Three-dimensional medical image super-resolution reconstruction method and system

Abstract

The application provides a three-dimensional medical image super-resolution reconstruction method and system, relates to the technical field of image super-resolution reconstruction, and comprises the following steps: collecting low-resolution medical volume data of the same part, and performing isotropic interpolation to obtain an axial slice sequence, a coronal slice sequence and a sagittal slice sequence; performing feature extraction, and imposing a projection consistency constraint on each feature extraction result to form fused hidden space features; sending the feature slice sequences in each direction into an up-sampling decoding head to perform sub-pixel convolution up-sampling and residual convolution refinement; performing loss calculation on the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence and the sagittal super-resolution two-dimensional slice sequence based on training true values, and outputting network parameters after training is completed; and mapping the super-resolution two-dimensional slice sequence back to a three-dimensional physical coordinate space and stacking in sequence, so that the multi-plane feature space alignment accuracy and the super-resolution image texture fidelity are improved.

Description

Technical Field

[0001] This invention relates to the field of image super-resolution reconstruction technology, specifically to a method and system for super-resolution reconstruction of three-dimensional medical images. Background Technology

[0002] With the widespread clinical application of medical imaging technology, equipment such as computed tomography (CT) and magnetic resonance imaging (MRI) have become core carriers for disease screening, preoperative planning, and efficacy evaluation. However, due to limitations in radiation dose control and scanning hardware bandwidth, low-resolution medical images acquired clinically generally exhibit characteristics such as inter-slice resolution being far lower than intra-slice resolution and voxel anisotropy. This directly leads to blurred multi-plane reconstruction images and poor 3D visualization. 3D medical image super-resolution reconstruction technology, by restoring high-frequency details and improving spatial resolution based on low-resolution data, has become a research hotspot in the field of medical image post-processing. Current technological development is evolving from traditional interpolation algorithms to deep learning-driven multi-plane joint modeling.

[0003] In the prior art, CN112669209A discloses a method and system for super-resolution reconstruction of three-dimensional medical images. By decomposing the three-dimensional super-resolution task into two sub-problems, namely single-slice super-resolution and correlation modeling of adjacent slice content, a single-channel reconstruction network is first trained based on easily obtainable two-dimensional high-resolution medical images. Then, it is extended to a multi-channel structure through transfer learning. Only the output layer weights need to be optimized to complete the adaptation using a small amount of three-dimensional high-resolution data. In the application stage, after low-rank constraint alignment of scan layer misalignment and isotropic interpolation, the data is input into the network along the coronal plane through a sliding window and stacked to obtain the final three-dimensional super-resolution volume data. The principle is to use the strong correlation between adjacent coronal slices to make up for the lack of spatial continuity of single-slice super-resolution, while avoiding the dependence of three-dimensional convolutional networks on high-volume three-dimensional labeled data.

[0004] While the above-mentioned scheme improves the spatial continuity of 3D super-resolution to some extent through multi-channel structure and low-rank preprocessing, and lowers the threshold for training data, it still has limitations: the scheme only uses slice sequences in the coronal plane as the processing object, and the anatomical information in the axial and sagittal orthogonal directions is not included in the feature extraction process, which limits the resolution improvement effect on non-scanning planes; it does not establish a shared encoding and spatial alignment constraint mechanism at the three-plane feature level, and the feature responses of the same anatomical structure in the three orthogonal directions are heterogeneous, and cross-plane misalignment artifacts and edge blurring are still easy to occur after back-projection stacking; the loss function only depends on pixel-level grayscale error and lacks multi-dimensional constraints on frequency domain energy distribution and wavelet high-frequency subbands, resulting in insufficient recovery ability for high-frequency pathological features such as small lesions and fine textures.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for super-resolution reconstruction of three-dimensional medical images to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method and system for super-resolution reconstruction of three-dimensional medical images, the specific steps of which include: S1. Collect low-resolution medical body data from the same location, perform isotropic interpolation, and output isotropic medical body data. Using the geometric center of the isotropic medical body data as the origin, segment along the three orthogonal directions of axial, coronal, and sagittal to obtain axial slice sequences, coronal slice sequences, and sagittal slice sequences. S2. A multi-plane shared encoder is used to extract features from axial slice sequences, coronal slice sequences, and sagittal slice sequences, and projection consistency constraints are applied to the feature extraction results to form fused latent space features. S3. The fused latent space features are split into feature slice sequences according to the axial direction, coronal direction, and sagittal direction. The feature slice sequences in each direction are sent to the upsampling decoder to perform sub-pixel convolution upsampling and residual convolution refinement, and output the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence, and the sagittal super-resolution two-dimensional slice sequence. S4. Obtain high-resolution medical body data of the same location as training ground values. Based on the training ground values, calculate the loss of axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences. Train and adjust the three-plane synchronous super-resolution network according to the loss calculation results, and output the network parameters after training. S5. Configure the three-plane synchronous super-resolution network using the network parameters after training. Input the low-resolution medical body data to be reconstructed into the three-plane synchronous super-resolution network and output the super-resolution two-dimensional slice sequence in each direction. Select the super-resolution two-dimensional slice sequence in any direction, map the super-resolution two-dimensional slice sequence back to the three-dimensional physical coordinate space and stack them in sequence to complete the super-resolution reconstruction of the three-dimensional medical image.

[0008] Further, isotropic interpolation is performed to output isotropic medical volume data. The data is then segmented along three orthogonal directions—axial, coronal, and sagittal—with the geometric center of the isotropic medical volume data as the origin. The specific steps are as follows: Read the intra-layer pixel physical spacing and inter-layer slice physical spacing of the low-resolution medical volume data, select the minimum value of the two as the target voxel physical size, and perform interpolation processing on the low-resolution medical volume data according to the target voxel physical size to obtain isotropic medical volume data. The interpolation process refers to adjusting the voxel distribution of low-resolution medical volume data through trilinear interpolation or spline interpolation algorithms, so that the physical size of voxels in the output isotropic medical volume data is equal in three orthogonal directions. Using the geometric center of the isotropic medical volume data as the origin, a unified segmentation interval in three orthogonal directions is determined based on the physical dimensions of the target voxel. Along the axial direction perpendicular to the z-axis, the coronal direction perpendicular to the y-axis, and the sagittal direction perpendicular to the x-axis, samples are taken layer by layer at a uniform segmentation interval. The number of samples in each direction is consistent with the number of voxels in the corresponding axis of the isotropic medical body data. The axial slice sequence, coronal slice sequence, and sagittal slice sequence are output.

