Pyramid condition normalization stream method for low-quality eye fundus image enhancement

By utilizing the Pyramid Conditional Normalized Flow (PCFlow) model and retinal structural constraints, the characteristic frequency components of the retina are identified, which solves the problem of existing fundus image enhancement algorithms ignoring lesion areas and achieves simultaneous improvement in visual quality and clinical information.

CN119027332AActive Publication Date: 2024-11-26ANHUI MEDICAL UNIV

Patent Information

Application Number
CN202411063588.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-26
Estimated Expiration
2044-08-05

AI Technical Summary

Technical Problem

While existing fundus image enhancement algorithms improve the visual quality of images, they may overlook lesion areas that are important for clinical applications, leading to reduced diagnostic accuracy. Furthermore, traditional methods may result in the loss of detailed information, while deep learning-based methods may produce artifacts.

Method used

The Pyramid Conditional Normalized Flow (PCFlow) model is adopted, using retinal structure as a conditional factor. Through a reversible coupling layer, various frequency components of retinal features are identified, and combined with retinal anatomy and pathological details, an image enhancement network model is constructed.

Benefits of technology

It performs well in terms of visual quality and clinical information enhancement, with improvements in both quantitative indicators and visual quality. It can better preserve retinal anatomy and pathological features, thereby improving diagnostic accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027332A_ABST
    Figure CN119027332A_ABST
Patent Text Reader

Abstract

The invention discloses a pyramid condition normalization streaming method for low-quality eye fundus image enhancement, and the method comprises the steps: S1, preparing a training data set, and randomly selecting 10, 10, 11, 12, 13, 14, 16, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, corresponding low-quality fundus images are generated by using image quality degradation methods such as light transmission interference, image blurring, retina artifacts and the like to form an image pair to serve as a training data set; s2, establishing a network model, mapping a high-quality eye fundus image to a hidden variable and combining the hidden variable with a low-quality eye fundus image as input, decoupling different frequency components of the eye fundus image by a reversible pyramid flow module, extracting an eye fundus structure by a retina structure condition module, and outputting the extracted eye fundus structure; the extracted fundus structure is used as a condition to guide reversible pyramid flow module training; and S3, training a network model, inputting the collected low-quality retina image into a constructed network method, loading pre-trained network parameters, and carrying out image enhancement processing to obtain an enhanced image.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and in particular to an enhancement network model (PCFlow) applied to low-quality fundus images. Background Art

[0002] Retinal imaging enables visualization of fundus capillaries, arteries, and veins, which makes them essential in the diagnosis and assessment of a range of ophthalmic pathologies. However, due to complex lighting conditions, dust, and blinking movements during fundus imaging, the quality of captured retinal images is easily corrupted by over- and underexposure, uneven light distribution, and low contrast, leading to serious misdiagnosis of eye diseases. Therefore, image enhancement technology for low-quality fundus images is of great research significance.

[0003] At present, fundus image enhancement algorithms can be divided into two categories: traditional fundus image enhancement algorithms and fundus image enhancement algorithms based on deep learning.

[0004] 1. Traditional fundus image enhancement algorithms mainly include methods based on histogram equalization and retinex theory. The enhancement methods based on histogram equalization are simple and efficient, and can effectively enhance the contrast of fundus images. However, they rely heavily on the global histogram of the image, which may lead to the loss of detailed information. The enhancement methods based on retinex theory generally include SSR (single-scale Retinex), MSR (multi-scale Retinex) and other improved algorithms based on Retinex theory. This type of method decomposes the image into a brightness map and a reflectance map, where the reflectance map reflects the intrinsic properties of the image and can be used to generate the final enhanced image. However, such methods focus on image brightness rather than features that are conducive to clinical disease diagnosis, such as fundus structure and detail features, which may lead to undesirable color and structural distortion.

