A PET image processing method, device, equipment and storage medium

Abstract

The present application discloses a PET image processing method, device, equipment and storage medium, including: acquiring matched PET image data pairs through a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images; processing the target image set according to a preset preprocessing method to obtain a target training set, constructing a neural network through a preset code, and determining a preset loss function; inputting the preset quality image into the neural network to obtain a corresponding output image, calculating the difference between the output image and the standard quality image through a preset loss function to obtain a target loss value; after back-propagating the target loss value, adjusting the neural network parameters based on the target loss value to obtain stable parameters and thus determine the target neural network. The present application determines stable parameters by calculating the target loss value between the output image and the standard quality image, determines the target neural network, improves the image signal-to-noise ratio and image quality, and eliminates artifacts.

Description

Technical Field

[0001] The present invention relates to the technical field of PET image reconstruction, and in particular to a PET image processing method, device, equipment and storage medium. Background Art

[0002] PET (Positron Emission Computed Tomography) is an imaging technology that collects decay positron signals for imaging during biological metabolism. Since the time required to collect positron signals depends on the decay half-life of the radioactive isotope and the effective time interval of the signal collector, and the subsequent reconstruction algorithm has a huge amount of computation, the system faces the problems of long imaging time and low image spatial resolution during use. Currently, the imaging time is shortened by shortening the signal acquisition time, but this method has problems such as low signal-to-noise ratio and artifacts.

[0003] The existing imaging algorithms that can solve the above-mentioned problems of low signal-to-noise ratio and artifacts and improve image quality can be divided into three categories: (1) iterative reconstruction algorithms, (2) image filtering and post-processing, and (3) machine learning. Traditional machine learning methods are data-driven and use methods such as mapping-based sparse representation, semi-supervised triple dictionary, and multi-polar canonical correlation analysis to predict matching standard-duration PET images from fast PET images. In recent years, due to the great potential of deep learning in computer vision applications, this technology has been applied to image analysis, such as image segmentation and image registration. Deep learning performs well in the task of generating standard images from PET images with short data acquisition times. However, due to different radioactive half-lives, different intake doses, and different absorption sites of different contrast agents, the generated images have large feature differences. Existing deep learning methods are only applicable to one contrast agent. For images generated by multiple contrast agents, they can only be adapted to different clinical scenarios by training different models. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a PET image processing method, device, equipment and storage medium, which can adjust the neural network parameters based on the target loss value to obtain stable parameters, thereby determining the target neural network, improving the image signal-to-noise ratio and image quality, eliminating artifacts, and is suitable for imaging with a variety of contrast agents. The specific scheme is as follows:

[0005] In a first aspect, the present application discloses a PET image processing method, comprising:

[0006] Acquiring matched PET image data pairs by a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images;

[0007] Performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set, and constructing a neural network through a preset code to determine a preset loss function;

[0008] Inputting the preset quality image into the neural network to obtain a corresponding output image, and calculating a target loss value through the preset loss function based on the output image and the standard quality image;

[0009] Back-propagating the target loss value, and adjusting the neural network parameters based on the target loss value to obtain stable parameters, and determining the target neural network based on the stable parameters;

[0010] Wherein, determining the preset loss function includes:

[0011] Determine any one function in the first loss function set as the first target loss function; the first loss function set includes an L1 loss function, an L2 loss function and a Bce loss function;

[0012] Determine any one or a combination of several functions in the second loss function set as the second target loss function; the second loss function set includes an SSIM loss function, a GAN loss function, a Branch loss function, and a Deepsupervision loss function;

[0013] Determine any one or a combination of several functions in the third loss function set as the third target loss function; the third loss function set includes a lesion attention loss function, a filtered lesion attention loss function, a weighted lesion attention loss function and a SUV feature loss function;

[0014] The first objective loss function, the second objective loss function and the third objective loss function are combined according to a preset weight ratio to obtain a preset loss function.

[0015] Optionally, the matched PET image data pairs are acquired by a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images, including:

[0016] Acquire standard quality images and preset quality images generated based on different doses of preset medicines; wherein the image generated based on the standard dose of preset medicine is the standard quality image, and the image generated based on the preset dose of preset medicine is the preset quality image, and the preset dose is less than the standard dose;

[0017] Or, collecting standard quality images and preset quality images generated based on different acquisition times; wherein the image generated based on the standard acquisition time is the standard quality image, and the image generated based on the preset acquisition time is the preset quality image;

[0018] The standard quality image and the preset quality image are determined as a matched PET image data pair to obtain a target image set.

