Lightweight Fourier Stack Reconstruction System and Method

Abstract

This invention belongs to the field of image processing technology, and particularly relates to a lightweight Fourier layered reconstruction system and method. The system includes: a feature extraction module, which performs shallow feature extraction on the input low-resolution pathological slide image to obtain shallow features; a multi-level frequency domain transformation module, which performs multi-level frequency domain information extraction and enhancement on the input low-resolution pathological slide image to obtain multi-level frequency domain enhanced features; an encoding and decoding module, which uses the multi-level frequency domain enhanced features to perform multi-level feature encoding on the shallow features, and then decodes the encoded features level by level to obtain deep features; and a reconstruction module, which performs upsampling operations on the deep features and outputs a target high-resolution amplitude image of the pathological slide. This invention overcomes the problems of excessive image smoothing and artifact generation, resulting in reconstructed images with clearer tissue contours and more realistic cell details.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a lightweight Fourier layer reconstruction system and method. Background Technology

[0002] With the development of digital pathology and clinical medical diagnostic technologies, microscopic imaging systems need to simultaneously handle rapid scanning of large field-of-view tissue samples and high-resolution observation of fine structures such as cell nuclei and chromatin. Traditional optical microscopes, limited by their spatial bandwidth product, struggle to achieve both a large field of view and high resolution. Fourier transform microscopy (FPM) is a computational imaging technique that integrates synthetic aperture imaging and phase retrieval. By acquiring low-resolution intensity images under multi-angle LED illumination and performing frequency domain spectral stitching, it can increase the system's spatial bandwidth product without reducing the field of view. With its advantages of high throughput and low cost, it has enormous application potential in the fields of digital scanning of pathological sections and biomedical diagnostics.

[0003] Current methods for high-resolution pathological amplitude image reconstruction using FPM (Focus-Performance Model) mainly fall into two categories: traditional physical iterative algorithms and data-driven deep learning algorithms. The former, represented by the alternating projection method, constructs a forward optical propagation model and iterates repeatedly between the spatial and Fourier frequency domains to recover high-frequency information. However, this approach suffers from long computation times, weak robustness, and difficulty adapting to high-throughput imaging scenarios. The latter includes three mainstream models: first, Convolutional Neural Networks (CNNs), which focus on local feature extraction but easily lose high-frequency details, resulting in overly smoothed images; second, Generative Adversarial Networks (GANs), which can generate rich visual details but easily introduce harmful artifacts, affecting the reliability of clinical diagnosis; and third, Visual Transformer models, which possess long-distance dependency modeling and global feature capture capabilities, but have high computational complexity and large memory overhead, making rapid deployment on clinical hardware difficult. Overall, both existing reconstruction algorithms have significant shortcomings and cannot simultaneously meet the comprehensive requirements of speed, accuracy, and imaging security for high-throughput, high-resolution pathological slide amplitude reconstruction. Summary of the Invention

[0004] In view of this, the present invention aims to provide a lightweight Fourier layered reconstruction system and method to solve the problems of cumbersome and time-consuming iterations of traditional Fourier layered reconstruction algorithms in large-field-of-view, high-resolution imaging of pathological sections, as well as the problems of large parameter numbers, difficulty in deployment on computationally limited devices, and insufficient high-frequency cell texture recovery in existing deep learning reconstruction networks. This invention is based on a lightweight wavelet-Fourier reconstruction model, combining the multi-scale spatial-frequency separation capability of wavelet transform with the lightweight advantage of Fourier networks in global frequency domain interaction. The lightweight wavelet-Fourier reconstruction model extracts deep features from the input low-resolution pathological section image sequence, decouples and jointly optimizes high-frequency cell texture and low-frequency tissue structural contours, and develops a rapid reconstruction method for high-resolution amplitude images, providing optional technical support and scientific assurance for clinical high-throughput digital pathological microscopy imaging.

[0005] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A Fourier layer-by-layer microscopic reconstruction system includes: The feature extraction module performs shallow feature extraction on the input low-resolution pathological slide image to obtain shallow features; The multi-level frequency domain transformation module performs multi-level frequency domain information extraction and enhancement on the input low-resolution pathological slice image to obtain multi-level frequency domain enhancement features; The encoding and decoding module uses multi-level frequency domain enhancement features to encode shallow features at multiple levels, and then decodes the encoded features level by level to obtain deep features. The reconstruction module performs upsampling on deep features and outputs a high-resolution amplitude image of the target pathological section.

[0006] Furthermore, the frequency domain transformation module includes a discrete wavelet transform module and a wavelet feature fusion module, wherein: The discrete wavelet transform module performs spatial frequency decomposition on the input features by high-frequency details and low-frequency approximations to obtain multi-band sub-features. The multi-band sub-features include low-frequency approximation sub-bands and three high-frequency detail sub-bands corresponding to the horizontal, vertical and diagonal directions, respectively. The wavelet feature fusion module concatenates high-frequency detail subbands in three directions along the channel dimension and then performs dimensionality reduction through pointwise convolution to obtain high-frequency features. These high-frequency features are then input in parallel into the spatial convolution branch and the global Fourier branch for processing. The low-frequency approximate subbands, the output features of the spatial convolution branch, and the output features of the global Fourier branch are concatenated and nonlinearly aggregated to obtain frequency domain enhanced features.

[0007] Furthermore, the encoding / decoding module includes an encoding branch and a decoding branch: The coding branch includes multiple cascaded wavelet-Fourier coding submodules. In each wavelet-Fourier coding submodule, the input is the coding features of the previous wavelet-Fourier coding submodule and the frequency domain enhancement features of the same scale. After multi-scale spatial frequency decomposition and frequency domain interaction processing, the coding features of the current wavelet-Fourier coding submodule are obtained. The decoding branch includes multiple cascaded wavelet-Fourier decoding submodules. In each wavelet-Fourier decoding submodule, the input is the decoding features of the previous wavelet-Fourier decoding submodule and the coding features at the same scale. After multi-scale spatial frequency decomposition and frequency domain interaction processing, the decoding features of the current wavelet-Fourier decoding submodule are obtained.