[0009] Furthermore, the step of using a multi-plane shared encoder to extract features from axial slice sequences, coronal slice sequences, and sagittal slice sequences, and applying projection consistency constraints to each feature extraction result, specifically involves the following steps: The axial, coronal, and sagittal slice sequences are input into a multi-plane shared encoder, which is composed of stacked two-dimensional convolutional blocks and channel attention blocks. The axial, coronal, and sagittal inputs call the same set of convolutional kernels and attention parameters. The two-dimensional convolutional blocks are used to extract local texture and edge structure features of the slice sequences, and the channel attention blocks are used to adaptively allocate weights according to the differences in the importance of feature channels to enhance the response of key pathological features. Hierarchical feature extraction is performed by using two-dimensional convolutional blocks and channel attention blocks within a multi-plane shared encoder, outputting axial two-dimensional feature map sequences, coronal two-dimensional feature map sequences, and sagittal two-dimensional feature map sequences. Based on the spatial coordinate mapping relationship of isotropic medical body data, the axial two-dimensional feature map sequence, coronal two-dimensional feature map sequence and sagittal two-dimensional feature map sequence are back-projected to the same three-dimensional grid coordinate system. The feature vectors of the corresponding axes are extracted at the overlapping voxel positions of the three-dimensional grid to obtain the axial feature vector, coronal feature vector and sagittal feature vector. The cosine distance between the axial feature vector, coronal feature vector and sagittal feature vector at the same overlapping voxel position is calculated. The cosine distance is used as the projection consistency loss. During backpropagation, the network parameters of the multi-plane shared encoder are updated synchronously using the projection consistency loss to form the fused latent space features.

[0010] Furthermore, the specific steps for feeding the feature slice sequences from each direction into the upsampling decoding head to perform subpixel convolutional upsampling and residual convolutional refinement are as follows: The fused latent space features are split and sorted according to the axial, coronal, and sagittal directions to form axial feature slice sequences, coronal feature slice sequences, and sagittal feature slice sequences; The sub-pixel convolutional layer in the upsampling decoder performs channel augmentation convolution on each input feature slice sequence, expanding the number of feature channels to the product of the number of output channels and the square of the magnification factor. The amplified feature map is then reorganized into pixels according to a periodic filtering method, rearranging the feature pixels of adjacent channels to the spatial dimension to form high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences. High-frequency texture information is extracted from high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences through multiple cascaded convolutional layers within the residual convolutional block. The high-frequency texture information extraction results are then superimposed with the high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences to output axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences.

[0011] Furthermore, the upsampling decoding head consists of cascaded subpixel convolutional layers and residual convolutional blocks. The subpixel convolutional layers are used to perform channel amplification and pixel recombination to improve the feature space resolution, while the residual convolutional blocks are used to perform high-frequency texture extraction and feature superposition to suppress amplification artifacts.

[0012] Furthermore, the specific steps for calculating the loss based on the training ground truth for axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences are as follows: Single super-resolution two-dimensional slices are extracted from axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences, and corresponding single slices are extracted from the slice sequences of high-resolution medical body data in the corresponding directions as training ground value slices. Calculate the mean square error of a single super-resolution 2D slice and its corresponding training ground slice in terms of pixel grayscale values; Perform two-dimensional fast Fourier transform on a single super-resolution two-dimensional slice and its corresponding training ground value slice, and calculate the mean square error of the full frequency domain amplitude in the transformed frequency domain space. Perform multi-level discrete wavelet decomposition on a single super-resolution two-dimensional slice and the corresponding training ground value slice, extract the horizontal high-frequency sub-band, vertical high-frequency sub-band and diagonal high-frequency sub-band obtained after decomposition, and calculate the mean square error of the corresponding high-frequency sub-band. The spatial domain pixel error, frequency domain global error, and wavelet high-frequency subband error are weighted and summed according to preset weighting coefficients to obtain the total loss function value.

[0013] Furthermore, the specific steps for training and adjusting the three-plane synchronous super-resolution network according to the loss calculation results, and outputting the network parameters after training, are as follows: With minimizing the total loss function value as the optimization objective, the current values ​​of the learnable parameters in the three-plane synchronous super-resolution network are updated through the backpropagation algorithm. The iterative process of three-plane slice extraction, error calculation, weighted summation and parameter update is repeated until the total loss function value converges or the preset number of iterations is reached. The three-plane synchronous super-resolution network is constructed by cascading a multi-plane shared encoder and an upsampling decoder.

[0014] Furthermore, the step of selecting a super-resolution two-dimensional slice sequence in any direction, mapping the super-resolution two-dimensional slice sequence back to three-dimensional physical coordinate space, and stacking them sequentially to complete the super-resolution reconstruction of the three-dimensional medical image, specifically involves the following steps: Axial super-resolution two-dimensional slice sequence is selected as the reconstruction benchmark. Based on the geometric center origin of the isotropic medical volume data and the spatial coordinate mapping relationship of the isotropic medical volume data, the normal vector and offset of each slice in the axial super-resolution two-dimensional slice sequence in the three-dimensional physical coordinate space are determined. The geometric center origin is the geometric center origin of the isotropic medical volume data obtained after isotropic interpolation of the low-resolution medical volume data to be reconstructed. Based on the determined normal vector and offset, each slice in the axial super-resolution two-dimensional slice sequence is mapped to the corresponding position in the three-dimensional physical coordinate space. The mapped slices are stacked sequentially along the slice normal vector direction, and trilinear interpolation is performed to fill the gaps generated during the stacking process to complete the super-resolution reconstruction of the three-dimensional medical image.

