Diversified epiphyseal development map generation method and device based on potential diffusion model

Abstract

The invention discloses a diversified epiphyseal development map generation method and device based on a potential diffusion model. The method comprises the following steps: making a data set, covering and collecting an original X-ray film image, preprocessing data, labeling a skeleton maturity level and desensitizing the image; an original image is encoded to a potential space by using a variational auto-encoder, and the image detail generation capability is improved by optimizing a loss function; a fuzzy learning module is introduced in the de-noising stage of the diffusion model, and the diversity of generated skeleton features is enhanced through operations such as batch normalization and random disturbance; a progressive alignment strategy is adopted, skeleton basic structure learning is emphasized in the initial stage, a skeleton grade evaluation network is introduced in the later stage, and skeleton detail features are emphasized; optimizing the generation model by using the skeleton recognition model; and inputting development levels, ages and genders of 14 skeletons, and generating a corresponding left-hand skeleton X-ray film image. According to the method, diversified and high-quality skeletal development images are generated through a potential diffusion model framework in combination with VAE optimization, an FLM module and a progressive alignment strategy, skeletal development features are truly reflected, the problems of detail loss and insufficient diversity in the prior art are solved, the method is particularly excellent in performance when sparse data and long-tail distribution are processed, and the method is suitable for large-scale popularization and application. And the authenticity of the generated image and the practicability of medical diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing and artificial intelligence technology, and in particular to a method and device for generating a diversified epiphyseal development atlas based on a potential diffusion model. Background Art

[0002] The development of the epiphysis is influenced by multiple factors, including genetics, nutrition, endocrine system, and social environment. Therefore, there may be significant differences in overall developmental trends and local details. Traditional bone age assessment methods are usually based on discrete bone development levels, which makes it difficult to capture subtle differences between individuals. Existing technologies have problems such as loss of details and lack of diversity when generating bone images, especially when dealing with sparse data and long-tail distributions. Therefore, developing a method that can generate diverse and high-quality bone images has important clinical and scientific research value.

[0003] The development of the epiphysis is influenced by multiple factors, including genetics, nutrition, endocrine system, and social environment. Therefore, there may be significant differences in overall developmental trends and local details. Traditional bone age assessment methods are usually based on discrete bone development levels, which makes it difficult to capture subtle differences between individuals. Existing technologies have problems such as loss of details and lack of diversity when generating bone images, especially when dealing with sparse data and long-tail distributions. Therefore, developing a method that can generate diverse and high-quality bone images has important clinical and scientific research value. Summary of the Invention

[0004] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provides a method for generating a diversified epiphyseal development map based on a potential diffusion model.

[0005] The present invention solves the technical problem by adopting the following technical solutions:

[0006] A method for generating a diversified epiphyseal development map based on a latent diffusion model comprises the following steps:

[0007] Step 1: Dataset preparation, including collecting original X-ray images, data preprocessing, annotating skeletal maturity grades, and image desensitization;

[0008] Step 2: Use variational autoencoder (VAE) to encode the original image into the latent space, and make the latent representation conform to the Gaussian distribution by minimizing the reconstruction loss and KL divergence loss;

[0009] Step 3: In the denoising stage of the diffusion model, a fuzzy learning module (FLM) is introduced to enhance the diversity of generated bone features;

[0010] Step 4: Through a progressive alignment strategy, the basic structure of the skeleton is learned in the early stage, and the skeleton grade evaluation network is introduced in the later stage to focus on generating detailed features of the skeleton;

[0011] Step 5: Use the skeleton recognition model to optimize the generated model to ensure the structural accuracy and detail quality of the generated image.

[0012] Step 6: Input conditions include the developmental level of 14 bones, age, and gender, and generate an X-ray image of the left hand bones.

[0013] The specific steps of step 1 are as follows:

[0014] (1) Data collection: Data covering skeletal development of children aged 3 to 16 years were collected, and high-quality X-ray images were selected to ensure the integrity and availability of the data. This provided a rich sample base for generating a diverse epiphyseal development atlas.

[0015] (2) Data preprocessing: Clean and organize the X-ray film scoring data, handle missing values ​​and outliers, and perform manual corrections to ensure data consistency and accuracy. This step ensures the quality of data input to the model and lays the foundation for subsequent model training.

