel value deviation between each pixel in an ASL difference image is referred to as the frame deviation information corresponding to the ASL difference image). The frame deviation information includes the pixel value deviation between each pixel in the corresponding ASL difference image and the corresponding pixel in the initial aggregate image.

[0058] Determine frame offset information The formula is:

[0059] Indicates the first Frame offset information of each ASL differential image, This represents the initial aggregated image.

[0060] Then, the computing device determines the logarithmic variance corresponding to each pixel of the ASL difference image based on the pixel value deviation. The expression is:

[0061]

[0062] in, Indicates the first The element in the frame offset information of each ASL difference image that corresponds to the pixel in the h-th row and w-th column of the corresponding ASL difference image. This indicates taking the logarithmic variance of the elements in parentheses.

[0063] Finally, the computing device can perform a weighted summation of the pixel values ​​of the corresponding pixels in each ASL difference image based on the logarithmic variance of the corresponding pixels in each ASL difference image, generating an ASL aggregate image. The pixel values ​​in the h-th row and w-th column of the ASL aggregate image are... It can be represented as:

[0064] In other words, the computing device can use the logarithmic variance of each pixel in the corresponding ASL difference image as the weight associated with that ASL difference image. That is, for an ASL difference image, the weight associated with that ASL difference image is pixel-by-pixel. Then, the computing device weights the pixel values ​​of the corresponding pixels in each ASL difference image according to the logarithmic variance of the corresponding pixels in each ASL difference image to generate an ASL aggregate image.

[0065] By aggregating the data in the above manner, the pixel-by-pixel deviation between each ASL differential image and the initial aggregated image is used as the weight of the corresponding ASL differential image. This enables the elimination of motion artifacts caused by local motion in the ASL differential image sequence at the pixel level. Furthermore, since the weights are independent of the original pixels of the ASL differential images, the weighted aggregation of ASL differential images will not suppress the representation of real lesions in the reconstructed ASL perfusion images.

[0066] Alternatively, the softmax function can be used to determine the corresponding weights for each pixel position in the ASL difference image based on the logarithmic variance of the corresponding pixels in each ASL difference image. :

[0067] Then, based on the corresponding weights, the pixel values ​​of corresponding pixels in the ASL difference image are weighted and summed to generate the ASL aggregate image, where the pixel value in the h-th row and w-th column of the ASL aggregate image is... It can be represented as:

[0068] in, Indicates the first t The first ASL difference image h Line 1 w The pixel values ​​of the column.

[0069] This paper presents another method for aggregating ASL difference images in an ASL difference image sequence. Specifically, the computing device can use a self-attention mechanism to generate weights corresponding to each ASL difference image in the ASL difference image sequence. Then, according to the weights corresponding to each ASL difference image, the ASL difference images are weighted and fused to generate an ASL aggregated image. In this method, the weights of the entire frame of the same ASL difference image are uniform.

[0070] Optionally, the first encoding module includes a residual unit and at least one downsampling unit, and the first decoding module includes at least one upsampling unit and a convolutional unit. The operation of encoding the ASL aggregated image through the first encoding module to generate a first feature map includes: generating a first intermediate feature map from the ASL aggregated image using the residual unit; and generating the first feature map from the first intermediate feature map using the at least one downsampling unit. The operation of generating a reconstructed ASL infusion image from a third feature map using the first decoding module includes: generating a second intermediate feature map from the third feature map using the at least one upsampling unit; and generating the reconstructed ASL infusion image from the second intermediate feature map using the convolutional unit.

[0071] In this embodiment, the first encoding module and the first decoding module form a structure similar to a U-net network.

[0072] For details, please refer to Figure 4A As shown, the first encoding module includes a residual unit and at least one downsampling unit, and the first decoding module includes at least one upsampling unit and a convolution unit. Figure 4A The image shows an example of a first encoding module including a residual unit and three downsampling units, and a first decoding module including three upsampling units and a convolution unit.

[0073] Figure 4A In the ASL aggregated image, [1,128,128], 1 indicates that the number of channels is 1, and 128 and 128 represent the image size.

[0074] The computing device generates a first intermediate feature map corresponding to the ASL aggregated image through residual units, referencing... Figure 4A As shown, the first intermediate feature map is in the form of a 32×128×128 matrix.