[0015] The present invention also provides a three-dimensional medical image super-resolution reconstruction system, which is used to implement the above-mentioned three-dimensional medical image super-resolution reconstruction method, including: The voxel interpolation and segmentation module is used to acquire low-resolution medical body data from the same location, perform isotropic interpolation, and output isotropic medical body data. With the geometric center of the isotropic medical body data as the origin, it is segmented along three orthogonal directions: axial, coronal, and sagittal, to obtain axial slice sequences, coronal slice sequences, and sagittal slice sequences. The three-plane encoding fusion module is used to extract features from axial slice sequences, coronal slice sequences and sagittal slice sequences using a multi-plane shared encoder, and to apply projection consistency constraints to each feature extraction result to form fused latent space features; The super-resolution decoding and reconstruction module is used to split the fused latent space features into feature slice sequences according to the axial, coronal, and sagittal directions. The feature slice sequences in each direction are sent to the upsampling decoding head to perform sub-pixel convolution upsampling and residual convolution refinement, and output the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence, and the sagittal super-resolution two-dimensional slice sequence. The loss constraint optimization module is used to obtain high-resolution medical body data of the same site as training ground values. Based on the training ground values, loss calculation is performed on axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences. The three-plane synchronous super-resolution network is trained and adjusted according to the loss calculation results, and the network parameters after training are output. The 3D synthesis stacking module is used to configure a three-plane synchronous super-resolution network using the network parameters after training. The low-resolution medical body data to be reconstructed is input into the three-plane synchronous super-resolution network, which outputs super-resolution two-dimensional slice sequences in each direction. The super-resolution two-dimensional slice sequence in any direction is selected, mapped back to the three-dimensional physical coordinate space, and stacked sequentially to complete the super-resolution reconstruction of the three-dimensional medical image.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention extracts features from axial, coronal, and sagittal slice sequences using a multi-plane shared encoder and applies projection consistency constraints, achieving spatial alignment of the feature response of the same anatomical structure in three orthogonal directions. This enables cross-plane consistency in the three-plane feature representation, effectively eliminating the feature heterogeneity problem under the independent encoder architecture, avoiding cross-plane misalignment artifacts and edge blurring that occur after back-projection stacking, and improving the diagnostic reliability of multi-plane reconstructed image reading.

[0017] Meanwhile, this invention constructs a total loss function value by fusing spatial domain pixel error, frequency domain global error, and wavelet high-frequency subband error, thereby achieving multiple constraints on the super-resolution reconstruction results at the pixel structure, spectral distribution, and high-frequency texture levels. This allows the reconstructed slices to simultaneously match the structural features and energy distribution patterns of the real data, effectively enhancing the response intensity of high-frequency pathological features such as small lesions and fine textures, avoiding texture loss and artifact generation under single-pixel loss, and improving the clinical diagnostic value of the reconstruction results.

[0018] Furthermore, this invention completes the mapping and stacking of super-resolution slices by using the geometric center origin of isotropic medical volume data obtained after isotropic interpolation of the data to be reconstructed, and performs trilinear interpolation to fill the stacking gaps. This achieves geometric consistency between the reconstructed three-dimensional volume data and the original physical space, so that the final output three-dimensional super-resolution volume data meets the spatial positioning requirements of clinical image reading. It effectively avoids spatial distortion and inter-layer discontinuity problems introduced by the stacking process, and significantly improves the accuracy of three-dimensional visualization and multimodal registration. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall method flow of the present invention.

[0020] Figure 2 This is a comparison chart of reconstruction quality under the same combination of loss weighting coefficients.

[0021] Figure 3 This is a schematic diagram of the overall system modules of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0023] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0024] Example: Please see Figures 1-2 The present invention provides a technical solution: The method for super-resolution reconstruction of three-dimensional medical images includes the following steps: S1. Collect low-resolution medical body data from the same location, perform isotropic interpolation, and output isotropic medical body data. Using the geometric center of the isotropic medical body data as the origin, segment along the three orthogonal directions of axial, coronal, and sagittal to obtain axial slice sequences, coronal slice sequences, and sagittal slice sequences.

[0025] In this embodiment, the intra-layer pixel physical spacing and inter-layer slice physical spacing of the low-resolution medical volume data are read, and the minimum value of the two is selected as the target voxel physical size. Interpolation processing is then performed on the low-resolution medical volume data according to the target voxel physical size to obtain isotropic medical volume data. The low-resolution medical volume data is clinically desensitized CT or MRI volume data collected in compliance with ethical review. For example, in a case of abdominal CT data, the intra-layer pixel physical spacing is 0.8 mm, and the inter-layer slice physical spacing is 2.5 mm, so 0.8 mm is selected as the target voxel physical size. The interpolation processing prioritizes matching the data modal characteristics. When processing CT data, a trilinear interpolation algorithm is used to balance processing efficiency and interpolation accuracy. When processing MRI data and needing to preserve fine and soft tissue boundaries, a spline interpolation algorithm is used to reduce interpolation artifacts.

[0026] The interpolation process refers to adjusting the voxel distribution of low-resolution medical volume data using trilinear interpolation or spline interpolation algorithms, so that the physical size of the voxels in the output isotropic medical volume data is equal in the three orthogonal directions.

[0027] Using the geometric center of the isotropic medical volume data as the origin, a unified segmentation interval in three orthogonal directions is determined based on the physical size of the target voxels. The geometric center is calculated by statistically analyzing the total number of voxels in the x, y, and z axes of the isotropic medical volume data. For example, if the size of the isotropic medical volume data is 512×512×312 (corresponding to the x, y, and z axes), then the coordinates of the geometric center are (255.5, 255.5, 155.5). The unified segmentation interval is consistent with the physical size of the target voxels, ensuring that the segmented slice sequence has no scale scaling deviation.

[0028] Sampling is performed layer-by-layer along the axial direction perpendicular to the z-axis, the coronal direction perpendicular to the y-axis, and the sagittal direction perpendicular to the x-axis, with a uniform segmentation interval. The number of samples in each direction is consistent with the number of voxels in the corresponding axis of isotropic medical volume data. Axial, coronal, and sagittal slice sequences are output. Axial sampling is performed layer-by-layer along the z-axis, coronal sampling along the y-axis, and sagittal sampling along the x-axis. The coordinate systems of all three types of slice sequences are aligned with the medical image physical coordinate system defined by the DICOM standard, providing a unified benchmark for subsequent feature extraction and spatial mapping.

[0029] S2. A multi-plane shared encoder is used to extract features from axial slice sequences, coronal slice sequences, and sagittal slice sequences, and projection consistency constraints are applied to the feature extraction results to form fused latent space features.

[0030] In this embodiment, the axial slice sequence, coronal slice sequence, and sagittal slice sequence are respectively input into a multi-plane shared encoder. The multi-plane shared encoder is composed of stacked two-dimensional convolutional blocks and channel attention blocks. The three inputs of axial, coronal, and sagittal slices call the same set of convolutional kernels and attention parameters. The two-dimensional convolutional blocks are used to extract local texture and edge structure features of the slice sequence, and the channel attention blocks are used to adaptively allocate weights according to the differences in the importance of feature channels to enhance the response of key pathological features.

