Blood leukocyte staining style transfer method based on cyclic generative adversarial network

Abstract

The present invention relates to a blood leukocyte staining style conversion method based on a cyclic generative adversarial network. By introducing an unsupervised segmentation strategy, the present invention solves the expensive problem of obtaining label information of leukocyte images as a semantic condition constraint, and makes full use of existing image data; through the improved multi-scale discriminator, the deep and shallow feature information obtained by the multi-scale downsampling operation is subjected to feature fusion correction through a self-correction module (SCBlock), thereby ensuring the clarity and semantic accuracy of the converted image. Extensive experiments have shown that the converted leukocyte rapid staining image generated by the present invention is not only clear and accurate, but also the leukocyte rapid staining image generated by the present invention can greatly improve the performance of leukocyte image segmentation.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a blood leukocyte staining style conversion method based on a cyclic generative adversarial network. Background Art

[0002] White blood cells are an important type of blood cell in the blood, often also called immune cells. As part of the immune system, it helps the body fight infectious diseases and foreign pathogens. Abnormal white blood cell counts or morphology are essential for diagnosing diseases such as infections, autoimmune diseases, genetic diseases, and cancers that affect blood cells or bone marrow. For example, when the human body suffers from certain diseases, the total number and percentage of different types of white blood cells will change significantly. White blood cell differential count is an important step in routine blood tests. By detecting the number of white blood cells, the percentage of each type of white blood cell in the blood is determined to diagnose the disease. Pathologists widely use image analysis and pattern recognition methods in pathological analysis to study different cells in microscopic images. Therefore, automated microscopic image analysis of white blood cells is essential for clinical treatment and research.

[0003] The most commonly used white blood cell staining techniques are Wright staining and Gissam staining. These two staining techniques have good and stable staining effects, but the staining speed is slow, usually taking more than ten minutes, which cannot meet the needs of large-scale clinical applications, and blood cells often overlap frequently, causing inconvenience for subsequent image processing and analysis. In addition to these two staining techniques, there is another rapid staining method that can shorten the cell staining time to about ten seconds, with a fast staining speed, and will dissolve most of the red blood cells that interfere with diagnosis. This rapid staining method greatly reduces its negative impact on white blood cell positioning by dissolving red blood cells, and the images obtained are more suitable for research on the classification and segmentation of white blood cells. Therefore, by converting the white blood cell images with stable staining into corresponding rapid staining images with high quality, it is beneficial to the study of white blood cell images under rapid staining.

[0004] However, there are two main problems. First, the rapid staining process is unstable, which may lead to staining impurities and background contamination. Second, the rapid staining technology has not been popularized. Therefore, there are not many studies on white blood cell images under rapid staining, and there is also little related image data.

[0005] Under deep learning technology, most of the existing methods that can be used to complete the dyeing style transfer use the idea of ​​cyclic generative adversarial network [1] (CycleGAN). Using two sets of generator and discriminator networks, the mutual conversion between the two domains is learned through adversarial learning. However, due to the lack of semantic conditional constraints, irrelevant image noise is easily generated and the semantic consistency of the white blood cell image before and after conversion cannot be maintained. In medical images, obtaining additional label information as a conditional constraint is often very labor-intensive and resource-intensive. Therefore, how to complete the high-quality conversion of the dyeing style of white blood cell images in the absence of additional label information is an urgent problem to be solved. Summary of the invention

[0006] The purpose of the present invention is to provide a blood leukocyte staining style conversion method based on a cyclic generative adversarial network in order to convert the leukocyte image obtained by the Wright staining method into the corresponding rapid staining image with high quality, which integrates the unsupervised segmentation technology and the optimization method of the multi-scale discriminator (StainGAN).

[0007] To achieve the above object, the technical solution of the present invention is: a blood leukocyte staining style conversion method based on a cyclic generative adversarial network, comprising:

[0008] Introducing unsupervised segmentation strategies to obtain semantic guidance information and achieve white blood cell staining style transfer, thereby overcoming the expensive problem of obtaining label information of white blood cell images as a semantic condition restriction and making full use of existing image data;

[0009] The multi-scale discriminator is improved, and the deep and shallow feature information obtained by the multi-scale downsampling operation is fused and corrected through a self-correction module to ensure the clarity and semantic accuracy of the converted image.

