Pathological image super-resolution generation method fusing residual error and attention mechanism

Abstract

The invention discloses a residual error and attention mechanism fused pathological image super-resolution generation method, which comprises the following steps: acquiring a pathological section digital image, extracting a low-resolution image pair and a high-resolution image pair in the same visual field, preprocessing, and dividing a training set and a verification set; a generator for constructing a residual error-attention generation network, wherein the generator is used for reconstructing a super-resolution image according to the low-resolution image; constructing a discriminator network for distinguishing the authenticity of the generated image and the authenticity of the real high-resolution image; adversarial training is carried out on the generator and the discriminator based on the training set, network parameters are optimized through the verification set, and a trained generator network is obtained; and inputting a to-be-processed low-resolution pathological image into the trained generator to obtain a super-resolution reconstructed image with a clear structure. Through the combination of the residual structure and the attention mechanism, the reconstruction quality and the detail reduction capability of the pathological image are remarkably improved, and improvement of the accuracy of pathological diagnosis and the reliability of digital pathological analysis is facilitated.

Description

Technical Field

[0001] This invention relates to the field of pathology and medical image computing, and in particular to a method for generating super-resolution pathology images by integrating residual and attention mechanisms. Background Technology

[0002] With the popularization of digital pathology slide scanning technology, its resolution directly affects the accuracy of disease diagnosis and the identification of early lesions. However, due to limitations in hardware resolution, scanning time costs, and storage pressure, it is practically difficult to acquire high-resolution pathology images in real time. Therefore, reconstructing low-resolution pathology images using algorithms to generate high-resolution images has become one of the key technologies in precision medicine. Early super-resolution methods for pathological images mainly relied on interpolation algorithms, but these methods can lead to blurred image edges. With the development of deep learning, models based on convolutional neural networks have gradually become mainstream. Among them, generative adversarial networks (GANs) have been widely introduced into medical image reconstruction tasks due to their ability to generate high-frequency details. GANs enable the generator to output images with higher visual quality through adversarial games between the generator and the discriminator. On this basis, researchers have introduced modules such as attention mechanisms, dense connections, and residual learning to enhance the network's ability to extract features of complex textures and cell morphologies in pathological images. Despite significant advancements in existing technologies, several challenges remain in practical applications: First, many CNN-based models suffer from "averaging effects," resulting in insufficient texture detail and blurred edges in reconstructed images, failing to meet the cellular-level precision requirements of pathological diagnosis. Second, generative adversarial networks exhibit instability during training, prone to mode collapse and gradient vanishing issues, and their ability to maintain the consistency of image spatial structure needs improvement. Third, most existing methods directly borrow from the super-resolution framework of natural images, failing to adequately optimize for the uneven cell distribution, diverse morphologies, and complex tissue structures in pathological images, potentially leading to artifacts or distortions when reconstructing high-frequency details. Furthermore, obtaining high-quality training data is difficult, and the multi-center heterogeneity of pathological images and the scarcity of labeled data further restrict the improvement of model performance and generalization ability. Summary of the Invention

[0003] This invention provides a method for generating super-resolution pathological images that integrates residual and attention mechanisms. This method overcomes the problem that the reconstructed images suffer from insufficient texture details and blurred edges due to the "averaging effect" of CNN models. It also fails to fully optimize for the uneven distribution of cells, diverse morphologies, and complex tissue structures in pathological images, which may lead to artifacts or distortions when reconstructing high-frequency details.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows: A method for generating super-resolution pathological images that integrates residual and attention mechanisms includes: S1. Acquire digital images of pathological slides and extract low-resolution and high-resolution image pairs from the same field of view of the digital images; preprocess the low-resolution and high-resolution image pairs and divide the preprocessed low-resolution and high-resolution image pairs into training and validation sets according to a specified number. S2. Construct a residual-attention generation network as a generator network to generate a super-resolution reconstructed image based on a low-resolution image; S3. Construct a discriminator network to obtain confidence levels characterizing the authenticity of the super-resolution reconstructed image and the high-resolution image, so as to determine the authenticity of the super-resolution reconstructed image and the high-resolution image. S4. Based on the training set, perform adversarial training between the residual-attention generator network and the discriminator network, and adjust the parameters using the validation set to obtain the trained generator network. S5. Input the low-resolution pathological image to be processed into the trained generator network to obtain a super-resolution reconstructed image with clear pathological structure.

