An ai training image algorithm for micro-dose ct

Abstract

This invention discloses an AI training image algorithm for CT scans, comprising: inputting an initial image Image, wherein the initial image is a clinical routine CT noise reduction image or a medical model CT image; performing orthographic projection processing on the initial image to generate 180° multi-channel attenuation values; determining the target photon count according to the required noise level, obtaining the maximum projection value among the multi-channel attenuation values, converting the attenuation values ​​of each channel into photon counts, and generating Poisson noise and Gaussian noise; performing modulation processing on the Poisson noise and Gaussian noise corresponding to each channel to generate the comprehensive noise of that channel, adding the comprehensive noise of each channel to the photon count of the corresponding channel to obtain the noisy photon count; converting the noisy photon count of each channel to the projection domain to obtain the noisy attenuation value, and performing back projection processing on the noisy attenuation value to generate a simulated CT noise image for AI training, which can reduce clinical image noise, improve image diagnostic reliability and equipment performance.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and more specifically, to an AI training image algorithm for micro-dose CT. Background Technology

[0002] Image noise refers to information that interferes with normal signals received during image acquisition, transmission, and processing. Noise affects image quality, reduces the signal-to-noise ratio, and makes features difficult to identify. Currently, there are traditional image denoising methods based on signal processing principles and AI-based image denoising techniques based on deep learning. Compared to traditional image denoising methods (such as filtered backprojection), which are prone to losing image details under low-dose conditions, deep learning-based image denoising methods utilize AI self-supervised learning to effectively remove low-dose, especially micro-dose, image noise while preserving image features. Deep learning requires input CT noisy images and the desired denoised images for AI training to obtain the AI ​​denoising model. This method requires a large number of CT noisy images that accurately reflect the noise characteristics, along with paired denoised images.

[0003] Random noise in CT imaging is divided into two categories: one is signal-related noise, which is related to the number of photons detected by the detector during CT scanning; the other is noise related to the device and imaging process, but not to the signal magnitude, and is usually modeled as Poisson noise and Gaussian noise.

[0004] Currently, the common method for obtaining noisy CT images for AI training is: The medical simulation imaging model is used to generate simulated CT noise images and paired denoised images. That is, based on Geant4 simulated medical imaging, CT phantoms and CT noise images are generated by using MDCT and preset CT data. A sine wave is generated from the input initial image, and then a noisy sine wave is added to the initial sine wave to produce a high-noise image. A method was used to obtain a noisy CT image by adding Gaussian noise to a real CT image.

[0005] Noisy CT images and paired denoised images generated using medical models lack real clinical tissue structures and differ from actual clinical images. Therefore, they cannot obtain good denoising models and effects when used for AI training.

[0006] Although real clinical images are input, the noise is simulated on the sine wave, i.e., the projection domain data. The photon number noise simulation is not performed on the non-CT image reconstruction data source - the data domain. Therefore, the generated CT noisy images lack the process of back-projecting from the photon number into projection data. The photon number-related noise cannot be simulated, which reduces the realism and reliability of the CT noisy images.

[0007] Adding Gaussian noise to the image domain for simulation results in a single noise model that cannot accurately simulate the noise characteristics in actual clinical settings. Furthermore, the modulation effect in the image domain is not as good as that in the projection domain or the data domain. Summary of the Invention

[0008] To address at least one of the aforementioned technical problems, this invention proposes an AI-based image training algorithm for CT scans, comprising: S1, Input the initial image Image, which is a clinical routine CT noise-reduced image or a medical model CT image; S2, Perform orthographic projection processing on the initial image to generate 180° multi-channel attenuation values; S3, determine the target photon count based on the required noise level, obtain the maximum projection value among the multi-channel attenuation values, and convert the attenuation values ​​of each channel into the photon count; S4, each channel generates Poisson noise and Gaussian noise respectively; S5. Modulate the Poisson noise and Gaussian noise corresponding to each channel to generate the comprehensive noise of that channel. Add the comprehensive noise of each channel to the photon count of the corresponding channel to obtain the noisy photon count. S6. The number of noisy photons in each channel is converted to the projection domain to obtain the noise attenuation value. The noise attenuation value is then back-projected to generate a simulated CT noise image.

[0009] Specifically, this invention is also applicable to low-dose CT, ultra-low-dose CT, photon counting CT, conventional CT, etc.