[0008] Furthermore, in each wavelet-Fourier coding submodule, or each wavelet-Fourier decoding submodule, the specific operations for multi-scale spatial frequency decomposition and frequency domain interaction processing are as follows: The discrete wavelet transform module performs spatial frequency decomposition on the input features using high-frequency details and low-frequency approximations to obtain multi-band sub-features. These multi-band sub-features are then preprocessed and mapped via group normalization and pointwise convolution before being input to the dual-frequency feature fusion module for frequency domain feature interaction to obtain interactive features. After processing with channel attention and pointwise convolution, the interactive features are residually concatenated with the multi-band sub-features to obtain fused features. The fused features are then processed via group normalization, pointwise convolution, depthwise convolution, GELU nonlinear activation function, and pointwise convolution before being residually concatenated with the fused features to obtain the output features.

[0009] Furthermore, in the dual-frequency feature fusion module, the input features are subjected to pointwise convolution and GELU activation operations to obtain the first feature. The first feature is then processed by local Fourier transform and global Fourier transform respectively to obtain global and local features. The corresponding elements of the global and local features are added together to obtain the second feature. The second feature is then processed by global average pooling, pointwise convolution, GELU activation function, pointwise convolution and channel separation operations to obtain the third and fourth features. The third and fourth features are then processed by the Softmax function to obtain the fifth and sixth features respectively. The fifth feature is added to the local features to obtain the seventh feature. The sixth feature is added to the global features to obtain the eighth feature. The seventh and eighth features are then added together to obtain the ninth feature. The ninth feature is processed by pointwise convolution to obtain the output feature.

[0010] A Fourier layer-by-layer microscopic reconstruction method includes: S1: Obtain low-resolution pathological slide image sequences under different lighting angles, and use the high-resolution real amplitude images corresponding to the low-resolution pathological slide image sequences as labels to construct a training set; S2: Constructing a Fourier layered microscopic reconstruction system; S3: Use the training set obtained in step S1 to train the Fourier stacked microscopic reconstruction system constructed in step S2. Use the multi-domain sensing loss function to calculate the difference between the network output and the high-resolution real amplitude image and update the network parameters to obtain the microscopic reconstruction model. S4: Input the low-resolution image sequence of the pathological slice to be reconstructed into the microscopic reconstruction model trained in step S3 to obtain a high-resolution amplitude reconstruction image.

[0011] Furthermore, in step S3, the multi-domain sensing loss function The expression is: ; ; ; ; ; in, For pixel-level reconstruction loss, For frequency domain amplitude constraint loss, To deeply perceive quality loss, For multi-level wavelet subband loss, High-resolution reconstructed amplitude images output by a Fourier layered microscopy reconstruction system. For high-resolution true amplitude images, and These represent the height and width of the high-resolution true amplitude image, respectively. This is a two-dimensional fast Fourier transform operation. To calculate the magnitude of a complex number, These are specific weight coefficients for the l-th layer in the pre-trained VGG-19 model. Features extracted from the first layer of the pre-trained VGG-19 model. These are the target coefficients corresponding to the first-level discrete wavelet transform. These are the target coefficients corresponding to the second-level discrete wavelet transform. The decoded features are output by the first-level wavelet-Fourier decoding submodule. This refers to the encoded features output by the last-level wavelet-Fourier coding submodule.

[0012] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The lightweight Fourier stacked reconstruction system and method described in this invention utilizes Fourier frequency domain calculation to efficiently replace the complex spatial self-attention mechanism, saving a large number of model parameters and memory overhead for processing high-throughput scanning tasks with large field of view of pathological slides. Compared with traditional physical iterative algorithms and existing large-scale deep learning networks, it significantly shortens the reconstruction inference time.

[0013] (2) The lightweight Fourier stacked reconstruction system and method described in this invention does not rely on traditional network space downsampling operations that easily lead to loss of high-frequency information when performing multi-scale feature extraction. Instead, it introduces discrete wavelet transform and fusion, which not only effectively preserves the high-fidelity extraction of micro-textures such as cell edges, but also overcomes the problems of excessive image smoothing and artifact generation by benefiting from the synergistic interaction of global and local Fourier branches and the joint constraints of multi-domain perceptual loss functions, so that the reconstructed image has clearer tissue contours and more realistic cell details.

[0014] (3) The lightweight Fourier layer reconstruction system and method described in this invention provide a high-resolution amplitude image rapid reconstruction scheme with a small number of parameters and better objective evaluation indicators. Users can deploy it in clinical high-throughput digital pathology microscopy imaging equipment, providing practical and feasible technical support for actual engineering implementation.

[0015] (4) The lightweight Fourier stacked reconstruction system and method described in this invention introduces a discrete wavelet transform and wavelet feature fusion module to achieve a multi-scale spatial-frequency domain separation and compensation mechanism for input features. This enables the system to effectively avoid the loss of high-frequency information caused by traditional network downsampling and achieve high-fidelity extraction of micro-textures such as cell edges. In addition, this invention benefits from the synergistic interaction of global and local Fourier branches in the dual-frequency feature fusion module and the joint constraints of the multi-domain perception loss function at the pixel, spectrum and perception quality levels. This makes the reconstructed high-resolution amplitude image not only have better objective evaluation indicators, but also effectively overcome the problems of over-smoothing and artifact generation. Subjectively, it has clearer tissue contours and more realistic cell details. Finally, this application uses Fourier frequency domain calculation to efficiently replace the complex spatial self-attention mechanism, which greatly compresses the number of model parameters and memory overhead. In the high-throughput scanning scenario of large field of view of pathological slides, compared with traditional physical iterative algorithms and existing large-scale deep learning networks, the reconstruction inference time of this invention is significantly shortened, providing feasibility for the practical engineering implementation of rapid microscopic imaging in clinical digital pathology. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the lightweight Fourier layer reconstruction system described in the embodiments of the present invention; Figure 2 A schematic diagram of the FPM device and lightweight Fourier stacked reconstruction system described in the embodiments of the present invention; Figure 2 (a) A schematic diagram of the structure of the FPM device according to an embodiment of the present invention; Figure 2 (b) A schematic diagram of the lightweight Fourier stacked reconstruction system described in an embodiment of the present invention; Figure 3 A schematic diagram of the processing flow of each wavelet-Fourier coding submodule or each wavelet-Fourier decoding submodule as described in the embodiments of the present invention; Figure 4 A schematic diagram of the structure of the dual-frequency domain feature fusion module described in an embodiment of the present invention; Figure 5 A schematic diagram of the wavelet feature fusion module described in the embodiment of the present invention; Figure 6 A schematic flowchart of the lightweight Fourier layer reconstruction method described in the embodiments of the present invention; Figure 7 Comparative reconstruction results of different algorithms described in the embodiments of the present invention; Figure 8 A schematic diagram of quantitative reconstruction indicators for different algorithms described in the embodiments of the present invention.