[0031] It should be noted that the two-dimensional convolutional block consists of three cascaded 3×3 two-dimensional convolutional layers, with each convolutional layer followed by a batch normalization layer and a ReLU activation function. The first convolutional layer has 1 input channel, corresponding to a single-channel medical grayscale slice, and 64 output channels. Subsequent convolutional layers maintain 64 input and output channels. The shallow edge texture and deep semantic features of the slice are extracted layer by layer through hierarchical convolutional operations. The channel attention block first performs global average pooling on the input feature map, compressing the H×W spatial dimension of each feature channel into a 1×1 channel statistic. Then, it models the inter-channel dependencies through two fully connected layers. The first fully connected layer compresses the number of channels to 1 / 16 of the input. After ReLU activation, the second fully connected layer restores the original number of channels. Finally, the Sigmoid activation function generates channel weight coefficients in the range of 0 to 1. The weight coefficients are multiplied with the original feature map channel by channel to achieve adaptive weight allocation, strengthen the channel responses corresponding to key pathological features such as blood vessels, micronodules, and lesion boundaries, and suppress the weights of background noise channels.

[0032] Hierarchical feature extraction is performed using two-dimensional convolutional blocks and channel attention blocks within a multi-plane shared encoder, outputting axial two-dimensional feature map sequences, coronal two-dimensional feature map sequences, and sagittal two-dimensional feature map sequences.

[0033] Based on the spatial coordinate mapping relationship of isotropic medical body data, the axial two-dimensional feature map sequence, the coronal two-dimensional feature map sequence, and the sagittal two-dimensional feature map sequence are back-projected to the same three-dimensional grid coordinate system. The feature vectors of the corresponding axes are extracted at the overlapping voxel positions of the three-dimensional grid to obtain the axial feature vector, the coronal feature vector, and the sagittal feature vector.

[0034] The spatial coordinate mapping relationship is established based on the geometric center origin of isotropic medical body data, the physical size of the target voxel, and a unified segmentation interval. The physical coordinates of each voxel in the three-dimensional grid are converted into the index positions of the corresponding axial slice sequence, coronal slice sequence, and sagittal slice sequence through a unified segmentation interval, ensuring that the spatial positions of the axial two-dimensional feature map sequence, the coronal two-dimensional feature map sequence, and the sagittal two-dimensional feature map sequence correspond one-to-one. At the overlapping voxel positions of the three-dimensional grid, the multi-dimensional feature vectors of the corresponding axial feature map at the index position are extracted respectively, and the dimensions of the three are consistent with the number of channels of the feature map.

[0035] The cosine distance between the axial feature vector, coronal feature vector and sagittal feature vector at the same overlapping voxel position is calculated. The cosine distance is used as the projection consistency loss. During backpropagation, the network parameters of the multi-plane shared encoder are updated synchronously using the projection consistency loss to form the fused latent space features.

[0036] The cosine distance is obtained by calculating the cosine of the angle between the three axial feature vectors at the same overlapping voxel position. The closer the cosine value is to 1, the higher the consistency of the three axial features. The projection consistency loss is the difference between 0 and the cosine value. The smaller the difference, the better the spatial alignment of the three axial features. During the training phase, the gradient of the projection consistency loss is simultaneously backpropagated to the feature extraction paths of the three axes. Since the inputs of the three axes call the same set of convolution kernels and attention parameters, the gradients are accumulated at the shared parameter nodes and updated synchronously, forcing the feature responses of the same anatomical structure in different planes to tend to be consistent. Finally, the latent space features that take into account the spatial consistency of the three orthogonal planes are fused.

[0037] The formula for the cosine similarity of eigenvectors along three axes is: The formula for projection consistency loss is: In the formula, , This represents any two axial eigenvectors at the same overlapping voxel position, which in this scheme correspond to the axial eigenvectors respectively. Coronal feature vector Sagittal eigenvectors ; This represents the vector transpose operator; The cosine similarity between two feature vectors is represented, and its value ranges from 1 to 2. The closer the value is to 1, the higher the consistency of the directions of the two vectors, and the higher the matching degree of the feature responses of the same anatomical structure on different planes. This represents the loss of projection consistency.

[0038] S3. The fused latent space features are split into feature slice sequences according to the axial direction, coronal direction, and sagittal direction. The feature slice sequences in each direction are sent to the upsampling decoder to perform sub-pixel convolution upsampling and residual convolution refinement, and output the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence, and the sagittal super-resolution two-dimensional slice sequence. In this embodiment, the fused latent space features are split and sorted according to the axial, coronal, and sagittal directions to form axial feature slice sequences, coronal feature slice sequences, and sagittal feature slice sequences. The splitting and sorting strictly follows the spatial coordinate mapping relationship of the isotropic medical volume data. The arrangement order of the axial feature slice sequences corresponds one-to-one with the voxel indexes in the z-axis direction of the isotropic medical volume data, the arrangement order of the coronal feature slice sequences corresponds one-to-one with the voxel indexes in the y-axis direction, and the arrangement order of the sagittal feature slice sequences corresponds one-to-one with the voxel indexes in the x-axis direction. This ensures that the spatial order of the split three-dimensional feature slice sequences is completely aligned with the spatial order of the axial, coronal, and sagittal slice sequences, providing a sequential reference for subsequent upsampling and three-dimensional physical coordinate space mapping.

[0039] The voxel index refers to the position number obtained by traversing and sorting the three-dimensional mesh along the x, y, and z axes with the geometric center of the isotropic medical volume data as the origin and a uniform segmentation interval as the step size. It is used to uniquely identify the coordinate position of each voxel in three-dimensional space.

[0040] The sub-pixel convolutional layer in the upsampling decoder performs channel augmentation convolution on each input feature slice sequence, expanding the number of feature channels to the product of the number of output channels and the square of the magnification factor. The amplified feature map is then reorganized into pixels according to a periodic filtering method, rearranging the feature pixels of adjacent channels to the spatial dimension to form high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences.