[0016] (3) Skeletal maturity grade annotation: Multiple bone age experts grade the 14 reference bones in each hand X-ray, using a voting mechanism to determine the final grade to ensure data reliability and consistency. This collaborative annotation method reduces subjective bias and improves data credibility.

[0017] (4) Image desensitization: The Segment Everything model is used to adjust the image background to black and remove the identity information in the upper right corner to ensure the protection of patient privacy. This step ensures data privacy and compliance and is suitable for actual clinical application scenarios.

[0018] The specific steps of step 2 are as follows:

[0019] (1) Encoder training: VAE is used to encode the original high-resolution X-ray image into a low-dimensional latent space vector, significantly reducing the computational burden of the subsequent diffusion denoising process. This encoding process not only reduces the computational complexity but also preserves the key feature information of the image.

[0020] (2) Loss function optimization: Optimize the loss function of VAE and add perceptual loss and structural similarity loss (SSIMLoss) to improve the ability to generate image details. The optimization process includes the following formula:

[0021] Reconstruction loss: Maintain pixel-level consistency between the input image and the reconstructed image to ensure that the generated image is similar in appearance to the original image. It is calculated as:

[0022]

[0023] KL divergence loss: constrains the distribution of latent space vectors to be close to the standard Gaussian distribution, enhancing the diversity and authenticity of generated images. The calculation method is:

[0024]

[0025] Perceptual loss: Use a pre-trained feature extraction network (such as the VGG network) to extract high-level features of the image and calculate the difference between the input and reconstructed images in the feature space to improve the semantic consistency of the image. The calculation method is:

[0026]

[0027] Structural similarity loss: The structural similarity between local regions is calculated through the SSIM sliding window, emphasizing the preservation of local structure and details. The calculation method is:

[0028]

[0029] The specific steps of step 3 are as follows:

[0030] (1) Introduction of FLM module: The FLM module is introduced in the middle layer of UNet. It fuzzifies the latent vector through fuzzy membership function and noise injection mechanism, enhancing its diversity and preventing model overfitting. This fuzzification can introduce diversity of local details while maintaining the consistency of macro structure.

[0031] (2) The operation steps of the FLM module are as follows:

[0032] Batch normalization: Batch normalization is performed on the feature map to standardize the feature distribution and ensure training stability. The calculation method is:

[0033]

[0034] Introducing random perturbation: Adding Gaussian noise perturbation to the normalized feature map to obtain the fuzzy features after perturbation and enhance the generalization ability of the model. The calculation method is:

[0035]

[0036] Output features: The processed feature map is used as output and passed to the subsequent network layer to complete the denoising task, ensuring the richness and diversity of the generated image in details.

[0037] The specific steps of step 4 are as follows:

[0038] (1) Initial stage: Diffusion reconstruction loss is used to ensure that the generated image can accurately reconstruct the input latent vector. The mean square error (MSE) is used as the main loss function to enable the model to learn the basic structure of the skeleton. The calculation method is:

[0039]

[0040] (2) Later stage: Perceptual alignment loss, using the pre-trained YOLOv8 model to evaluate the maturity level of each bone in the generated image and compare it with the input condition to ensure that the detailed features of the generated image are consistent with the real data. The calculation method is:

[0041]

[0042] The specific steps of step 5 are as follows:

[0043] (1) Skeleton detection: The generated latent vector is input into the skeleton recognition model to detect and identify the developmental levels of 14 major bones, ensuring the structural accuracy and detail quality of the generated image.

[0044] (2) Classification loss calculation: By calculating the classification loss, the generative model is optimized so that the generated image is consistent with the real data at the bone level, thereby improving the accuracy and reliability of the model. The calculation method is:

[0045]

[0046] The specific steps of step 6 are as follows:

[0047] (1) Input condition preparation: Prepare input conditions including the developmental level, age, and gender of the 14 bones;

[0048] (2) Condition encoding: Encoding the input conditions into a format understandable to the model, including numerical representations of bone grade, age, and gender;

[0049] (3) Latent vector generation: The encoded conditional information is used to generate a latent vector as the input of the diffusion model;

[0050] (4) Image generation: Through the denoising process of the diffusion model, the left hand bone X-ray image that meets the input conditions is gradually generated;

[0051] (5) Output image: The generated latent vector is restored to a high-resolution left hand skeletal X-ray image through the VAE decoder to ensure the structural accuracy and detail quality of the generated image.