[0017] Explanation of reference numerals in the attached figures: 1. Programmable LED illumination array; 2. Stage; 3. Microscope objective; 4. Lens; 5. Image sensor. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] like Figures 1-2 As shown, a lightweight Fourier layer reconstruction system includes: The feature extraction module performs shallow feature extraction on the input low-resolution pathological slide image to obtain shallow features; The multi-level frequency domain transformation module performs multi-level frequency domain information extraction and enhancement on the input low-resolution pathological slice image to obtain multi-level frequency domain enhancement features; The encoding and decoding module uses multi-level frequency domain enhancement features to encode shallow features at multiple levels, and then decodes the encoded features level by level to obtain deep features. The reconstruction module performs upsampling on deep features and outputs a high-resolution amplitude image of the target pathological section.

[0024] It should be noted that the Fourier layered microscopic reconstruction system includes: a feature extraction module: performing initial convolution operations on the input low-resolution pathological slide image to extract shallow features, i.e., extracting basic spatial visual features, providing information carriers for subsequent depth processing; and an encoding / decoding module using a wavelet-Fourier feature extraction backbone, performing multi-scale spatial frequency decomposition and frequency domain interaction on the shallow features to obtain deep features. The wavelet-Fourier feature extraction backbone uses a symmetrical U-shaped network (UNET network), containing multiple cascaded wavelet-Fourier encoding sub-modules and wavelet-Fourier decoding sub-modules. In the feature downsampling stage, discrete wavelet transform is used to replace the traditional spatial pooling operation, explicitly decomposing the features into low-frequency approximate sub-bands representing macroscopic tissue structures, and three high-frequency detail sub-bands representing microscopic cell edges (corresponding to the horizontal, vertical, and diagonal directions). Each wavelet-Fourier coding submodule or wavelet-Fourier decoding submodule contains a dual-frequency feature fusion module for lightweight extraction of deep features in the frequency domain. The wavelet feature fusion module, during the upsampling skip connection stage, performs spatial alignment and nonlinear fusion of high and low frequency components of the multi-band sub-band features extracted in the downsampling stage with the deep features of the corresponding level wavelet-Fourier decoding submodule. It also performs cross-band and cross-level dimensional recalibration of the wavelet sub-bands and backbone features to compensate for lost high-frequency details. The image reconstruction module upsamples and performs channel dimensionality reduction on the extracted deep features to complete the reconstruction and output the final high-resolution amplitude image.

[0025] In some embodiments, the low-resolution pathological slide images are obtained after preprocessing, and the low-resolution pathological slide image sequence is as follows: ,in, and These represent the height and width of a low-resolution pathological slide image, respectively. This indicates the number of images taken at different illumination angles, and the corresponding reconstructed images output are... Furthermore, in specific embodiments of the present invention, a single Convolutional layers, acting as feature extraction modules, extract basic features (such as spatial structure and edges) to obtain shallow features. , The number of channels representing shallow features, i.e.:

[0026] in, express Convolution operation.

[0027] In some embodiments, the frequency domain transformation module includes a discrete wavelet transform module and a wavelet feature fusion module, wherein: The discrete wavelet transform module performs spatial frequency decomposition on the input features by high-frequency details and low-frequency approximations to obtain multi-band sub-features. The multi-band sub-features include low-frequency approximation sub-bands and three high-frequency detail sub-bands corresponding to the horizontal, vertical and diagonal directions, respectively. The wavelet feature fusion module concatenates high-frequency detail subbands in three directions along the channel dimension and then performs dimensionality reduction through pointwise convolution to obtain high-frequency features. These high-frequency features are then input in parallel into the spatial convolution branch and the global Fourier branch for processing. The low-frequency approximate subbands, the output features of the spatial convolution branch, and the output features of the global Fourier branch are concatenated and nonlinearly aggregated to obtain frequency domain enhanced features.

[0028] It should be noted that in the discrete wavelet transform module: One-dimensional wavelet transform is applied sequentially to the input features along the horizontal and vertical directions; The input features are decomposed into a low-frequency approximate subband that characterizes the macroscopic tissue structure of a pathological section, and three high-frequency detail subbands that correspond to the horizontal, vertical, and diagonal directions, respectively.

[0029] In the wavelet feature fusion module: The three high-frequency detail subbands are concatenated along the channel dimension and then dimensionality is reduced by pointwise convolution to obtain high-frequency features. High-frequency features are input in parallel to the spatial convolution branch and the global Fourier branch for processing. The output features of the low-frequency approximate subband, the spatial convolution branch, and the global Fourier branch are then spliced ​​and nonlinearly aggregated to obtain enhanced wavelet features. The enhanced wavelet features and wavelet-Fourier features are combined and fused to extract the features of the corresponding layer of the backbone.

[0030] In the spatial convolution branch, depthwise separable convolution is used to process high-frequency features and capture local texture patterns; In the global Fourier branch, the high-frequency features are modeled globally in the frequency domain using the fast Fourier transform and its inverse transform.

[0031] The wavelet-Fourier feature extraction backbone is a symmetrical U-shaped skeleton structure; In the U-shaped skeleton structure, discrete wavelet transform and its corresponding inverse wavelet transform are used to replace the traditional spatial pooling downsampling and convolution upsampling operations, so as to preserve spatial resolution and multi-band information during feature extraction.

[0032] In some embodiments, the encoding / decoding module includes an encoding branch and a decoding branch: The coding branch includes multiple cascaded wavelet-Fourier coding submodules. In each wavelet-Fourier coding submodule, the input is the coding features of the previous wavelet-Fourier coding submodule and the frequency domain enhancement features of the same scale. After multi-scale spatial frequency decomposition and frequency domain interaction processing, the coding features of the current wavelet-Fourier coding submodule are obtained. The decoding branch includes multiple cascaded wavelet-Fourier decoding submodules. In each wavelet-Fourier decoding submodule, the input is the decoding features of the previous wavelet-Fourier decoding submodule and the coding features at the same scale. After multi-scale spatial frequency decomposition and frequency domain interaction processing, the decoding features of the current wavelet-Fourier decoding submodule are obtained.