[0075] Then, the computing device generates a third intermediate feature map in the form of a 64×64×64 matrix based on the first intermediate feature map through the first downsampling unit; and generates a third intermediate feature map in the form of a 128×32×32 matrix based on the third intermediate feature map in the form of a 64×64×64 matrix through the second downsampling unit.

[0076] Then, the computing device generates the first feature map based on the third intermediate feature map in the form of a 128×32×32 matrix through the third downsampling unit.

[0077] Then, the computing device performs cross-modal fusion of the first and second feature maps through the bottleneck layer, generating a third feature map in the form of a 256×16×16 matrix.

[0078] That is, the computing device generates a first feature map in the form of a 256×16×16 matrix based on the first intermediate feature map through at least one downsampling unit.

[0079] The computing device performs cross-modal fusion of the first and second feature maps through the bottleneck layer, and generates a third feature map in the form of a 256×16×16 matrix.

[0080] Then, the computing device generates a second intermediate feature map in the form of a 32×128×128 matrix based on the third feature map using at least one upsampling unit.

[0081] Then, the computing device generates an ASL perfusion image in the form of a 1×128×128 matrix based on the second intermediate feature map through a convolutional unit.

[0082] Optionally, the operation of generating a second intermediate feature map from the third feature map using at least one upsampling unit includes: A second intermediate feature map is generated by at least one upsampling unit based on a third feature map and a third intermediate feature map received from a first encoding module via a skip connection layer, wherein the operation of generating a reconstructed ASL perfusion image based on the second intermediate feature map by a convolution unit includes: generating a reconstructed ASL perfusion image by a convolution unit based on the second intermediate feature map and a first intermediate feature map received from a first encoding module via a skip connection layer.

[0083] Specifically, after generating the third feature map, the computing device generates a fourth intermediate feature map in the form of a 128×32×32 matrix based on the third feature map through the first upsampling unit.

[0084] Then, the computing device fuses the third intermediate feature map in the form of a 128×32×32 matrix, which is received from the first encoding module via the skip connection layer, and the fourth intermediate feature map in the form of a 128×32×32 matrix generated by the first upsampling unit to obtain an input feature map in the form of a 128×32×32 matrix, and generates the fourth intermediate feature map in the form of a 64×64×64 matrix through the second upsampling unit.

[0085] Then, the computing device fuses the third intermediate feature map in the form of a 64×64×64 matrix, which is received from the first encoding module via the skip connection layer, and the fourth intermediate feature map in the form of a 64×64×64 matrix generated by the second upsampling unit to obtain an input feature map in the form of a 64×64×64 matrix, and generates a second intermediate feature map in the form of a 32×128×128 matrix through the third upsampling unit.

[0086] Then, the computing device fuses the first intermediate feature map in the form of a 32×128×128 matrix, which is received from the first coding module via the skip connection layer, and the second intermediate feature map in the form of a 32×128×128 matrix generated by the third upsampling unit to obtain an input feature map in the form of a 32×128×128 matrix, and generates an ASL infusion image through a convolution unit.

[0087] exist Figure 4A In this example, 'f' represents the fusion module, which performs the fusion operation on the feature maps. Taking the generation of the second feature map input to the second upsampling unit as an example, the fusion is performed... Figure 4AThe fusion (f) operation is explained as follows: When the computing device fuses the 128×32×32 matrix form of the third intermediate feature map received from the first encoding module via the skip connection layer and the 128×32×32 matrix form of the fourth intermediate feature map generated by the first upsampling unit, the corresponding third intermediate feature map can be concatenated with the corresponding fourth intermediate feature map to obtain a concatenated feature map in the form of a 256×32×32 matrix. Then, the concatenated feature map is convolved with a 1×1 convolution to obtain a convolutional feature map in the form of a 128×32×32 matrix. The convolutional feature map is then normalized (GroupNorm) and then activated by the SiLU activation function to generate an input feature map in the form of a 128×32×32 matrix. Before concatenating the third intermediate feature map and the fourth intermediate feature map, dropout (e.g., dropping out 30% of the feature points in the corresponding feature map) can be performed on the two feature maps respectively.