[0005] Furthermore, the residual-attention generation network includes an initial feature extraction module, a deep feature extraction module, a feature fusion module, and an upsampling reconstruction module; The initial feature extraction module is used to perform convolution operations on the input low-resolution image to obtain shallow features, and output them to the deep feature extraction module. The deep feature extraction module is used to perform initial deep feature extraction and channel weight recalibration on the shallow features to obtain deep features, and output them to the feature fusion module. The feature fusion module is used to fuse deep features and shallow features through residual connections to obtain fused features; The upsampling reconstruction module performs at least one upsampling operation on the fused features. After each upsampling, the features are corrected through a convolutional layer, and finally, a super-resolution reconstructed image is output.

[0006] Furthermore, the deep feature extraction module includes multiple stacked RRDB-CA sub-modules, wherein each RRDB-CA sub-module includes three cascaded residual density blocks, channel attention blocks, and cross-module residual connections; The residual density block includes k serially connected density convolutional layers and local residual connection layers; among them, the input of the first density convolutional layer is the shallow features output by the initial feature extraction module; the input of the j-th density convolutional layer, 1 < j ≤ k, is composed of the output feature maps of the first to the j-1-th density convolutional layers; its output is weighted and fused with the input shallow features through the local residual connection layer to obtain the initial deep features; The channel attention module is used to perform global average pooling on the initial deep features through a compression operation to obtain a channel-level global descriptor; through an excitation operation, a gating mechanism composed of two fully connected layers is used to perform a non-linear transformation on the channel-level global descriptor to generate the normalized weights of each channel; the normalized weights of each channel are multiplied by the initial deep features channel by channel to obtain the channel-weighted features; The cross-module residual connection is used to fuse the channel-weighted features with the shallow features to obtain the deep features.

[0007] Furthermore, the execution steps of the discriminator network include: S31. Extract the features of the super-resolution reconstructed image and the high-resolution image through the basic convolutional layer to obtain the primary features; S32. Through multiple sequentially connected downsampling stages, gradually expand the receptive field and increase the feature dimension of the primary features to obtain the downsampled features; S33. Aggregate the downsampled features through the global pooling layer to obtain the global features; S34. Integrate the global features through at least one fully connected equivalent layer to obtain the confidence representing the authenticity of the image.

[0008] Furthermore, in the adversarial training of the residual-attention generation network and the discriminator network, the residual-attention generation network is optimized through the L1 loss function, perceptual loss function, adversarial loss function, and brightness consistency loss function; the discriminator network is optimized through the following loss function:

[0009] In the formula, is the loss of the discriminator network; is the real image loss; is the generated image loss; Among them, the real image loss can be expressed as:

[0010] In the formula, is the binary cross-entropy loss; is the output of the discriminator; is the high-resolution image; Labels indicating whether an image is real or not; The loss in generating the image can be expressed as:

[0011] In the formula, This is the output of the generator; Low-resolution image; Is it a label for the generated image? The binary cross-entropy loss is expressed as follows:

[0012] In the formula, Input image; For target labels.