[0033] It should be noted that the first-level wavelet-Fourier coding submodule has no preceding wavelet-Fourier coding submodule. Therefore, the input of the first-level wavelet-Fourier coding submodule is shallow features. In addition, the first-level wavelet-Fourier decoding submodule has no preceding wavelet-Fourier decoding submodule. Therefore, the input of the first-level wavelet-Fourier decoding submodule is the encoded features output by the last-level wavelet-Fourier coding submodule and encoded features of the same scale.

[0034] Furthermore, assuming the encoding / decoding module includes a first-level wavelet-Fourier encoding submodule, a second-level wavelet-Fourier encoding submodule, a third-level wavelet-Fourier encoding submodule, a first-level wavelet-Fourier decoding submodule, and a second-level wavelet-Fourier decoding submodule, the multi-level frequency domain transformation module specifically consists of two-level frequency domain transformation modules. The first-level frequency domain transformation module extracts and enhances frequency domain information from the input low-resolution pathological slice image to obtain first-level frequency domain enhancement features. The second-level frequency domain transformation module extracts and enhances frequency domain information from the output first-level frequency domain enhancement features to obtain second-level frequency domain enhancement features. The first-level wavelet-Fourier coding submodule processes the shallow features to obtain coded feature A1. The second-level wavelet-Fourier coding submodule processes coded feature A1 and the first-level frequency domain enhancement features to obtain coded feature A2. The third-level wavelet-Fourier coding submodule processes coded feature A2 and the second-level frequency domain enhancement features to obtain coded feature A3. The first-level wavelet-Fourier decoding submodule processes coded feature A3 and coded feature A2 to obtain decoded feature B1. The second-level wavelet-Fourier decoding submodule processes decoded feature B1 and coded feature A1 to obtain deep features.

[0035] In some embodiments, the specific operations of multi-scale spatial frequency decomposition and frequency domain interaction processing in each wavelet-Fourier coding submodule or each wavelet-Fourier decoding submodule are as follows: The discrete wavelet transform module performs spatial frequency decomposition on the input features using high-frequency details and low-frequency approximations to obtain multi-band sub-features. These multi-band sub-features are then preprocessed and mapped via group normalization and pointwise convolution before being input to the dual-frequency feature fusion module for frequency domain feature interaction to obtain interactive features. After processing with channel attention and pointwise convolution, the interactive features are residually concatenated with the multi-band sub-features to obtain fused features. The fused features are then processed via group normalization, pointwise convolution, depthwise convolution, GELU nonlinear activation function, and pointwise convolution before being residually concatenated with the fused features to obtain the output features.

[0036] It should be noted that feature preprocessing and channel mapping are channel-separated operations within the dual-frequency feature fusion module: The input features are processed using a parallel dual-branch structure, including a global Fourier branch and a local Fourier branch. The macroscopic structural features extracted by the global Fourier branch and the local high-frequency features extracted by the local Fourier branch are dynamically weighted, fused using an adaptive channel attention mechanism, and then concatenated with the input feature residuals to obtain the fused features. Furthermore, the process of obtaining the macroscopic structural features and local high-frequency features includes: In the global Fourier branch, the input features are mapped to the frequency domain through a two-dimensional fast Fourier transform, and then in the frequency domain... After the convolution operation, the spatial domain is restored by inverse transformation, capturing the macroscopic structure and global contextual dependencies of pathological slide samples in a lightweight manner; In the local Fourier branch, the input features are divided into non-overlapping local blocks. Within each local block, local frequency domain filtering is applied to extract high-frequency texture details such as cell nucleus edges and chromatin, and then the blocks are reassembled to obtain local high-frequency features. Finally, an adaptive channel attention mechanism is used to dynamically assign weights and selectively fuse the features output from the global and local branches.

[0037] Specifically, such as Figure 3As shown, each wavelet-Fourier coding submodule or each wavelet-Fourier decoding submodule includes a discrete wavelet transform module and a dual-frequency feature fusion module. These two modules work together to perform multi-scale spatial frequency decomposition, targeted feature enhancement, and deep frequency domain fusion of image features. Specifically: the discrete wavelet transform module performs spatial frequency decomposition of the input features using high-frequency details and low-frequency approximations, effectively avoiding the loss of high-frequency information caused by traditional spatial downsampling, preserving the spatial resolution of the features, and obtaining multi-frequency sub-band features; in the dual-frequency feature fusion module, a local Fourier branch is used to perform local frequency domain filtering on the input features to extract the microscopic texture details of pathological slices, obtaining local features; a global Fourier branch is used to perform fast Fourier transform and frequency domain convolution operations on the input features to capture the global contextual dependencies and macroscopic structure of the input features, obtaining global features; finally, adaptive channel attention weight allocation and dynamic fusion are applied to the global and local features, and then combined with the input features to obtain the output features.

[0038] In some embodiments, such as Figure 4 As shown, in the dual-frequency feature fusion module, the input features are subjected to pointwise convolution and GELU activation operations to obtain the first feature. The first feature is then processed by local Fourier transform and global Fourier transform respectively to obtain global features and local features. The corresponding elements of the global features and local features are added together to obtain the second feature. The second feature is then processed by global average pooling, pointwise convolution, GELU activation function, pointwise convolution and channel separation operations to obtain the third feature and the fourth feature. The third feature and the fourth feature are then processed by the Softmax function to obtain the fifth feature and the sixth feature respectively. The fifth feature is added to the local feature to obtain the seventh feature. The sixth feature is added to the global feature to obtain the eighth feature. The seventh feature and the eighth feature are then added together to obtain the ninth feature. The ninth feature is then processed by pointwise convolution to obtain the output feature.

