Method for reconstructing chest CT (Computed Tomography) based on orthogonal biplane X-ray film
By constructing feature enhancement blocks in the generator and introducing perceived consistency loss function, the problem of two-dimensional image overlap in X-ray imaging is solved, and high-quality CT images are realized from X-rays to reduce radiation dose and improve diagnostic accuracy.
Patent Information
- Application Number
- CN202510646470.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-20
- Publication Date
- 2025-08-26
AI Technical Summary
The existing X-ray imaging technology has two-dimensional image tissue overlap, which makes the diagnosis difficult. Although CT imaging can solve this problem, the radiation dose is high and the cost is high, which limits its popularity in resource-limited areas. The existing methods of reconstructing CT images from two-dimensional X-rays fail to effectively combine the advantages of X-ray and CT.
Using an orthogonal biplanar X-ray method, the reconstruction process is optimized to generate high-quality CT images by building feature enhancement blocks in the generator and embedding the encoding stage.
It realizes the generation of high-quality CT images under the premise of reducing radiation dose, improves the reliability and accuracy of diagnosis, and provides reliable support for clinical diagnosis.
Smart Images
Figure CN120543676A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing based on artificial intelligence, and in particular to a method for reconstructing chest CT based on orthogonal biplane X-ray films. Background Art
[0002] X-rays and CT are commonly used non-invasive medical imaging techniques in clinical practice. X-rays utilize the penetrating properties of electromagnetic waves to produce two-dimensional projection images of the human body. Their low radiation dose (approximately 0.1 ± 0.01 mSv for a chest exam) and low cost have led to their widespread use in hospitals. However, since the images are two-dimensional, overlapping tissue structures may occur, increasing diagnostic complexity and posing the risk of misdiagnosis. Therefore, X-rays are primarily used to provide preliminary diagnostic evidence.
[0003] In contrast, CT imaging generates high-resolution three-dimensional images by rotating the X-ray source and detector around the patient, acquiring multi-angle two-dimensional projections and using reconstruction algorithms (such as analytical reconstruction or iterative reconstruction). This process effectively solves the problem of tissue overlap in X-ray images and can present more complex anatomical structures, and is therefore considered the "gold standard" for lesion analysis. Although CT imaging provides higher diagnostic value, its radiation dose is relatively high (literature indicates that a single chest scan is approximately 61 times that of an X-ray) and is accompanied by high costs, limiting its popularity in resource-limited areas.
[0004] In recent years, researchers have attempted to reconstruct CT images using only two-dimensional X-rays, aiming to combine the advantages of X-rays and CT, reducing radiation risks while producing high-quality CT images. This approach has attracted widespread attention in the field of medical image generation and reconstruction, and, driven in particular by deep learning technology, a variety of methods for mapping X-rays to CT images have emerged. Summary of the Invention
[0005] To reconstruct CT images from X-rays, this paper provides a method for chest CT reconstruction based on orthogonal biplane X-rays. By constructing a feature enhancement block within the generator and embedding it into the generator's encoding phase, the model's ability to extract key anatomical features is enhanced, enabling more efficient adaptation and completion of reconstruction tasks. Furthermore, by introducing a perceptual consistency loss, the enhanced reconstructed CT images maintain perceptual consistency with the real images in terms of brightness, contrast, and structure, providing more reliable support for clinical diagnosis.
[0006] A method for reconstructing chest CT based on orthogonal biplane X-ray films, the invention includes the following steps:
[0007] Step 1: Preprocess the acquired CT data to mitigate interference caused by data heterogeneity and remove irrelevant noise. Subsequently, orthogonal biplane X-rays are generated using digital radiographic reconstruction techniques. Finally, the paired datasets are appropriately partitioned.
[0008] Step 2: Build a model for reconstructing chest CT from orthogonal biplane X-rays, which consists of two parts: a generator and a discriminator. Because it is different from the same-dimensional reconstruction task (for example, reconstructing MRI from CT), the task of the present invention involves a cross-dimensional conversion from two dimensions to three dimensions. This conversion places higher demands on the model, requiring it to retain key low-level details, minimize information loss, and enhance the ability to extract important anatomical features from X-ray images. To meet these challenges, the present invention adds a feature enhancement block to the generator. Similar to previous CT reconstruction studies, the present invention also introduces a 3D image block discriminator with strong generalization capabilities. The discriminator consists of three Conv3D-BN-ReLU modules, which are used to evaluate the input 3D image blocks to effectively distinguish between real CT images and generated pseudo CT images.