[0010] In a preferred embodiment of the present invention, the multi-channel attenuation value in step S2 is calculated using the following formula: ; in: Sinogram represents the attenuation value of each channel in the initial image; Image represents the initial input image; It is a set of angles, uniformly distributed in [0, 180).

[0011] In a preferred embodiment of the present invention, the formula for converting the attenuation value into the photon number I in step S3 is as follows: ; in: I refers to the number of photons in each channel; Indicates the target photon number; This represents the maximum projection value of the initial image; Sinagram represents the attenuation value of each channel of the initial image.

[0012] In a preferred embodiment of the present invention, the combined noise is added to the photon count of the corresponding channel, as shown in the following formula: ; in: I n This represents the number of photons in each channel of the simulated noisy image; I represents the number of photons in each channel; T n This represents the combined noise of each channel after modulation.

[0013] In a preferred embodiment of the present invention, the photon number is converted to the projection domain. Sinogram noise The formula is as follows: ; in: This represents the attenuation value of each channel in the simulated noisy image; This represents the maximum projection value of the initial image; This represents the number of photons in each channel of the simulated noisy image.

[0014] In a preferred embodiment of the present invention, step S6 involves processing the Sinogram. noise The formula for generating the noisy image through back projection is as follows: ; in: Represents a simulated noise image; It is an angle set, uniformly distributed in [0, 180); The filter is the filter kernel used when reconstructing an image; it can be a ramp, cosine, Hanning window, etc.

[0015] In a preferred embodiment of the present invention, the 180° multi-channel attenuation value mentioned in step 2 has the number of channels set according to the resolution requirements of CT imaging, and the angle θ is taken at uniform intervals within the range of [0, 180).

[0016] In a preferred embodiment of the present invention, the Poisson noise Pn simulates signal-type noise related to the number of photons detected by the detector during CT imaging, and the Gaussian noise Gn simulates non-signal-type noise related to the device and imaging process and independent of the signal magnitude during CT imaging.

[0017] The technical solution of the present invention has the following advantages compared with the prior art: 1. High realism of noise simulation: This algorithm starts from the CT image data domain and modulates Poisson noise and Gaussian noise at the same time, which correspond to signal noise related to the number of photons and non-signal noise related to the device and imaging process in the CT imaging process, respectively. It accurately restores the real noise characteristics of clinical CT imaging and solves the problem of single noise model and large difference between simulation effect and actual clinical image in the existing technology.

[0018] 2. High matching accuracy and strong adaptability: Using high-quality clinical CT noise reduction images or medical model images as initial images, the generated simulated CT noise images are precisely aligned with the initial images in terms of image size, pixel resolution, and spatial position. They can be directly used as matching samples for training self-supervised learning noise reduction models without the need for multiple CT scans on the same patient. This reduces clinical scanning costs and avoids radiation damage to patients caused by repeated scans.

[0019] 3. Flexible and adjustable, adaptable to various scenarios: The number of target photons can be adjusted. CNT target The values ​​of can be used to simulate CT noise images at different dose levels. At the same time, the modulation weights and generation parameters of the two types of noise can be adjusted according to the noise distribution characteristics of actual CT imaging and the parameters of imaging devices, so as to adapt to different clinical CT equipment, different resolution requirements and different machine learning training scenarios.

[0020] 4. Reliable operation and strong practicality: Through optimization measures such as preset threshold correction and floating-point precision operation, abnormal problems that occur during operation are avoided, improving the stability of the algorithm and the accuracy of noise simulation. The generated simulated CT noise images can be output in DICOM standard format, and the algorithm can be embedded in CT equipment imaging simulation system or machine learning noise reduction model training platform, supporting batch generation and adapting to the actual needs of clinical application and model training.

