A CT image quality assessment method based on pre-noise reduction

Abstract

The present invention discloses a CT image quality evaluation method based on pre-denoising, comprising: S1, using an improved denoising variational autoencoder to perform denoising on a CT image to be evaluated to obtain a pseudo reference image; S2, performing global evaluation and local evaluation on the CT image to be evaluated and the pseudo reference image to obtain a global evaluation result and a local evaluation result; S3, performing weighted fusion on the global evaluation result and the local evaluation result to obtain a CT image quality evaluation result. The method maintains high performance while maintaining low computational cost, thereby ensuring usability and efficiency in practical applications.

Description

Technical Field

[0001] The present application relates to the field of image processing, and in particular to a CT image quality evaluation method based on pre-noise reduction. Background Art

[0002] The CT imaging process can be divided into four stages according to the data flow: ① data acquisition; ② data processing; ③ data reconstruction; ④ image display and post-processing. Combined with the above process, traditional CT noise reduction can be roughly divided into: sine wave filtering algorithm, iterative reconstruction algorithm and other noise reduction algorithms based on projection data and post-processing algorithms based on reconstructed CT images.

[0003] The sinusoidal image filtering algorithm is based on the sinusoidal projection data before CT reconstruction. This method is mostly based on traditional filtering methods, such as using nonlinear anisotropic diffusion filters to smooth data noise, using adaptive trimmed average filters to reduce streak artifacts, bilateral filtering, non-stationary filters, etc. Since it is based on frequency domain data noise reduction, it reduces the specific insensitivity of spatial noise and has a good noise reduction effect. However, most of these methods require sacrificing resolution, so they need to be combined with sinusoidal domain interpolation. After noise reduction is completed in the sinusoidal domain, FBP can be used to convert CT reconstruction into the image domain. The iterative reconstruction method synthesizes the statistical characteristics of sinusoidal filtering, prior information in the image domain, and parameters of the imaging system into the objective function. Common ones include compressed sensing technology, total variation technology, and dictionary learning. Although the algorithm of the iterative reconstruction method has relatively ideal reconstruction accuracy, its efficiency is generally low and requires a large amount of resource consumption, which limits the application of this method. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a CT image quality assessment method based on pre-noise reduction, which solves the problem that the existing CT image quality assessment has high computational cost and consumes a lot of resources.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: A CT image quality evaluation method based on pre-noise reduction, comprising:

[0006] S1. Use the improved denoising variational autoencoder to denoise the CT image to be evaluated to obtain a pseudo reference image;

[0007] S2, performing global evaluation and local evaluation on the CT image to be evaluated and the pseudo reference image to obtain global evaluation results and local evaluation results;

[0008] S3. Perform weighted fusion on the global evaluation result and the local evaluation result to obtain the CT image quality evaluation result.

[0009] Further: In S1, the improved denoising variational autoencoder includes an encoder, a decoder and a discriminator.

[0010] Further: S1 includes:

[0011] S11, using an encoder to encode the CT image x to be evaluated to obtain a high-dimensional feature representation c;

[0012] S12, using a decoder to decode the high-dimensional feature representation c to obtain a reconstructed image z;

[0013] S13. The CT image x to be evaluated and the reconstructed image z are used as inputs of the discriminator to perform generative adversarial training to obtain a pseudo reference image.

[0014] The beneficial effect of the above further scheme is: when an ordinary autoencoder performs a generation task, the generated image will introduce new noise due to the model; in this scheme, a discriminator is introduced to distinguish whether the CT image is natural or generated. The result of adversarial training is that the generated image can deceive the past on this issue, thereby reducing reconstruction artifacts.

[0015] Further: S2 includes:

[0016] S21, using the VGG16 model to extract the multi-scale features Fc1 of the CT image to be evaluated and the multi-scale features Fc2 of the pseudo reference image;

[0017] S22, calculating the corresponding difference between the multi-scale feature Fc1 and the multi-scale feature Fc2 at each scale, and obtaining the weight w of the features at different scales;

[0018] S23, performing a global evaluation on the CT image to be evaluated and the pseudo reference image based on the weights w of the features of different scales, and obtaining a global evaluation result;

[0019] S24. Based on the region of interest mask and the artifact active region, locally evaluate the CT image to be evaluated and the pseudo reference image to obtain a local evaluation result.