[0088] In addition, refer to Figure 4B As shown, three skip connection layers are set between the first encoding module and the first decoding module. The shallowest layer is the third skip connection layer, the deepest layer is the first skip connection layer, and the middle layer is the second skip connection layer. In this embodiment, the fusion operation of the fusion module corresponding to the first skip connection layer (or the first skip connection layer and the second skip connection layer) can be the same cross-modal fusion operation as the bottleneck layer. That is, the fusion module of the first skip connection layer (or the first skip connection layer and the second skip connection layer) can use the same cross-attention mechanism (or the cross-attention mechanism combined with the residual gating mechanism) as the bottleneck layer to fuse the feature maps of the two modalities corresponding to the ASL difference image and the T1 image, respectively, thereby further avoiding the "illusion" problem caused by the guidance of the T1 image in the first decoding module.

[0089] For example, for the first skip connection layer, when the computing device fuses the third intermediate feature map in the form of a 128×32×32 matrix received from the first encoding module via the skip connection layer and the fourth intermediate feature map in the form of a 128×32×32 matrix generated by the first upsampling unit, firstly, it can generate the corresponding query matrix and value matrix based on the third intermediate feature map in the form of a 128×32×32 matrix, and generate the key matrix based on the fourth intermediate feature map in the form of a 128×32×32 matrix. Then, the computing device can use a cross-attention mechanism to generate the corresponding fourth intermediate feature map (128×32×32) for input to the second upsampling unit based on the query matrix, value matrix, and key matrix. Further, the computing device can generate the corresponding attention feature based on the query matrix, value matrix, and key matrix, and further use a residual gating mechanism to generate the corresponding fourth intermediate feature map for input to the second upsampling unit based on the aforementioned gating parameters and the attention feature. The specific calculation methods and formulas of the cross-attention mechanism and residual gating mechanism mentioned here are consistent with the relevant content in the description of the bottleneck layer above.

[0090] The above description uses the first hop connection layer as an example. For the second hop connection layer, the fusion operation of the fusion module adopts the same cross-modal fusion operation as the bottleneck layer, and the specific method is the same as that of the first hop connection layer, so it will not be repeated here.

[0091] Optionally, before acquiring the ASL differential image sequence corresponding to the target site of the patient, the method further includes: A T1 image segmentation model is constructed, comprising a second encoding module and a second decoding module to be trained. The first encoding module and the second encoding module have the same structure, and the first decoding module and the second decoding module have the same structure. The T1 image segmentation model is trained in a supervised manner using first training samples, which include T1 sample images and corresponding segmented image samples. An ASL image reconstruction model is constructed, comprising a first encoding module, a first decoding module, a bottleneck layer, and a second encoding module. The ASL image reconstruction model is trained using second training samples while freezing the parameters of the second encoding module. The second training samples include ASL differential image sample sequences and corresponding T1 sample images.

[0092] In this embodiment, reference Figure 5A and Figure 5BAs shown, the training process is divided into two stages. The first stage is mainly used to train the second encoding module. The second stage treats the first encoding module, the second encoding module, the bottleneck layer between the first and second encoding modules, and the second encoding module trained in the first stage as a whole network (ASL image reconstruction model) and trains this network. However, in this stage, the network parameters of the second encoding module are frozen, and the network parameters of the other network modules in the ASL image reconstruction model are updated.

[0093] The first encoding module and the second encoding module have the same network structure, and the first decoding module and the second decoding module have the same network structure.

[0094] Specifically, in the first stage, the computing device acquires a first training sample, which includes a T1 sample image and corresponding segmentation image samples. The segmentation image samples corresponding to the T1 sample images can represent four image regions in the T1 sample images: gray matter, white matter, cerebrospinal fluid, and extracerebral regions, respectively. The computing device can then construct a T1 image segmentation model. The T1 image segmentation model includes a second encoding module and a second decoding module to be trained. Subsequently, the computing device performs supervised training on the T1 image segmentation model based on the first training sample. That is, the computing device can input the T1 sample images into the T1 image segmentation model, generate corresponding segmentation prediction images, and train the T1 image segmentation model with the training objective of minimizing the difference between the segmentation prediction image and the segmentation image samples.

[0095] After the T1 image segmentation model is trained, the trained second encoding module, the first encoding module to be trained, and the first decoding module to be trained can be treated as a whole network (referred to as the ASL image reconstruction model in this embodiment). Then, the ASL image reconstruction model is trained based on the second training samples, and the parameters of the second encoding module are frozen during the training process. The second training samples include ASL differential image sample sequences and corresponding T1 sample images (i.e., ASL differential image sample sequences and T1 sample images corresponding to the same subject).