[0041] The magnification factor is a preset scaling factor for the super-resolution reconstruction target. In this embodiment, the preset magnification factor is 2, meaning that the side length of the output super-resolution slice is twice the side length of the input feature slice. Therefore, the number of output channels of the channel augmentation convolution needs to be set as the product of the number of input feature channels and the square of the magnification factor. The number of input feature channels is consistent with the number of output channels of the multi-plane shared encoder, which is 64. Therefore, the number of output channels of the channel augmentation convolution is 256. The pixel recombination by the periodic filtering method is as follows: the 256-channel feature map is rearranged according to the periodicity of spatial position. The corresponding pixels of every 4 adjacent channels are mapped to 4 pixels in the same 2×2 neighborhood in the high-resolution feature map. After recombination, the height and width of the feature map are both expanded to twice that of the original feature map, and the number of channels is restored to 64, thus achieving the improvement of feature space resolution without introducing additional computational complexity.

[0042] High-frequency texture information is extracted from high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences through multiple cascaded convolutional layers within the residual convolutional block. The high-frequency texture information extraction results are then superimposed with the high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences to output axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences.

[0043] The residual convolutional block consists of three cascaded 3×3 two-dimensional convolutional layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The number of input channels and output channels are both 64. Multiple cascaded convolutional layers are specifically used to extract high-frequency texture information from high-resolution feature sequences, including the edges of small lesions, the texture of blood vessel branches, and the details of tissue interfaces. The feature superposition is achieved by element-wise addition. The high-frequency texture feature map output by the residual convolutional block is added one by one with the pixel values ​​of the corresponding positions of the high-resolution feature map output by the sub-pixel convolutional layer. This preserves the basic structural information restored by the sub-pixel convolution and supplements the high-frequency texture details, effectively suppressing magnification artifacts and edge blurring problems. The output three-dimensional super-resolution two-dimensional slice sequence has one channel, corresponding to a single-channel medical grayscale image. The spatial resolution of the three-dimensional sequence matches the physical size of the target voxel, and the spatial order is completely aligned with the axial slice sequence, coronal slice sequence, and sagittal slice sequence.

[0044] The upsampling decoding head consists of cascaded subpixel convolutional layers and residual convolutional blocks. The subpixel convolutional layers are used to perform channel amplification and pixel recombination to improve the feature space resolution, while the residual convolutional blocks are used to perform high-frequency texture extraction and feature superposition to suppress amplification artifacts.

[0045] S4. Obtain high-resolution medical body data of the same location as training ground values. Based on the training ground values, calculate the loss of axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences. Train and adjust the three-plane synchronous super-resolution network according to the loss calculation results, and output the network parameters after training.

[0046] In this embodiment, a single super-resolution two-dimensional slice is extracted from the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence, and the sagittal super-resolution two-dimensional slice sequence, and a corresponding single slice is extracted from the slice sequence corresponding to the high-resolution medical body data as a training ground value slice.

[0047] The high-resolution medical volume data is clinical desensitized CT or MRI volume data from the same source and location as the low-resolution medical volume data collected in step S1. Before training, the pixel values ​​of axial super-resolution 2D slices, coronal super-resolution 2D slices, sagittal super-resolution 2D slices, and corresponding training ground value slices are normalized to the [0,1] interval to eliminate illumination differences acquired by different devices. The slice extraction process strictly matches the voxel index, that is, the z-axis index of the extracted axial super-resolution 2D slice is consistent with the z-axis index of the corresponding training ground value slice. The extraction of coronal and sagittal slices also follows the rule of one-to-one correspondence between the y-axis and x-axis voxel indices to ensure pixel-level spatial alignment of error calculation.

[0048] The mean square error of a single super-resolution 2D slice and its corresponding ground truth slice in terms of pixel grayscale values ​​is calculated. This mean square error is the spatial domain pixel error, and the calculation formula is as follows: In the formula, Indicates the super-resolution 2D slice in the th... line, number The normalized pixel values ​​of the column; This indicates that the corresponding training ground slices are at the same position. line, number The normalized pixel values ​​of the column; This represents the total number of pixels in the vertical direction of a super-resolution 2D slice. This represents the total number of pixels in the horizontal direction of a super-resolution 2D slice; This represents the spatial pixel error, with a value range of [value range missing]. The smaller the value, the higher the pixel-level structural matching between the super-resolution slice and the ground truth slice.

[0049] Perform two-dimensional fast Fourier transforms on a single super-resolution two-dimensional slice and its corresponding training ground truth slice, respectively, and calculate the mean square error of the full-frequency domain amplitude in the transformed frequency space; the mean square error of the full-frequency domain amplitude in the frequency space after the two-dimensional fast Fourier transform is the global error in the frequency domain, and the calculation formula is as follows: In the formula, This represents the two-dimensional Fast Fourier Transform operator. Represents frequency domain coordinate points; Indicates the global error in the frequency domain. and These represent the super-resolution 2D slice and its corresponding ground truth slice in [the context of the training data]. The normalized pixel value at that location.

[0050] Multi-level discrete wavelet decomposition is performed on a single super-resolution 2D slice and its corresponding training ground truth slice to extract the horizontal, vertical, and diagonal high-frequency sub-bands obtained after decomposition. The mean square error of the corresponding high-frequency sub-bands is calculated. This step uses a three-level discrete wavelet decomposition, with Daubechies 1 (db1) as the wavelet basis function. The formula for calculating the mean square error of the high-frequency sub-bands is as follows: In the formula, This represents the mean square error of the high-frequency subband; Identifiers representing horizontal high-frequency sub-bands, vertical high-frequency sub-bands, and diagonal high-frequency sub-bands; This represents the third level of discrete wavelet decomposition after a super-resolution two-dimensional slice. Subband in the first line, number The pixel values ​​of the column; This represents the corresponding training ground slice. Subband in the first line, number The pixel values ​​of the column; These represent the total number of pixels in the vertical and horizontal directions of the high-frequency subband after the third-level discrete wavelet decomposition, respectively. The spatial domain pixel error, frequency domain global error, and wavelet high-frequency subband error are weighted and summed according to preset weight coefficients to obtain the total loss function value; the preset weight coefficients are preset according to the modal characteristics of medical images.

