Medical image compression method, device, equipment, medium and product
By decomposing medical images into structural and texture information subimages, and using Transformer feature alignment and parallel autoregressive coding modules for compression, the problem of slow compression speed of high-bit depth images in the prior art is solved, and more efficient compression performance is achieved.
Patent Information
- Application Number
- CN202510564780.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-08-05
AI Technical Summary
The existing medical image compression framework fails to effectively utilize the characteristics of high-deep medical images, resulting in slow compression speed and low throughput, which cannot meet the needs of efficient storage and transmission.
The target medical image is decomposed into sub-images of structure information and texture information, and the feature alignment module based on Transformer and the parallel autoregressive coding module are used for compression. Sliced images of different bit depths are processed separately, and the dependencies between slices are used to reduce redundancy.
Through bit-deep division and joint compression, memory usage is reduced, compression performance is improved, compression speed and efficiency are improved.
Smart Images
Figure CN120434397A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data compression technology, and in particular to a medical image compression method, device, equipment, medium and product. Background Art
[0002] Medical images, such as computed tomography (CT) and magnetic resonance imaging (MRI) scans, utilize tomography technology to generate a series of cross-sectional images (slices) at varying depths, constructing a comprehensive three-dimensional (3D) representation of internal body structures. A single CT scan can generate a 3D image of a patient's body consisting of dozens or even hundreds of two-dimensional (2D) slices, each with a depth of 16 bits, for a total data volume of hundreds of megabytes or even a gigabyte. Therefore, the slice images need to be compressed for transmission. Relevant medical image compression frameworks are not specifically designed for the high bit-depth characteristics of medical images, resulting in slow compression speeds and low throughput when compressing high-bit-depth medical images. Therefore, there is an urgent need to develop more advanced compression algorithms to improve data storage and transmission efficiency. Summary of the Invention
[0003] The purpose of this application is to provide a medical image compression method, device, equipment, medium and product, which reduces the redundancy between high and low bitmaps and improves compression performance.
[0004] To achieve the above objectives, this application provides the following solutions:
[0005] In a first aspect, the present application provides a medical image compression method, comprising:
[0006] Decomposing a target medical image into a first sub-image and a second sub-image according to a preset bit division position; the first sub-image represents structural information of the target medical image, the second sub-image represents texture information of the target medical image, the first sub-image includes a plurality of first slice images arranged in chronological order, and the second sub-image includes a plurality of second slice images arranged in chronological order;
[0007] compressing the first sub-image using a first codec;
[0008] A second codec is used to sequentially compress the second slice images in the second sub-image; the second encoder includes a Transformer-based feature alignment module and a parallel autoregressive encoding module; the Transformer-based feature alignment module is used to extract a first feature aligned in time from the first slice image and the second slice image; the parallel autoregressive encoding module is used to extract a second feature from the first feature and the second slice image at the time corresponding to the first feature, and the first feature and the second feature are fused to obtain a compressed second slice image.
[0009] Optionally, decomposing the target medical image into a first sub-image and a second sub-image according to a preset bit division position specifically includes:
[0010] The first sub-image is represented as
[0011] The second sub-image is represented by X L =X mod 2 d =(X1 L , X2 L ,…,X T L );
[0012] Wherein, X represents the target medical image, X=(X1, X2, ..., X T ), X t represents the t-th original slice, 1≤t≤T, T is the number of original slices, X M represents the first sub-image, X L represents the second sub-image, d represents the preset bit division position, d is an integer, represents the tth first slice image, Represents the t-th second slice image.
[0013] Optionally, the Transformer-based feature alignment module is expressed as:
[0014]
[0015] in, represents the t-th first feature, represents the Transformer-based feature alignment module, represents the t-1th first slice image, represents the t-1th second slice image.