[0009] Step 3: Using the training and test sets constructed in Step 1, we feed the orthogonal biplane X-rays into the chest CT reconstruction model. The generator generates CT images, which are then fed into the discriminator along with the real CT images. We calculate the differences using perceptual consistency loss, projection loss, and generative adversarial loss functions, and optimize the chest CT reconstruction model from Step 2 until the generated CT images are indistinguishable from the discriminator.
[0010] Step 4: Obtain orthogonal biplane X-rays of the patient and input them into the chest CT reconstruction model trained in Step 3. The model performs a 3D reconstruction of the chest anatomy based on the input X-ray images and generates the corresponding CT images.
[0011] A further improvement of the present invention is that in step 1, the acquired raw CT data is preprocessed (e.g., image size unification, irrelevant noise removal, etc.). Subsequently, digital radiographic reconstruction techniques are used to generate orthogonal biplane X-ray images in the anteroposterior and lateral directions, thereby constructing a registered X-CT dataset.
[0012] A further improvement of the present invention is that in step 2, in order to minimize information loss and enhance the ability to extract key anatomical features from X-ray images, a feature enhancement block is added to the generator, thereby effectively improving the reconstruction quality of the CT image.
[0013] The present invention is further improved in that: the specific design of the feature enhancement block in step 2 is that in the Sth encoding stage of the generator, the two-dimensional feature map (denoted as, ) is first downsampled through the convolutional layer. Then, is passed into the dense block, where dense connections and feature reuse mechanisms are used to generate output feature maps This design ensures the effective preservation of low-level features and their propagation to higher levels.
[0014]
[0015] in Represents the output feature map of the Nth dense layer in the Sth encoding stage, and Concat represents the connection of the output feature maps of all dense layers along the channel dimension.
[0016] Then, in order to emphasize important features and suppress minor features, global average pooling and global maximum pooling are first used to transform Converted into two eigenvectors, denoted as v avg and v max These vectors are then processed through a shared fully connected layer, summed, and passed into a Sigmoid activation function to obtain the final channel attention weight W ch :
[0017]
[0018] W ch =σ(FC(v avg )+FC(v max ))(11)
[0019] Where σ represents the sigmoid activation function and the fully connected layer is represented by FC.
[0020] The calculated channel attention weight W ch With the input feature map Multiply element-wise to generate optimized feature maps
[0021]
[0022] Similarly, for Extracting spatial attention weights W sp and apply it to Concentrate features on key areas within the image to generate optimized feature maps
[0023]
[0024] Finally, yes Perform convolution layer compression processing to reduce the number of output channels to half the number of input feature map channels, thereby generating the final output feature map
[0025]
[0026] A further improvement of the present invention is that in step three, based on the projection loss and the generative adversarial loss, in order to enhance the perceptual consistency of the reconstructed CT image in terms of brightness, contrast and structure, the present invention creatively proposes a perceptual consistency loss function.
[0027]
[0028] Where N represents the total number of CT slices, α is the weighting coefficient, and M represents the maximum number of scales for multiscale analysis. represents the contrast and structural similarity of the i-th slice of the p-th anatomical plane compared with its corresponding position in the real CT at the k-th scale. Indicates the brightness similarity at the largest scale. and They correspond to the i-th slice of the p-th anatomical plane in the real CT and the reconstructed CT, respectively.
[0029] The present invention has the following benefits: By constructing a feature enhancement block within the generator and embedding it into the generator's encoding phase, the model's ability to extract key anatomical features is enhanced, allowing it to more effectively adapt to and complete reconstruction tasks. Furthermore, by introducing a perceptual consistency loss, the reconstructed CT images maintain perceptual consistency with the real images in terms of brightness, contrast, and structure, providing more reliable support for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 It is a schematic diagram of the process of the present invention.
[0031] Figure 2 This is a structural diagram of the chest CT reconstruction model.
[0032] Figure 3 Figure 2 is the structural diagram of the feature enhancement block in the generator.