[0051] Table 1: Ablation Comparison of Projection Consistency Constraints on Cross-Plane Feature Alignment Accuracy Based on the 16 sets of control experimental data shown in Table 1, combined with Figure 2 It can be seen that with the frequency domain weight ( ) and wavelet domain weights ( The reconstructed image quality exhibits a single-peak trend of first increasing and then decreasing due to changes in frequency domain weights and wavelet domain weights. When both frequency domain weights and wavelet domain weights are at relatively low levels (e.g., in experimental group 1), =0.0、 =0.0), the model is only constrained by spatial domain pixel error, resulting in insufficient high-frequency texture recovery, with a peak signal-to-noise ratio (PSNR) of only 32.15dB and a structural similarity (SSIM) of 0.921; as the weights of the frequency domain and wavelet domain gradually increase (experimental groups 2-5), the model's ability to capture the texture of small lesion edges and vascular branches is significantly enhanced, and PSNR and SSIM continue to rise, in experimental group 5 ( =0.1、 The optimal balance is achieved at (=0.5), where PSNR reaches 33.24dB and SSIM reaches 0.952, resulting in the best reconstruction quality. When the weighting coefficients deviate further from this optimal range (such as in experimental groups 6-13 and 15), excessively high frequency domain or wavelet domain constraints introduce spectral shift and ringing artifacts, causing a significant drop in objective indicators. PSNR drops to a minimum of 30.21dB and SSIM drops to 0.892. Experimental groups 14-16 are the reproduction verification of the optimal parameter combination. The data shows that all indicators are highly consistent and stable.

[0052] The aforementioned data variation patterns verify that the composite loss function designed in this invention can effectively balance the global structural fidelity and local high-frequency detail restoration of the reconstructed image. Through multi-dimensional collaborative constraints in the spatial, frequency, and wavelet domains, it precisely adapts to the stringent diagnostic requirements of medical imaging for high-fidelity pathological textures. When the loss weights are optimally configured, the system can suppress magnification artifacts and edge blurring to the greatest extent, achieving high-quality super-resolution reconstruction of clinical CT / MRI images and ensuring that doctors can clearly identify minute lesions and complex tissue interfaces.

[0053] With minimizing the total loss function value as the optimization objective, the current values ​​of the learnable parameters in the three-plane synchronous super-resolution network are updated through the backpropagation algorithm. The iterative process of three-plane slice extraction, error calculation, weighted summation and parameter update is repeated until the total loss function value converges or the preset number of iterations is reached.

[0054] In this embodiment, the Adam optimizer is used to execute the backpropagation algorithm during the training phase. The initial learning rate is set to 1e-4. For each training batch, 8 slices are randomly selected from each of the three directions to calculate the total loss function value. During gradient update, the gradient of the total loss is accumulated at the shared parameter nodes of the multi-plane shared encoder to ensure consistency with the projection loss gradient of the S2 step. All learnable parameters of the three-plane synchronous super-resolution network are updated synchronously, including the convolutional kernels and attention parameters of the multi-plane shared encoder, as well as the weights of the sub-pixel convolutional layers and residual convolutional blocks of the upsampling decoder. The preset number of iterations is 200 training epochs. When the total loss function value decreases by less than 1e-4 for 10 consecutive epochs, it is considered converged, training is stopped, and the network parameters after training are output and saved as a binary file for loading and calling during the inference phase.

[0055] The three-plane synchronous super-resolution network is constructed by cascading a multi-plane shared encoder and an upsampling decoder.

[0056] S5. Configure the three-plane synchronous super-resolution network using the network parameters after training. Input the low-resolution medical body data to be reconstructed into the three-plane synchronous super-resolution network and output the super-resolution two-dimensional slice sequence in each direction. Select the super-resolution two-dimensional slice sequence in any direction, map the super-resolution two-dimensional slice sequence back to the three-dimensional physical coordinate space and stack them in sequence to complete the super-resolution reconstruction of the three-dimensional medical image.

[0057] In this embodiment, a three-plane synchronous super-resolution network is configured using the network parameters after training. The low-resolution medical body data to be reconstructed is input into the three-plane synchronous super-resolution network, and the super-resolution two-dimensional slice sequence in each direction is output.

[0058] Before loading network parameters during the inference phase, isotropic interpolation and triaxial slice sequence generation operations, which are completely consistent with step S1, are performed on the low-resolution medical body data to be reconstructed. This ensures that the data format and spatial reference of the input triplane synchronous super-resolution network are completely consistent with those in the training phase, and avoids spatial deviations caused by differences in preprocessing.

[0059] Axial super-resolution two-dimensional slice sequences are selected as the reconstruction benchmark (the axial direction is consistent with the orientation of the original scanning plane of the clinical CT and MRI equipment, and the z-axis arrangement order of the axial slice sequence strictly corresponds to the z-axis voxel index of the isotropic medical volume data, which can reduce the computational complexity of subsequent spatial mapping). Based on the geometric center origin of the isotropic medical volume data and the spatial coordinate mapping relationship of the isotropic medical volume data, the normal vector and offset of each slice in the axial super-resolution two-dimensional slice sequence in the three-dimensional physical coordinate space are determined. The geometric center origin is the geometric center origin of the isotropic medical volume data obtained after isotropic interpolation of the low-resolution medical volume data to be reconstructed.

[0060] It should be noted that although coronal super-resolution 2D slice sequences and sagittal super-resolution 2D slice sequences can be used as alternative reconstruction benchmarks during the inference stage, their core value lies in the training stage: on the one hand, they provide supervision signals for the calculation of the three-dimensional composite loss, and on the other hand, they provide a basis for multi-plane feature alignment for the projection consistency constraint, allowing the shared encoder to learn spatially consistent feature representations.

[0061] The normal vector is the perpendicular direction vector of the plane containing the axial slice, i.e., the z-axis unit vector in the three-dimensional physical coordinate space; the offset is the displacement of each axial super-resolution two-dimensional slice along the normal vector direction, calculated based on the target voxel physical size and corresponding z-axis voxel index of the isotropic medical volume data. The formula for calculating the offset of a axial slice is: In the formula, Indicates the first Offset of the axial slice The z-axis coordinates represent the geometric center origin. This represents the total number of voxels along the z-axis in isotropic medical volume data. Indicates the physical size of the target voxel.

[0062] According to the determined normal vector and offset, each slice in the axial super-resolution two-dimensional slice sequence is mapped to the corresponding position in the three-dimensional physical coordinate space. During the mapping process, the two-dimensional coordinates of each pixel on the slice are converted into three-dimensional physical coordinates by combining the offset and binding them one by one with the voxel position of the three-dimensional mesh. The mapped slices are stacked sequentially along the slice normal vector direction, and trilinear interpolation is performed to fill the gaps generated during the stacking process to complete the super-resolution reconstruction of the three-dimensional medical image.