[0005] 2. Image enhancement algorithms based on deep learning have been widely used in image processing tasks such as image deblurring, image denoising and image enhancement. At present, methods based on deep learning have also been widely used in the field of fundus image enhancement. For example, a bidirectional generative adversarial network with structural and illumination constraints (StillGAN) treats low-quality and high-quality images as two different domains, and introduces local structure and illumination constraints to learn overall features and local details. However, this method only focuses on generating realistic images, while ignoring the lesion areas that are important for clinical applications. To address this problem, clinical-oriented fundus enhancement networks, importance-guided semi-supervised learning frameworks, Transformer-based adversarial networks, and end-to-end optimized teacher-student frameworks have subsequently appeared. These methods have made significant progress compared to traditional methods. However, most of these methods focus mainly on improving the visual quality of images and aim to learn deterministic mappings from pixel to pixel, which may produce artifacts and reduce the accuracy of diagnosis. Summary of the invention

[0006] In view of the defects of the above-mentioned existing methods, the present invention proposes a pyramid conditional normalized flow model, which uses the retinal structure as a conditional factor to constrain model optimization, and adopts a pyramid structure to construct a reversible coupling layer to identify various frequency components of retinal characteristics while retaining the retinal anatomical structure and pathological details.

[0007] In order to solve the shortcomings of the prior art, the present invention proposes an image enhancement network model that learns the complex relationship between low-quality fundus images and their corresponding high-quality versions. The technical solution adopted is:

[0008] Step S1: prepare a training dataset, randomly select 10,000 high-quality images with a “Good” label from the EyeQ dataset as label images, and use image quality degradation methods such as light transmission interference, image blur, and retinal artifacts to generate corresponding low-quality fundus images as input images;

[0009] Step S2: Establish a network model, map the high-quality fundus image to a latent variable and combine it with the low-quality fundus image as input, use the high-frequency filter in the retinal structure condition module to extract the retinal structure, and then use it as a condition to guide the training of the reversible pyramid flow module after passing through a network composed of 8 RRDB modules; at the same time, the reversible pyramid flow module decomposes the input image into feature sets of different scales, and the coupling layer in the pyramid flow module fuses multi-scale pyramid features by using multi-scale affine parameters, and uses a reversible 1×1 convolutional layer for feature fusion, and finally activates the standard layer to normalize the activation to obtain the enhanced image;

[0010] Step S3: training the network model, combining the data set obtained in step S1 according to the batch size as input, setting the initial learning rate, learning rate decay rate, number of training, weight initialization method, loss function of the network model, and optimizing the network parameters using the Adam optimizer by minimizing the negative log-likelihood loss function;

[0011] Step S4: Image enhancement: input the collected low-quality fundus images into the constructed network model, load the pre-trained network parameters, perform image enhancement processing, and obtain the enhanced image.

[0012] Beneficial effects:

[0013] (1) This paper proposes a new pyramid conditional flow (PCFlow) model that learns the complex relationship between low-quality fundus images and their corresponding high-quality versions, rather than a deterministic mapping of pixels to pixels. PCFlow can restore images with comfortable visual quality, focusing on enhancing valuable clinical information in the results, and achieving better performance in both quantitative indicators and visual quality.

[0014] (2) The present invention proposes a retinal structure conditional module, which uses the retinal structure as a conditional factor to constrain the optimization of PCFlow, and adopts a pyramid structure to construct a reversible coupling layer to identify the various frequency components of retinal characteristics, while retaining the anatomical structure and pathological characteristics of the retina for clinical use.

[0015] (3) The network model proposed in this paper has reached an advanced level in the task of fundus image enhancement. Compared with other different state-of-the-art methods (Setiawan, DCP, RetinexNet, COFNet, SCRNet, StillGAN, T-MAGE-Net, ARCNet, PCENet, CEFNet) on the DRIVE dataset, REFUGE dataset, and Subtrain-EyeQ dataset, the index values ​​SSIM and PSNR are ranked first. At the same time, compared with other methods (COFNet, SCRNet, RealSR, ACRNet, RFormer, PCENet) on the RETFORM dataset, which includes real low-quality and high-quality paired fundus images, the index values ​​SSIM and PSNR are still ranked first. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of an image enhancement system for low-quality fundus images according to the present invention;

[0017] Figure 2 It is a structure-pyramid conditional flow processing flow chart for low-quality fundus image enhancement of the present invention;