[0039] Specifically, the encoding and decoding modules are used for deep feature extraction and frequency domain interaction. The wavelet-Fourier encoding module and the wavelet-Fourier decoding module, hereinafter referred to as the WF encoding module and WF decoding module respectively, both employ a dual residual structure, sequentially containing a feature fusion subnetwork and a feedforward feature transformation subnetwork. Specifically, in the feature fusion subnetwork, the input features first undergo group normalization and pointwise convolution for feature preprocessing and channel mapping; subsequently, after segmentation, they are input to the dual-frequency feature fusion module for frequency domain feature interaction; the interacted features then sequentially pass through the channel attention module and pointwise convolution, and are residually connected to the input features of the feature fusion subnetwork. In the feedforward feature transformation subnetwork, the output features undergo group normalization, pointwise convolution, depthwise convolution, GELU nonlinear activation function, and pointwise convolution, and are finally residually connected to the input of the feedforward subnetwork to obtain the output features of the WF encoding module or the WF decoding module.

[0040] In the WF encoding or decoding module, the dual-frequency feature fusion module is the core unit for achieving collaborative perception of global context and local texture. Because pathological sections are rich in high-frequency textures such as cell nuclei at the microscopic level, and exhibit specific periodic arrangements such as glands at the macroscopic level, single-scale feature extraction is difficult to achieve simultaneously. Therefore, the dual-frequency feature fusion module adopts a parallel dual-branch structure, specifically composed of a global Fourier branch and a local Fourier branch.

[0041] Specifically, the global Fourier branch aims to capture the macroscopic structure and global contextual information of pathological sections in a lightweight manner. Input features... Mapping to the frequency domain via a two-dimensional Fast Fourier Transform, according to the convolution theorem, dot product in the frequency domain is equivalent to global circular convolution in the spatial domain. The calculation formula for this branch is: ; in, and These represent the two-dimensional fast Fourier transform and its inverse transform, respectively; The weights applied in the frequency domain are... Spectral weighting operation of convolution. O(·) denotes the asymptotic upper bound of the algorithm's time complexity, where H and W represent the height and width of the feature map, respectively. This operation... The increased complexity allows for a larger sensory field. () is the symbol for the asymptotic upper bound of time complexity.

[0042] In Local Fourier Branch (LFB), the aim is to compensate for the insufficient representation of local high-frequency cell edge structures by the global Fourier branch. The input feature map is divided into a set of non-overlapping local blocks. Frequency domain filtering is applied within each local block to accurately extract high-frequency details: ; in, Element-by-element multiplication For learnable frequency domain filter weights, This is the i-th local block. The processed i-th local block. The parts were reassembled to obtain local features. .

[0043] Subsequently, to balance the representation of global morphology and local texture, the dual-frequency feature fusion module introduces an adaptive channel attention mechanism to dynamically fuse the dual-branch features. First, the global features... and local features Adding the corresponding elements together yields Subsequently, channel adaptive weights are calculated using average pooling, multilayer perceptron, and the Softmax function, ultimately outputting the fused features of the dual-frequency feature fusion module. The mathematical formula for the fusion process is as follows: ; ; ; in, and These are the fusion weights dynamically assigned to the global branch and the local branch, respectively. ( ) represents average pooling. It is a multilayer perceptron. () represents the Softmax function.

[0044] Furthermore, in the feature downsampling stage of the network skeleton, to avoid the irreversible loss of high-frequency micro-textures in pathological slides caused by traditional spatial pooling (such as max pooling or stride convolution), this invention uses a discrete wavelet transform module to replace the conventional downsampling operation. Given an input feature map... First, apply a one-dimensional wavelet transform along the horizontal direction to calculate the low-frequency components. and high frequency components Its mathematical expression is: ; ; in, and These represent the pixel or feature value of the input feature map x at the i-th row, 2j-th column, and 2j+1-th column, respectively; i represents the row index of the input feature map, and j represents the group index of adjacent sampling positions in the horizontal direction; This represents the low-frequency near-component obtained through horizontal wavelet transform. This represents the corresponding high-frequency component.

[0045] Next, wavelet transform is performed again on the above components along the vertical direction, explicitly decomposing the input features into four frequency sub-bands with half the spatial resolution. The specific calculation is as follows: ; ; in, The low-frequency approximate subband (LL) is used to represent the macroscopic tissue structure of pathological sections. These are high-frequency detail subbands representing the horizontal (LH), vertical (HL), and diagonal (HH) dimensions, respectively, which carry texture information such as microscopic cell nucleus edges and chromatin. The extracted multi-scale frequency subbands are then injected into the backbone network during the upsampling stage via a wavelet feature fusion module.

[0046] In this embodiment of the invention, in order to fully utilize the multi-band representation extracted by the discrete wavelet transform module and deeply fuse it with the decoder features of the backbone network, a wavelet feature fusion module is designed. The wavelet feature fusion module aims to explicitly inject multi-scale prior information from the wavelet domain into the deep feature space, thereby effectively alleviating the texture smoothing problem commonly encountered in deep networks during pathological slide reconstruction.

[0047] Specifically, such as Figure 5 As shown, the feature processing procedure of the wavelet feature fusion module is as follows: First, the three high-frequency detail subbands generated during the downsampling stage are concatenated along the channel dimension, and then... Convolution is used for dimensionality reduction and aggregation to obtain high-frequency features. Meanwhile, for the low-frequency approximate sub-band Also through independent branches Convolution is used for dimensionality reduction to obtain low-frequency features. Its mathematical expression is: ; ; in, express Convolution operation, This indicates a channel splicing operation.

[0048] To further explore the structural information in high-frequency details, high-frequency features will be used... The inputs are fed in parallel to the global Fourier branch and the local Fourier branch. The global Fourier branch uses depthwise separable convolutions to capture local texture patterns, while the local Fourier branch utilizes the Fast Fourier Transform and its inverse to capture global context dependencies. ; ; in, and The features extracted are from the global Fourier branch and the local Fourier branch, respectively. Represents a non-linear activation function. () indicates depthwise separable convolution.