[0016] Optionally, the Transformer-based feature alignment module includes a first feature extraction layer, a second feature extraction layer, a first embedding layer, a second embedding layer, a third embedding layer, a fourth embedding layer, a fifth embedding layer, a sixth embedding layer, a first attention module and a second attention module, wherein the first input of the first feature extraction layer is used to input The second input of the first feature extraction layer is used to input The second feature extraction layer is used to input The first output of the first feature extraction layer is connected to the input of the first embedding layer, the second output of the first feature extraction layer is connected to the input of the second embedding layer, the output of the second feature extraction layer is connected to the input of the third embedding layer, the output of the first embedding layer, the output of the second embedding layer and the output of the third embedding layer are all connected to the input of the first attention module, the first output of the first feature extraction layer is connected to the input of the fourth embedding layer, the second output of the first feature extraction layer is connected to the input of the fifth embedding layer, the output of the first attention module is connected to the input of the sixth embedding layer, the output of the fourth embedding layer, the output of the fifth embedding layer and the output of the sixth embedding layer are all connected to the input of the second attention module, and the second attention module is used to output the first feature.
[0017] Optionally, the first attention module and the second attention module have the same structure, both comprising a cross-attention block, a conditional position embedding and a self-attention block connected in sequence.
[0018] Optionally, the parallel autoregressive encoding module includes a masked convolution block, a parameter prediction network, an arithmetic encoder, an arithmetic decoder and a logical distribution function, the masked convolution block is used to input the second slice image, the arithmetic encoder is used to input the second slice image, the masked convolution block is used to output the second feature, the parameter prediction network is used to input the second feature and the first feature, the output of the parameter prediction network is input to the arithmetic encoder and the arithmetic decoder respectively after passing through the logical distribution function, the output of the arithmetic encoder is connected to the input of the arithmetic decoder, and the output of the arithmetic encoder is the compressed second slice image; the parameter prediction network is used to output the mixed logical model parameters of the second slice image, and the mixed logical model parameters include mixing weights, mean and variance.
[0019] In a second aspect, the present application provides a medical image compression device, comprising:
[0020] a medical image decomposition module, configured to decompose a target medical image into a first sub-image and a second sub-image according to a preset bit division position; the first sub-image represents structural information of the target medical image, the second sub-image represents texture information of the target medical image, the first sub-image includes a plurality of first slice images arranged in chronological order, and the second sub-image includes a plurality of second slice images arranged in chronological order;
[0021] A first sub-image compression module, configured to compress the first sub-image using a first codec;
[0022] The second sub-image compression module is used to sequentially compress the second slice image in the second sub-image using a second codec; the second encoder includes a Transformer-based feature alignment module and a parallel autoregressive encoding module; the Transformer-based feature alignment module is used to extract the first feature aligned in time from the first slice image and the second slice image; the parallel autoregressive encoding module is used to extract the second feature from the first feature and the second slice image at the corresponding time of the first feature, and fuse the first feature and the second feature to obtain a compressed second slice image.
[0023] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any one of the above-described medical image compression methods.
[0024] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the above-mentioned medical image compression methods.
[0025] In a fifth aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any one of the above-mentioned medical image compression methods.
[0026] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0027] The present application provides a medical image compression method, apparatus, device, medium and product, which decomposes a target medical image into a first sub-image and a second sub-image according to a preset bit division position, reduces the bit depth of each slice image, reduces memory usage, and jointly compresses slice images of different bit depths separately. Based on the dependency relationship between the slice images, the redundancy between high- and low-bit-depth slice images can be reduced, thereby improving compression performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0029] Figure 1 A flowchart of a medical image compression method provided in one embodiment of the present application;
[0030] Figure 2 A schematic diagram of a 3D lung CT image provided in one embodiment of the present application;
[0031] Figure 3 A schematic diagram of a slice image provided in an embodiment of the present application;
[0032] Figure 4 A schematic diagram of the second codec structure provided in an embodiment of the present application;
[0033] Figure 5 A schematic diagram of the Transformer-based feature alignment module (TFAM) structure provided in one embodiment of the present application;
[0034] Figure 6 A schematic diagram of the structure of a parallel autoregressive coding module (PACM) provided in one embodiment of the present application;
[0035] Figure 7 A schematic diagram of the feature extraction layer structure provided in one embodiment of the present application;
[0036] Figure 8 A schematic diagram of the cross-attention block structure provided in one embodiment of the present application;
[0037] Figure 9 A schematic diagram of the structure of an efficient multi-head self-attention block provided in one embodiment of the present application;
[0038] Figure 10 A schematic diagram of the embedded layer structure provided in one embodiment of the present application;
[0039] Figure 11 A schematic diagram of the structure of a self-attention module provided in one embodiment of the present application;
[0040] Figure 12 A schematic diagram of the structure of a convolutional feed-forward network (ConvFFN) provided in one embodiment of the present application;
[0041] Figure 13A schematic diagram of a parameter prediction network structure provided in one embodiment of the present application;
[0042] Figure 14 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION
[0043] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0044] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0045] This application provides a medical image compression method, such as Figure 1 As shown, the medical image compression method includes:
[0046] Step 101: Decompose a target medical image into a first sub-image and a second sub-image according to a preset bit division position; the first sub-image represents structural information of the target medical image, the second sub-image represents texture information of the target medical image, the first sub-image includes a plurality of first slice images arranged in chronological order, and the second sub-image includes a plurality of second slice images arranged in chronological order.
