A CT imaging method with constant image quality

Abstract

The present invention discloses a CT imaging method with constant image quality, which relates to the technical field of medical CT imaging, and comprises the following steps: step 1, preparing a data set, including high-dose CT projections or images of a large number of patients, and corresponding patient prior information; step 2, constructing a model, including a dose estimation model, a scanning simulation model, and an image reconstruction model; step 3, training the model constructed in step 2; step 4, obtaining patient prior information; step 5, using the dose estimation model to obtain the patient's optimal examination dose; step 6, performing CT scanning to obtain patient projection data; step 7, using the image reconstruction model to obtain a patient CT image with constant quality. The present invention provides a CT imaging method with constant image quality, which realizes the "As Low As Reasonably Possible (ALARA)" principle in radiation protection, can obtain CT images with constant quality, and is beneficial to the clinical diagnosis of CT images.

Description

Technical Field

[0001] The invention relates to the technical field of medical CT imaging, and in particular to a CT imaging method with constant image quality. Background Art

[0002] Medical computed tomography (CT) has become an indispensable tool in clinical diagnosis and treatment due to its excellent three-dimensional structure representation capabilities. The physical mechanism of CT imaging stems from the penetration attenuation of X-rays in human tissues. Excessive X-ray exposure will cause symptoms such as hair loss and burning, and even increase the risk of cancer. However, reducing the radiation dose of CT examinations will cause the X-ray signal to weaken, causing an increase in noise in the CT detection signal, which will lead to the degradation of CT image quality and ultimately affect the CT diagnosis and treatment effect. How to maintain or even improve image quality while reducing the examination dose is a hot issue in the field of CT imaging.

[0003] On the other hand, different patients have different attenuation of X-rays. Under the same X-ray intensity, for obese patients, X-rays attenuate more when penetrating the body of such patients, resulting in weak CT detector signals, more data noise, and more serious image quality degradation. In contrast, for thin patients, X-rays attenuate less when penetrating the body of such patients, resulting in stronger CT detector signals, fewer data noise, and general image quality degradation. Therefore, each individual patient should use the most appropriate radiation dose when undergoing CT examination. The image examined at this dose can just meet the diagnosis and treatment needs of CT examination. A lower dose cannot meet clinical needs. In traditional methods, the CT examination dose is adjusted by tube current control technology, but the existing tube current control technology cannot be adapted to nonlinear advanced reconstruction algorithms such as iterative reconstruction and deep learning reconstruction, and therefore cannot effectively control the quality of CT images. In addition, the development process of the existing advanced reconstruction algorithm is independent of tube current control, and therefore cannot adapt to the detection data generated by tube current control. The dose reduction effect brought by these algorithms can be further improved. Summary of the invention

[0004] The purpose of the present invention is to provide a CT imaging method with constant image quality, so as to solve the problems in the prior art that it is difficult to output CT images with constant quality and the minimum radiation dose of an individual patient under constant image quality cannot be estimated.

[0005] To achieve the above object, the present invention provides a CT imaging method with constant image quality, comprising the following steps:

[0006] Step 1: Prepare a data set, including high-dose CT projections or images of a large number of patients and corresponding patient prior information;

[0007] Step 2: construct a model, including a dose estimation model, a scanning simulation model, and an image reconstruction model;

[0008] Step 3: Train the model constructed in step 2;

[0009] Step 4: Obtain patient prior information;

[0010] Step 5: Use the dose estimation model to obtain the optimal examination dose for the patient;

[0011] Step 6: Perform CT scanning to obtain patient projection data;

[0012] Step 7: Use the image reconstruction model to obtain a patient CT image of constant quality.

[0013] Preferably, in step 1, if the data set only contains high-dose CT projections, the high-dose CT images are obtained by a reconstruction algorithm; if the data set only contains high-dose CT images, the high-dose CT projections are obtained by forward projection.

[0014] Preferably, the patient's prior information includes the patient's positioning image, the patient's previous CT images, and information obtained by optimizing these two types of prior information. For the optimization of the two types of prior information, if the prior information is a positioning image, a neural network is used to generate a prior CT image from the positioning image; if the prior information is a previous CT image, the previous CT image is aligned to the patient's current CT examination position.