[0020] Further: the S24 includes:

[0021] S241, calculating the difference between the CT image to be evaluated and the pseudo reference image as the artifact active area;

[0022] S242, obtaining a region of interest in the CT image to be evaluated as a region of interest mask;

[0023] S243, calculating the convolution sum of the artifact active area and the mask of the region of interest as the first local evaluation result;

[0024] S244, calculating the contrast-to-noise ratio of the region of interest to the background according to the CT image to be evaluated and the mask of the region of interest as a second local evaluation result;

[0025] S245. The first local evaluation result and the second local evaluation result are taken together as a local evaluation result.

[0026] Further: the S244 includes:

[0027] S2441, taking the area outside the body in the CT image to be evaluated as a background area;

[0028] S2442, respectively calculate the average value of the signal strength in the mask area of ​​the region of interest and the background area, and the expression is:

[0029]

[0030] Among them, SI ROI is the average signal intensity of the mask in the region of interest, SI ROI,all is the intensity value of all pixels in the mask of the region of interest, N ROI is the number of pixels of the mask of the region of interest; SI b is the average signal intensity of the background area, SI b,all is the intensity value of all pixels in the background area, N b is the number of pixels in the background area;

[0031] S2443. Calculate the contrast-to-noise ratio of the region of interest to the background according to the average value of the signal strength in the region of interest mask and the background region as the second local evaluation result.

[0032] Further: In S2443, the expression for calculating the contrast-to-noise ratio CNR of the region of interest to the background is:

[0033]

[0034] Among them, SD b is the standard deviation of the signal intensity in the background area.

[0035] The beneficial effects of the present invention are:

[0036] 1. Use an improved denoising variational autoencoder with a discriminator to perform denoising. In the training phase, further distinguish the input image before denoising from the reconstructed image after denoising, reduce the reconstruction artifacts caused by the image generation process, and thus improve the credibility of the pseudo-reference;

[0037] 2. When conducting global evaluation, an adaptive feature fusion method is used to learn the gain weight of each layer of features on the evaluation results, and the features of each layer are retained and calculated in a targeted manner to reduce the scale of calculation and the operation time of the program;

[0038] 3. Perform local evaluation based on obvious regions of interest in medical images to reduce the quality evaluation deviation caused by metal artifacts. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the CT image quality assessment method based on pre-noise reduction.

[0040] Figure 2 Schematic diagram of the structure of the denoising variational autoencoder in the embodiment.

[0041] Figure 3 Schematic diagram of the discriminator structure in the embodiment.

[0042] Figure 4 This is a comparison chart of CT images with different quality levels in the simulation experiment.

[0043] Figure 5 This is a graph showing the pseudo reference images generated by different methods in the simulation experiment.

[0044] Figure 6 Comparison of CT images of patients with different qualities. DETAILED DESCRIPTION

[0045] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0046] like Figure 1 As shown, in one embodiment of the present invention, a CT image quality evaluation method based on pre-noise reduction includes:

[0047] S1. Use the improved denoising variational autoencoder to denoise the CT image to be evaluated to obtain a pseudo reference image;

[0048] S2, performing global evaluation and local evaluation on the CT image to be evaluated and the pseudo reference image to obtain global evaluation results and local evaluation results;

[0049] S3. Perform weighted fusion on the global evaluation result and the local evaluation result to obtain the CT image quality evaluation result.

[0050] In S1, the improved denoising variational autoencoder includes an encoder, a decoder and a discriminator.

[0051] The S1 includes:

[0052] S11, using an encoder to encode the CT image x to be evaluated to obtain a high-dimensional feature representation c;

[0053] S12, using a decoder to decode the high-dimensional feature representation c to obtain a reconstructed image z;

[0054] S13. The CT image x to be evaluated and the reconstructed image z are used as inputs of the discriminator to perform generative adversarial training to obtain a pseudo reference image.