[0013] Beneficial effects: This invention provides a method for generating super-resolution pathological images that integrates residual and attention mechanisms. By constructing a residual-attention generation network, residual learning and attention mechanisms are deeply integrated. The residual structure effectively alleviates the gradient vanishing and information decay problems in deep network training, ensuring the stability and efficiency of network training. Meanwhile, the attention mechanism enables the network to autonomously focus on key cell morphology, tissue boundaries, and other fine structural regions in pathological images, significantly enhancing the model's ability to extract and reconstruct key pathological features, thereby generating super-resolution images with richer details, clearer textures, and higher visual quality. By employing adversarial training between a residual-attention generation network and a discriminator network, the discriminator network continuously drives the residual-attention generation network to output reconstruction results that are visually indistinguishable from real high-resolution pathological images. This not only improves accuracy at the pixel level but also effectively suppresses artifacts and blurring at the perceptual level, generating higher-quality images that are more realistic and in line with pathological observation habits, providing a reliable foundation for subsequent quantitative analysis and visual diagnosis. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the method for generating super-resolution pathological images according to the present invention. Figure 2 This is a flowchart illustrating the super-resolution generation process of pathological images in an embodiment of the present invention. Figure 3This is a structural diagram of the residual-attention generation network in an embodiment of the present invention; Figure 4 This is a structural diagram of the residual dense block in an embodiment of the present invention; Figure 5 This is a structural diagram of the channel attention block in an embodiment of the present invention; Figure 6 This is a structural diagram of the discriminator network in an embodiment of the present invention; Figure 7 This is a batch loss result diagram of the generator and discriminator in an embodiment of the present invention; Figure 8 This is a graph showing the average loss results of the generator and discriminator in an embodiment of the present invention; Figure 9 This is a graph showing the PSNR and SSIM index results in an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the fidelity of pathological structures in an embodiment of the present invention; Figure 11 These refer to the low-resolution image, the generated super-resolution image, and the true high-resolution image in the embodiments of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This embodiment provides a method for generating super-resolution pathological images by fusing residual and attention mechanisms, such as... Figure 1 As shown, it includes: S1. Acquire digital images of pathological slides and extract low-resolution and high-resolution image pairs from the same field of view of the digital images; preprocess the low-resolution and high-resolution image pairs and divide the preprocessed low-resolution and high-resolution image pairs into training and validation sets according to a specified number. S2. Construct a residual-attention generation network as a generator network to generate a super-resolution reconstructed image based on a low-resolution image; S3. Construct a discriminator network to obtain confidence levels characterizing the authenticity of the super-resolution reconstructed image and the high-resolution image, so as to determine the authenticity of the super-resolution reconstructed image and the high-resolution image. S4. Based on the training set, perform adversarial training between the residual-attention generator network and the discriminator network, and adjust the parameters using the validation set to obtain the trained generator network. S5. Input the low-resolution pathological image to be processed into the trained generator network to obtain a super-resolution reconstructed image with clear pathological structure.

[0018] Specifically, such as Figure 2 As shown, firstly, digital images of pathological sections are acquired, and low-resolution and high-resolution image pairs from the same field of view are extracted as basic data. After preprocessing the image pairs, they are divided into training and validation sets to provide sufficient and registered samples for subsequent model learning. Secondly, a residual-attention generation network is constructed as the generator network: this network introduces an attention mechanism module, enabling it to adaptively focus on key tissue structure regions in the pathological images, enhancing its ability to recover subtle features such as cell boundaries and nucleocytoplasmic ratio. Thirdly, a discriminator network is constructed to distinguish between the generated super-resolution images and real high-resolution images. Its output confidence drives the generator to continuously optimize and generate more realistic images. Subsequently, the generator and discriminator are trained adversarially based on the training set. The overall performance is improved through dynamic game between the two, and the parameters are adjusted using the validation set to prevent overfitting, ultimately resulting in a fully trained generator network. Finally, the low-resolution pathological image to be processed is input into the generator, which outputs a super-resolution reconstructed image with significantly enhanced tissue structure and texture details.

[0019] Specifically, the digitized images of the pathological slides are several pathological slides from multiple patients diagnosed with clear cell renal cell carcinoma. The slides are digitized by scanning them using a Macody or other high-definition pathological slide scanner. Using the professional software Motic DSAssistant, 100x and 400x images of the same field of view were exported sequentially, ensuring that the coordinates of the two were precisely aligned. The 100x image was cropped into a 256×256 pixel block as a low-resolution image, and the 400x image was cropped into a 1024×1024 pixel block as a high-resolution image. All defective samples in the low-resolution and high-resolution images were removed. The defective samples include: images lacking effective tissue areas, images with localized blurring due to scanning failures, images with stains or abnormal staining; Finally, coordinate alignment checks are performed on all low-resolution and high-resolution image pairs to ensure that the spatial positions of the low-resolution and high-resolution images are accurately matched.

[0020] In this embodiment, there are a total of 7238 pairs of low- and high-resolution images in the training set and a total of 25 pairs of low- and high-resolution image blocks in the validation set.