[0021] 5. Improve the training effect of noise reduction model: The simulated CT noise image generated by this algorithm is accurately paired with the initial high-quality noise reduction image for training the self-supervised learning noise reduction model. This can effectively improve the accuracy, reliability and adaptability of the noise reduction model, enabling the trained noise reduction model to better cope with complex clinical noise scenarios, reduce clinical CT image noise, improve the reliability of image diagnosis, and help optimize and upgrade CT medical diagnostic technology. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, some of the drawings in the following description are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart of an AI training image algorithm for CT according to an embodiment of the present invention; Figure 2 This is an analysis block diagram of the AI ​​training image algorithm for CT according to an embodiment of the present invention. Detailed Implementation

[0024] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0026] Example 1 See Figures 1-2 As shown, this invention proposes an AI training image algorithm for CT scans, comprising: S1, Input the initial image Image, which is a clinical routine CT noise-reduced image or a medical model CT image; S2, Perform orthographic projection processing on the initial image to generate 180° multi-channel attenuation values; S3, determine the target photon count based on the required noise level, obtain the maximum projection value among the multi-channel attenuation values, and convert the attenuation values ​​of each channel into the photon count; S4, each channel generates Poisson noise and Gaussian noise respectively; S5. Modulate the Poisson noise and Gaussian noise corresponding to each channel to generate the comprehensive noise of that channel. Add the comprehensive noise of each channel to the photon count of the corresponding channel to obtain the noisy photon count. S6. The number of noisy photons in each channel is converted to the projection domain to obtain the noise attenuation value. The noise attenuation value is then back-projected to generate a simulated CT noise image.

[0027] It should be noted that by simulating different doses of CT noise in the data domain of routine clinical images or medical model images to generate CT noise images, the noise images are precisely paired with the clinical images or medical model images for AI training to obtain deep learning denoising algorithms. This eliminates the need for multiple scans of the same patient to obtain highly matched and accurate training images, improving the accuracy, reliability, and adaptability of the deep learning denoising algorithm.

[0028] In a preferred embodiment of the present invention, the multi-channel attenuation value in step S2 is calculated using the following formula: ; in: Sinogram represents the attenuation value of each channel in the initial image; Image represents the initial input image; It is a set of angles, uniformly distributed in [0, 180).

[0029] In a preferred embodiment of the present invention, the formula for converting the attenuation value into the photon number I in step S3 is as follows: ; in: I refers to the number of photons in each channel; Indicates the target photon number; This represents the maximum projection value of the initial image; Sinagram represents the attenuation value of each channel of the initial image.

[0030] In a preferred embodiment of the present invention, the combined noise is added to the photon count of the corresponding channel, as shown in the following formula; ; in: I n This represents the number of photons in each channel of the simulated noisy image; I represents the number of photons in each channel; T n This represents the combined noise of each channel after modulation.

[0031] In a preferred embodiment of the present invention, the photon number is converted to a projection domain sinogram. noise The formula is as follows: ; in: This represents the attenuation value of each channel in the simulated noisy image; This represents the maximum projection value of the initial image; This represents the number of photons in each channel of the simulated noisy image.

[0032] In a preferred embodiment of the present invention, step S6 involves processing the Sinogram. noise The formula for generating the noisy image through back projection is as follows:

[0033] in: Represents a simulated noise image; It is an angle set, uniformly distributed in [0, 180); The filter is the filter kernel used when reconstructing an image; it can be a ramp, cosine, Hanning window, etc.

[0034] In a preferred embodiment of the present invention, the 180° multi-channel attenuation value mentioned in step 2 has the number of channels set according to the resolution requirements of CT imaging, and the angle θ is taken at uniform intervals within the range of [0, 180).

[0035] In a preferred embodiment of the present invention, the Poisson noise Pn simulates signal-type noise related to the number of photons detected by the detector during CT imaging, and the Gaussian noise Gn simulates non-signal-type noise related to the device and imaging process and independent of the signal magnitude during CT imaging.

[0036] The technical solution of the present invention has the following advantages compared with the prior art: 1. High realism of noise simulation: This algorithm starts from the CT image data domain and modulates Poisson noise and Gaussian noise at the same time, which correspond to signal noise related to the number of photons and non-signal noise related to the device and imaging process in the CT imaging process, respectively. It accurately restores the real noise characteristics of clinical CT imaging and solves the problem of single noise model and large difference between simulation effect and actual clinical image in the existing technology.

[0037] 2. High matching accuracy and strong adaptability: Using high-quality clinical CT noise reduction images or medical model images as initial images, the generated simulated CT noise images are precisely aligned with the initial images in terms of image size, pixel resolution, and spatial position. They can be directly used as matching samples for AI training of deep learning noise reduction models without the need for multiple CT scans on the same patient. This reduces clinical scanning costs and avoids radiation damage to patients caused by repeated scans.