[0010] In one embodiment of the present invention, the method is specifically implemented as follows:

[0011] The image of white blood cells stained with Wright's stain is defined as domain A, and the image of white blood cells stained with rapid stain is defined as domain B. Two semantic generators G are defined. A2B and G B2A They are used to convert the white blood cell images in domain A to domain B and the white blood cell images in domain B to domain A, as well as two multi-scale discriminators D A and D B They are used to identify the authenticity of the generated white blood cell images in domain A and domain B, and output a probability matrix M of size N×N prob ; The real Wright staining image of leukocytes and the real rapid staining image of leukocytes are represented as and The generated leukocyte Wright staining image and leukocyte rapid staining image are represented as and in and Will be mapped back to the image on the original domain, and they are represented as and

[0012] When training the semantic generator G A2B When As input, we learn the mapping from domain A to domain B under the semantic guidance information generated by the image segmentation network, generating Next, G. B2A Under the condition of semantic guidance information, the generated image in domain B will be Mapping back to domain A, that is, generating Train the semantic generator G B2A The process is to reverse the conversion direction described above; training G A2B With G B2A The process can be expressed as:

[0013]

[0014]

[0015] When training a multi-scale discriminator, for D A , and Enter D A For each input image, the multi-scale discriminator outputs the corresponding N×N matrix, where each element of the matrix represents the probability that the corresponding image block is a real image or a fake image. The size of N is determined by the structure of the multi-scale discriminator; D B Same reason.

[0016] In one embodiment of the present invention, the overall loss L of the semantic generator G It is expressed as:

[0017] L G =α·L adv +β·L cyc +γ·L idt (3)

[0018] Among them, α, β and γ are weight parameters, L adv , L cyc and L idt They represent adversarial loss, cycle consistency loss and identity loss respectively.

[0019] In one embodiment of the present invention, the overall loss L of the multi-scale discriminator isD It is expressed as:

[0020]

[0021] Among them, α and β are weight parameters, and Respectively expressed as:

[0022]

[0023]

[0024]

[0025] Among them, M true and M false represents a Boolean matrix whose size is consistent with the size of the prediction matrix output by the multi-scale discriminator, Y i and Represent the true label and predicted value respectively, and i is the element index.

[0026] In one embodiment of the present invention, the process of introducing an unsupervised segmentation strategy and obtaining semantic guidance information is implemented as follows: first, a segmentation network G is defined. For a given image x, the process of generating a segmentation network prediction graph is expressed as follows:

[0027] Prediction=argmax(BN(Conv(G(x)))) (8)

[0028] The prediction map of the segmentation network has a size of 1×H×W, and each value on the prediction map of the segmentation network represents a predicted category. For pixels belonging to the same semantic category, the corresponding area of ​​the original image can be located, and then the mean value on each channel of the corresponding area is calculated to obtain semantic guidance information. The image is then filtered by mean, so that the white blood cells in the semantic guidance information image are morphologically closer to real white blood cells.

[0029] In one embodiment of the present invention, the semantic generator follows an encoder-decoder architecture and is composed of a residual encoder and a Spade residual decoder; the semantic generator uses the generated semantic guidance information as a conditional constraint to generate a high-quality rapid leukocyte staining image.

[0030] In one embodiment of the present invention, the multi-scale discriminator can extract and fuse feature information from receptive fields of different sizes.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention uses deep learning technology to propose a method for achieving leukocyte staining style conversion. The specific innovations are as follows: the existing methods that can be used to complete the staining style conversion lack semantic conditional restrictions, easily generate irrelevant image noise, cannot maintain the semantic consistency of leukocyte images before and after conversion, and it is expensive to obtain additional labels for medical images. The present invention uses an unsupervised learning strategy to obtain semantic guidance information and make full use of existing data sets.

[0033] In addition, the present invention also proposes a multi-scale discriminator structure, which generates a clearer and more accurate rapid staining image of white blood cells by extracting multi-scale information and fusing deep and shallow features. Through experimental verification, the rapid staining image of white blood cells generated by the present invention can greatly improve the performance of white blood cell segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is the overall network model of StainGAN of the present invention.