[0052] The second aspect of the present invention relates to a device for generating a diversified epiphyseal development map based on a potential diffusion model, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the method for generating a diversified epiphyseal development map based on a potential diffusion model of the present invention.

[0053] Compared with the existing technology, the present invention uses the latent diffusion model framework, combined with VAE optimization, FLM module and progressive alignment strategy, to generate diverse and high-quality bone development images, truly reflecting the characteristics of bone development, and solving the problems of detail loss and insufficient diversity in the existing technology. It performs well in processing sparse data and long-tail distribution, improving the authenticity of the generated images and the practicality of medical diagnosis.

[0054] The present invention has the following beneficial effects:

[0055] (1) This method generates a variety of reasonable epiphyseal development images from a single skeletal development level input through a generative model, covering the skeletal development atlas of children aged 3 to 16 years, providing high-quality and diverse data support for bone age assessment;

[0056] (2) It solves the problems of detail loss and insufficient diversity in existing technologies, performs well in the case of sparse data and long-tail distribution, and improves the authenticity of generated images and the practicality of medical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is the overall flow chart of the present invention.

[0058] Figure 2 It is a data preprocessing flow chart of the present invention.

[0059] Figure 3 This is the overall architecture diagram of the multidimensional skeletal development generation model of the present invention.

[0060] Figure 4 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0062] Example 1

[0063] Reference Figure 1 This embodiment relates to a method for detecting illegal wearing of a safety belt for high-altitude work, comprising the following steps:

[0064] Step 1: Dataset preparation, including collecting original X-ray images, data preprocessing, annotating skeletal maturity grades, and image desensitization;

[0065] Step 2: Use variational autoencoder (VAE) to encode the original image into the latent space, and make the latent representation conform to the Gaussian distribution by minimizing the reconstruction loss and KL divergence loss;

[0066] Step 3: In the denoising stage of the diffusion model, a fuzzy learning module (FLM) is introduced to enhance the diversity of generated bone features;

[0067] Step 4: Through a progressive alignment strategy, the basic structure of the skeleton is learned in the early stage, and the skeleton grade evaluation network is introduced in the later stage to focus on generating detailed features of the skeleton;

[0068] Step 5: Use the skeleton recognition model to optimize the generated model to ensure the structural accuracy and detail quality of the generated image.

[0069] Step 6: Input conditions include the developmental level of 14 bones, age, and gender, and generate an X-ray image of the left hand bones.

[0070] Step 1 specifically includes:

[0071] (1) Data collection: More than 100,000 data items were collected from the Zhejiang Province “Campus Charity Tour” project, and 10,052 high-quality X-ray images were screened to ensure the integrity and availability of the data.

[0072] (2) Data preprocessing: Clean and organize the X-ray film scoring data, handle missing values ​​and outliers, and perform manual corrections to ensure data consistency and accuracy.

[0073] (3) Marking of skeletal maturity grade: Multiple bone age experts grade the 14 reference bones in each hand X-ray, and a voting mechanism is used to determine the final grade to ensure the reliability and consistency of the data.

[0074] (4) Image desensitization: The Segment Everything model is used to adjust the image background to black and remove the identity information in the upper right corner to ensure the protection of patient privacy.

[0075] Step 2 specifically includes:

[0076] (1) VAE training: VAE is used to encode the original high-resolution X-ray images into low-dimensional latent space vectors, significantly reducing the computational burden in the subsequent diffusion denoising process.

[0077] (2) Loss function optimization: Perceptual loss and structural similarity loss (SSIM Loss) are added to improve the ability to generate image details. The optimization process includes the following formula:

[0078] Reconstruction loss: Maintain pixel-level consistency between the input image and the reconstructed image to ensure that the generated image is similar in appearance to the original image. It is calculated as:

[0079]

[0080] KL divergence loss: constrains the distribution of latent space vectors to be close to the standard Gaussian distribution, enhancing the diversity and authenticity of generated images. The calculation method is:

[0081]

[0082] Perceptual loss: Use a pre-trained feature extraction network (such as the VGG network) to extract high-level features of the image and calculate the difference between the input and reconstructed images in the feature space to improve the semantic consistency of the image. The calculation method is:

[0083]

[0084] Structural similarity loss: The structural similarity between local regions is calculated through the SSIM sliding window, emphasizing the preservation of local structure and details. The calculation method is:

[0085]

[0086] Step 3 specifically includes:

[0087] (1) Introduction of FLM module: The FLM module is introduced in the middle layer of UNet. It fuzzifies the latent vector through fuzzy membership function and noise injection mechanism, enhancing its diversity and preventing model overfitting. This fuzzification can introduce diversity of local details while maintaining the consistency of macro structure.