[0038] 3. Flexible and adjustable, adaptable to various scenarios: The number of target photons can be adjusted. CNT target The values ​​of can simulate CT noise images at different dose levels. At the same time, the modulation weights and generation parameters of the two types of noise can be adjusted according to the noise distribution characteristics of actual CT imaging and the parameters of imaging devices, so as to adapt to different clinical CT equipment, different resolution requirements and different deep learning AI training scenarios.

[0039] 4. Reliable operation and strong practicality: Through optimization measures such as preset threshold correction and floating-point precision operation, abnormal problems that occur during operation are avoided, improving the stability of the algorithm and the accuracy of noise simulation. The generated simulated CT noise images can be output in DICOM standard format, and the algorithm can be embedded in CT equipment imaging simulation system or deep learning noise reduction model AI training platform, supporting batch generation and adapting to the actual needs of clinical application and model training.

[0040] 5. Improve the training effect of noise reduction model: The simulated CT noise image generated by this algorithm is accurately paired with the initial high-quality noise reduction image for training the deep learning noise reduction model. This can effectively improve the accuracy, reliability and adaptability of the noise reduction model, enabling the trained noise reduction model to better cope with complex clinical noise scenarios, reduce clinical CT image noise, improve the reliability of image diagnosis, and help optimize and upgrade CT medical diagnostic technology.

[0041] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to the above embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An AI training image algorithm for CT scans, characterized in that, include: S1, Input the initial image Image, which is a clinical routine CT noise-reduced image or a medical model CT image; S2, Perform orthographic projection processing on the initial image to generate 180° multi-channel attenuation values; S3, determine the target photon count based on the required noise level, obtain the maximum projection value among the multi-channel attenuation values, and convert the attenuation values ​​of each channel into the photon count; S4 generates Poisson noise and Gaussian noise for each channel respectively; S5. Modulate the Poisson noise and Gaussian noise corresponding to each channel to generate the comprehensive noise of that channel. Add the comprehensive noise of each channel to the photon count of the corresponding channel to obtain the noisy photon count. S6. The number of noisy photons in each channel is converted to the projection domain to obtain the noise attenuation value. The noise attenuation value is then back-projected to generate a simulated CT noise image.

2. The AI ​​training image algorithm for CT according to claim 1, characterized in that, The multi-channel attenuation value in step S2 is calculated using the following formula: in: Sinogram represents the attenuation value of each channel in the initial image; Image represents the initial input image; It is a set of angles, uniformly distributed in [0, 180).

3. The AI ​​training image algorithm for CT according to claim 2, characterized in that, In step S3, the attenuation value is converted into the photon number I using the following formula: 。 in: I refers to the number of photons in each channel; Indicates the target photon number; This represents the maximum projection value of the initial image; Sinagram represents the attenuation values ​​of each channel in the initial image.

4. The AI ​​training image algorithm for CT according to claim 3, characterized in that, The combined noise is added to the photon count of the corresponding channel, as shown in the following formula. ； in: I n This represents the number of photons in each channel of the simulated noisy image; I represents the number of photons in each channel; T n This represents the combined noise of each channel after modulation.

5. The AI ​​training image algorithm for CT according to claim 2, characterized in that, Convert the photon number to the projection domain sinogram noise The formula is as follows: ； in: This represents the attenuation value of each channel in the simulated noisy image; This represents the maximum projection value of the initial image; This represents the number of photons in each channel of the simulated noisy image.

6. The AI ​​training image algorithm for CT according to claim 2, characterized in that, In step S6, the sinogram is... noise The formula for generating the noisy image through back projection is as follows: ； in: Represents a simulated noise image; It is an angle set, uniformly distributed in [0, 180); `filter` is the filter kernel used when reconstructing an image.

7. The AI ​​training image algorithm for CT according to claim 1, characterized in that, The 180° multi-channel attenuation value mentioned in step 2 has the number of channels set according to the resolution requirements of CT imaging, and the angle θ is taken at uniform intervals within the range of [0, 180).

8. The AI ​​training image algorithm for CT according to claim 1, characterized in that, The Poisson noise Pn simulates signal-type noise related to the number of photons detected by the detector during CT imaging, while the Gaussian noise Gn simulates non-signal-type noise related to the device and imaging process and independent of the signal magnitude during CT imaging.

Images