[0047] Step 102: Compress the first sub-image using a first codec.
[0048] Step 103: Use a second codec to sequentially compress the second slice image in the second sub-image; the second encoder includes a Transformer-based feature alignment module and a parallel autoregressive encoding module; the Transformer-based feature alignment module is used to extract a first feature aligned in time from the first slice image and the second slice image; the parallel autoregressive encoding module is used to extract a second feature from the first feature and the second slice image at the time corresponding to the first feature, and fuse the first feature and the second feature to obtain a compressed second slice image.
[0049] The present application decomposes the target medical image into a first sub-image and a second sub-image according to a preset bit division position, thereby reducing the bit depth of each slice image and reducing memory usage. By jointly compressing slice images of different bit depths, based on the dependency relationship between the slice images, the redundancy between high and low bit depth slice images can be reduced, thereby improving compression performance.
[0050] In an exemplary embodiment, the target medical image in this application is a high-depth medical image, such as Figure 2 As shown, the xy axis is the slice cross section, the z axis is the slice stacking direction, and the first codec is a traditional codec, for example, JPEG-XL is used to compress the first sub-image.
[0051] The second codec is a learning-based codec that compresses LSBV slice by slice with MSBV as the condition. Specifically, the Transformer-based Feature Alignment Module (TFAM) is used to extract the aligned first feature from MSBV and LSBV. Extract local features using Parallel Autoregressive Coding Module (PACM) That is the second feature, and Fusion to estimate the probability distribution of the current LSBV.
[0052] The first sub-image is represented as The first sub-image is a significance bit sub-volume picture (MSBV).
[0053] The second sub-image is represented by X L =X mod 2 d =(X1 L , X2 L ,…,X T L ), the second sub-image is a LSBV image.
[0054] Wherein, X represents the target medical image, X=(X1, X2, ..., X T ), X t represents the tth original slice, such as Figure 3 As shown in part (a), 1≤t≤T, T is the number of original slices, X M represents the first sub-image, X L represents the second sub-image, d represents the preset bit division position, d is an integer, 1≤d≤16, Represents the tth first slice image, such as Figure 3 As shown in part (b), Represents the t-th second slice image, such as Figure 3As shown in part (c), X t+1 and X t+2 The residual graph between Figure 3 As shown in part (f), X t+1 The t+1th original slice, such as Figure 3 As shown in part (d), X t+2 The t+2th original slice, such as Figure 3 As shown in part (e).
[0055] The first slice image contains sparse structural information. This sparsity indicates that there is similarity between pixels in neighboring areas and is easy to compress. MSBV can be compressed using any traditional codec. Based on speed and performance considerations, this application uses JPEG-XL to compress MSBV to reduce the computational complexity of the entire compression framework.
[0056] LSBS contains additional texture information and exhibits more random noise, which brings more challenges to compression. This application constructs a slice-by-slice autoregressive model to exploit the redundancy between slices.
[0057] The distribution of LSBV can be factored as:
[0058] Among them, p θ (X L ) is the distribution of LSBV, To pass estimate The probability distribution of are all second slice images before the t-th second slice image.
[0059] For volume data with many slices (slice images), it is impractical to use all the coded slices as conditions. This application can significantly reduce the redundancy between slices and greatly reduce the complexity of calculation. The probability is simplified to the Markov chain model To pass estimate To further exploit the redundancy within a slice, we use and As an additional condition, the The conditional probability distribution of . The probability distribution of is thus expressed as Indicates passing and estimate The probability distribution of . Assume that the true distribution is p(X L ), then the cross entropy H(p, p θ), that is, the estimated value of the compression code rate is:
[0060]
[0061] Where p=p(X L ), represents p(X L ) expectations.