[0021] Preferably, as Figure 3 shown, the residual-attention generation network includes an initial feature extraction module, a deep feature extraction module, a feature fusion module, and an upsampling reconstruction module; Among them, the initial feature extraction module is used to perform a convolution operation on the input low-resolution image to obtain shallow features and output them to the deep feature extraction module; The deep feature extraction module is used to perform initial deep feature extraction and channel weight recalibration on the shallow features to obtain deep features and output them to the feature fusion module; The feature fusion module is used to fuse the deep features and shallow features through a residual connection to obtain fused features; The upsampling reconstruction module performs at least one upsampling operation on the fused features, and performs feature correction through a convolutional layer after each upsampling, and finally outputs a super-resolution reconstructed image.

[0022] Preferably, the deep feature extraction module includes multiple stacked Residual in Residual Dense Block-Channel Attention (RRDB-CA) sub-modules. Among them, each RRDB-CA sub-module includes three cascaded residual density blocks, a channel attention block, and a cross-module residual connection; As Figure 4 shown, the residual density block includes k sequentially connected density convolutional layers and a local residual connection layer; among them, the input of the first density convolutional layer is the shallow feature output by the initial feature extraction module; the input of the jth density convolutional layer, 1 < j ≤ k, is composed of the output feature maps of the first to the j-1th density convolutional layers; its output is weighted and fused with the input shallow feature through the local residual connection layer to obtain an initial deep feature; As Figure 5 shown, the channel attention module is used to perform global average pooling on the initial deep feature through a compression operation to obtain a channel-level global descriptor; through an excitation operation, a gating mechanism composed of two fully connected layers is used to perform a non-linear transformation on the channel-level global descriptor to generate a normalized weight for each channel; the normalized weight of each channel is multiplied by the initial deep feature channel by channel to obtain a channel-weighted feature; The cross-module residual connection is used to fuse the channel-weighted feature with the shallow feature to obtain a deep feature.

[0023] Specifically, the upsampling reconstruction module employs a progressive, enhanced structure to achieve 4x super-resolution magnification. The core of the upsampling reconstruction module consists of two consecutive 2x upsampling operations, each performed via sub-pixel convolution (PixelShuffle). Compared to traditional deconvolution or interpolation methods, sub-pixel convolution constructs a high-resolution space by periodically reorganizing the channels of the feature map. This is a parameter-free and efficient operation that effectively reduces the generation of "checkerboard" artifacts. To ensure the spatial accuracy and detail quality of the feature map after upsampling, a standard 3x3 convolutional layer is immediately applied as a correction convolution after each PixelShuffle operation. This design aims to learn and correct subtle pixel shifts or distortions that may be introduced by the upsampling process, while smoothing and refining the newly generated pixels to generate feature maps with more photometric continuity. Throughout the upsampling path, the LeakyReLU activation function with a negative slope of 0.2 is used. This choice introduces moderate non-linearity into the network, enabling it to learn more complex mappings. On the other hand, unlike the ReLU function, LeakyReLU allows small gradients to pass through the negative region. This characteristic can effectively prevent the "neuron death" problem common in deep networks, thereby significantly improving the stability and convergence of model training.

[0024] In this embodiment, the method of convolution + LeakyReLU is used to gradually reduce the number of channels, from 96 to 48 and then to 3, and outputs a super-resolution (SR) image with the same red-green-blue (RGB) 3 channels as the input low-resolution image.

[0025] Preferred, such as Figure 6 As shown, the execution steps of the discriminator network include: S31. By using basic convolutional layers, features of the super-resolution reconstructed image and the high-resolution image are extracted to obtain primary features; S32. Through multiple sequentially connected downsampling stages, the receptive field of the primary features is gradually expanded and the feature dimensions are increased to obtain downsampled features. Specifically, each stage includes a convolutional layer with a stride of 2 to halve the spatial dimension, and a batch normalization layer to stabilize the training process and accelerate convergence. At the same time, a LeakyReLU activation function with a negative slope of 0.01 is used to introduce nonlinearity and prevent gradient vanishing. S33. The downsampled features are aggregated through a global pooling layer to obtain global features; S34. By using at least one fully connected equivalent layer, global features are integrated to obtain the confidence level that represents the authenticity of the image.

[0026] In this embodiment, the hyperparameters for adversarial training are set as follows: (1) Scaling factor 4, batch size 16, iteration count 1000; (2) The optimizers for the generator and discriminator use Adaptive Moment Estimation (Adam), with initial learning rates set to 1e-4 and 5e-5, respectively; (3) Set up a learning rate scheduler, and halve the learning rate every 100 training epochs; (4) The early stop threshold is 0.001 and the patience value is 50.