[0049] Finally, the wavelet feature fusion module performs cross-level aggregation of multi-frequency band features. This includes low-frequency features... High-frequency spatial characteristics and high-frequency characteristics in the frequency domain By performing splicing and nonlinear mapping, enhanced wavelet features are obtained. ; and then compared it with the coding features corresponding to the downsampling stage of the backbone network. The frequency domain enhancement features of this module are fused and projected to output the signal. The specific formula is as follows: ; ; like Figure 6 As shown, the present invention also provides a lightweight Fourier stack reconstruction method, comprising: S1: Obtain low-resolution pathological slide image sequences under different lighting angles, and use the high-resolution real amplitude images corresponding to the low-resolution pathological slide image sequences as labels to construct a training set; S2: Constructing a Fourier layered microscopic reconstruction system; S3: Use the training set obtained in step S1 to train the Fourier stacked microscopic reconstruction system constructed in step S2. Use the multi-domain sensing loss function to calculate the difference between the network output and the high-resolution real amplitude image and update the network parameters to obtain the microscopic reconstruction model. S4: Input the low-resolution image of the central pathological section from the low-resolution image sequence of the pathological sections to be reconstructed into the microscopic reconstruction model trained in step S3 to obtain a high-resolution amplitude reconstruction image.

[0050] It should be noted that the purpose of this invention is to provide a lightweight Fourier layered microscopic reconstruction system and method, which aims to achieve efficient and high-fidelity reconstruction of high-resolution amplitude images of pathological sections with a low number of model parameters by using frequency domain global interaction and wavelet multi-scale feature fusion.

[0051] Further, step S1: Data acquisition and preprocessing of the FPM device. A programmable LED array is used as the illumination source for the microscope. LEDs at different positions are lit sequentially to expose the pathological slide samples at different illumination angles, acquiring a sequence of low-resolution intensity images of the pathological slides containing low-frequency contours and high-frequency diffraction information. Simultaneously, corresponding high-resolution true amplitude images are acquired as supervision labels. After registration and normalization preprocessing, training and testing sets are constructed. S2: Construction of a lightweight Fourier layered reconstruction system. A Fourier layered microscopic reconstruction system is constructed to establish a nonlinear mapping relationship between the "low-resolution intensity image sequence of pathological slides" and the "high-resolution true amplitude image." S3: Model training and joint parameter optimization. The Fourier layered reconstruction system is trained end-to-end using the training set obtained in step S1. A multi-domain perceptual loss function is used to comprehensively calculate the difference between the network output amplitude and the high-resolution true amplitude image. The network parameters are updated through backpropagation to obtain the trained microscopic reconstruction model. S4: Rapid reconstruction of high-resolution amplitude of pathological slides. The low-resolution image of the central pathological section in the low-resolution image sequence of the pathological section to be reconstructed is input into the trained microscopic reconstruction model. Through the forward inference of the model, a high-resolution amplitude image with clear cell texture and edge details is directly output. Step S1 specifically includes: S11: Using an LED array as the illumination source for the microscope, place the pathological slide sample on stage 2 and illuminate it sequentially. LED beads located at different positions expose pathological slides under different illumination angles to obtain images within the same field of view. Low-resolution intensity images of pathological sections were obtained, and a sequence of low-resolution intensity images of pathological sections was acquired. The FPM device includes a programmable LED illumination array 1, a stage 2, a lens 4, a microscope objective 3, and an image sensor 5. The programmable LED illumination array 1 is placed below the sample to provide partially coherent plane waves at different illumination angles; the pathological slide sample is placed on the stage 2; after passing through the 4f system formed by the lens 4, the microscope objective 3 and the image sensor 5 together constitute a low-resolution, large-field-of-view optical acquisition module.

[0052] Let the ideal high-resolution complex amplitude function of the pathological section sample be... It includes amplitude and phase components. When the first... When each LED is lit, it uses wave vector The sample is illuminated by a tilted plane wave, and the emitted light field is truncated by a low-pass filter of the microscope objective 3 before being received by the image sensor 5.

[0053] S12: Low-resolution intensity image of the i-th pathological slide The mathematical expression is:

[0054] in, Represents spatial domain coordinates, Represents frequency domain coordinates, Indicates the first Frequency domain offset corresponding to each LED This indicates the inverse Fourier transform operation. The spectral function representing a pathological slide sample. The coherence transfer function (i.e., pupil function) of microscope objective 3 is represented. S13: Replace the low-resolution intensity image of the i-th pathological slide with the low-resolution intensity image of the (i+1)-th pathological slide until the mathematical expression of the low-resolution intensity image of all pathological slides is obtained. By lighting them one by one Different LED beads can be used to obtain the same field of view. A sequence of low-resolution intensity images of pathological sections.

[0055] S14: Based on the frequency domain offset of each LED, the sub-apertures are located in the frequency domain. Each sub-aperture is independent and does not overlap with the others. Each sub-aperture covers different local frequency band information of the pathological slide sample spectrum. The amplitude information of the corresponding sub-aperture is replaced by the square root of the low-resolution intensity image of each pathological slide to obtain the stitched high-resolution spectrum. The stitched high-resolution spectrum is iterated using the physical iterative phase recovery algorithm to obtain the global wideband high-resolution spectrum. The global wideband high-resolution spectrum is subjected to inverse Fourier transform to obtain the high-resolution complex amplitude distribution of the pathological slide sample. Finally, the absolute value of the high-resolution complex amplitude distribution of the pathological slide sample is taken to obtain the high-resolution true amplitude image.

[0056] To train a deep learning reconstruction network, corresponding high-resolution real amplitude images are needed as supervision labels. The specific steps are as follows: using traditional physical iterative phase retrieval algorithms such as alternating projection or Gaussian-Newton, multiple iterative optimizations are performed between the spatial and frequency domains. In the frequency domain, the square root of the low-resolution intensity image of the real pathological slide acquired in the spatial domain is used. The amplitude information of the corresponding sub-aperture is replaced to continuously update and stitch together the global wideband high-resolution spectrum of the sample. After iterative convergence, an inverse Fourier transform is performed on the global wideband high-resolution spectrum to obtain the high-resolution complex amplitude distribution of the pathological slide sample. Its formula is: ; Finally, the absolute values ​​of the complex amplitude distribution are extracted as the labels for the final high-resolution true amplitude image. : ; In some embodiments, in step S3, the multi-domain sensing loss function The expression is: ; ; ; ; ; in, For pixel-level reconstruction loss, For frequency domain amplitude constraint loss, To deeply perceive quality loss, For multi-level wavelet subband loss, High-resolution reconstructed amplitude images output by a Fourier layered microscopy reconstruction system. For high-resolution true amplitude images, and These represent the height and width of the high-resolution true amplitude image, respectively. This is a two-dimensional fast Fourier transform operation. To calculate the magnitude of a complex number, These are specific weight coefficients for the l-th layer in the pre-trained VGG-19 model. Features extracted from the first layer of the pre-trained VGG-19 model. These are the target coefficients corresponding to the first-level discrete wavelet transform. These are the target coefficients corresponding to the second-level discrete wavelet transform. The decoded features are output by the first-level wavelet-Fourier decoding submodule. This refers to the encoded features output by the last-level wavelet-Fourier coding submodule.