[0033] Figure 4 This is the rendering of the reconstructed CT image achieved by the chest CT reconstruction model using orthogonal biplane X-ray films. DETAILED DESCRIPTION
[0034] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.
[0035] The process of the present invention is as follows Figure 1 As shown, this embodiment provides a method for reconstructing chest CT based on orthogonal biplane X-ray films, including the following steps:
[0036] Step 1: The present invention uses the public lung nodule dataset LIDC-IDRI, which contains 1018 chest CT image data. In order to ensure the stability and accuracy of subsequent model training, the original chest CT data is first standardized and preprocessed to reduce the heterogeneity problem caused by differences in image sources, scanning parameters or reconstruction algorithms. The preprocessing process includes image resampling, normalization, and threshold processing to remove irrelevant structures and background interference. Subsequently, digital reconstruction radiographic imaging technology is used to generate orthogonal biplane X-rays (anteroposterior and lateral) from three-dimensional CT volume data along different perspectives. Finally, the generated X-ray images are paired with the corresponding CT slices to construct a high-quality image pair dataset. In the data partitioning stage, the training set and test set are strictly divided scientifically and randomly to ensure the independence of the data and avoid inaccurate model evaluation due to sample leakage.
[0037] Step 2: When constructing a model for reconstructing chest CT, in order to preserve key low-level details and minimize information loss during the cross-dimensional conversion process from two-dimensional to three-dimensional, the present invention proposes a feature enhancement block and integrates it into the encoding stage of the generator. It assigns attention weights to the features extracted by each dense block in the channel and spatial dimensions, thereby further enhancing the feature representation and improving the quality of the reconstructed CT image. The structure of the chest CT reconstruction model is shown in the attached manual. Figure 2 shown.
[0038] The present invention innovatively proposes a feature enhancement block module, as shown in the attached specification. Figure 3 Specifically, in the Sth encoding stage of the generator, the two-dimensional feature map (denoted as, ) is first downsampled through the convolutional layer. Then, is passed into the dense block, where dense connections and feature reuse mechanisms are used to generate output feature maps This design ensures the effective preservation of low-level features and their propagation to higher levels.
[0039]
[0040] in Represents the output feature map of the Nth dense layer in the Sth encoding stage, and Concat represents the connection of the output feature maps of all dense layers along the channel dimension.
[0041] Then, in order to emphasize important features and suppress minor features, global average pooling and global maximum pooling are first used to transform Converted into two eigenvectors, denoted as v avg and v max These vectors are then processed through a shared fully connected layer, summed, and passed into a Sigmoid activation function to obtain the final channel attention weight Wch :
[0042]
[0043] W ch =σ(FC(v avg )+FC(v max ))(19)
[0044] Where σ represents the sigmoid activation function and the fully connected layer is represented by FC.
[0045] The calculated channel attention weight W ch With the input feature map Multiply element-wise to generate optimized feature maps
[0046]
[0047] Similarly, for Extracting spatial attention weights W sp and apply it to Concentrate features on key areas within the image to generate optimized feature maps
[0048]
[0049] Finally, yes Perform convolution layer compression processing to reduce the number of output channels to half the number of input feature map channels, thereby generating the final output feature map
[0050]
[0051] Step 3: In the process of optimizing the chest CT reconstruction model, based on the projection loss and the generative adversarial loss, in order to enhance the perceptual consistency of the reconstructed CT image in terms of brightness, contrast and structure, the present invention creatively proposes a perceptual consistency loss function. This loss function combines multi-resolution scales, captures features slice by slice and direction by direction, and applies a Gaussian filter. To smooth the pixel-level error. The perceptual consistency loss function is defined as follows:
[0052]
[0053] Where N represents the total number of CT slices, α is the weighting coefficient, and M represents the maximum number of scales for multiscale analysis. represents the contrast and structural similarity of the i-th slice of the p-th anatomical plane compared with its corresponding position in the real CT at the k-th scale. Indicates the brightness similarity at the largest scale. and They correspond to the i-th slice of the p-th anatomical plane in the real CT and the reconstructed CT, respectively.