[0018] Figure 3 is a schematic diagram of qualitative comparison results between the method of the present invention and other advanced methods on the DRIVE dataset;

[0019] Figure 4 is a schematic diagram of qualitative comparison results between the method of the present invention and other advanced methods on the REFGUE dataset;

[0020] Figure 5 is a schematic diagram of qualitative comparison results between the method of the present invention and other advanced methods on the eyeQ dataset;

[0021] Figure 6 is a schematic diagram of qualitative comparison results between the method of the present invention and other advanced methods on the RETFORM dataset;

[0022] Figure 7 It is a schematic diagram for intuitively comparing the enhancement effect of the method of the present invention and other advanced methods on the results of blood vessel segmentation. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0024] The present invention proposes a pyramid conditional normalized flow method for low-quality fundus image enhancement. Specifically, the retinal structure is used as a conditional factor to constrain the optimization of PCFlow, and a pyramid structure is used to construct a reversible coupling layer for each frequency component for identifying retinal features, while retaining the retinal anatomical structure and pathological features for clinical use. The structure of the network model is as follows: Figure 1 As shown:

[0025] According to an embodiment of the present invention, the network mainly includes an Invertible Pyramid Coupling Module and a Retinal Structure Conditional Module. The purpose of the Invertible Pyramid Coupling Module is to independently identify each frequency component of the retinal feature. The pyramid layer decomposes the input into different feature sets. The pyramid coupling layer first fuses the multi-scale pyramid features by using multi-scale affine parameters, and then uses a reversible 1×1 convolution layer for further feature fusion. Finally, the normalization layer is activated for normalized activation; the retinal structure conditional module uses a reasonable retinal structure as a conditional factor of the normalization flow to constrain the final image enhancement result, which specifically includes the following steps:

[0026] S1. Prepare training data set;

[0027] A total of 29,032 retinal images were collected from public sources and hospitals, and then 10,000 images with a "Good" label were randomly selected from the EyeQ dataset as a training dataset, and high-quality fundus images were captured. For high-quality fundus images without corresponding low-quality labels, we synthesized degraded images using light transmission interference, image blur, and retinal artifacts.

[0028] S2, build a network model;

[0029] The network of this example is mainly divided into two parts: the reversible pyramid coupling module and the retinal structure condition module.

[0030] (1) Reversible pyramid coupling module: In this example, the present invention first gives an input image I and a decomposition level L. The pyramid layer decomposes the input into different feature sets, as shown in Formula 1. D() and U() are downsampling and upsampling operators, P i Represents the i-th layer pyramid features. After obtaining these decomposed features, they are used as the input of the pyramid coupling layer. The pyramid coupling layer first fuses the multi-scale pyramid features by using multi-scale affine parameters, and then adopts reversible 1×1 convolution for further feature fusion, while the activation normalization layer is used as a normalization layer within the process step to normalize the activations, promoting stable and effective learning through consistent statistical properties between data. The process is shown in Equation 2, where A represents the affine coupling operator, C represents the reversible 1×1 convolution layer, and N is the activation normalization layer.

[0031] P i =D i (P)-U(D i+1 (P) )(1)

[0032] 0=N(C(A(P i))) (2)

[0033] The affine coupling operator is used to decompose the input features along the channel dimension to achieve reversible linear unmapping. One of the decomposed sub-features retains its original properties, while the other undergoes an affine transformation under its control. The transformation can be formally written as Equation (3), where x i0 and x i1 is the segmentation feature, y i0 and i1 is the output of level i, s() and t() are affine parameters. The logarithmic determinant of the Jacobian matrix of equation (4) can be estimated as shown in equation (5), and the affine parameters are estimated by two linear layers. m and n represent the position of s. The pyramid coupling layer is a pyramid flow with K layers stacked, and K=3 is set.