[0057] The multi-domain perceptual loss function is composed of the following four weighted components: Pixel-level reconstruction loss: Calculates the difference in spatial domain between the amplitude image of the network output (i.e., the output of the Fourier stacked microscopic reconstruction system) and the high-resolution real amplitude image. Absolute pixel difference; Frequency domain amplitude constraint loss: Calculate the difference in spectral amplitude between the network output amplitude image and the high-resolution real amplitude image after Fast Fourier Transform, which is used to constrain the spectral distribution of the reconstructed image; Depth perception quality loss: Calculate the perception distance based on the high-level features extracted by the pre-trained network to improve the visual realism of pathological slide texture; Multi-level wavelet subband loss: Generate target coefficients of corresponding scales by multi-level dynamic wavelet decomposition of the high-resolution real amplitude image, and calculate the difference between the target coefficients and the output features of the corresponding level decoder of the network, which is used to constrain the output of the intermediate layers of the network and implement depth supervision.

[0058] In this embodiment of the invention, in order to maintain the overall structural consistency during the reconstruction process while enhancing the microscopic texture and high-frequency details of pathological tissues, a multi-domain perceptual loss function is designed to perform end-to-end joint optimization of the network. The multi-domain perceptual loss function is a composite loss that integrates spatial domain pixel consistency constraints, frequency domain amplitude constraints, depth-aware quality supervision, and multi-level wavelet coefficient supervision, guiding the update of network parameters from multiple dimensions.

[0059] First, adopt Loss as pixel-level reconstruction loss This is used to ensure the overall spatial consistency between the network output image and the real image. Compared to loss, The loss function is more robust to outliers and helps generate sharper edges. Its definition is as follows: ; in, For high-resolution reconstructed amplitude images output by the network, For the corresponding high-resolution true amplitude image, and These represent the height and width of the high-resolution true amplitude image, respectively.

[0060] However, relying solely on spatial domain loss often leads to over-smoothing of high-frequency details. Therefore, this invention further introduces frequency domain amplitude constraint loss. This loss function calculates the difference in spectral amplitude between the reconstructed image and the high-resolution ground-real image after Fast Fourier Transform, explicitly supervising the spectral distribution of the reconstructed image to help suppress blur artifacts and sharpen textures. ; in, This represents the two-dimensional Fast Fourier Transform operation. This indicates the magnitude of the complex number.

[0061] Furthermore, to improve the subjective visual quality of reconstructed pathological slide images and make them more consistent with the observation habits of clinicians, this invention introduces LPIPS perceptual quality loss based on deep features. This is known as depth-perceptual quality loss. Depth-perceptual quality loss utilizes high-level features extracted by a pre-trained network to measure the perceptual distance between images, thereby better evaluating the realism of textures. ; in, Indicates the first in the pre-trained network Feature extractor of the layer, These are the specific weight coefficients for the corresponding layer.

[0062] Finally, to explicitly guide the model to learn multi-scale frequency information in deep networks, this invention designs a multi-level wavelet subband loss. "Deep supervision" is applied to the intermediate layer features of the wavelet-Fourier decoding submodule. Specifically, this is applied to high-resolution true amplitude images. Perform two-level discrete wavelet transform to dynamically generate target coefficients. and : ; in, These are the target coefficients corresponding to the first-level discrete wavelet transform. Subsequently, the target coefficients and the decoded features output by the first-level wavelet-Fourier decoding submodule are calculated. and the coding features output by the last-stage wavelet-Fourier coding submodule Between loss: ; This deep supervision strategy forces the network to learn the correct frequency subband representation during the intermediate feature extraction stage, effectively alleviating the gradient vanishing problem and guiding the model to achieve high-fidelity feature reconstruction from coarse to fine.

[0063] In summary, the total loss function used in the embodiments of this invention... Defined as a weighted combination of the above four losses: Based on parameter sensitivity analysis experiments, this embodiment determines the optimal weight configuration of the loss function as follows: , This configuration maximizes the visual reconstruction quality of microscopic tissue images of pathological sections while ensuring leading performance in all objective numerical indicators.

[0064] To verify the superiority of the Fourier layered reconstruction effect of the method provided in this invention, the reconstruction results of this invention are compared with those of existing methods (CNN (U-Net, MDSR), Transformer (SwinIR, Uformer), and GAN (WM-FPM), etc.) to perform high-resolution reconstruction of pathological slide images in the FPM-BioCell dataset. Peak signal-to-noise ratio (PSNR) (dB), structural similarity (SSIM), multi-scale structural similarity (MS-SSIM), and learned perceptual patch similarity (LPIPS) are used for full-reference quality assessment to obtain the corresponding high-resolution amplitude reconstruction results. In addition, Methods refers to the algorithm model name, Ours refers to the algorithm proposed in this paper, Params refers to the number of parameters, FLOPS refers to floating-point operations, and Inference time refers to the inference time per image. Figure 7 As can be seen, this invention significantly outperforms other methods in terms of clarity, continuity, and realism in reconstructing fine structures (such as cell nuclear outlines), microscopic textures (such as chromatin distribution), and complex tissue morphologies (such as layered squamous epithelium and Ascaris eggs). The comparison results of reconstruction indices are shown below. Figure 8 As shown. From Figure 8 As can be seen, this invention achieves optimal values ​​across all objective evaluation metrics. The average PSNR, SSIM, and MS-SSIM comprehensively outperform all comparative models, particularly excelling in the LPIPS metric, which measures perceptual similarity. Furthermore, it achieves approximately 40 times faster inference speed compared to traditional iterative algorithms, indicating that the reconstruction results better align with the visual judgment standards of clinical pathologists and meet the needs of practical engineering applications.