[0096] Therefore, this embodiment adopts a two-stage training method. Compared with training the second encoding module, the first encoding module, and the first decoding module simultaneously, the second encoding module can have better image semantic understanding ability after the first stage of training. That is, the second encoding module can perform better semantic understanding of the four image regions corresponding to the gray matter, white matter, cerebrospinal fluid, and extracerebral regions (i.e., image regions not belonging to the brain) in the image. This ensures the feature quality of the second feature map generated by the second encoding module in the second stage, thereby improving the training effect of the overall network of the ASL image reconstruction model.

[0097] Optionally, the operation of training the ASL image reconstruction model using the second training samples includes: The ASL differential image sample sequence is split into a first subset and a second subset; the ASL differential images in the second subset are fused by averaging to determine the mean image corresponding to the second subset; a predicted image is generated based on the first subset and the T1 sample image using the first encoding module, the first decoding module, and the second encoding module to be trained; and the ASL image reconstruction model is trained in a self-supervised manner based on the predicted image and the mean image of the second subset.

[0098] Specifically, in this embodiment, the ASL image reconstruction model is trained using a second training sample and a self-supervised approach.

[0099] The computing device can acquire the ASL differential image sample sequence and the corresponding T1 sample image for the same subject, and split the ASL differential image sample sequence into a first subset ( Figure 5B set_a) and the second subset ( Figure 5B (set_b in the example). For example, if the ASL differential image sample sequence contains 12 ASL differential image samples, then the first subset can contain the first 6 ASL differential image samples, and the second subset can contain the last 6 ASL differential image samples.

[0100] During the second phase of training, in accordance with... Figure 5B The illustrated process involves aggregating the first subset using an aggregator to generate an aggregated image corresponding to the first subset. This aggregated image is then encoded by a first encoding module to generate a first feature map. A second encoding module encodes the T1 sample image to generate a second feature map corresponding to the T1 sample image. A bottleneck layer fuses the first and second feature maps to generate a third feature map. Finally, a first decoding module generates a prediction image. Based on the prediction image and the mean image of the second subset, the computing device updates the parameters of the first encoding module, the first decoding module, and the bottleneck layer between the first encoding and first decoding modules using a self-supervised approach. The parameters of the second encoding module are frozen.

[0101] Optionally, the operation of training the ASL image reconstruction model in a self-supervised manner based on the predicted image and the mean image of the second subset includes: determining the task loss based on the predicted image and the mean image of the second subset, wherein the task loss includes at least one of the main supervision loss, edge gradient loss, structural similarity loss, and contrast loss; jointly training the first encoding module, the first decoding module, and the second encoding module to be trained with minimizing the task loss as the optimization objective, and freezing the parameters of the second encoding module during the training process; wherein the operation of determining the main supervision loss includes: determining the brain region images of the predicted image and the brain region images of the mean image respectively using the T1 image segmentation model; determining the main supervision loss based on the degree of difference between the brain region images of the predicted image and the brain region images of the mean image; and determining the edge gradient loss includes: determining the edge gradient loss based on the degree of difference in gradient between the brain region images of the predicted image and the brain region images of the mean image.

[0102] In the second phase of training, the computing device can determine the task loss based on the predicted image and the mean image of the second subset, and update the parameters of the first encoding module and the first decoding module with the goal of minimizing the task loss. The task loss can include four types of losses: master supervision loss, edge gradient loss, structural similarity loss, and contrast-weighted loss.

[0103] The master supervision loss is used to represent the degree of difference between the brain region image of the predicted image and the brain region image of the mean image.

[0104] Brain region images refer to images from which regions other than the brain have been removed, leaving only the brain regions, based on either a prediction image or a mean image. The computing device can use a T1 image segmentation model to determine the brain region images for both the prediction and mean images, thereby determining the primary supervised loss.

[0105] Main supervisor loss The expression is as follows:

[0106] in, This represents the predicted image after preprocessing. This represents the preprocessed mean image. Preprocessing here can refer to extracting the brain mask region (the brain mask region refers to the portion of the image that belongs to the brain) from the predicted / mean image.

[0107] Edge gradient loss is used to represent the degree of difference in edges between the brain region image of the predicted image and the brain region image of the mean image.

[0108] Marginal gradient loss The expression is:

[0109] in, This indicates calculating the gradient.

[0110] Structural similarity loss is used to represent the degree of difference between the predicted and mean images by the structural similarity between the brain region images of the predicted image and the brain region images of the mean image.