[0015] Preferably, the dose estimation model is specifically a deep neural network, the network structure of which is unrestricted, the input of which is the patient's prior information, and the output dose is the current or the amount of X-rays emitted when the patient's CT examination tube is exposed.

[0016] Preferably, the dose estimated by the dose estimation model is the total dose of the patient's axial range of the current CT examination, or the dose of the axial position corresponding to each layer of CT image in the expected image, or the dose of each exposure angle in the axial range of the current CT examination.

[0017] Preferably, the scanning simulation model is a differentiable model. The function of the scanning simulation model is to generate a low-dose projection from a high-dose projection. The implementation of this process includes the following two methods:

[0018] The first method is to estimate the variance of the target low-dose projection, then generate random numbers based on the variance and add them to the high-dose projection;

[0019] The second method is to use a deep neural network to generate low-dose projections from high-dose projections. The internal structure of the network is unrestricted. Its input is the patient's high-dose projection and the patient's CT examination dose. The output is the target low-dose projection or the variance of the target low-dose projection. If the output is the variance of the target low-dose projection, a random number is generated based on the variance and added to the high-dose projection.

[0020] Preferably, the expression for the variance of the target low-dose projection is as follows:

[0021] ;

[0022] In the formula, Pixels in the projection The variance of is the high-dose projection pixel value, is the variance of the measurable detector background random signal, , Projection pixels of photon quantity; Pixel The tube current of the projection, is the measurable projection pixel The current-to-photon conversion coefficient;

[0023] The process of adding random numbers to high-dose projections is expressed as:

[0024] ;

[0025] In the formula, is the target low-dose projection pixel value, Sample random numbers from a standard normal distribution.

[0026] Preferably, the image reconstruction model is a differentiable model, and the model parameters can be fixed or learned. The construction methods include: filtered back projection, iterative reconstruction, filtered back projection + image denoising, projection denoising + filtered back projection, projection denoising + filtered back projection + image denoising, and using a neural network to directly generate an image from the projection, wherein filtered back projection and iterative reconstruction can be expanded into a differentiable process, image denoising and projection denoising can use a deep neural network denoising model or expand the traditional denoising method into a differentiable process.

[0027] Preferably, the process of training the model constructed in step 2 in step 3 is as follows:

[0028] S31, dividing the data set prepared in step 1 into a training set, a validation set and a test set;

[0029] S32, constructing a loss function for model training;

[0030] S33, setting the hyperparameters for neural network model training, including learning rate, epoch iteration number, batch size, optimizer, and learning rate modulator;

[0031] S34, the model constructed in step 2 based on the gradient back propagation algorithm training.

[0032] Preferably, the loss function of the model training in S32 is as follows:

[0033] ;

[0034] In the formula, is the overall loss function, is the dose loss term, is the image loss term, is a manually set coefficient, where the dose loss term aims to reduce the radiation dose and is expressed as or , which means minimizing the amount of X-rays or the tube current; the image loss term is constructed in the form of hinge loss, expressed as: ,in, Indicates taking The larger value of is a differentiable image quality metric function, is the reconstructed CT image, is a pre-set image quality indicator; in model training, when the reconstructed image quality is higher than hour, Stop working, the model focuses on reducing radiation dose, when image quality is lower than hour, To participate in the work, the model needs to learn to reduce the radiation dose while maintaining image quality at least , and finally achieve the goal of outputting constant quality CT images. By training the dose estimation model through the above-mentioned overall loss function, it can output the optimal CT examination dose for patients, that is, the lowest radiation dose that can be set for individual patient CT scans under the constraint of meeting the constant image quality.