[0088] (2) The operation steps of the FLM module are as follows:

[0089] Batch normalization: Batch normalization is performed on the feature map to standardize the feature distribution and ensure training stability. The calculation method is:

[0090]

[0091] Introducing random perturbation: Adding Gaussian noise perturbation to the normalized feature map to obtain the fuzzy features after perturbation and enhance the generalization ability of the model. The calculation method is:

[0092]

[0093] Output features: The processed feature map is used as output and passed to the subsequent network layer to complete the denoising task, ensuring the richness and diversity of the generated image in details.

[0094] The specific steps of step 4 are as follows:

[0095] (1) Initial stage: Diffusion reconstruction loss is used to ensure that the generated image can accurately reconstruct the input latent vector. The mean square error (MSE) is used as the main loss function to enable the model to learn the basic structure of the skeleton. The calculation method is:

[0096]

[0097] (2) Later stage: Perceptual alignment loss, using the pre-trained YOLOv8 model to evaluate the maturity level of each bone in the generated image and compare it with the input condition to ensure that the detailed features of the generated image are consistent with the real data. The calculation method is:

[0098]

[0099] The specific steps of step 5 are as follows:

[0100] (1) Skeleton detection: The generated latent vector is input into the skeleton recognition model to detect and identify the developmental levels of 14 major bones, ensuring the structural accuracy and detail quality of the generated image.

[0101] (2) Classification loss calculation: By calculating the classification loss, the generative model is optimized so that the generated image is consistent with the real data at the bone level, thereby improving the accuracy and reliability of the model. The calculation method is:

[0102]

[0103] The specific steps of step 6 are as follows:

[0104] (1) Input condition preparation: Prepare input conditions including the developmental level, age, and gender of the 14 bones;

[0105] (2) Condition encoding: Encoding the input conditions into a format understandable to the model, including numerical representations of bone grade, age, and gender;

[0106] (3) Latent vector generation: The encoded conditional information is used to generate a latent vector as the input of the diffusion model;

[0107] (4) Image generation: Through the denoising process of the diffusion model, the left hand bone X-ray image that meets the input conditions is gradually generated;

[0108] (5) Output image: The generated latent vector is restored to a high-resolution left hand skeletal X-ray image through the VAE decoder to ensure the structural accuracy and detail quality of the generated image.

[0109] Example 2

[0110] Reference Figure 4 This embodiment relates to a device for generating a diversified epiphyseal development map based on a potential diffusion model, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the method for generating a diversified epiphyseal development map based on a potential diffusion model of Example 1.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating a diversified epiphyseal development map based on a latent diffusion model, characterized by The following steps are involved: Step 1: Dataset preparation, including collecting original X-ray images, data preprocessing, annotating skeletal maturity grades, and image desensitization; Step 2: Use variational autoencoder (VAE) to encode the original image into the latent space, and make the latent representation conform to the Gaussian distribution by minimizing the reconstruction loss and KL divergence loss; Step 3: In the denoising stage of the diffusion model, the fuzzy learning module FLM is introduced to enhance the diversity of generated bone features; Step 4: Through a progressive alignment strategy, the basic structure of the skeleton is learned in the early stage, and the skeleton grade evaluation network is introduced in the later stage to focus on generating detailed features of the skeleton; Step 5: Use the skeleton recognition model to optimize the generated model to ensure the structural accuracy and detail quality of the generated image; Step 6: Input conditions include the developmental level of 14 bones, age, and gender, and generate an X-ray image of the left hand bones.

2. The method for generating a diversified epiphyseal development map based on a potential diffusion model according to claim 1, characterized in that: The specific steps of step 1 are as follows: (11) Data collection: Collect data covering skeletal development information of children aged 3 to 16 years, screen high-quality X-ray images, and ensure the integrity and availability of the data; (12) Data preprocessing: Clean and organize the X-ray film scoring data, handle missing values ​​and outliers, and perform manual corrections to ensure data consistency and accuracy; (13) Marking skeletal maturity grade: Multiple bone age experts grade the 14 reference bones in each hand X-ray, and a voting mechanism is used to determine the final grade to ensure data reliability and consistency; (14) Image desensitization: The Segment Everything model is used to adjust the image background to black and remove the identity information in the upper right corner to protect the patient's privacy; ensuring the privacy and compliance of the data, and being suitable for actual clinical application scenarios.