[0111] Structural similarity loss The expression is:

[0112] in, This represents the similarity calculation function.

[0113] Contrast-weighted loss The expression for (contrast-weighted L1) is as follows:

[0114] in, This represents any pixel position corresponding to the predicted image and the mean image, respectively. Represents the predicted pixel location in the image. The corresponding pixels, Represents the pixel position in the mean image The corresponding pixels, This represents the tissue contrast weight corresponding to the respective pixel position. This represents the probability that the pixel at the corresponding pixel location belongs to gray matter. This represents the probability that a pixel at a given pixel location is white matter. Pixels at the same pixel location may contain both gray and white matter. and It can be determined from the mean image or from the predicted image. In determining... hour The purpose is to maximize the value at the junction of the two types of tissue (i.e., gray matter and white matter), that is, at the boundary between gray matter and white matter, thus marking the locations where edge preservation is crucial. Therefore, the contrast-weighted loss is used to apply a higher weight to the L1 loss of N2N (i.e., the master supervision loss) at the gray matter-white matter boundary to prevent the cortical boundary from being smoothed out under noisy targets.

[0115] Furthermore, in this embodiment, noise can be added to the second training samples to improve the denoising capability of the image reconstruction model. For example, the computing device can mask some pixel positions of any one or more ASL differential image samples in the first subset of the second training samples, so that the image reconstruction model can only predict the value of the corresponding pixel position through neighboring pixels. As another example, noisy frames can be added to the first subset.

[0116] In addition, refer to Figure 1 As shown, according to a second aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0117] Therefore, according to this embodiment, in the reconstruction of ASL perfusion images, a first feature map related to the ASL differential image sequence is generated by a first encoding module, and a second feature map related to the T1 image is generated by a second encoding module. The first feature map and the second feature map are fused through a bottleneck layer to generate a third feature map, thereby generating a reconstructed ASL perfusion image based on the third feature map. Thus, in this respect, the method guides the reconstruction of ASL perfusion images through the anatomical structure information carried in the T1 image, which can improve the quality of image reconstruction.

[0118] Furthermore, before generating the first feature map through the first encoding module, this method first determines the pixel-wise weights corresponding to each ASL difference image (i.e., the logarithmic variance corresponding to the corresponding pixel of the ASL difference image mentioned above). Then, through these pixel-wise weights, the ASL difference images are aggregated to generate an ASL aggregated image. The first encoding module encodes the ASL aggregated image to generate the first feature map, thereby suppressing the motion artifacts that may appear in the ASL difference image sequence and their impact on the final reconstructed ASL perfusion image.

[0119] Furthermore, when fusing the first feature map and the second feature map through the bottleneck layer, this method generates a query matrix and a value matrix through the first feature map, and a key matrix through the second feature map, and then generates a third feature map by combining the cross-attention mechanism, thereby avoiding the leakage of the anatomical structure information of the T1 image to the first decoding module.

[0120] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0121] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0122] Example 2

[0123] Figure 6 A T1 image-guided ASL perfusion image reconstruction apparatus according to a first aspect of this embodiment is shown, which corresponds to the method described according to the first aspect of Embodiment 1. Reference Figure 6 As shown, the device includes: an acquisition module 610 for acquiring an ASL differential image sequence corresponding to a target site on a patient; an aggregation module 620 for aggregating the ASL differential image sequence to generate a corresponding ASL aggregated image; an ASL encoding module 630 for encoding the ASL aggregated image through a first encoding module to generate a first feature map; a T1 encoding module 640 for encoding the T1 image corresponding to the target site through a second encoding module to generate a second feature map; a fusion module 650 for fusing the first feature map and the second feature map through a bottleneck layer between the first encoding module and the corresponding first decoding module to generate a third feature map; and a perfusion image generation module 660 for generating a reconstructed ASL perfusion image based on the third feature map through the first decoding module.

[0124] Therefore, according to this embodiment, in the reconstruction of ASL perfusion images, a first feature map related to the ASL differential image sequence is generated by a first encoding module, and a second feature map related to the T1 image is generated by a second encoding module. The first feature map and the second feature map are fused through a bottleneck layer to generate a third feature map, thereby generating a reconstructed ASL perfusion image based on the third feature map. Thus, in this respect, the device guides the reconstruction of ASL perfusion images through the anatomical structure information carried in the T1 image, which can improve the quality of image reconstruction.