[0063] The trilinear interpolation algorithm is completely consistent with the trilinear interpolation algorithm used in the isotropic interpolation in step S1. That is, for the voxel position where the gap is located, the pixel value of the gap position is calculated by weighted average according to the spatial distance based on the pixel values ​​of the eight adjacent voxels that have been assigned values, thus eliminating the problem of interlayer discontinuity after stacking.

[0064] Please see Figure 3 The present invention also provides a three-dimensional medical image super-resolution reconstruction system, which is used to implement the above-mentioned three-dimensional medical image super-resolution reconstruction method, including: The voxel interpolation and segmentation module is used to acquire low-resolution medical body data from the same location, perform isotropic interpolation, and output isotropic medical body data. With the geometric center of the isotropic medical body data as the origin, it is segmented along three orthogonal directions: axial, coronal, and sagittal, to obtain axial slice sequences, coronal slice sequences, and sagittal slice sequences. The three-plane encoding fusion module is used to extract features from axial slice sequences, coronal slice sequences and sagittal slice sequences using a multi-plane shared encoder, and to apply projection consistency constraints to each feature extraction result to form fused latent space features; The super-resolution decoding and reconstruction module is used to split the fused latent space features into feature slice sequences according to the axial, coronal, and sagittal directions. The feature slice sequences in each direction are sent to the upsampling decoding head to perform sub-pixel convolution upsampling and residual convolution refinement, and output the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence, and the sagittal super-resolution two-dimensional slice sequence. The loss constraint optimization module is used to obtain high-resolution medical body data of the same site as training ground values. Based on the training ground values, loss calculation is performed on axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences. The three-plane synchronous super-resolution network is trained and adjusted according to the loss calculation results, and the network parameters after training are output. The 3D synthesis stacking module is used to configure a three-plane synchronous super-resolution network using the network parameters after training. The low-resolution medical body data to be reconstructed is input into the three-plane synchronous super-resolution network, which outputs super-resolution two-dimensional slice sequences in each direction. The super-resolution two-dimensional slice sequence in any direction is selected, mapped back to the three-dimensional physical coordinate space, and stacked sequentially to complete the super-resolution reconstruction of the three-dimensional medical image.

[0065] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0066] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0067] 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; 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.

[0068] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for super-resolution reconstruction of three-dimensional medical images, characterized in that, The specific steps include: S1. Collect low-resolution medical body data from the same location, perform isotropic interpolation, and output isotropic medical body data. Using the geometric center of the isotropic medical body data as the origin, segment along the three orthogonal directions of axial, coronal, and sagittal to obtain axial slice sequences, coronal slice sequences, and sagittal slice sequences. S2. A multi-plane shared encoder is used to extract features from axial slice sequences, coronal slice sequences, and sagittal slice sequences, and projection consistency constraints are applied to the feature extraction results to form fused latent space features; the specific steps are as follows: The axial, coronal, and sagittal slice sequences are input into a multi-plane shared encoder, which is composed of stacked two-dimensional convolutional blocks and channel attention blocks. The axial, coronal, and sagittal inputs call the same set of convolutional kernels and attention parameters. The two-dimensional convolutional blocks are used to extract local texture and edge structure features of the slice sequences, and the channel attention blocks are used to adaptively allocate weights according to the differences in the importance of feature channels to enhance the response of key pathological features. Hierarchical feature extraction is performed by using two-dimensional convolutional blocks and channel attention blocks within a multi-plane shared encoder, outputting axial two-dimensional feature map sequences, coronal two-dimensional feature map sequences, and sagittal two-dimensional feature map sequences. Based on the spatial coordinate mapping relationship of isotropic medical body data, the axial two-dimensional feature map sequence, coronal two-dimensional feature map sequence and sagittal two-dimensional feature map sequence are back-projected to the same three-dimensional grid coordinate system. The feature vectors of the corresponding axes are extracted at the overlapping voxel positions of the three-dimensional grid to obtain the axial feature vector, coronal feature vector and sagittal feature vector. Calculate the cosine distance between the axial feature vector, coronal feature vector and sagittal feature vector at the same overlapping voxel position, and use the cosine distance as the projection consistency loss. During backpropagation, use the projection consistency loss to synchronously update the network parameters of the multi-plane shared encoder to form the fused latent space features. S3. The fused latent space features are split into feature slice sequences according to the axial direction, coronal direction, and sagittal direction. The feature slice sequences in each direction are sent to the upsampling decoder to perform sub-pixel convolution upsampling and residual convolution refinement, and output the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence, and the sagittal super-resolution two-dimensional slice sequence. The upsampling decoding head consists of cascaded subpixel convolutional layers and residual convolutional blocks. The subpixel convolutional layers are used to perform channel amplification and pixel recombination to improve the feature space resolution, while the residual convolutional blocks are used to perform high-frequency texture extraction and feature superposition to suppress amplification artifacts. S4. Obtain high-resolution medical body data of the same location as training ground values. Based on the training ground values, calculate the loss of axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences. Train and adjust the three-plane synchronous super-resolution network according to the loss calculation results, and output the network parameters after training. The three-plane synchronous super-resolution network is constructed by cascading a multi-plane shared encoder and an upsampling decoder head. S5. Configure the three-plane synchronous super-resolution network using the network parameters after training. Input the low-resolution medical body data to be reconstructed into the three-plane synchronous super-resolution network and output the super-resolution two-dimensional slice sequence in each direction. Select the super-resolution two-dimensional slice sequence in any direction, map the super-resolution two-dimensional slice sequence back to the three-dimensional physical coordinate space and stack them in sequence to complete the super-resolution reconstruction of the three-dimensional medical image.

2. The method for super-resolution reconstruction of three-dimensional medical images according to claim 1, characterized in that: The isotropic interpolation is performed to output isotropic medical volume data. The data is then segmented along three orthogonal directions—axial, coronal, and sagittal—with the geometric center of the isotropic medical volume data as the origin. The specific steps are as follows: Read the intra-layer pixel physical spacing and inter-layer slice physical spacing of the low-resolution medical volume data, select the minimum value of the two as the target voxel physical size, and perform interpolation processing on the low-resolution medical volume data according to the target voxel physical size to obtain isotropic medical volume data. The interpolation process refers to adjusting the voxel distribution of low-resolution medical volume data through trilinear interpolation or spline interpolation algorithms, so that the physical size of voxels in the output isotropic medical volume data is equal in three orthogonal directions. Using the geometric center of the isotropic medical volume data as the origin, a unified segmentation interval in three orthogonal directions is determined based on the physical dimensions of the target voxel. Along the axial direction perpendicular to the z-axis, the coronal direction perpendicular to the y-axis, and the sagittal direction perpendicular to the x-axis, samples are taken layer by layer at a uniform segmentation interval. The number of samples in each direction is consistent with the number of voxels in the corresponding axis of the isotropic medical body data. The axial slice sequence, coronal slice sequence, and sagittal slice sequence are output.