[0055] In this embodiment, the improved denoising variational autoencoder includes a denoising variational autoencoder and a generative adversarial network;

[0056] The encoder and the decoder together constitute a denoising variational autoencoder, and the denoising variational autoencoder is used to encode and then decode the CT image x to be evaluated to obtain a reconstructed image z;

[0057] In this embodiment, the structural diagram of the denoising variational autoencoder is as follows: Figure 2 As shown, in the denoising variational autoencoder, the encoder performs continuous residual blocks and downsampling operations on the CT image x to be evaluated after 2D convolution, and then obtains a high-dimensional feature representation c of the input image through 2D convolution;

[0058] The decoder performs continuous residual blocks and upsampling operations after 2D convolution to decode the high-dimensional feature representation c of the input image, and obtains the reconstructed image z through 2D convolution;

[0059] The decoder and the discriminator together constitute a generative adversarial network, which is used to further distinguish the CT image x and the reconstructed image z to be evaluated, reduce the reconstruction artifacts caused by the encoding and decoding process, and obtain a pseudo reference image;

[0060] In this embodiment, the specific structure of the discriminator is as follows: Figure 3 As shown, the discriminator is essentially a supervised two-classification network, which is used to distinguish whether the input image is the original real CT or the CT image generated by the encoder;

[0061] During the training process, the discriminator aims to identify the authenticity of the input image as accurately as possible, while the generator tries to generate images as realistic as possible to deceive the discriminator. The input of the discriminator is the original CT image and the CT image obtained by the encoder. The output is a 64*64 feature map, in which each element (corresponding to an 8*8 image block of the input image) indicates whether the image block is real or generated;

[0062] This structure enables it to focus on local details in the image rather than just the global structure. During the training process, the loss function of the generator and the loss function of the discriminator work together to ensure that the generator generates more realistic and practical CT images while performing denoising operations on the images to be detected.

[0063] The S2 includes:

[0064] S21, using the VGG16 model to extract the multi-scale features Fc1 of the CT image to be evaluated and the multi-scale features Fc2 of the pseudo reference image;

[0065] S22, calculating the corresponding difference between the multi-scale feature Fc1 and the multi-scale feature Fc2 at each scale, and obtaining the weight w of the features at different scales;

[0066] S23, performing a global evaluation on the CT image to be evaluated and the pseudo reference image based on the weights w of the features of different scales, and obtaining a global evaluation result;

[0067] S24. Based on the region of interest mask and the artifact active region, locally evaluate the CT image to be evaluated and the pseudo reference image to obtain a local evaluation result.

[0068] Unlike ordinary image datasets, medical images have obvious regions of interest (ROI). Taking lung CT as an example, the importance of the background area outside the body, the patient's peripheral tissue area, and the lung area where the lesion core may appear in diagnosis gradually increases. Metal artifacts, as one of the most common artifacts in CT images, often appear locally. When there are metal foreign bodies in or on the body of the subject, the metal foreign bodies will cause local image distortion, deformation or distortion. When metal artifacts or other local artifacts appear in non-interested areas, although it may cause a decrease in global quality, it is actually still considered to be a qualified film with usable quality.

[0069] The difference between the CT to be evaluated and the CT after pre-noise reduction is considered to be the artifact image of the CT. The larger the intersection range between the active artifact area and the region of interest, the greater the impact of the artifact on the original CT image. On the contrary, the smaller the intersection range, the less impact of the artifact on the original CT image.

[0070] The S24 includes:

[0071] S241, calculating the difference between the CT image to be evaluated and the pseudo reference image as the artifact active area;

[0072] S242, obtaining a region of interest in the CT image to be evaluated as a region of interest mask;

[0073] S243, calculating the convolution sum of the artifact active area and the mask of the region of interest as the first local evaluation result;

[0074] S244, calculating the contrast-to-noise ratio of the region of interest to the background according to the CT image to be evaluated and the mask of the region of interest as a second local evaluation result;

[0075] S245. The first local evaluation result and the second local evaluation result are taken together as a local evaluation result.