[0125] Furthermore, before generating the first feature map through the first encoding module, this device first determines the pixel-wise weights corresponding to each ASL difference image (i.e., the logarithmic variance corresponding to the corresponding pixel of the ASL difference image mentioned above). Then, through these pixel-wise weights, the ASL difference images are aggregated to generate an ASL aggregated image. The first encoding module encodes the ASL aggregated image to generate the first feature map, thereby suppressing the motion artifacts that may appear in the ASL difference image sequence and their impact on the final reconstructed ASL perfusion image.

[0126] Furthermore, when fusing the first feature map and the second feature map through the bottleneck layer, this device generates a query matrix and a value matrix through the first feature map, and generates a key matrix through the second feature map. Then, it generates a third feature map by combining the cross-attention mechanism, thereby avoiding the leakage of the anatomical structure information of the T1 image to the first decoding module.

[0127] Example 3

[0128] Figure 7 A T1 image-guided ASL perfusion image reconstruction apparatus according to a first aspect of this embodiment is shown, which corresponds to the method described according to the first aspect of Embodiment 1. Reference Figure 7 As shown, the device includes: a processor 710; and a memory 720 connected to the processor 710, for providing the processor 710 with instructions to process the following steps: acquiring an ASL differential image sequence corresponding to a target site of the patient; aggregating the ASL differential image sequence to generate a corresponding ASL aggregated image; encoding the ASL aggregated image through a first encoding module to generate a first feature map; encoding a T1 image corresponding to the target site through a second encoding module to generate a second feature map; fusing the first feature map and the second feature map through a bottleneck layer between the first encoding module and a corresponding first decoding module to generate a third feature map; and generating a reconstructed ASL perfusion image based on the third feature map through the first decoding module.

[0129] Therefore, according to this embodiment, in the reconstruction of ASL perfusion images, a first feature map related to the ASL differential image sequence is generated by a first encoding module, and a second feature map related to the T1 image is generated by a second encoding module. The first feature map and the second feature map are fused through a bottleneck layer to generate a third feature map, thereby generating a reconstructed ASL perfusion image based on the third feature map. Thus, in this respect, the device guides the reconstruction of ASL perfusion images through the anatomical structure information carried in the T1 image, which can improve the quality of image reconstruction.

[0130] Furthermore, before generating the first feature map through the first encoding module, this device first determines the pixel-wise weights corresponding to each ASL difference image (i.e., the logarithmic variance corresponding to the corresponding pixel of the ASL difference image mentioned above). Then, through these pixel-wise weights, the ASL difference images are aggregated to generate an ASL aggregated image. The first encoding module encodes the ASL aggregated image to generate the first feature map, thereby suppressing the motion artifacts that may appear in the ASL difference image sequence and their impact on the final reconstructed ASL perfusion image.

[0131] Furthermore, when fusing the first feature map and the second feature map through the bottleneck layer, this device generates a query matrix and a value matrix through the first feature map, and generates a key matrix through the second feature map. Then, it generates a third feature map by combining the cross-attention mechanism, thereby avoiding the leakage of the anatomical structure information of the T1 image to the first decoding module.

[0132] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0133] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0134] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0135] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0136] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0137] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0138] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for ASL perfusion image reconstruction based on T1 image guidance, characterized in that, include: Obtain ASL differential image sequences corresponding to the target sites on the patient; The ASL differential image sequences are aggregated to generate corresponding ASL aggregated images; The ASL aggregated image is encoded using the first encoding module to generate a first feature map; The second encoding module encodes the T1 image corresponding to the target region to generate a second feature map. as well as By using the bottleneck layer between the first encoding module and the corresponding first decoding module, the first feature map and the second feature map are fused to generate a third feature map; as well as The first decoding module generates a reconstructed ASL perfusion image based on the third feature map.

2. The method according to claim 1, characterized in that, The operation of fusing the first feature map and the second feature map to generate a third feature map through the bottleneck layer between the first encoding module and the corresponding first decoding module includes: The bottleneck layer generates a query matrix and a value matrix corresponding to the first feature map. A key matrix corresponding to the second feature map is generated through the bottleneck layer; and The bottleneck layer generates the third feature map using a cross-attention mechanism based on the query matrix, the value matrix, and the key matrix.