3. The method for super-resolution reconstruction of three-dimensional medical images according to claim 1, characterized in that: The specific steps for feeding the feature slice sequences from each direction into the upsampling decoding head to perform subpixel convolutional upsampling and residual convolutional refinement are as follows: The fused latent space features are split and sorted according to the axial, coronal, and sagittal directions to form axial feature slice sequences, coronal feature slice sequences, and sagittal feature slice sequences; The sub-pixel convolutional layer in the upsampling decoder performs channel augmentation convolution on each input feature slice sequence, expanding the number of feature channels to the product of the number of output channels and the square of the magnification factor. The amplified feature map is then reorganized into pixels according to a periodic filtering method, rearranging the feature pixels of adjacent channels to the spatial dimension to form high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences. High-frequency texture information is extracted from high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences through multiple cascaded convolutional layers within the residual convolutional block. The high-frequency texture information extraction results are then superimposed with the high-resolution axial feature sequences, high-resolution coronal feature sequences, and high-resolution sagittal feature sequences to output axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences.

4. The method for super-resolution reconstruction of three-dimensional medical images according to claim 1, characterized in that: The loss calculation for axial super-resolution 2D slice sequences, coronal super-resolution 2D slice sequences, and sagittal super-resolution 2D slice sequences based on training ground values ​​is performed in the following steps: Single super-resolution two-dimensional slices are extracted from axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences, and corresponding single slices are extracted from the slice sequences of high-resolution medical body data in the corresponding directions as training ground value slices. Calculate the mean square error of a single super-resolution 2D slice and its corresponding training ground slice in terms of pixel grayscale values; Perform two-dimensional fast Fourier transform on a single super-resolution two-dimensional slice and its corresponding training ground value slice, and calculate the mean square error of the full frequency domain amplitude in the transformed frequency domain space. Perform multi-level discrete wavelet decomposition on a single super-resolution two-dimensional slice and the corresponding training ground value slice, extract the horizontal high-frequency sub-band, vertical high-frequency sub-band and diagonal high-frequency sub-band obtained after decomposition, and calculate the mean square error of the corresponding high-frequency sub-band. The spatial domain pixel error, frequency domain global error, and wavelet high-frequency subband error are weighted and summed according to preset weighting coefficients to obtain the total loss function value.

5. The method for super-resolution reconstruction of three-dimensional medical images according to claim 4, characterized in that: The specific steps for training and adjusting the three-plane synchronous super-resolution network according to the loss calculation results, and outputting the network parameters after training are as follows: With minimizing the total loss function value as the optimization objective, the current values ​​of the learnable parameters in the three-plane synchronous super-resolution network are updated through the backpropagation algorithm. The iterative process of three-plane slice extraction, error calculation, weighted summation and parameter update is repeated until the total loss function value converges or the preset number of iterations is reached.

6. The method for super-resolution reconstruction of three-dimensional medical images according to claim 1, characterized in that: The process involves selecting a super-resolution two-dimensional slice sequence in any direction, mapping the super-resolution two-dimensional slice sequence back to three-dimensional physical coordinate space, and stacking them sequentially to complete the super-resolution reconstruction of the three-dimensional medical image. The specific steps are as follows: Axial super-resolution two-dimensional slice sequence is selected as the reconstruction benchmark. Based on the geometric center origin of the isotropic medical volume data and the spatial coordinate mapping relationship of the isotropic medical volume data, the normal vector and offset of each slice in the axial super-resolution two-dimensional slice sequence in the three-dimensional physical coordinate space are determined. The geometric center origin is the geometric center origin of the isotropic medical volume data obtained after isotropic interpolation of the low-resolution medical volume data to be reconstructed. Based on the determined normal vector and offset, each slice in the axial super-resolution two-dimensional slice sequence is mapped to the corresponding position in the three-dimensional physical coordinate space. The mapped slices are stacked sequentially along the slice normal vector direction, and trilinear interpolation is performed to fill the gaps generated during the stacking process to complete the super-resolution reconstruction of the three-dimensional medical image.

7. A three-dimensional medical image super-resolution reconstruction system, characterized in that: The three-dimensional medical image super-resolution reconstruction system is used to implement the three-dimensional medical image super-resolution reconstruction method according to any one of claims 1-6, including: The voxel interpolation and segmentation module is used to acquire low-resolution medical body data from the same location, perform isotropic interpolation, and output isotropic medical body data. With the geometric center of the isotropic medical body data as the origin, it is segmented along three orthogonal directions: axial, coronal, and sagittal, to obtain axial slice sequences, coronal slice sequences, and sagittal slice sequences. The three-plane encoding fusion module is used to extract features from axial slice sequences, coronal slice sequences and sagittal slice sequences using a multi-plane shared encoder, and to apply projection consistency constraints to each feature extraction result to form fused latent space features; The super-resolution decoding and reconstruction module is used to split the fused latent space features into feature slice sequences according to the axial, coronal, and sagittal directions. The feature slice sequences in each direction are sent to the upsampling decoding head to perform sub-pixel convolution upsampling and residual convolution refinement, and output the axial super-resolution two-dimensional slice sequence, the coronal super-resolution two-dimensional slice sequence, and the sagittal super-resolution two-dimensional slice sequence. The loss constraint optimization module is used to obtain high-resolution medical body data of the same site as training ground values. Based on the training ground values, loss calculation is performed on axial super-resolution two-dimensional slice sequences, coronal super-resolution two-dimensional slice sequences, and sagittal super-resolution two-dimensional slice sequences. The three-plane synchronous super-resolution network is trained and adjusted according to the loss calculation results, and the network parameters after training are output. The 3D synthesis stacking module is used to configure a three-plane synchronous super-resolution network using the network parameters after training. The low-resolution medical body data to be reconstructed is input into the three-plane synchronous super-resolution network, which outputs super-resolution two-dimensional slice sequences in each direction. The super-resolution two-dimensional slice sequence in any direction is selected, mapped back to the three-dimensional physical coordinate space, and stacked sequentially to complete the super-resolution reconstruction of the three-dimensional medical image.

Images