[0062] This application constructs two modules to efficiently model the cross entropy H(p, p θ ), are the feature alignment module and parallel autoregressive encoding module based on Transformer, and the second codec composed of the feature alignment module and parallel autoregressive encoding module based on Transformer is as follows Figure 4 As shown. The second codec is specifically a bit-splitting-based lossless medical volume image compression (BD-LVIC) framework. Each medical volume slice is decomposed into the most significant bitmap (MSBS) and the least significant bitmap (LSBS). First, all MSBS are compressed using a traditional codec. Then, each LSBS is encoded slice by slice. During this period, the Transformer-based Feature Alignment Module (TFAM) is used to generate alignment features And use the parallel autoregressive coding module (PACM) to extract local context and with Fusion to estimate The probability distribution of .
[0063] TFAM uses the Transformer module to simultaneously model intra-slice and inter-slice dependencies, generating Aligned Features
[0064] The Transformer-based feature alignment module is expressed as:
[0065] Therefore, the distribution Replaced by
[0066] in, represents the t-th first feature, represents the Transformer-based feature alignment module, represents the t-1th first slice image, represents the t-1th second slice image.
[0067] The parallel autoregressive encoding module uses the local autoregressive model to exploit the inherent spatial redundancy of LSBS. The process of the parallel autoregressive encoding module is as follows: Figure 6 As shown, masked convolution is first used to extract local spatial context, which is expressed as This context is then aligned with the features Connect and input to the Parameter Prediction Network (PPN) to estimate The distribution of Replaced by like Figure 5 As shown, PPN is generated to estimate The parameters of the mixed logistic model of the probability distribution include the mixing weight π, mean μ, and variance σ. This is achieved by stacking multiple 1×1 convolutional layers. and The use of 1×1 convolutional layers can avoid obtaining information beyond the local receptive field. Given π, μ, and σ, the probability distribution can be expressed as follows:
[0068]
[0069] Where logistic(·) refers to the logistic distribution and N represents the number of mixture components. Unlike color images which have three channels, medical images usually have only one channel. By formula (4), the cross entropy H(p, p θ ) can be expressed as:
[0070]
[0071] This result represents X L The compression code rate. In the framework of this application, formula (5) is used as the loss function, and formula (5) is minimized through end-to-end training. In actual encoding, an entropy encoder is needed to calculate the estimated probability distribution. Will Encoded into a bitstream.
[0072] like Figure 5 As shown, the Transformer-based feature alignment module includes a first feature extraction layer, a second feature extraction layer, a first embedding layer, a second embedding layer, a third embedding layer, a fourth embedding layer, a fifth embedding layer, a sixth embedding layer, a first attention module and a second attention module, and the first input of the first feature extraction layer is used to input The second input of the first feature extraction layer is used to input The second feature extraction layer is used to input The first output of the first feature extraction layer is connected to the input of the first embedding layer, the second output of the first feature extraction layer is connected to the input of the second embedding layer, the output of the second feature extraction layer is connected to the input of the third embedding layer, the output of the first embedding layer, the output of the second embedding layer and the output of the third embedding layer are all connected to the input of the first attention module, the first output of the first feature extraction layer is connected to the input of the fourth embedding layer, the second output of the first feature extraction layer is connected to the input of the fifth embedding layer, the output of the first attention module is connected to the input of the sixth embedding layer, the output of the fourth embedding layer, the output of the fifth embedding layer and the output of the sixth embedding layer are all connected to the input of the second attention module, and the second attention module is used to output the first feature.
[0073] The first feature extraction layer and the second feature extraction layer have the same structure. Figure 7 As shown in Figure 1, the feature extraction layer (FE) includes the first convolution, the first GELU activation function, the second convolution, the second GELU activation function and the third convolution connected in sequence. The first convolution, the second convolution and the third convolution are all convolution operations with a convolution kernel size of 3×3 and a stride of 1. The output of the first convolution is connected to the output of the third convolution. Figure 7 In the figure, k3s1 represents a convolution with a kernel size of 3×3 and a stride of 1.