[0035] Figure 2 This is the process of obtaining semantic guidance information of the present invention.

[0036] Figure 3 This is a network structure diagram of the semantic generator of the present invention.

[0037] Figure 4 This is a network structure diagram of the multi-scale discriminator of the present invention.

[0038] Figure 5 This is the result of the rapid staining image of white blood cells generated by the method proposed in the present invention.

[0039] Figure 6 The white blood cell segmentation results are obtained by combining TR and RT as the training set and validation set. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0041] The present invention provides a blood leukocyte staining style conversion method based on a cyclic generative adversarial network, comprising:

[0042] Introducing unsupervised segmentation strategies to obtain semantic guidance information and achieve white blood cell staining style transfer, thereby overcoming the expensive problem of obtaining label information of white blood cell images as a semantic condition restriction and making full use of existing image data;

[0043] The multi-scale discriminator is improved, and the deep and shallow feature information obtained by the multi-scale downsampling operation is fused and corrected through a self-correction module to ensure the clarity and semantic accuracy of the converted image.

[0044] The following is the specific implementation process of the present invention.

[0045] Figure 1 The overall network structure of the StainGAN model proposed in the present invention is shown. In the present invention, the white blood cell image under Wright staining is defined as domain A, and the white blood cell image under rapid staining is defined as domain B. The present invention realizes high-quality conversion from domain A to domain B. The present invention defines two semantic generators G A2B and G B2A They are used to convert images from domain A to domain B and from domain B to domain A, and two discriminators D A and D B . D A and D B All follow the proposed multi-scale discriminator structure to identify whether the generated white blood cell rapid staining image is real or fake, and output a probability matrix M of size N×N prob The present invention represents the real Wright staining image of leukocytes and the real rapid staining image of leukocytes as and The generated leukocyte Wright staining image and leukocyte rapid staining image are represented as and in and Will be mapped back to the image on the original domain, and they are represented as and

[0046] When training the generator G A2B When As input, the model learns the mapping from domain A to domain B under the guidance of semantics generated by the image segmentation network, generating Next, G. B2A Under the conditional constraints of semantic guidance, the generated image in domain B will be Mapping back to domain A, that is, generating Training G B2A The process of training G is to reverse the conversion direction described above. A2B With G B2A The process can be expressed as:

[0047]

[0048]

[0049] When training the discriminator, D A For example, and Enter D A For each input image, the discriminator outputs a corresponding N×N matrix, where each element of the matrix represents the probability that the corresponding image block is a real image or a fake image, where the size of N is determined by the definition structure of the discriminator.

[0050] This invention uses three loss functions: adversarial loss [1] , cycle consistency loss [1] and identity loss [1] The overall loss of the generator is L G It is expressed as:

[0051] L G =α·L adv +β·L cyc +γ·L idt (3)

[0052] Among them, α, β and γ are weight parameters, and in the present invention, these three parameters are all set to 1.000. adv , L cyc and L idt They represent adversarial loss, cycle consistency loss and identity loss respectively.

[0053] The overall loss L of the discriminator D It is expressed as:

[0054]

[0055] Wherein α and β are weight parameters, which are both set to 1.000 in the present invention. and Respectively expressed as:

[0056]

[0057]

[0058]

[0059] Among them, M true and M false Y represents a Boolean matrix whose size is consistent with the size of the prediction matrix output by the discriminator. iand Represent the true label and predicted value respectively, and i is the element index.

[0060] 1. Generate semantic guidance

[0061] In the present invention, in order to achieve pixel-level conversion accuracy, that is, to maintain the semantic consistency of the leukocyte Wright stain image before and after the conversion, the model needs to obtain semantic guidance information as a conditional restriction of the generator. Ideally, most supervised learning methods are used to obtain the image segmentation result map as semantic guidance information by relying on its good robustness and stability, but this requires additional label data, which is often difficult to obtain for medical images. In addition, in most cases, methods based on deep neural networks rely on a large amount of labeled data to achieve ideal results, which is even more difficult to achieve. In order to get rid of the dependence on labeled data and at the same time achieve relatively ideal results, the present invention adopts an improved method based on back propagation to achieve unsupervised segmentation [2]. The process of this method is as follows Figure 2 As shown:

[0062] Figure 2 The overall process architecture for obtaining semantic guidance information is shown, which mainly consists of two parts: feature extraction and pseudo label generation. Define a segmentation network G. For a given image x, the process of generating a network prediction graph is expressed as:

[0063] Prediction=argmax(BN(Conv(G(x)))) (8)

[0064] Prediction graph of the segmentation network and superpixel segmentation obtained using superpixel segmentation method Figure 1 The present invention adopts the efficient superpixel segmentation algorithm SLIC [3] Specifically, superpixel segmentation clusters pixels with similar colors into multiple clusters. A pixel cluster on the image is called a superpixel. After obtaining the superpixel segmentation map, for a superpixel area, calculate the area on the segmentation network prediction map corresponding to this area, calculate the mode of the network prediction category in the area, and use the predicted category represented by the mode as the pseudo label of the corresponding category of all pixels in the area.

[0065] For the semantic guidance acquisition module, the prediction map size of the segmentation network is 1×H×W, and each value on the prediction map represents the predicted category. For pixels belonging to the same semantic category, the corresponding area of ​​the original image can be located, and then the mean value of each channel of the corresponding area is calculated to obtain the semantic guidance information. Finally, in order to smooth the edge details of the acquired semantic guidance image and solve the anti-aliasing problem, the image is filtered by mean, so that the leukocytes in the semantic guidance image are closer to the real leukocytes in morphology.

[0066] 2. Semantic Generator

[0067] Figure 3 The network structure diagram of the semantic generator is shown. The generator follows the encoder-decoder architecture and consists of two parts: the residual encoder and the Spade residual decoder [4]. The semantic generator uses the generated semantic guidance information as a conditional constraint to generate high-quality white blood cell rapid staining images. Among them, the convolutional layer and the residual layer are used to extract image features, and the Spade residual layer can fuse the semantic guidance information with the feature information through spatial adaptive normalization.

[0068] 3. Multi-scale Discriminator

[0069] In the Generative Adversarial Network (GAN[5]), PatchGAN[6] is a typical discriminator. For an input image, it outputs an N×N matrix, where each value of the matrix represents the probability that the corresponding part of the image is real or fake. PatchGAN simply performs multiple downsampling operations on the input image. The size of the receptive field of each point on the feature map finally obtained depends on the number of downsampling operations. Therefore, the present invention combines feature information from receptive fields of different sizes to perform multi-scale discrimination on the input image. Inspired by the self-correcting convolution[7], the discriminator structure proposed in the present invention is as follows: Figure 4 shown.

[0070] The multi-scale discriminator proposed in the present invention can extract and fuse feature information from receptive fields of different sizes, wherein the SCBlock can fuse deep and shallow features.

[0071] 4. Experimental data

[0072] The data used in the experiment include 98 Wright-stained images of leukocytes with an image size of 256×256, and 200 rapid-stained images of leukocytes with an image size of 120×120.

[0073] 5. Evaluation Metrics

[0074] The experiment uses the mean Intersection over Union (mIoU) (Formula 9), an indicator of image segmentation, as the measurement standard, and indirectly measures the quality of the generated image through the result of image segmentation to improve the performance of the white blood cell segmentation task.

[0075]

[0076] Where p represents the segmented image, k is the number of pixels, and i and j are pixel indices. This experiment uses ResUNet-plus-plus [8] As an image segmentation network, different combinations of training sets and validation sets are used, and experiments are conducted using real white blood cell rapid staining images and converted white blood cell rapid staining images to measure the conversion quality and performance improvement. Table 1 shows the different combinations of training sets and validation sets used in this experiment. RR means using real white blood cell rapid staining images as training sets and validation sets; RT means using real white blood cell rapid staining images as training sets and using generated white blood cell rapid staining images as validation sets; TR means using generated white blood cell rapid staining images as training sets and using real white blood cell rapid staining images as validation sets; TR-R means using generated and real white blood cell rapid staining images as training sets and real white blood cell rapid staining images as validation sets.