[0035] Therefore, the present invention adopts the above-mentioned CT imaging method with constant image quality, which has the following beneficial effects:

[0036] (1) The method provided by the present invention estimates the minimum CT examination dose for individual patients under the constraint of constant image quality, realizing the “As Low As Reasonably Available (ALARA)” principle in radiological imaging;

[0037] (2) During the training process, the image loss term in the loss function is constructed based on the hinge loss, so that the image reconstruction model outputs CT images with constant quality, achieving the standardization of CT image quality between different patients, which is beneficial to clinical diagnosis and treatment based on CT images and the development of medical AI systems based on CT images;

[0038] (3) The constructed scanning simulation model is a differentiable model, which allows the gradient to be back-propagated from the image reconstruction model to the dose estimation model, thereby realizing radiation dose control guided by image quality optimization;

[0039] (4) The dose estimation model and the image reconstruction model can be jointly optimized to achieve mutual adaptation between the two and further reduce the radiation dose while maintaining constant image quality;

[0040] (5) The concept of the present invention can also be used in related imaging fields where scanning conditions are limited, such as positron emission tomography (PET) and magnetic resonance imaging (MRI), that is, to jointly optimize scanning technology and image reconstruction technology to achieve standardization of image quality and optimal scanning conditions for individual objects.

[0041] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is an overall flow chart of a CT imaging method with constant image quality according to the present invention;

[0043] Figure 2 This is a simplified workflow diagram of a conventional CT examination according to an embodiment of the present invention;

[0044] Figure 3 This is a structural diagram of a ResNet-18 dose estimation model according to an embodiment of the present invention;

[0045] Figure 4 This is a structural diagram of the Unet image denoising model according to an embodiment of the present invention;

[0046] Figure 5 The imaging results of three cases in the LoDoPa-CT test set of an embodiment of the present invention are numbered as case 1, case 2 and case 3, respectively. (a) is the imaging result of case 1, (b) is the imaging result of case 2, and (c) is the imaging result of case 3;

[0047] Figure 6: The MSE scatter plots of all cases in the LoDoPa-CT test set of an embodiment of the present invention are as follows: (a) is the MSE scatter plot of all cases in the test set of the Unet image denoising model trained alone when the X-ray dose is fixed to 5e4; (b) is the MSE scatter plot of all cases in the test set of the ResNet-18 dose estimation model and the Unet image denoising model jointly trained when the MSE is preset to 1e-4. DETAILED DESCRIPTION

[0048] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] See also Figure 1-6 , a CT imaging method with constant image quality, comprising the following steps:

[0050] Step 1: Prepare the dataset

[0051] The LoDoPa-CT public dataset is used as an example for implementation. The dataset contains high-dose CT images of 43,005 cases. The high-dose images of all cases in the dataset are forward projected to obtain high-dose projections. The forward projection settings are: the projection geometry is parallel beam geometry, the image size is 362*362, the FOV diameter is 260 mm, and the projection angle range is , the number of projection angles is 1000, the number of detector units is 513, and the detector unit size is 0.7168 mm. This dataset has no patient prior information, and high-dose images are used as prior information for implementation.

[0052] Step 2: Build the model

[0053] S21: Construct a dose estimation model, using the ResNet-18 neural network as the dose estimation model, whose input is the high-dose image and output is the X-ray quantity shared by all projection angles. Figure 3 The structure diagram of the ResNet-18 dose estimation model in this embodiment is shown, and its network weights are initialized to the ImageNet pre-trained weights officially provided by Pytorch.

[0054] S22: Construct a scanning simulation model to generate low-dose projections from high-dose projections. First, estimate the variance of the target low-dose projections, then generate random numbers based on the variance and add them to the high-dose projections. The variance of the target low-dose projections can be calculated as follows: ,in, Pixels in the projection The variance of is the high-dose projection pixel value, is the variance of the detector background random signal, which is set to ， is the amount of X-rays shared by all projection angles, i.e., the output of the ResNet-18 dose estimation model. The process of adding random numbers to high-dose projections can be expressed as: ,in, is the low-dose projection pixel value, is the high-dose projection pixel value, Sample random numbers from a standard normal distribution.