[0074] The first attention module and the second attention module have the same structure, both including a cross attention block (CAB), a conditional position embedding (CPE), and a self-attention block (SAB) connected in sequence. The structure of CAB and SAB includes the Transformer structure.
[0075] like Figure 8 As shown, the cross-attention block includes a first layer of normalization, a first efficient multi-head self-attention, a second layer of normalization and a first convolutional feedforward network connected in sequence. The first layer of normalization is used to input the output (Q) of the first embedding layer, the output (K) of the second embedding layer and the output (V) of the third embedding layer. The first layer of normalization is connected to the input of the first efficient multi-head self-attention, the output of the first efficient multi-head self-attention is connected to the input of the second layer of normalization, the output of the second layer of normalization is connected to the input of ConvFFN, the output of the third embedding layer is also connected to the output of the first efficient multi-head self-attention, and the input of the second layer of normalization is also connected to the output of the first convolutional feedforward network.
[0076] Efficient multi-head self-attention structure such as Figure 9As shown in the figure, the efficient multi-head self-attention includes the fourth convolution, the fifth convolution, the sixth convolution, the first depth convolution, the second depth convolution and the multi-head self-attention. The fourth convolution is used to perform a convolution operation on the dimension corresponding to V (the normalized output of the first layer), the fifth convolution is used to perform a convolution operation on the dimension corresponding to K (the normalized output of the first layer), and the sixth convolution is used to perform a convolution operation on the dimension corresponding to Q (the normalized output of the first layer). The outputs of the fourth and fifth convolutions are output to the multi-head self-attention after passing through the first depth convolution. The output of the sixth convolution is output to the multi-head self-attention after passing through the second depth convolution. The fourth, fifth and sixth convolutions are all convolution operations with a convolution kernel size of 1×1 and a stride of 1. Figure 9 In the above code, 1×1 represents convolution k1s1, k1s1 represents convolution with kernel size of 1×1 and stride of 1, H represents height, W represents width, C represents number of channels, and S represents downsampling multiple. The downsampling multiple is 1 less than the convolution kernel. This is mainly to ensure that only the edge of the field of view overlaps during downsampling, while ensuring performance while increasing the downsampling multiple of one-step convolution as much as possible.
[0077] The structures of the embedded layers (EL) (the first embedded layer, the second embedded layer, the third embedded layer, the fourth embedded layer, the fifth embedded layer and the sixth embedded layer) are the same. Figure 10 As shown, they all include a third depth convolution, a third GELU activation function, a fourth depth convolution, and a fifth depth convolution connected in sequence, and the output of the fourth depth convolution is also connected to the output of the fifth depth convolution, wherein, Figure 10 CPE stands for Conditional Position Embedding.
[0078] The structure of the self-attention block is as follows Figure 11 As shown, the self-attention block includes a third layer of normalization, a second efficient multi-head self-attention, a fourth layer of normalization, and a second convolutional feedforward network connected in sequence. The input of the third layer of normalization is also connected to the output of the second efficient multi-head self-attention, and the output of the second efficient multi-head self-attention is connected to the output of the second convolutional feedforward network.
[0079] The structures of the first convolutional feedforward network and the second convolutional feedforward network are as follows: Figure 12 The convolutional feed-forward network shown includes a sixth depthwise convolution, a fifth layer normalization, a seventh convolution, a fourth GELU activation function, and an eighth convolution connected in sequence, and the input of the sixth depthwise convolution is also connected to the output of the eighth convolution.
[0080] To optimize the computational efficiency of the Transformer-based feature alignment module and reduce memory usage, deep convolution and stride s are applied to Efficient Multi-Head Self-Attention (EMHSA) to reduce the spatial dimensions of K (key) and V (value) before the attention operation, where Q represents the query. In addition, unlike the original feedforward network (FFN) in the Visual Transformer (ViT), this application replaces the linear layer with deep convolution to better integrate local information.