[0077] Table 1 Different combinations of training sets and validation sets (RT, TR, RR and TR-R)

[0078]

[0079] When measuring the conversion quality of the generated white blood cell rapid staining images, two points are mainly considered: one is the domain similarity between the real rapid staining images and the generated rapid staining images; the other is the semantic consistency before and after the conversion. Therefore, this experiment uses RT mIoU and TR mIoU to measure the conversion quality. For the performance improvement of the white blood cell segmentation task, this experiment uses RR mIoU as the baseline, and then replaces the training set with the generated white blood cell rapid staining images to give T-RmIoU. In addition, the generated white blood cell rapid staining images are used as a data enhancement method to give TR-R mIoU.

[0080] 6. Qualitative Results

[0081] Figure 5 (The first row is the Wright staining image of the leukocytes, and the second row is the corresponding rapid staining image after conversion) shows the results of the Wright staining image of the leukocytes generated by the method proposed in the present invention. It can be seen that the method proposed in the present invention can convert the Wright staining image of the leukocytes into the corresponding rapid staining image with high quality, and ensure the clarity and accuracy before and after the conversion. Figure 6 The white blood cell segmentation results obtained by combining TR and RT as the training set and the validation set are shown. It can be seen that the method proposed in the present invention can generate high-quality white blood cell rapid staining images and maintain a high degree of semantic consistency.

[0082] 7. Quantitative results

[0083] Table 2 Comparison of different methods in terms of conversion quality and performance improvement on segmentation tasks. The results are given in mIoU (%).

[0084]

[0085] Table 2 shows the comparison of different methods in terms of conversion quality and performance improvement on segmentation tasks. In terms of conversion quality, the method proposed in this paper achieved the best results in both RT mIoU and TR mIoU, and achieved a performance improvement of 6.20%-8.64% in average mIoU. It is proved that the white blood cell rapid staining images generated by the method proposed in this paper have advantages in clarity and accuracy, and maintain a high degree of domain similarity with the real white blood cell rapid staining images and semantic consistency before and after conversion.

[0086] In order to give quantitative improvement results for the performance improvement of the white blood cell segmentation task, the following experiments were designed: 1. Use real white blood cell rapid staining images to train the segmentation network and make predictions; 2. Use the generated white blood cell rapid staining images to train the segmentation network, and use the real white blood cell rapid staining images for predictions; 3. Add the generated white blood cell rapid staining images as data augmentation to the real white blood cell rapid staining images to train the segmentation network, and use the real white blood cell rapid staining images for predictions.

[0087] From the results shown in Table 2, it can be seen that the method proposed in the present invention achieves 11.97% and 24.49% performance improvements in TR mIoU and TR-R mIoU, respectively. It can be seen that the white blood cell fast staining image generated by the method proposed in the present invention has a huge performance improvement for the white blood cell segmentation task.

[0088] References:

[0089] [1]J.-Y.Zhu,T.Park,P.Isola,and A.A.Efros.Unpaired image-to-imagetranslation using cycle-consistent adversarial networks.Proceeding of theIEEE International Conference on Computer Vision.2017;2223–2232.

[0090] [2]A.Kanezaki.Unsupervised image segmentation by backpropagation.2018IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).2018;1543–1547.

[0091] [3]R.Achanta,A.Shaji,K.Smith,A.Lucchi,P.Fua,and S.S¨usstrunk.Slicsuperpixels compared to state-of-the-art superpixel methods.IEEE Transactionson Pattern Analysis and Machine Intelligence.2012;2274–2282.

[0092] [4]T.Park,M.-Y.Liu,T.-C.Wang,and J.-Y.Zhu.Semantic image synthesiswith spatially-adaptive normalization.Proceedings of the IEEE / CVF Conferenceon Computer Vision and Pattern Recognition.2019;2337–2346.

[0093] [5]I.Goodfellow, J.Pouget-Abadie, M.Mirza, B.Xu, D.Warde-Farley, S.Ozair, A.Courville, and Y.Bengio.Generative adversarial nets.Advances in NeuralInformation Processing Systems, 2014; 27.

[0094] [6] P.Isola, J.-Y.Zhu, T.Zhou, and AAEfros.Image-to-image translation with conditional adversarial networks.in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2017; 1125–1134.

[0095] [7] J.-J. Liu, Q. Hou, M.-M. Cheng, C. Wang, and J. Feng. Improvingconvolutional networks with self-calibrated convolutions. Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2020; 96–105.