3. The method for generating a diversified epiphyseal development map based on a potential diffusion model according to claim 1, characterized in that: The specific steps of step 2 are as follows: (21) Encoder training: Encode the original high-resolution X-ray image into a low-dimensional latent space vector through VAE, significantly reducing the computational burden in the subsequent diffusion denoising process; (22) Loss function optimization: Optimize the loss function of VAE, add perceptual loss and structural similarity loss SSIM Loss, and improve the ability to generate image details. The optimization process includes: Reconstruction loss: Maintain pixel-level consistency between the input image and the reconstructed image to ensure that the generated image is similar in appearance to the original image. It is calculated as follows: KL divergence loss: constrains the distribution of latent space vectors to be close to the standard Gaussian distribution, enhancing the diversity and authenticity of generated images. It is calculated as follows: Perceptual loss: Use a pre-trained feature extraction network to extract high-level features of the image and calculate the difference between the input and reconstructed images in the feature space to improve the semantic consistency of the image. The calculation method is as follows: Structural similarity loss: The structural similarity between local regions is calculated through the SSIM sliding window, emphasizing the preservation of local structure and details. The calculation method is as follows:

4. The method for generating a diversified epiphyseal development map based on a potential diffusion model according to claim 1, characterized in that: The specific steps of step 3 are as follows: (31) The FLM module is introduced into the middle layer of UNet to fuzzify the latent vector through fuzzy membership function and noise injection mechanism, thereby enhancing its diversity and preventing model overfitting; (32) The operation steps of the FLM module are as follows: Batch normalization: Batch normalization is performed on the feature map to standardize the feature distribution and ensure training stability. The calculation method is as follows: Introducing random perturbation: Adding Gaussian noise perturbation to the normalized feature map to obtain the perturbed fuzzy features and enhance the generalization ability of the model. The calculation method is as follows: Output features: The processed feature map is used as output and passed to the subsequent network layer to complete the denoising task, ensuring the richness and diversity of the generated image in details.

5. The method for generating a diversified epiphyseal development map based on a potential diffusion model according to claim 1, characterized in that: The specific steps of step 4 are as follows: (41) Initial stage: diffusion reconstruction loss is used to ensure that the generated image can accurately reconstruct the input latent vector. The mean square error (MSE) is used as the main loss function to enable the model to learn the basic structure of the skeleton. The calculation method is as follows: (42) Later stage: Perceptual alignment loss, using the pre-trained YOLOv8 model to evaluate the maturity level of each bone in the generated image and compare it with the input condition to ensure that the detailed features of the generated image are consistent with the real data. The calculation method is as follows:

6. The method for generating a diversified epiphyseal development map based on a potential diffusion model according to claim 1, characterized in that: The specific steps of step 5 are as follows: (51) Skeleton detection: The generated latent vector is input into the skeleton recognition model to detect and identify the developmental levels of 14 major bones, ensuring the structural accuracy and detail quality of the generated image; (52) Classification loss calculation: By calculating the classification loss, the generative model is optimized so that the generated image is consistent with the real data at the bone level, thereby improving the accuracy and reliability of the model. The calculation method is as follows:

7. The method for generating a diversified epiphyseal development map based on a potential diffusion model according to claim 1, characterized in that: The specific steps of step 6 are as follows: (61) Input condition preparation: Prepare input conditions including the developmental level, age, and gender of 14 bones; (62) Condition encoding: Encoding input conditions into a format understandable to the model, including numerical representations of bone grade, age, and gender; (63) Latent vector generation: Generate a latent vector using the encoded conditional information as the input of the diffusion model; (64) Image generation: Through the denoising process of the diffusion model, the left hand bone X-ray image that meets the input conditions is gradually generated; (65) Output image: The generated latent vector is restored to a high-resolution left hand skeletal X-ray image through the VAE decoder to ensure the structural accuracy and detail quality of the generated image.

8. A device for generating a diversified epiphyseal development map based on a potential diffusion model, characterized in that: The method comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the method is used to implement the method for generating a diversified epiphyseal development map based on a potential diffusion model according to any one of claims 1 to 7.

Images