[0055] S23: Construct an image reconstruction model using the "filtered back projection + image denoising" method. First, use the filtered back projection algorithm to reconstruct the image, and then use the Unet neural network to denoise the image. Figure 4 The structure diagram of the Unet image denoising model in this embodiment is shown. Compared with the original Unet, the number of feature channels at each scale is halved, and the output adopts the residual output mode: ,in For the image reconstructed by the filtered back projection algorithm, the Unet network weights are initialized to random numbers sampled from a truncated normal distribution with a mean of 0, a standard deviation of 0.02, and a truncation range of [-0.06, 0.06].

[0056] Step 3: Train the model

[0057] S31: The data set is divided into a training set, a validation set, and a test set. The training set contains 35,839 cases, the validation set contains 3,583 cases, and the test set contains 3,583 cases.

[0058] S32: Construct loss function: ,in, is the overall loss function, is the dose loss term, is the image loss term, The coefficient is manually set, which is set as follows in this embodiment: The dose loss term is constructed as: The image loss term is constructed as: ,in, Indicates taking The larger value of is the mean square error, expressed as: , is the batch size, is the image size, and They are the output images of the Unet image denoising model respectively. With high-dose images Medium Pixels The value of For pre-set The numerical value is set to: .

[0059] S33: Set the hyperparameters for neural network model training. The learning rate of the Unet image denoising model is 5e-4, the learning rate of the ResNet-18 dose estimation model is 0.1, the number of epoch iterations is 50, the batch size is 64, the optimizer is the AdamW optimizer with default parameters, and the learning rate modulation method is: linearly increase from 5e-7 to 5e-4 in the first 5 iterations, and then decay from 5e-4 to 1e-6 in the "CosineAnnealing" mode.

[0060] S34: Based on the model constructed in step 2 of the gradient back-propagation algorithm training, the automatic differentiation module in the Pytorch deep learning platform calculates the gradient information of the dose estimation model and the image denoising model to update the parameters.

[0061] Step 4: Obtain patient prior information. In this embodiment, the high-dose images of all cases in the test set are used as patient prior information.

[0062] Step 5: Use the dose estimation model to obtain the optimal examination dose for the patient. In this embodiment, the trained ResNet-18 dose estimation model is used to estimate the X-ray dose shared by all projection angles of the high-dose projection of each case in the test set. After the ResNet-18 dose estimation model is trained using the loss function constructed by S32, the X-ray dose output by the model is the minimum X-ray dose that can be set under the condition that the MSE value of the reconstructed image is constant at 1e-4 in the implementation environment of this case, that is, the optimal examination dose for individual patients.

[0063] Step 6: Perform a CT scan to obtain patient projection data. In this embodiment, the scanning simulation model constructed in S22 is used to simulate the CT scan, and the corresponding low-dose projections are generated from the high-dose projections of all cases in the test set.

[0064] Step 7: Use the image reconstruction model to obtain patient CT images of constant quality. In this embodiment, the filtered back projection algorithm is used to reconstruct the image, and then the trained Unet image denoising model is used to obtain constant quality CT images, including CT images of all cases in the test set.

[0065] Figure 5The imaging results of three cases in the LoDoPa-CT test set are shown, numbered as case 1, case 2 and case 3, respectively. (a) is the imaging result of case 1, (b) is the imaging result of case 2, and (c) is the imaging result of case 3. This result shows that with the support of the method of the present invention, for different anatomical structures, the expected MSE constant of 1e-4 images can be obtained, and the inspection dose required for the corresponding anatomical structure can be adaptively estimated. Figure 6 The MSE scatter plots of all cases in the LoDoPa-CT test set are shown. (a) is the MSE scatter plot of all cases in the test set of the Unet image denoising model trained alone when the X-ray dose is fixed to 5e4; (b) is the MSE scatter plot of all cases in the test set of the ResNet-18 dose estimation model and the Unet image denoising model jointly trained when the MSE is preset to 1e-4. This result shows that when the X-ray dose is fixed, the quality difference of CT images of different patients is large. Using the method of the present invention, the quality difference of CT images of different patients is small and can be stabilized near the predetermined image quality index.