[0076] The S244 includes:

[0077] S2441, taking the area outside the body in the CT image to be evaluated as a background area;

[0078] S2442, respectively calculate the average value of the signal strength in the mask area of ​​the region of interest and the background area, and the expression is:

[0079]

[0080] Among them, SI ROI is the average signal intensity of the mask in the region of interest, SI ROI,all is the intensity value of all pixels in the mask of the region of interest, N ROI is the number of pixels of the mask of the region of interest; SI b is the average signal intensity of the background area, SI b,all is the intensity value of all pixels in the background area, N b is the number of pixels in the background area;

[0081] S2443. Calculate the contrast-to-noise ratio of the region of interest to the background according to the average value of the signal strength in the region of interest mask and the background region as the second local evaluation result.

[0082] In S2443, the expression for calculating the contrast-to-noise ratio CNR of the region of interest to the background is:

[0083]

[0084] Among them, SD b is the standard deviation of the signal intensity in the background area.

[0085] In one embodiment of the present invention, a simulation experiment is conducted based on the open dataset MAYO2016 for medical image processing research, which contains a large number of high-quality high-dose X-ray computed tomography (CT) images from multiple patients, covering different anatomical structures and pathological features.

[0086] In this example, we mainly focus on CT images of the lung and abdomen. Based on the "ASimple Low-Dose X-Ray CT Simulation From High-Dose Scan" method, we simulated 7 different levels of low-quality CT data, a total of 540 images, and 80 images of high-quality CT data at normal doses, for a total of 640 simulated data. Examples of CT images of different quality levels are shown below: Figure 4 As shown, Figure 4 (a) is a high-quality image with a general dose. Figure 4 (b) is the image with the scanning dose reduced to 1 / 4 of the normal dose. Figure 4 In (b), a clearly visible streak artifact appears.

[0087] In the process of generating pseudo reference images, the experiment compares the direct generation of denoised CT and the generation of CT artifacts and then the difference to obtain denoised CT. At present, the experimental part of this method mainly implements CT denoising based on the encoding network, such as Figure 5 As shown in the figure, the generation result of the CT image “pseudo reference image” based on the improved denoising variational autoencoder on the simulated CT dataset is displayed; Figure 5 In the traditional method, the denoising results are not ideal, or the tissue edges in the obtained image are blurred, or there are obvious unreasonable "artifacts". The generated results of this method have clearer tissue edges, and both the overall and local aspects are more in line with the expectations of generating a "pseudo reference image" of the CT image after denoising the original image.

[0088] The verification results of common quality evaluation algorithms based on PLCC and SRCC indicators on the simulation data set are shown in Table 1:

[0089] Table 1 Comparison of common quality evaluation algorithm indicators based on different true values ​​on simulation datasets

[0090]

[0091] According to Table 1, this method has excellent overall performance and shows good stability under various test indicators. In other words, this type of objective evaluation indicator is more consistent with the evaluation results based on "people".

[0092] In addition, this embodiment selected 600 real data from 12 patients, of which 300 images were evaluated as "good quality" by doctors and 300 images were evaluated as "poor quality";

[0093] like Figure 6 (a) shows an example of a CT image that is evaluated as “good quality”. The image contrast and image smoothness are more in line with human observation preferences. Figure 6(b) is an example of a CT scan that is rated as “poor quality”. The image is gray as a whole, and the soft tissue and bone parts lack details, making it impossible to observe possible lesions.

[0094] As shown in Table 2, the classification results on the real patient dataset are shown;

[0095] Among them, True Positive (TP) indicates the number of samples predicted to be “good quality” and actually are “good quality”, False Positive (FP) indicates the number of samples predicted to be “good quality” and actually are “poor quality”, True Negative (TN) indicates the number of samples predicted to be “poor quality” and actually are “poor quality”, and False Negative (FN) indicates the number of samples predicted to be “poor quality” and actually are “good quality”.