[0034] y i0 =x i0

[0035]

[0036] (2) Retinal structure condition module: In this example, the present invention first extracts the high-frequency components of the fundus image based on a high-pass filter with a Gaussian kernel, and embeds it into the model to extract the retinal structure. The process can be expressed as formula (5). H represents a high-pass filter, G represents a Gaussian filter, and m and σ represent the mean and variance of the Gaussian kernel, respectively. After extracting the retinal structure, it will be used as part of the input of the affine coupling layer in the pyramid coupling module to guide model optimization. The specific process can be expressed as formula (6). p represents the input feature, A represents the affine coupling operator, and θ is a feature converter that provides scaling and shifting for the affine coupling layer during forward and backward propagation. The network consists of 8 RRDB modules:

[0037] H(I)=IG(m,σ) (5)

[0038] f=A(θ(H(I),p)) (6)

[0039] (3) Loss function: In this example, the present invention uses the minimization of the negative log-likelihood loss function. During the back propagation process, the L1 loss and the VGG loss are applied to the enhanced fundus image. The L1 loss is used to enhance the feature details of the enhanced fundus image result, which is defined as formula (7), where x is the real fundus image and g(x) is the generated image:

[0040] Loss l1 =|g(x)-x| (7)

[0041] VGG loss is used to constrain the high-level structural features of the obtained results, defined as formula (8), VGG represents the VGG-16 network, which is pre-trained on the ImageNet dataset, and i represents the number of layers of the selected feature output:

[0042] Loss vgg =|VGG i (g(x))-VGG i (x)| (8)

[0043] The comprehensive total loss function is (9), λ 1 , 2 , 3 is a hyperparameter used to balance these loss functions:

[0044] Loss total =λ 1 (-log(p(x)))+λ 2 Loss l1 +λ 3 Loss vgg (9)

[0045] S3, training network model;

[0046] The model is implemented using the PyTorch framework trained on an NVIDIA RTX 3090Ti GPU. In addition, we use an initial learning rate of 2×10 -5 The Adam optimizer is used to optimize the model. The batch size and patch size are set to 24 and 512×512 respectively. The weight λ 1 , 2 , 3 and λ 4 Set to 1, 0.1, 0.5 and 0.01 respectively.

[0047] S4. Prepare test dataset;

[0048] In this example, the DRIVE dataset, REFUGE test dataset, Subtrain-EyeQ dataset, and RETFORM dataset are used to evaluate the model. The DRIVE dataset contains 40 fundus images with a resolution of 512×512, the REFUGE test dataset contains 400 images with a resolution of 1024×1024, the SubtestEyeQ dataset contains another 500 images from the EyeQ dataset, and the RETFORM dataset consists of 120 pairs of real low-quality and high-quality fundus images. For datasets that only contain high-quality fundus images, the degradation model is used to generate the corresponding degraded low-quality fundus images.

[0049] Result analysis;

[0050] (1) Results on the DRIVE dataset, REFUGE test dataset, and Subtrain-EyeQ dataset. The qualitative comparison results between the method of the present invention and the most advanced methods on the DRIVE dataset, REFUGE test dataset, and Subtrain-EyeQ dataset are shown in Figure 2 , Figure 3 and Figure 4 In the figure, input refers to the original image input, and the quantitative comparison results are shown in Table 1 (the best value of each comparison index is shown in bold). From the qualitative comparison point of view, compared with other advanced methods, the method proposed in the present invention can ensure better image quality, fewer artifacts, clearer vascular details, closer to high-quality fundus images, and also shows advantages in detail enhancement and retaining more complete retinal blood vessels. From the quantitative results, compared with other advanced methods, the present invention achieves the best values ​​in the two comparison indicators of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The leading in these quantitative comparison indicators corresponds to the leading in the qualitative comparison results, which reflects the superiority of this method.

[0051] Table 1 Quantitative comparison of different methods on DRIVE, REFUGE and Subtest-EyeQ datasets

[0052]

[0053] (2) Results on the RETFORM dataset. The qualitative comparison results of the method of the present invention and other current advanced methods on the RETFORM dataset, which includes real low-quality and high-quality paired fundus images, are shown in Figure 5 The quantitative comparison results are shown in Table 2 (the best value of each comparison index is shown in bold). From the qualitative comparison results, the method of the present invention can restore finer and clearer details of the retinal structure, which is beneficial for clinical diagnosis. From the quantitative comparison results, the method of the present invention achieves the best performance on the RETFORM dataset, and obtains the best values ​​in the two comparison indicators of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The leading of these quantitative comparison results shows the excellent performance of the method of the present invention in the task of fundus image enhancement.