[0081] The parallel autoregressive coding module includes a masked convolution block, a parameter prediction network (PPN), an arithmetic encoder, an arithmetic decoder and a logical distribution function. The masked convolution block is used to input the second slice image, the arithmetic encoder is used to input the second slice image, the masked convolution block is used to output the second feature, the parameter prediction network is used to input the second feature and the first feature, and the output of the parameter prediction network is input to the arithmetic encoder and the arithmetic decoder respectively after passing through the logical distribution function. The output of the arithmetic encoder is connected to the input of the arithmetic decoder, and the output of the arithmetic encoder is the compressed second slice image; the parameter prediction network is used to output the mixed logical model parameters of the second slice image, and the mixed logical model parameters include mixing weights, mean and variance.
[0082] The structure of the parameter prediction network is as follows Figure 13 The figure includes the concatenation (Concat) operation, the ninth convolution, the tenth convolution, the eleventh convolution, the twelfth convolution, the thirteenth convolution, the fourteenth convolution and Chunk, which are connected in sequence. The ninth convolution and the tenth convolution are connected by a GELU activation function, the eleventh convolution and the twelfth convolution are connected by a GELU activation function, and the thirteenth convolution and the fourteenth convolution are connected by a GELU activation function. Chunk is used to output the mixed logic model parameters.
[0083] Chunk represents the segmentation of data. Usually, the deep learning matrix structure is (BC HW), where B represents the batch and C represents the number of channels. If C is divisible by 3, then chunk(3, 1) means cutting into three equal parts along the second dimension.
[0084] In the data transmission scenario, the first codec compresses the first sub-image into a code stream ( Figure 4 The code stream ( Figure 4 After the code stream (output by the arithmetic encoder) is transmitted, it is decoded at the receiving end.
[0085] This application efficiently compresses high-bit-depth medical images. By using bit-slicing, the bit depth of each sub-image is reduced, avoiding the slow encoding speed and high video memory usage caused by building a large probability table. Furthermore, by jointly optimizing high- and low-bit-depth sub-images, redundancy between them is effectively reduced, significantly improving the model's compression performance.
[0086] This application designs a Transformer-based feature alignment module (TFAM) and an efficient multi-head self-attention mechanism, so that TFAM can efficiently model the dependencies between slices and between high and low subgraphs, which helps to improve the compression performance of the model.
[0087] Based on the same inventive concept, embodiments of the present application also provide a medical image compression device for implementing the aforementioned medical image compression method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations in one or more embodiments of the medical image compression device provided below can be found in the above-described limitations on the medical image compression method and will not be further elaborated here.
[0088] In an exemplary embodiment, the present application provides a medical image compression device comprising:
[0089] a medical image decomposition module, configured to decompose a target medical image into a first sub-image and a second sub-image according to a preset bit division position; the first sub-image represents structural information of the target medical image, the second sub-image represents texture information of the target medical image, the first sub-image includes a plurality of first slice images arranged in chronological order, and the second sub-image includes a plurality of second slice images arranged in chronological order;
[0090] A first sub-image compression module, configured to compress the first sub-image using a first codec;
[0091] The second sub-image compression module is used to sequentially compress the second slice image in the second sub-image using a second codec; the second encoder includes a Transformer-based feature alignment module and a parallel autoregressive encoding module; the Transformer-based feature alignment module is used to extract the first feature aligned in time from the first slice image and the second slice image; the parallel autoregressive encoding module is used to extract the second feature from the first feature and the second slice image at the corresponding time of the first feature, and fuse the first feature and the second feature to obtain a compressed second slice image.
[0092] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 14 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store medical image compression data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a medical image compression method is implemented.
[0093] Those skilled in the art will understand that Figure 14 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application and does not constitute a limitation on the computer device to which the solution of the present application is applied. A specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the above-mentioned method embodiments when executing the computer program.
[0094] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.
[0095] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0096] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.
[0097] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0098] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, data processing logic of programmable logic devices, and the like.
[0099] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0100] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A medical image compression method, characterized in that: The medical image compression method comprises: Decomposing a target medical image into a first sub-image and a second sub-image according to a preset bit division position; the first sub-image represents structural information of the target medical image, the second sub-image represents texture information of the target medical image, the first sub-image includes a plurality of first slice images arranged in chronological order, and the second sub-image includes a plurality of second slice images arranged in chronological order; compressing the first sub-image using a first codec; A second codec is used to sequentially compress the second slice images in the second sub-image; the second encoder includes a Transformer-based feature alignment module and a parallel autoregressive encoding module; the Transformer-based feature alignment module is used to extract a first feature aligned in time from the first slice image and the second slice image; the parallel autoregressive encoding module is used to extract a second feature from the first feature and the second slice image at the time corresponding to the first feature, and the first feature and the second feature are fused to obtain a compressed second slice image.