[0027] Preferably, such as Figure 7 and Figure 8 As shown, in the adversarial training of the residual-attention generation network and the discriminator network, the residual-attention generation network is optimized using the L1 loss function, perceptual loss function, adversarial loss function, and brightness consistency loss function; the discriminator network is optimized using the following loss function:

[0028] In the formula, The loss of the discriminator network; Loss to the true image; To generate image loss; The real image loss can be expressed as:

[0029] In the formula, For binary cross-entropy loss; This is the output of the discriminator; High-resolution images; Labels indicating whether an image is real or not; The loss in generating the image can be expressed as:

[0030] In the formula, This is the output of the generator; Low-resolution image; Is it a label for the generated image? The binary cross-entropy loss is expressed as follows:

[0031] In the formula, Input image; For target labels.

[0032] Specifically, the adversarial training process is as follows: S41. When reading image data, the image needs to be preprocessed. Preprocessing includes random image cropping, image enhancement, and data conversion. Image random cropping involves randomly cropping a 48×48 pixel region from a low-resolution image, while simultaneously cropping a 192×192 pixel region with a matching center position from the corresponding high-resolution image. This aims to save computational resources, prevent overfitting, and encourage the model to learn local details and the mapping relationship from the local to the larger region, respectively. Image enhancement involves applying random data augmentation to cropped image pairs, including discrete random rotations (0°, 90°, 180°, or 270°), random brightness and contrast fine-tuning (with an amplitude of ±0.1), and random horizontal or vertical flipping, to improve the model's generalization ability. Data conversion involves converting the enhanced image into tensor format and normalizing the pixel values ​​to the [-1,1] interval for model input; S42. Train the input model with the preprocessed images. Training is performed in batches, with 16 pairs of images input per batch. The dataset is divided into multiple epochs, and each complete epoch requires traversing 453 batches of data. S43. In the training loop, in order to solve the problem that the discriminator may converge too quickly and inhibit the learning of the generator, a differentiated training frequency is adopted: the generator is trained once for each batch, but the discriminator is trained only once for every 3 batches, so as to stabilize the dynamic balance of adversarial training. S44. During training, a performance evaluation is performed on a separate validation set after every 5 training epochs. This validation set contains 20 pairs of images. Figure 9 and Figure 10 As shown, the evaluation calculates the average of three metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and the crucial Pathology Structure Fidelity (PSF). Early termination of training is determined by the PSF metric: if the PSF improvement does not exceed 0.001 within 50 consecutive validation cycles, an early termination mechanism is triggered. S45. Training ended at 816 epochs due to early stopping. Reviewing the validation logs throughout the training process, the model reached its optimal performance at 565 epochs, with the following evaluation metrics: PSNR 25.27 dB, SSIM 0.6879, PSF 0.7848. This model was saved as the final optimal model, and three images were randomly selected from the training process for validation. The validation images are as follows: Figure 11 As shown.

[0033] The present invention has the following beneficial effects: This invention presents a method for generating super-resolution pathological images that integrates residual and attention mechanisms. By constructing a residual-attention generation network, residual learning and attention mechanisms are deeply integrated. The residual structure effectively alleviates the gradient vanishing and information decay problems in deep network training, ensuring the stability and efficiency of network training. Meanwhile, the attention mechanism enables the network to autonomously focus on key cell morphology, tissue boundaries, and other fine structural regions in pathological images, significantly enhancing the model's ability to extract and reconstruct key pathological features, thereby generating super-resolution images with richer details, clearer textures, and higher visual quality. By employing adversarial training between a residual-attention generation network and a discriminator network, the discriminator network continuously drives the residual-attention generation network to output reconstruction results that are visually indistinguishable from real high-resolution pathological images. This not only improves accuracy at the pixel level but also effectively suppresses artifacts and blurring at the perceptual level, generating higher-quality images that are more realistic and in line with pathological observation habits, providing a reliable foundation for subsequent quantitative analysis and visual diagnosis.