[0054] Table 2 Quantitative comparison of different methods on the RETFORM dataset

[0055]

[0056] (3) Results in clinical applications. The quantitative comparison results of the method of the present invention and other currently advanced methods in the task of blood vessel segmentation are shown in Table 3 (the best value of each comparison indicator is shown in bold). From the results, the method of the present invention achieves the best values ​​in the three indicators of AUC, accuracy (ACC) and intersection over union (IoU), and can achieve the best effect in the blood vessel segmentation of the image, obtaining finer and more complete blood vessels. The qualitative comparison results of the method of the present invention and other advanced methods in the task of optic disc / cup detection are shown in Figure 6 The quantitative comparison results are shown in Table 4. From the qualitative comparison results, the image enhanced by the method of the present invention is conducive to the detection of the optic disc and the optic cup, the fundus structure is clearer and the artifact area is smaller. From the quantitative comparison results, the method of the present invention achieves the best values ​​in both mIoU and Dice. The leading in these quantitative comparison indicators corresponds to the leading in the qualitative comparison results, which reflects the superiority of the method and proves the effectiveness of the method for clinical observation.

[0057] Table 3 Quantitative comparison of vessel segmentation results of different methods on DRIVE training dataset

[0058]

[0059] Table 4 Quantitative comparison of optic disc / cup detection results of different methods on REFGUE training dataset

[0060]

[0061] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, and it should be clear that the present invention is not limited to the scope of the specific embodiments, for those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A pyramid conditional flow method for low-quality fundus image enhancement. It is characterized in that The steps include: Step S1: Prepare a training dataset, randomly select 10,000 high-quality images with a "Good" label from the EyeQ dataset as label images, and use image quality degradation methods such as light transmission interference, image blur, and retinal artifacts to generate corresponding low-quality fundus images as input images; Step S2: Establish a network model, map the high-quality fundus image to a latent variable and combine it with the low-quality fundus image as input, use the high-frequency filter in the retinal structure condition module to extract the retinal structure, and then use it as a condition to guide the training of the reversible pyramid flow module after passing through a network composed of 8 RRDB modules; at the same time, the reversible pyramid flow module decomposes the input image into feature sets of different scales, and the coupling layer in the pyramid flow module fuses multi-scale pyramid features by using multi-scale affine parameters, and uses a reversible 1×1 convolutional layer for feature fusion, and finally activates the standard layer to normalize the activation to obtain the enhanced image; Step S3: training the network model, combining the data set obtained in step S1 according to the batch size as input, setting the initial learning rate, learning rate decay rate, number of training, weight initialization method, loss function of the network model, and optimizing the network parameters using the Adam optimizer by minimizing the negative log-likelihood loss function; Step S4: Image enhancement: input the collected low-quality fundus images into the constructed network model, load the pre-trained network parameters, perform image enhancement processing, and obtain the enhanced image.

2. The pyramid conditional normalized flow method for low-quality fundus image enhancement according to claim 1, It is characterized in that The network model in step 2 includes a reversible pyramid flow module and a retinal structure condition module; The reversible pyramid flow module aims to independently identify each frequency component of the retinal features. The pyramid layer decomposes the input into different feature sets. The pyramid coupling layer first fuses the multi-scale pyramid features by using multi-scale affine parameters, and then uses a reversible 1×1 convolution layer for further feature fusion. Finally, the activation normalization layer is used to normalize the activation. The retinal structure condition module adopts a reasonable retinal structure as a conditional factor of the normalization flow to constrain the final image enhancement result.