[0066] Therefore, the present invention adopts the above-mentioned constant image quality CT imaging method, uses a deep neural network to estimate the examination dose from the patient's prior information, and jointly optimizes the dose estimation and image reconstruction, ultimately generating a CT image with constant quality, while achieving the estimation of the minimum examination dose for individual patients at a constant quality level.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A CT imaging method with constant image quality, characterized in that: The following steps are involved: Step 1: Prepare a data set, including high-dose CT projections or images of a large number of patients and the corresponding patient prior information; Step 2, construct a model, including a dose estimation model, a scanning simulation model, and an image reconstruction model; wherein the dose estimation model is specifically a deep neural network, whose input is the patient's prior information, and the output dose is the current or the amount of X-rays emitted when the patient's CT examination tube is exposed; the dose estimated by the dose estimation model is the overall dose of the patient's axial range of the CT examination, the dose of the axial site corresponding to each layer of CT image in the expected image, or the dose of each exposure angle in the axial range of the CT examination; The scanning simulation model is a differentiable model. The function of the scanning simulation model is to generate low-dose projections from high-dose projections. There are two ways to achieve this: The first method is to estimate the variance of the target low-dose projection, then generate random numbers based on the variance and add them to the high-dose projection; The second method is to use a deep neural network to generate low-dose projections from high-dose projections. The input is the patient's high-dose projection and the patient's CT examination dose. The output is the target low-dose projection or the variance of the target low-dose projection. If the output is the variance of the target low-dose projection, a random number is generated based on the variance and added to the high-dose projection. The image reconstruction model is a differentiable model, and the construction methods include: filtered back projection, iterative reconstruction, filtered back projection + image denoising, projection denoising + filtered back projection, projection denoising + filtered back projection + image denoising, or using a neural network to directly generate images from projections; Step 3: Train the model constructed in step 2; the specific process is as follows: S31, dividing the data set prepared in step 1 into a training set, a validation set and a test set; S32. Construct the loss function for model training. The expression is as follows: ； In the formula, is the overall loss function, is the dose loss term, is the image loss term, is a manually set coefficient, where the dose loss term aims to reduce the radiation dose and is expressed as or ; The image loss term is constructed in the form of hinge loss, expressed as: ,in, Indicates taking The larger value of is a differentiable image quality metric function, is the reconstructed CT image, is a pre-set image quality indicator; S33, setting the hyperparameters for neural network model training, including learning rate, epoch iteration number, batch size, optimizer, and learning rate modulator; S34, a model constructed in step 2 based on the gradient back propagation algorithm training; Step 4: Obtain patient prior information; Step 5: Use the dose estimation model to obtain the optimal examination dose for the patient; specifically: use a differentiable image quality metric function Approaching the pre-set image quality index is the constraint, the dose loss term When it is minimum, the optimal examination dose for the patient is obtained through the dose estimation model; Step 6: Perform CT scanning to obtain patient projection data; Step 7: Using the image reconstruction model to obtain a patient CT image of constant quality; The patient's prior information includes the patient's positioning image, the patient's previous CT images, and the information obtained by optimizing these two prior information. For the optimization of the two prior information, if the prior information is the positioning image, a neural network is used to generate a prior CT image from the positioning image; if the prior information is a previous CT image, the previous CT image is aligned to the patient's current CT examination position.

2. The CT imaging method with constant image quality according to claim 1, characterized in that: In step 1, if the data set contains only high-dose CT projections, the high-dose CT images are obtained through a reconstruction algorithm; if the data set contains only high-dose CT images, the high-dose CT projections are obtained through forward projection.

3. A CT imaging method with constant image quality according to claim 2, characterized in that: The expression for the variance of the target low-dose projection is as follows: ； In the formula, Pixels in the projection The variance of is the high-dose projection pixel value, is the variance of the measurable detector background random signal, , Projection pixels The amount of X-rays; Pixel The tube current of the projection, is the measurable projection pixel The current-photon conversion coefficient; The process of adding random numbers to high-dose projections is expressed as: ； In the formula, is the target low-dose projection pixel value, Sample random numbers from a standard normal distribution.

Images