[0034] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating super-resolution pathological images by integrating residual and attention mechanisms, characterized in that, Including: S1. Collect the digital images of pathological sections, and extract the low-resolution and high-resolution image pairs of the same field of view of the digital images; Preprocess the low-resolution and high-resolution image pairs, and divide the preprocessed low-resolution and high-resolution image pairs into a training set and a validation set according to a specified quantity; S2. Construct a residual-attention generation network as a generator network for generating a super-resolution reconstructed image according to the low-resolution image; S3. Construct a discriminator network for obtaining the confidence representing the authenticity of the image according to the super-resolution reconstructed image and the high-resolution image, so as to determine the authenticity of the super-resolution reconstructed image and the high-resolution image; S4. According to the training set, perform adversarial training on the residual-attention generation network and the discriminator network, and adjust the parameters through the validation set to obtain a trained generator network; S5. Input the low-resolution pathological image to be processed into the trained generator network to obtain a super-resolution reconstructed image with clear pathological structure.

2. The method for generating super-resolution pathological images by fusing residual and attention mechanisms according to claim 1, characterized in that, The residual-attention generation network includes an initial feature extraction module, a deep feature extraction module, a feature fusion module, and an upsampling reconstruction module; Among them, the initial feature extraction module is used to perform a convolution operation on the input low-resolution image to obtain shallow features and output them to the deep feature extraction module; The deep feature extraction module is used to perform initial deep feature extraction and channel weight recalibration on the shallow features to obtain deep features and output them to the feature fusion module; The feature fusion module is used to fuse the deep features and the shallow features through a residual connection to obtain fused features; The upsampling reconstruction module performs at least one upsampling operation on the fused features, and performs feature correction through a convolutional layer after each upsampling, and finally outputs a super-resolution reconstructed image.

3. The method for generating super-resolution pathological images by fusing residual and attention mechanisms according to claim 2, characterized in that, The deep feature extraction module includes multiple stacked RRDB-CA sub-modules. Among them, each RRDB-CA sub-module includes three cascaded residual density blocks, a channel attention block, and a cross-module residual connection; The residual density block includes k sequentially connected density convolutional layers and a local residual connection layer; among them, the input of the first density convolutional layer is the shallow feature output by the initial feature extraction module; the input of the jth density convolutional layer, 1 < j ≤ k, is composed of the output feature maps of the first to the j-1th density convolutional layers; its output is weighted and fused with the input shallow feature through the local residual connection layer to obtain initial deep features; The channel attention module is used to perform global average pooling on the initial deep features through a compression operation to obtain a channel-level global descriptor; through an excitation operation, a gating mechanism composed of two fully connected layers is used to perform a non-linear transformation on the channel-level global descriptor to generate the normalized weights of each channel; multiply the normalized weights of each channel with the initial deep features channel by channel to obtain channel-weighted features; The cross-module residual connection is used to fuse the channel-weighted features with the shallow features to obtain deep features.

4. The method for generating super-resolution pathological images by fusing residual and attention mechanisms according to claim 1, characterized in that, The execution steps of the discriminator network include: S31. By using basic convolutional layers, features of the super-resolution reconstructed image and the high-resolution image are extracted to obtain primary features; S32. Through multiple sequentially connected downsampling stages, the receptive field of the primary features is gradually expanded and the feature dimensions are increased to obtain downsampled features. S33. The downsampled features are aggregated through a global pooling layer to obtain global features; S34. By using at least one fully connected equivalent layer, global features are integrated to obtain the confidence level that represents the authenticity of the image.

5. The method for generating super-resolution pathological images by fusing residual and attention mechanisms according to claim 1, characterized in that, In the adversarial training of the residual-attention generation network and the discriminator network, the residual-attention generation network is optimized using L1 loss function, perceptual loss function, adversarial loss function, and brightness consistency loss function; the discriminator network is optimized using the following loss function: In the formula, The loss of the discriminator network; Loss to the true image; To generate image loss; The real image loss can be expressed as: In the formula, For binary cross-entropy loss; This is the output of the discriminator; High-resolution images; Labels indicating whether an image is real or not; The loss in generating the image can be expressed as: In the formula, This is the output of the generator; Low-resolution image; Is it a label for the generated image? The binary cross-entropy loss is expressed as follows: In the formula, Input image; For target labels.

Images