High-resolution chest X-ray image bone suppression method based on GL-LCM model

Abstract

The invention discloses a high-resolution chest X-ray image bone suppression method based on a GL-LCM model. The method comprises the steps that image data are collected and preprocessed; building a high-resolution chest X-ray image bone suppression network model; respectively and repeatedly training a vector quantization generative adversarial network two-way condition consistency model in the high-resolution chest X-ray image bone suppression network model, and optimizing network parameters; and inputting the preprocessed chest X-ray image into the trained high-resolution chest X-ray image bone suppression network model based on the GL-LCM model, and finally generating a soft tissue image. According to the method, a high-resolution soft tissue image is automatically generated based on an input chest X-ray image.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly relates to a bone suppression method for high-resolution chest X-ray images based on the GL-LCM model. Background Art

[0002] Chest X-Ray (CXR) imaging plays an important role in the diagnosis of lung diseases due to its high accessibility, low cost, and low radiation exposure. However, the overlap of bone structures in CXR images obscures key diagnostic information, leading to difficulties in diagnosis. Currently, the commonly used bone suppression technique in clinics is Dual-Energy Subtraction (DES) imaging, but this technique requires special equipment and increases radiation exposure, limiting its application in resource-limited environments. Therefore, it is of great significance to develop a bone suppression method that can effectively suppress bone structures and retain the details of lung textures.

[0003] Traditional statistically based bone suppression methods for CXR imaging provide an alternative to dedicated DES equipment. However, these methods require precise segmentation and boundary annotation, which hinders their practical application. In contrast, recent advances in deep learning enable models to automatically learn complex image features, providing more accurate bone suppression. Therefore, advanced bone suppression techniques have emerged. For example, Gusarev et al. used an autoencoder (AE) to generate soft tissue images from CXR images. Similarly, Zhou et al. drew inspiration from generative adversarial networks (GANs) and proposed a multi-scale conditional adversarial network (MCA-Net). In addition, Wang et al. introduced an improved U-Net with self-attention function. Despite these efforts, limitations in performance still exist, restricting their application in actual clinical settings.

[0004] Recently, diffusion models have shown powerful capabilities in various generative tasks. Although some studies, such as the method of Chen et al., have successfully applied diffusion models to bone suppression in CXR imaging, there are still challenges in applying these models for fast high-resolution bone suppression considering the key requirements in the clinical environment. In particular, although these methods are end-to-end, they are unable to balance global suppression of bone structures and local detail retention, mainly due to the difficulty of processing global bone residuals and local detail features through a single network. In addition, existing bone suppression diffusion models have high computational requirements and require a large amount of processing time, which is not practical for clinical applications.

[0005] Generally speaking, the current clinical and scientific challenges are as follows:

[0006] From a clinical perspective, it mainly includes: in the soft tissue image obtained by DES, the bones are almost completely suppressed, and appropriate data inclusion or exclusion criteria are formulated.

[0007] From a scientific perspective, it mainly includes: the trained model should have efficient bone suppression ability, be able to accurately remove the bone structure when processing medical images, so as to generate high-quality soft tissue images; the trained model should not introduce or reduce other substances while suppressing bones (that is, just suppress bones); the trained model can well preserve the detail textures, such as blood vessel structure, clarity, etc.; the trained model can remove the motion artifacts generated by the patient's heartbeat, breathing, etc. during the DES shooting. Summary of the Invention

[0008] In view of the deficiencies of the prior art, the present invention proposes a bone suppression method for high-resolution chest X-ray images based on GL-LCM (Global-Local Latent Consistency Model), which automatically generates high-resolution soft tissue images including texture details and spatial features based on the input chest X-ray images.

[0009] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0010] A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model, comprising the following steps:

[0011] S1. Collect image data and preprocess it;

[0012] S1-1. Use a DES subtraction device to collect chest X-ray images of the same patient and the matching soft tissue images;

[0013] Data outside the inclusion criteria will interfere with the prediction effect of the model. The inclusion criteria include: age > 18 years old, no previous chest surgery history or trauma history; use dual-energy photography conditions for posterior-anterior chest X-ray; the photography positioning meets the standard requirements of posterior-anterior chest; the patient's chest cage is normal; the thoracic cavity diagnosis is normal; the emphysema diagnosis is normal.