3. The pyramid conditional normalized flow method for enhancing low-quality fundus images according to claim 2, Features: The reversible pyramid coupling module: Given an input image I and a decomposition level L, the pyramid layer decomposes the input into different feature sets, as shown in Equation 1, where D( ) and U( ) are downsampling and upsampling operators, P i Represents the i-th layer pyramid features. After obtaining these decomposed features, they are used as the input of the pyramid coupling layer. The pyramid coupling layer first fuses the multi-scale pyramid features by using multi-scale affine parameters, and then adopts reversible 1×1 convolution for further feature fusion. The activation normalization layer is used as a normalization layer within the process step to normalize the activation and promote stable and effective learning through consistent statistical characteristics between data. The process is shown in Equation 2, where A represents the affine coupling operator, C represents the reversible 1×1 convolution layer, and N is the activation normalization layer. P i =D i (P)-U(D i+1 (P)) (1) O=N(C(A(P i )))(2)。 4. The pyramid conditional normalized flow method for low-quality fundus image enhancement according to claim 3, Features: The affine coupling operator is used to decompose the input features along the channel dimension to achieve reversible linear unmapping, where one of the decomposed sub-features retains its original properties while the other undergoes an affine transformation under its control. The transformation can be formally written as Equation (3), where x i0 and x i1 is the segmentation feature, y i0 and i1 is the output of level i, s() and t() are affine parameters, the logarithmic determinant of the Jacobian matrix of equation (4) can be estimated as shown in equation (5), the affine parameters are estimated by two linear layers, m and n represent the position of s, the pyramid coupling layer is a pyramid flow with K layers stacked, and K=3 is set. and i0 =x i0 5. The pyramid conditional normalized flow method for low-quality fundus image enhancement according to claim 2, Features: The retinal structure condition module: The high-pass filter based on the Gaussian kernel can extract the high-frequency components of the fundus image and embed it into the model to extract the retinal structure. The process can be expressed as formula (5), H represents the high-pass filter, G represents the Gaussian filter, m and σ represent the mean and variance of the Gaussian kernel respectively. After extracting the retinal structure, it will be used as part of the input of the affine coupling layer in the pyramid coupling module to guide the model optimization. The specific process can be expressed as formula (6), p represents the input feature, A represents the affine coupling operator, and θ is a feature converter that provides scaling and shifting for the affine coupling layer during forward and backward propagation. The network consists of 8 RRDB modules: H(I)=IG(m,σ)(5) f=A(θ(H(I),p))(6).

6. The pyramid conditional normalized flow method for low-quality fundus image enhancement according to claim 1, Features: The step 3 optimizes the network parameters by minimizing the negative log-likelihood loss function using the Adam optimizer, and specifically includes: Loss function: Using the minimized negative log-likelihood loss function, during the back propagation process, L1 loss and VGG loss are applied to the enhanced fundus image. L1 loss is used to enhance the feature details of the enhanced fundus image result, which is defined as formula (7), x is the real fundus image, g(x) is the generated image: Loss l1 =|g(x)-x| (7) VGG loss is used to constrain the high-level structural features of the obtained results, defined as formula (8), VGG represents the VGG-16 network, which is pre-trained on the ImageNet dataset, and i represents the number of layers of the selected feature output: Loss vgg =|VGG i (g(x))-VGG i (x)| (8) The comprehensive total loss function is (9), λ 1 , 2 , 3 is a hyperparameter used to balance these loss functions: Loss total =λ 1 (-log(p(x)))+λ 2 Loss l1 +λ 3 Loss vgg (9)。

Citation Information

Patent Citations

  • Structure-texture perception method for weak light image enhancement

    CN112712481A

  • Wide-area fundus camera image fusion method and system and storage medium

    CN115775210A

  • Feature consistency-based eye fundus image enhancement method, system and device, and medium

    CN115829885A

  • CONDITIONAL GENERATIVE ADVERSARIAL NETWORK (cGAN) FOR POSTERIOR SAMPLING AND RELATED METHODS

    US20240169611A1

Cited By

  • Image processing method and device, equipment, storage medium and chip

    CN120852220A

  • Low-dose PET image denoising reconstruction method based on flow model

    CN121639514A

  • Low dose pet image denoising reconstruction method based on flow model

    CN121639514B