[0096] [8] D.Jha, PHSmedsrud, MARiegler, D.Johansen, T.De Lange, P.Halvorsen, and HD Johansen. Resunet++: An advanced architecture for medical imagesegmentation. IEEE International Symposium on Multimedia (ISM). 2019; 225–2255.

[0097] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A blood leukocyte staining style transfer method based on a cyclic generative adversarial network, characterized in that: include: Introducing unsupervised segmentation strategies to obtain semantic guidance information and achieve white blood cell staining style conversion; The method of introducing the unsupervised segmentation strategy and obtaining the semantic guidance information process is as follows: first, a segmentation network G is defined. For a given image x, the process of generating a segmentation network prediction graph is expressed as: Prediction=argmax(BN(Conv(G(x)))) The prediction map size of the segmentation network is 1×H×W, and each value on the prediction map of the segmentation network represents the predicted category; For pixels belonging to the same semantic category, the corresponding area of ​​the original image can be located, and then the mean value of each channel of the corresponding area is calculated to obtain semantic guidance information; then the image is filtered by mean, so that the leukocytes in the semantic guidance information image are closer to the real leukocytes in morphology; The semantic generator follows the encoder-decoder architecture and consists of a residual encoder and a Spade residual decoder. The semantic generator uses the generated semantic guidance information as a conditional constraint to generate high-quality white blood cell rapid staining images. The convolutional layer and the residual layer are used to extract image features, and the Spade residual layer fuses the semantic guidance information with the feature information through spatial adaptive normalization. The multi-scale discriminator is improved, and the deep and shallow feature information obtained by the multi-scale downsampling operation is fused and corrected through a self-correction module to ensure the clarity and semantic accuracy of the converted image.

2. The blood leukocyte staining style conversion method based on cyclic generative adversarial network according to claim 1 is characterized in that: The specific implementation of this method is as follows: The image of white blood cells stained with Wright's stain is defined as domain A, and the image of white blood cells stained with rapid stain is defined as domain B. Two semantic generators G are defined. A2B and G B2A They are used to convert the white blood cell images in domain A to domain B and the white blood cell images in domain B to domain A, as well as two multi-scale discriminators D A and G B They are used to identify the authenticity of the generated white blood cell images in domain A and domain B, and output a probability matrix M of size N×N prob ; The real Wright staining image of leukocytes and the real rapid staining image of leukocytes are represented as and The generated leukocyte Wright staining image and leukocyte rapid staining image are represented as and in and Will be mapped back to the image on the original domain, and they are represented as and When training the semantic generator G A2B When As input, we learn the mapping from domain A to domain B under the semantic guidance information generated by the image segmentation network, generating Next, G. B2A Under the condition of semantic guidance information, the generated image in domain B will be Mapping back to domain A, that is, generating Train the semantic generator G B2A The process is to reverse the conversion direction described above; training G A2B With G B2A The process can be expressed as: When training a multi-scale discriminator, for D A , and Enter D A , for each input image, the multi-scale discriminator outputs the corresponding N×N matrix, each element of the matrix represents the probability that the corresponding image block is a real image or a fake image, where the size of N is determined by the structure of the multi-scale discriminator; D B Same reason.

3. The blood leukocyte staining style conversion method based on cyclic generative adversarial network according to claim 2 is characterized in that: The overall loss L of the semantic generator G It is expressed as: L G =α·L adv +β·L cyc +γ·L idt (3) Among them, α, β and γ are weight parameters, L adv , L cyc and L idt They represent adversarial loss, cycle consistency loss and identity loss respectively.

4. The blood leukocyte staining style transfer method based on cyclic generative adversarial network according to claim 2 is characterized in that: The overall loss L of the multi-scale discriminator D It is expressed as: Among them, α and β are weight parameters, and Respectively expressed as: Among them, M true and M false represents a Boolean matrix whose size is consistent with the size of the prediction matrix output by the multi-scale discriminator, Y i and Represent the true label and predicted value respectively, and i is the element index.

5. The blood leukocyte staining style conversion method based on cyclic generative adversarial network according to claim 2 is characterized in that: The multi-scale discriminator can extract and fuse feature information from receptive fields of different sizes.

Images