[0014] S2. Build a bone suppression network model for high-resolution chest X-ray images. The high-resolution chest X-ray image bone suppression network model includes a pre-trained lung region segmentation network based on Dense-U-Net, a vector quantization generative adversarial network, and a dual-path conditional consistency model. Among them, the core of the high-resolution chest X-ray image bone suppression network model is to perform the forward and backward processes of the consistency model in the latent space. The vector quantization generative adversarial network maps the image in the pixel space to the latent space through the encoder and maps the latent variables in the latent space back to the pixel space through the decoder. The forward process of GL-LCM (high-resolution chest X-ray image bone suppression network model) can be defined as:

[0015]

[0016] where ∈ is the noise component sampled from the Gaussian distribution, here, α t = 1 - β t is a differentiable function of the time step t determined by the denoising LCM sampler. The training loss of GL-LCM is expressed as:

[0017]

[0018] where ∈ θ represents the noise predicted by the noise estimation network with parameters θ at time step t, is the given image condition.

[0019] In the backward process of GL-LCM, starting from the random noise , the final result is predicted through a multi-step denoising process:

[0020]

[0021] where T is the total number of sampling time steps, c out (t) and c skip (t) are differentiable functions such that c out (0) = 0 and c skip (0) = 1.

[0022] S3. Use the preprocessed chest X-ray image as both the input and the label for the self-reconstruction task, repeatedly train the vector quantization generative adversarial network, optimize the network parameters, and continuously perform iterative optimization to minimize the difference between the real-value image (label) and the model output image;

[0023] S4. Use the preprocessed chest X-ray image as an input condition, perform noise addition and denoising tasks on the matched soft tissue image obtained by DES, repeatedly train the dual-path conditional consistency model, optimize the network parameters, and continuously perform iterative optimization to minimize the difference between the real noise and the model output noise;

[0024] S5. Use the DES device to collect the chest X-ray image of the patient, preprocess it, and input it into the trained high-resolution chest X-ray image bone suppression network model based on the GL-LCM model. First, pass through the lung region segmentation network, the encoder of the vector quantization generative adversarial network, the dual-path conditional consistency model, and the decoder of the vector quantization generative adversarial network to finally generate the global soft tissue image and the local (lung) soft tissue image. Among them, the vector quantization generative adversarial network maps the image in the pixel space to the latent space through the encoder, and maps the latent variable in the latent space back to the pixel space through the decoder. In the conditional consistency model, the model accepts the splicing of Gaussian noise and the chest X-ray image as input, obtains the latent variable of the predicted soft tissue image after multiple sampling denoising, and decodes to obtain the final global soft tissue image and local soft tissue image;

[0025] S6. Use the Poisson fusion algorithm to fuse the global soft tissue image and the local (lung) soft tissue image to obtain the final soft tissue image output. Among them, Poisson fusion is achieved by solving the following equation:

[0026]

[0027] Among them, R, S l , S g and M(I g ) respectively represent the fusion result, the local path result, the global path result, and the lung region mask calculated by Dense-U-Net based on the CXR image. The boundary condition is defined as R = S g .

[0028] Preferably, in step S1, the preprocessed image is uniformly adjusted to 1024×1024.

[0029] Preferably, the vector quantization generative adversarial network includes an encoder and a decoder. The image in the pixel space is mapped to the latent space through the encoder, and the latent variable in the latent space is mapped back to the pixel space through the decoder.

[0030] Preferably, in step S3, the loss function for model training is:

[0031]

[0032] Among them, is the mean absolute error loss; For quantization loss; For the perceptual loss of a pre-trained Visual Geometry Neural Network; For the adversarial loss on the patch discriminator of a pixel-to-pixel super-resolution model; Is the weight of the mean absolute error loss, λ Qua = 1 is the weight of the quantization loss, λ Per = 0.001 is the weight of the perceptual loss of the pre-trained Visual Geometry Neural Network, λ Adv = 0.01 is the weight of the adversarial loss on the patch discriminator of the pixel-to-pixel super-resolution model.

[0033] Preferably, in step S4, the loss function for model training is the mean square error; the dual-path conditional consistency model includes a global path and a local path, where the global path selects a chest X-ray image as the guiding condition, and the local path is selected from that obtained by a pre-trained Dense-U-Net-based lung region segmentation network.

[0034] Preferably, in step S5, the dual-path conditional consistency model includes a global path and a local path, where the global path selects a chest X-ray image as the guiding condition, and the local path is selected from that obtained by a pre-trained Dense-U-Net-based lung region segmentation network; the dual-path sampling applies local enhancement guidance (LEG) to avoid potential boundary artifacts and detail blurring, and the local enhancement guidance can be expressed as:

[0035]

[0036] where z l，t represents the sample at a given time step t. The terms and refer to the local and global conditions respectively, while α l represents the weight related to the local condition.