[0054] The chest CT reconstruction model was trained and optimized on an NVIDIA RTX 4090 GPU with 24GB of video memory, using Python 3.8 and PyTorch 2.2.0. Due to GPU memory limitations, the training batch size was set to 2. Training lasted 100 epochs, with the learning rate held constant for the first 50 epochs and then linearly decayed to zero for the next 50 epochs. During training, the generator and discriminator were alternately optimized until Nash equilibrium was reached. Training took approximately 16 hours, resulting in a parameter size of 29.38M for the generator and 11.05M for the discriminator.
[0055] Step 4: Obtain the patient's orthogonal biplane X-ray film and input it into the chest CT reconstruction model trained in step 3. The model reconstructs the chest anatomical structure in three dimensions based on the input X-ray image and generates the corresponding CT image, as shown in the attached manual. Figure 4 As shown, the output results of the reconstructed model are presented in axial, coronal and sagittal slices respectively.
Claims
1. A method for reconstructing chest CT based on orthogonal biplane X-ray films, characterized in that: The following steps are involved: (1) The acquired CT data are preprocessed to avoid interference caused by data heterogeneity and remove irrelevant noise. Subsequently, orthogonal biplane X-rays are generated by digital radiographic reconstruction technology. Finally, the paired data sets are reasonably divided. (2) Constructing a chest CT reconstruction model, which consists of two parts: a generator and a discriminator. Since the task involves cross-dimensional conversion from 2D to 3D, it is necessary to extract key anatomical features while retaining low-level details. To this end, the present invention designs a feature enhancement block in the generator to reduce information loss. (3) The constructed training set and test set are input into the model, and the orthogonal biplane X-ray film is used as the reconstruction input. The difference is calculated by combining the perceptual consistency loss, projection loss and generative adversarial loss until the model reaches Nash equilibrium; (4) Obtain the patient's orthogonal biplane X-ray film and input it into the trained CT reconstruction model. The model performs three-dimensional reconstruction of the chest anatomical structure based on the input X-ray image and generates the corresponding CT image.
2. The method for reconstructing chest CT based on orthogonal biplane X-ray films according to claim 1, characterized in that: In step 1, the acquired raw CT data are preprocessed (unifying the image size, removing irrelevant noise, etc.), and then, digitally reconstructing radiographic imaging technology is used to generate orthogonal biplane X-ray images from the anteroposterior and lateral directions, thereby constructing a registered X-CT dataset.
3. The method for reconstructing chest CT based on orthogonal biplane X-ray films according to claim 1, characterized in that: In order to minimize information loss in step 2 and enhance the ability to extract key anatomical features from X-ray images, the present invention proposes a feature enhancement block in the generator to improve image quality.
4. The feature enhancement block according to claim 3, wherein: In the Sth encoding stage of the generator, the two-dimensional feature map (denoted as, ) is first downsampled, and then, is passed into the dense block, where dense connections and feature reuse mechanisms are used to generate output feature maps This design ensures the effective preservation of low-level features and their propagation to higher levels; in Represents the output feature map of the Nth dense layer in the Sth encoding stage, and Concat represents the concatenation of the output feature maps of all dense layers along the channel dimension; Then, in order to emphasize important features and suppress minor features, GAP and GMP were first used to Converted into two eigenvectors, denoted as v avg and v max These vectors are then processed, summed, and passed to Sigmoid to obtain the final channel attention weight W ch ; W ch =σ(FC(v avg )+FC(in max )) (3) The calculated channel attention weight W ch With the input feature map Multiply element-wise to generate optimized feature maps Similarly, for Extracting spatial attention weights W sp and apply it to Concentrate features on key areas within the image to generate optimized feature maps Finally, yes Compression is performed to reduce the number of output channels to half the number of input feature map channels, thereby generating the final output feature map 5. The method for reconstructing chest CT based on orthogonal biplane X-ray films according to claim 1, characterized in that: In step 3, in order to enhance the perceptual consistency of the reconstructed CT image in terms of brightness, contrast and structure, the present invention creatively proposes a perceptual consistency loss function; Where N represents the total number of CT slices, α is the weighting coefficient, and M represents the maximum number of scales for multi-scale analysis. represents the contrast and structural similarity of the i-th slice of the p-th anatomical plane at the k-th scale compared with its corresponding position in the real CT, represents the brightness similarity at the maximum scale, and They correspond to the i-th slice of the p-th anatomical plane in the real CT and the reconstructed CT, respectively.