2. The medical image compression method according to claim 1, wherein: Decomposing the target medical image into a first sub-image and a second sub-image according to a preset bit division position specifically includes: The first sub-image is represented as The second sub-image is represented by X L =X mod 2 d =(X1 L , X2 L ,…,X T L ); Wherein, X represents the target medical image, X=(X1, X2, ..., X T ), X t represents the t-th original slice, 1≤t≤T, T is the number of original slices, X M represents the first sub-image, X L represents the second sub-image, d represents the preset bit division position, d is an integer, represents the tth first slice image, Represents the t-th second slice image.
3. The medical image compression method according to claim 2, wherein: The Transformer-based feature alignment module is expressed as: in, represents the t-th first feature, represents the Transformer-based feature alignment module, represents the t-1th first slice image, represents the t-1th second slice image.
4. The medical image compression method according to claim 3, wherein: The Transformer-based feature alignment module includes a first feature extraction layer, a second feature extraction layer, a first embedding layer, a second embedding layer, a third embedding layer, a fourth embedding layer, a fifth embedding layer, a sixth embedding layer, a first attention module and a second attention module. The first input of the first feature extraction layer is used to input The second input of the first feature extraction layer is used to input The second feature extraction layer is used to input The first output of the first feature extraction layer is connected to the input of the first embedding layer, the second output of the first feature extraction layer is connected to the input of the second embedding layer, the output of the second feature extraction layer is connected to the input of the third embedding layer, the output of the first embedding layer, the output of the second embedding layer and the output of the third embedding layer are all connected to the input of the first attention module, the first output of the first feature extraction layer is connected to the input of the fourth embedding layer, the second output of the first feature extraction layer is connected to the input of the fifth embedding layer, the output of the first attention module is connected to the input of the sixth embedding layer, the output of the fourth embedding layer, the output of the fifth embedding layer and the output of the sixth embedding layer are all connected to the input of the second attention module, and the second attention module is used to output the first feature.
5. The medical image compression method according to claim 4, characterized in that: The first attention module and the second attention module have the same structure, both including a cross-attention block, a conditional position embedding and a self-attention block connected in sequence.
6. The medical image compression method according to claim 2, wherein: The parallel autoregressive encoding module includes a masked convolution block, a parameter prediction network, an arithmetic encoder, an arithmetic decoder and a logical distribution function, wherein the masked convolution block is used to input the second slice image, the arithmetic encoder is used to input the second slice image, the masked convolution block is used to output the second feature, the parameter prediction network is used to input the second feature and the first feature, and the output of the parameter prediction network is input to the arithmetic encoder and the arithmetic decoder respectively after passing through the logical distribution function, the output of the arithmetic encoder is connected to the input of the arithmetic decoder, and the output of the arithmetic encoder is the compressed second slice image; The parameter prediction network is used to output the mixed logic model parameters of the second slice image, and the mixed logic model parameters include a mixed weight, a mean, and a variance.
7. A medical image compression device, characterized in that: The medical image compression device comprises: a medical image decomposition module, configured to decompose a target medical image into a first sub-image and a second sub-image according to a preset bit division position; the first sub-image represents structural information of the target medical image, the second sub-image represents texture information of the target medical image, the first sub-image includes a plurality of first slice images arranged in chronological order, and the second sub-image includes a plurality of second slice images arranged in chronological order; A first sub-image compression module, configured to compress the first sub-image using a first codec; The second sub-image compression module is used to sequentially compress the second slice image in the second sub-image using a second codec; the second encoder includes a Transformer-based feature alignment module and a parallel autoregressive encoding module; the Transformer-based feature alignment module is used to extract the first feature aligned in time from the first slice image and the second slice image; the parallel autoregressive encoding module is used to extract the second feature from the first feature and the second slice image at the corresponding time of the first feature, and fuse the first feature and the second feature to obtain a compressed second slice image.
8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the medical image compression method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the medical image compression method according to any one of claims 1 to 6 is implemented.
10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the medical image compression method according to any one of claims 1 to 6 is implemented.