[0037] The present invention has the following characteristics and beneficial effects:

[0038] With the above technical solution, the present invention truly realizes the medical requirement of automatically removing bones and generating soft tissues based on chest X-ray images, specifically including: 1. The deboned X-ray chest photos obtained through relevant deep learning algorithms can make the soft tissue structures (such as the heart, lungs, blood vessels, etc.) in the chest X-ray images more clearly visible, which can help diagnose lung lesions overlapping with the rib area, reduce misdiagnosis or missed diagnosis, and effectively enable doctors to more easily observe and evaluate the nature, size and location of some lung diseases (such as lung tumors, pneumonia, nodules), providing assistance and intervention for further diagnosis and treatment. 2. Patients do not need to undergo the currently commonly used high-dose radiation examination equipment DES in clinical practice, and at the same time, they can also avoid the image artifacts caused by respiratory movement and heartbeat of this equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0040] Figure 1 It is a flowchart of the present invention;

[0041] Figure 2 It is the overall architecture of the high-resolution chest X-ray image bone suppression network model based on the GL-LCM model in the embodiment of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0043] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be understood as a limitation to the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first", "second", etc. can explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, the meaning of "a plurality" is two or more.

[0044] The present invention provides a bone suppression method for high - resolution chest X - ray images based on the GL - LCM model, as Figure 1 and Figure 2 shown. The specific operations are as follows:

[0045] S1. Use a dual - energy subtraction device to collect chest X - ray images of the same patient and the matching soft - tissue images, both with a size of 2021×2021. Data outside the inclusion criteria will interfere with the prediction effect of the model. Therefore, according to the inclusion criteria (as follows):

[0046] (1) Age > 18 years old, no previous history of chest surgery or trauma;

[0047] (2) Perform posterior - anterior chest X - ray photography;

[0048] (3) Use dual - energy imaging conditions for posterior - anterior chest radiography;

[0049] (4) There is no situation where the radiographic positioning does not meet the standard requirements of posterior - anterior chest or obvious chest deformities such as S - shaped scoliosis of the spine;

[0050] (5) There is no situation where either side of the chest is diagnosed with pneumothorax, pleural effusion, or hydropneumothorax;

[0051] (6) There is no situation where either side is diagnosed with emphysema.

[0052] Screen the collected paired images according to the inclusion criteria, and perform pre - processing operations such as automatic registration and local adaptive enhancement of the images.

[0053] The dataset used in this embodiment is 741 pairs of anterior - posterior DES chest X - ray images collected from cooperative hospitals. These images were taken by a digital radiography (DR) machine equipped with a dual - exposure DES device (Discovery XR656, GE Healthcare). The images were initially stored in DICOM format with a 14 - bit depth, but for convenience, they were later converted into PNG files. The pixel size of all chest X - ray images is 2021×2021, and the pixel size range is from 0 to 0.1943 millimeters. Paired X - ray films including operation errors, obvious motion artifacts, and visible pleural effusion and pneumothorax have been excluded. The entire dataset is divided into a training set, a validation set, and a test set, with a ratio of 8:1:1. To save memory, all images are adjusted to 1024×1024 pixels. Subsequently, all image pixel values are normalized to [-1,1].

[0054] S2. Build a bone suppression network model for high - resolution chest X - ray images, as Figure 2As shown, the high-resolution chest X-ray image bone suppression network model includes a pre-trained lung region segmentation network based on Dense-U-Net, a vector quantization generative adversarial network, and a dual-path conditional consistency model;

[0055] It should be noted that the high-resolution chest X-ray image bone suppression network model is based on the latent consistency model (LCM). Among them, the core of the latent consistency model (LCM) is to perform the forward and backward processes of the consistency model in the latent space.

[0056] In this embodiment,

[0057] The forward process of the high-resolution chest X-ray image bone suppression network model can be defined as:

[0058]

[0059] where ∈ is the noise component sampled from the Gaussian distribution, α t = 1 - β t is a differentiable function of the time step t determined by the denoising LCM sampler, and z t represents the sample of the latent consistency model at time step t;

[0060] In the reverse process of the high-resolution chest X-ray image bone suppression network model, starting from the random noise and predicting the final result through a multi-step denoising process:

[0061]

[0062] where T is the total number of sampling time steps, and c out (t) and c skip (t) are differentiable functions such that c out (0) = 0 and c skip (0) = 1.