3. The method according to claim 2, characterized in that, The operation of generating the third feature map through the bottleneck layer, based on the query matrix, the value matrix, and the key matrix, using a cross-attention mechanism, includes: Through the bottleneck layer, attention information corresponding to the third feature map is generated using a cross-attention mechanism based on the query matrix, the value matrix, and the key matrix; and Through the bottleneck layer, the third feature map is generated based on the attention information using a residual gating mechanism.

4. The method according to claim 1, characterized in that, The operation of aggregating the ASL difference image sequences to generate corresponding ASL aggregated images includes: The average value of pixels at the same position in each ASL differential image in the ASL differential image sequence is calculated to generate an initial aggregated image; Determine the pixel value deviation between each pixel in the ASL difference image and the corresponding pixel in the initial aggregate image, and determine the logarithmic variance corresponding to each pixel in the ASL difference image based on the pixel value deviation; For each pixel position in the ASL difference image, the pixel values ​​of the corresponding pixels in each ASL difference image are weighted and summed according to the logarithmic variance of the corresponding pixels in each ASL difference image to generate the ASL aggregate image.

5. The method according to claim 1, characterized in that, The first encoding module includes a residual unit and at least one downsampling unit, and the first decoding module includes at least one upsampling unit and a convolution unit. The operation of encoding the ASL aggregated image through the first encoding module to generate the first feature map includes: A first intermediate feature map is generated from the ASL aggregated image using the residual unit; and The first feature map is generated based on the first intermediate feature map using the at least one downsampling unit, and wherein... The operation of generating a reconstructed ASL perfusion image based on the third feature map through the first decoding module includes: A second intermediate feature map is generated based on the third feature map using the at least one upsampling unit; and The reconstructed ASL perfusion image is generated using the convolutional unit based on the second intermediate feature map.

6. The method according to claim 5, characterized in that, The operation of generating a second intermediate feature map from the third feature map using the at least one upsampling unit includes: The second intermediate feature map is generated by the at least one upsampling unit based on the third feature map and the third intermediate feature map received from the first encoding module through the skip connection layer, wherein... The operation of generating the reconstructed ASL perfusion image based on the second intermediate feature map through the convolutional unit includes: generating the reconstructed ASL perfusion image based on the second intermediate feature map and the first intermediate feature map received from the first encoding module through the skip connection layer through the convolutional unit.

7. The method according to claim 1, characterized in that, Before acquiring the ASL differential image sequence corresponding to the target site on the patient, the following steps are also included: A T1 image segmentation model is constructed, which includes a second encoding module to be trained and a second decoding module to be trained. The first encoding module and the second encoding module have the same structure, and the first decoding module and the second decoding module have the same structure. The T1 image segmentation model is trained in a supervised manner using the first training sample, which includes T1 sample images and corresponding segmented image samples. An ASL image reconstruction model is constructed, comprising the first encoding module, the first decoding module, the bottleneck layer, and the second encoding module. The ASL image reconstruction model is trained using second training samples while the parameters of the second encoding module are frozen. The second training samples include ASL differential image sample sequences and corresponding T1 sample images.

8. The method according to claim 7, characterized in that, The operation of training the ASL image reconstruction model using the second training sample includes: The ASL differential image sample sequence is split into a first subset and a second subset; The ASL difference images in the second subset are fused by averaging to determine the mean image corresponding to the second subset; A predicted image is generated based on the first subset and the T1 sample image using the first encoding module to be trained, the first decoding module to be trained, and the second encoding module. The ASL image reconstruction model is trained using a self-supervised approach based on the predicted image and the mean image of the second subset.

9. The method according to claim 8, characterized in that, The operation of training the ASL image reconstruction model in a self-supervised manner based on the predicted image and the mean image of the second subset includes: Based on the predicted image and the mean image of the second subset, a task loss is determined, wherein the task loss includes at least one of the following: main supervision loss, edge gradient loss, structural similarity loss, and contrast-weighted loss: The ASL image reconstruction model is trained with the goal of minimizing the task loss, and the parameters of the second encoding module are frozen during the training process; and wherein, The operation of determining the primary supervisory loss includes: The brain region images of the predicted image and the brain region images of the mean image are determined using the T1 image segmentation model. The main supervision loss is determined based on the degree of difference between the brain region images of the predicted image and the brain region images of the mean image; The operation of determining the marginal gradient loss includes: The edge gradient loss is determined based on the degree of difference in gradient between the brain region image of the predicted image and the brain region image of the mean image.

Images