[0096] Table 2 Classification results on real datasets

[0097]

[0098] Based on the above results, the final classification accuracy is 0.8514 and the F1 score is 0.8452, which shows that this method has certain applicability in quality scoring problems.

[0099] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A CT image quality evaluation method based on pre-noise reduction, characterized in that: include: S1. Use the improved denoising variational autoencoder to denoise the CT image to be evaluated to obtain a pseudo reference image; S2, performing global evaluation and local evaluation on the CT image to be evaluated and the pseudo reference image to obtain global evaluation results and local evaluation results; S3, performing weighted fusion on the global evaluation result and the local evaluation result to obtain the CT image quality evaluation result; The S2 includes: S21, using the VGG16 model to extract the multi-scale features Fc1 of the CT image to be evaluated and the multi-scale features Fc2 of the pseudo reference image; S22, calculating the corresponding difference between the multi-scale feature Fc1 and the multi-scale feature Fc2 at each scale, and obtaining the weight w of the features at different scales; S23, performing a global evaluation on the CT image to be evaluated and the pseudo reference image based on the weights w of the features of different scales, and obtaining a global evaluation result; S24, based on the region of interest mask and the artifact active region, locally evaluate the CT image to be evaluated and the pseudo reference image to obtain a local evaluation result; The S24 includes: S241, calculating the difference between the CT image to be evaluated and the pseudo reference image as the artifact active area; S242, obtaining a region of interest in the CT image to be evaluated as a region of interest mask; S243, calculating the convolution sum of the artifact active area and the mask of the region of interest as the first local evaluation result; S244, calculating the contrast-to-noise ratio of the region of interest to the background according to the CT image to be evaluated and the mask of the region of interest as a second local evaluation result; S245. The first local evaluation result and the second local evaluation result are taken together as a local evaluation result.

2. The CT image quality assessment method based on pre-noise reduction according to claim 1, characterized in that: In S1, the improved denoising variational autoencoder includes an encoder, a decoder and a discriminator.

3. The CT image quality assessment method based on pre-noise reduction according to claim 2, characterized in that: The S1 includes: S11, using an encoder to encode the CT image x to be evaluated to obtain a high-dimensional feature representation c; S12, using a decoder to decode the high-dimensional feature representation c to obtain a reconstructed image z; S13. The CT image x to be evaluated and the reconstructed image z are used as inputs of the discriminator to perform generative adversarial training to obtain a pseudo reference image.

4. The CT image quality assessment method based on pre-noise reduction according to claim 1, characterized in that: The S24 includes: S241, calculating the difference between the CT image to be evaluated and the pseudo reference image as the artifact active area; S242, obtaining a region of interest in the CT image to be evaluated as a region of interest mask; S243, calculating the convolution sum of the artifact active area and the mask of the region of interest as the first local evaluation result; S244, calculating the contrast-to-noise ratio of the region of interest to the background according to the CT image to be evaluated and the mask of the region of interest as a second local evaluation result; S245. The first local evaluation result and the second local evaluation result are taken together as a local evaluation result.

5. The CT image quality assessment method based on pre-noise reduction according to claim 4, characterized in that: The S244 includes: S2441, taking the area outside the body in the CT image to be evaluated as a background area; S2442, respectively calculate the average signal strength in the mask region of interest and the background region, and the expression is: Among them, SI ROI is the average signal intensity of the mask in the region of interest, SI ROI,all is the intensity value of all pixels in the mask of the region of interest, N ROI is the number of pixels of the mask of the region of interest; SI b is the average signal intensity of the background area, SI b,all is the intensity value of all pixels in the background area, N b is the number of pixels in the background area; S2443. Calculate the contrast-to-noise ratio of the region of interest to the background according to the average value of the signal strength in the region of interest mask and the background region as the second local evaluation result.

6. The CT image quality assessment method based on pre-noise reduction according to claim 5, characterized in that: In S2443, the expression for calculating the contrast-to-noise ratio CNR of the region of interest to the background is: Among them, SD b is the standard deviation of the signal intensity in the background area.

Images