[0063] Furthermore, the training loss of GL-LCM is expressed as:

[0064]

[0065] where ∈ θ represents the noise predicted by the noise estimation network with parameters θ at time step t, is the given image condition.

[0066] S3. During the training phase of the vector quantization generative adversarial network, the preprocessed chest X-ray images are used as both the input and the label for the self-reconstruction task, and the vector quantization generative adversarial network is repeatedly trained to optimize the network parameters. Iterative optimization is continuously performed to minimize the difference between the real value image (label) and the model output image. The loss function is as follows:

[0067]

[0068] where is the mean absolute error loss; is the quantization loss; is the perceptual loss of the pre-trained Visual Geometry Group neural network; is the adversarial loss on the patch discriminator of the pixel-to-pixel super-resolution model; is the weight of the mean absolute error loss, λ Qua = 1 is the weight of the quantization loss, λ Per = 0.001 is the weight of the perceptual loss of the pre-trained Visual Geometry Group neural network, λ Adv = 0.01 is the weight of the adversarial loss on the patch discriminator of the pixel-to-pixel super-resolution model.

[0069] Furthermore, the vector quantization generative adversarial network includes an encoder and a decoder. The encoder maps the image in the pixel space to the latent space, and the decoder maps the latent variable in the latent space back to the pixel space.

[0070] S4. Using the preprocessed chest X-ray image as the input condition, a noise addition and denoising task is performed on the matched soft tissue image obtained by DES, and the dual-path conditional consistency model is repeatedly trained to optimize the network parameters. Iterative optimization is continuously performed to minimize the difference between the real noise and the model output noise. Among them, the loss function for model training is the mean square error; the dual-path conditional consistency model includes a global path and a local path. The global path selects the chest X-ray image as the guiding condition, and the local path is selected from the pre-trained lung region segmentation network based on Dense-U-Net.

[0071] S5. In the model inference phase, the chest X-ray image of the patient collected by the DES device is preprocessed and then input into the trained high-resolution chest X-ray image bone suppression network model based on the GL-LCM model. It passes through the lung region segmentation network, the encoder of the vector quantization generative adversarial network, the dual-path conditional consistency model, and the decoder of the vector quantization generative adversarial network in sequence, and finally generates the global soft tissue image and the local (lung) soft tissue image.

[0072] Among them, the lung region segmentation network mainly includes a U-Net with dense connections; the vector quantization generative adversarial network mainly includes an encoder and a decoder, which map the image in the pixel space to the latent space through the encoder and map the latent variables in the latent space back to the pixel space through the decoder; in the conditional consistency model, the model accepts the concatenation of Gaussian noise and the chest X-ray image as input and obtains the latent variables of the predicted soft tissue image after multiple sampling denoising; among them, the dual-path conditional consistency model includes a global path and a local path, where the global path selects the chest X-ray image as the guiding condition and the local path selects the one obtained by the pre-trained Dense-U-Net-based lung region segmentation network; the dual-path sampling applies local enhancement guidance (LEG) to avoid potential boundary artifacts and detail blurring, and the local enhancement guidance can be expressed as:

[0073]

[0074] where z l，t represents the sample at a given time step t. The terms and refer to the local and global conditions respectively, while α l represents the weight related to the local condition.

[0075] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, without departing from the principle and spirit of the present invention, various changes, modifications, substitutions, and variations to these embodiments including components still fall within the protection scope of the present invention.

Claims

1. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model, characterized in that, It includes the following steps: S1. Collect image data and preprocess it. The collected images include the chest X-ray image of the same patient and the matching soft tissue image; S2. Build a high-resolution chest X-ray image bone suppression network model. The high-resolution chest X-ray image bone suppression network model includes a pre-trained lung region segmentation network based on Dense-U-Net, a dual-path conditional consistency model, and a vector quantization generative adversarial network; S3. Use the preprocessed chest X-ray image as both the input and the label for the self-reconstruction task, and repeatedly train the vector quantization generative adversarial network to optimize the network parameters; S4. Use the preprocessed chest X-ray image as the input condition, and perform the noise addition and denoising tasks on the matching soft tissue image, and repeatedly train the dual-path conditional consistency model to optimize the grid parameters; S5. Input the preprocessed chest X-ray image of the patient into the trained high-resolution chest X-ray image bone suppression network model, and finally generate the global soft tissue image and the local soft tissue image; S6. Use the Poisson fusion algorithm to fuse the global soft tissue image and the local soft tissue image to obtain the final soft tissue image output.

2. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that, In step S1, inclusion criteria are set when collecting images. The inclusion criteria include: age > 18 years old, no previous chest surgery history or trauma history; use dual-energy imaging conditions for posterior-anterior chest X-ray; the imaging position meets the standard requirements for posterior-anterior chest; the patient's chest cage is normal; the thoracic cavity diagnosis is normal; the emphysema diagnosis is normal.

3. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 2, characterized in that, In step S1, the preprocessing method is: remove the images outside the inclusion criteria, and uniformly adjust the images meeting the inclusion criteria to 1024×1024.

4. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 2, characterized in that, The core of the high-resolution chest X-ray image bone suppression network model is to perform the forward and backward processes of the consistency model in the latent space. The forward process of the high-resolution chest X-ray image bone suppression network model can be defined as: where ∈ is the noise component sampled from a Gaussian distribution, α t = 1 - β t is a differentiable function of the time step t determined by the denoising LCM sampler, and z t represents the sample of the latent consistency model at time step t; During the reverse process of the bone suppression network model for high-resolution chest X-ray images, starting from random noise and predicting the final result through a multi-step denoising process: where T is the total number of sampling time steps, c out (t) and c akip (t) are differentiable functions such that c out (0) = 0 and c skip (0) = 1.

5. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that, The vector quantization generative adversarial network includes an encoder and a decoder. The image in the pixel space is mapped to the latent space through the encoder, and the latent variable in the latent space is mapped back to the pixel space through the decoder.

6. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 4, characterized in that The training loss of the high-resolution chest X-ray image bone suppression network model is expressed as: where, ∈ θ denotes the noise predicted by the noise estimation network with parameter θ at time step t, is the given image condition, T is the total number of sampling time steps, z t represents the sample of the latent consistency model at time step t.

7. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that, In step S3, when training the vector quantization generative adversarial network, iterative optimization is continuously performed to minimize the difference between the real value image and the image output by the vector quantization generative adversarial network. The loss function is: Among them, is the mean absolute error loss; is the quantization loss; is the perceptual loss of the pre-trained Visual Geometry Network; is the adversarial loss on the patch discriminator of the pixel-to-pixel based super-resolution model; is the weight of the mean absolute error loss, λ Qua is the weight of the quantization loss, λ Per is the weight of the perceptual loss of the pre-trained Visual Geometry Network, λ Adv is the weight of the adversarial loss on the patch discriminator of the pixel-to-pixel based super-resolution model.

8. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that, In step S4, when training the dual-path conditional consistency model, iterative optimization is continuously performed to minimize the difference between the real noise and the noise output by the dual-path conditional consistency model, and the mean square error is used as the loss function.

9. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that The dual-path conditional consistency model includes a global path and a local path. Among them, the global path selects the chest X-ray image through channel splicing as the guiding condition to guide the generation process, and the local path selects the lung region obtained by the pre-trained lung region segmentation network based on Dense-U-Net through channel splicing as the guiding condition to guide the generation process.

10. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that In the step S5, a global soft tissue image and a local soft tissue image are finally generated through a lung region segmentation network, an encoder of a vector quantization generative adversarial network, a dual-path conditional consistency model, and a decoder of the vector quantization generative adversarial network. Among them, the vector quantization generative adversarial network maps the image in the pixel space to the latent space through the encoder, and maps the latent variable in the latent space back to the pixel space through the decoder. In the conditional consistency model, the model accepts the concatenation of Gaussian noise and a chest X-ray image as input, and obtains the latent variable of the predicted soft tissue image after multiple sampling and denoising.

11. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that, In the step S5, the dual-path conditional consistency model applies local enhancement guidance for dual-path sampling to avoid potential boundary artifacts and detail blurring. The local enhancement guidance can be expressed as: Among them, z l，t represents the sample at a given time step t, and respectively refer to local and global conditions, while α l represents the weight related to the local condition.

12. A bone suppression method for high-resolution chest X-ray images based on the GL-LCM model according to claim 1, characterized in that, In the step S6, Poisson fusion is achieved by solving the following equation: Among them, R and S l , S g and M(I g ) represent the fusion result, the local path result, the global path result, and the lung region mask calculated by Dense-U-Net from the CXR image respectively. The boundary condition is defined as R = S g .

Images