[0065] This invention is applicable to, but not limited to, high-resolution amplitude reconstruction of pathological sections, as well as: dynamic observation of living cells: for large-field-of-view, high-throughput label-free imaging of living cells in culture dishes, assisting researchers in observing cell division and growth dynamics; hematological microscopic detection: for improving the resolution of microscopic images such as peripheral blood smears, enhancing the ability to identify key information such as red blood cell morphology and white blood cell classification; industrial microscopic detection: for high-precision industrial microscopic scenarios such as semiconductor wafer defect detection and material surface microstructure analysis.

[0066] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0067] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A lightweight Fourier layer reconstruction system, characterized in that, include: The feature extraction module performs shallow feature extraction on the input low-resolution pathological slide image to obtain shallow features; The multi-level frequency domain transformation module performs multi-level frequency domain information extraction and enhancement on the input low-resolution pathological slice image to obtain multi-level frequency domain enhancement features; The encoding and decoding module uses multi-level frequency domain enhancement features to encode shallow features at multiple levels, and then decodes the encoded features level by level to obtain deep features. The reconstruction module performs upsampling on deep features and outputs a high-resolution amplitude image of the target pathological section.

2. The lightweight Fourier layer reconstruction system according to claim 1, characterized in that: The frequency domain transformation module includes a discrete wavelet transform module and a wavelet feature fusion module, wherein: The discrete wavelet transform module performs spatial frequency decomposition on the input features by high-frequency details and low-frequency approximations to obtain multi-band sub-features. The multi-band sub-features include low-frequency approximation sub-bands and three high-frequency detail sub-bands corresponding to the horizontal, vertical and diagonal directions, respectively. The wavelet feature fusion module concatenates high-frequency detail subbands in three directions along the channel dimension and then performs dimensionality reduction through pointwise convolution to obtain high-frequency features. These high-frequency features are then input in parallel into the spatial convolution branch and the global Fourier branch for processing. The low-frequency approximate subbands, the output features of the spatial convolution branch, and the output features of the global Fourier branch are concatenated and nonlinearly aggregated to obtain frequency domain enhanced features.

3. The lightweight Fourier layer reconstruction system according to claim 1, characterized in that: The encoding / decoding module includes an encoding branch and a decoding branch: The coding branch includes multiple cascaded wavelet-Fourier coding submodules. In each wavelet-Fourier coding submodule, the input is the coding features of the previous wavelet-Fourier coding submodule and the frequency domain enhancement features of the same scale. After multi-scale spatial frequency decomposition and frequency domain interaction processing, the coding features of the current wavelet-Fourier coding submodule are obtained. The decoding branch includes multiple cascaded wavelet-Fourier decoding submodules. In each wavelet-Fourier decoding submodule, the input is the decoding features of the previous wavelet-Fourier decoding submodule and the coding features at the same scale. After multi-scale spatial frequency decomposition and frequency domain interaction processing, the decoding features of the current wavelet-Fourier decoding submodule are obtained.

4. The lightweight Fourier layer reconstruction system according to claim 3, characterized in that: In each wavelet-Fourier coding submodule or each wavelet-Fourier decoding submodule, the specific operations of multi-scale spatial frequency decomposition and frequency domain interactive processing are as follows: The discrete wavelet transform module performs spatial frequency decomposition on the input features using high-frequency details and low-frequency approximations to obtain multi-band sub-features. These multi-band sub-features are then preprocessed and mapped via group normalization and pointwise convolution before being input to the dual-frequency feature fusion module for frequency domain feature interaction to obtain interactive features. After processing with channel attention and pointwise convolution, the interactive features are residually concatenated with the multi-band sub-features to obtain fused features. The fused features are then processed via group normalization, pointwise convolution, depthwise convolution, GELU nonlinear activation function, and pointwise convolution before being residually concatenated with the fused features to obtain the output features.

5. The lightweight Fourier layer reconstruction system according to claim 4, characterized in that: In the dual-frequency feature fusion module, the input features are subjected to pointwise convolution and GELU activation operations to obtain the first feature. The first feature is then processed by local Fourier transform and global Fourier transform respectively to obtain global and local features. The corresponding elements of the global and local features are added together to obtain the second feature. The second feature is then processed by global average pooling, pointwise convolution, GELU activation function, pointwise convolution and channel separation operations to obtain the third and fourth features. The third and fourth features are then processed by the Softmax function to obtain the fifth and sixth features respectively. The fifth feature is added to the local features to obtain the seventh feature. The sixth feature is added to the global features to obtain the eighth feature. The seventh and eighth features are then added together to obtain the ninth feature. The ninth feature is then processed by pointwise convolution to obtain the output feature.

6. A lightweight Fourier layer reconstruction method, characterized in that, include: S1: Obtain low-resolution pathological slide image sequences under different lighting angles, and use the high-resolution real amplitude images corresponding to the low-resolution pathological slide image sequences as labels to construct a training set; S2: Construct the Fourier layered microscopic reconstruction system as described in any one of claims 1 to 5; S3: Use the training set obtained in step S1 to train the Fourier stacked microscopic reconstruction system constructed in step S2. Use the multi-domain sensing loss function to calculate the difference between the network output and the high-resolution real amplitude image and update the network parameters to obtain the microscopic reconstruction model. S4: Input the low-resolution image of the central pathological section from the low-resolution image sequence of the pathological sections to be reconstructed into the microscopic reconstruction model trained in step S3 to obtain a high-resolution amplitude reconstruction image.

7. The lightweight Fourier layer reconstruction method according to claim 6, characterized in that, In step S3, the multi-domain sensing loss function The expression is: ； ； ； ； ； in, For pixel-level reconstruction loss, For frequency domain amplitude constraint loss, To deeply perceive quality loss, For multi-level wavelet subband loss, High-resolution reconstructed amplitude images output by a Fourier layered microscopy reconstruction system. For high-resolution true amplitude images, and These represent the height and width of the high-resolution true amplitude image, respectively. This is a two-dimensional fast Fourier transform operation. To calculate the magnitude of a complex number, These are specific weight coefficients for the l-th layer in the pre-trained VGG-19 model. Features extracted from the first layer of the pre-trained VGG-19 model. These are the target coefficients corresponding to the first-level discrete wavelet transform. These are the target coefficients corresponding to the second-level discrete wavelet transform. The decoded features are output by the first-level wavelet-Fourier decoding submodule. This refers to the encoded features output by the last-level wavelet-Fourier coding submodule.

Images