[0019] Optionally, performing a data preprocessing operation on the target image set according to a preset preprocessing method to obtain a target training set includes:

[0020] Calculating a signal mean of the preset quality image, and calculating the preset quality image and the signal mean in the target image set by using a first preset normalization method to obtain a normalized preset quality image;

[0021] Calculating the standard quality image and the signal mean in the target image set by using a second preset normalization method to obtain a normalized standard quality image;

[0022] Using a preset image adjustment algorithm, the sizes of the normalized preset quality image and the normalized standard quality image are adjusted to obtain an adjusted image, and a target training set is obtained based on the adjusted image;

[0023] The calculation formula of the first preset normalization method is:

[0024] X′=X / X_Mean;

[0025] The calculation formula of the second preset normalization method is:

[0026] Y′=Y / X_Mean;

[0027] Wherein, X is the preset quality image in the target image set, Y is the standard quality image in the target image set, X′ is the normalized preset quality image, Y′ is the normalized standard quality image, and X_Mean is the signal mean of the preset quality image X;

[0028] The calculation formula of the preset image adjustment algorithm is:

[0029] H1=ceil(h / pixel_unit)*pixel_unit;

[0030] H2=floor(h / pixel_unit)*pixel_unit;

[0031]

[0032] Wherein, ceil is a rounding function, floor is a rounding function, pixel_unit is a preset network structure characteristic value, H1 is the length of the adjusted image, H2 is the width of the adjusted image, h is the length of the normalized image, and H′ is the calculated optimal length.

[0033] Optionally, any one or a combination of several functions in the third loss function set is determined as the third target loss function; the third loss function set includes a lesion attention loss function, a filtered lesion attention loss function, a weighted lesion attention loss function and a SUV feature loss function, including:

[0034] The calculation formula of the lesion attention loss function is:

[0035]

[0036] Where L1 is the mean absolute error, lesion mask is the label value indicating whether the current part is a lesion, and x i Represents the pixel value of the image output by the neural network, y i It is represented as the pixel value of the standard quality image, n is the number of pixels in the lesion, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, ∑ pixel i mlesion |x i -y i | is represented by the sum of the pixel value errors corresponding to all the pixel points in the lesion; L lesion Loss of attention for the lesion;

[0037] The calculation formula of the filtering lesion attention loss function is:

[0038]

[0039] Where, L1 is the mean absolute error, L lesion is the lesion attention loss, filter represents the weight of each lesion determined by the filter, lesion mask is the label value showing whether the current part is the lesion, x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, n is the number of pixels in the lesion, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, |xi -y i |*filter is the first processed error obtained after filtering the pixel value error, ∑ pixel i inlesion |x i -y i |*filter is the sum of the first processed errors corresponding to all pixels in the lesion;

[0040] The calculation formula of the weighted lesion attention loss function is:

[0041]

[0042] Where, L1 is the mean absolute error, L lesion is the lesion attention loss, lesion mask is the label value showing whether the current part is the lesion, filter is the weight of each part of the lesion determined by the filter, pixel is the pixel value of the standard quality image, and x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, y i ·|x i -y i |*filter is the weight of the pixel value error and the pixel value of the standard quality image, and then the second processed error is obtained by filtering the result, ∑ pixel i in lesion y i ·|x i -y i |*filter is the sum of the second processed errors corresponding to all pixels in the lesion;

[0043] The calculation formula of the SUV feature loss function is:

[0044] L SUVmax =|max(x i )-max(y i )| lesion SUVmap ;

[0045] Where, L SUVmax is the SUVmax error loss; SUVmax is the maximum value of the standard absorbed dose, x iis the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, |max(x i )-max(y i )| is the maximum error between the maximum pixel value of the image output by the neural network in the lesion and the maximum pixel value of the standard quality image, |max(x i )-max(y i )| lesion SUVmap is the maximum error after conversion into SUV image.

[0046] Optionally, constructing a neural network by using a preset code includes:

[0047] A neural network including a feature extraction network, an image optimization network and an image discrimination network is constructed by a preset code; the input of the neural network is a single-layer preset quality image or multiple layers of the preset quality images.

[0048] Optionally, the inputting the preset quality image into the neural network to obtain a corresponding output image, and obtaining a target loss value by calculating the preset loss function based on the output image and the standard quality image, comprises:

[0049] Inputting the preset quality image in the target training set into the feature extraction network to obtain a feature image and a noise feature map of the preset quality image;

[0050] Inputting the characteristic image and the noise characteristic image into the image discrimination network to obtain the medium type corresponding to the preset quality image;

[0051] Using the medium type to train the feature extraction network and the image discrimination network to obtain a trained feature extraction network and a trained image discrimination network;

[0052] Channel-superimpose the feature image, the noise feature map and the preset quality image to obtain a superimposed image, and input the superimposed image into the image optimization network so that the image optimization network processes the superimposed image by a processing method corresponding to a preset network structure to obtain the corresponding output image;

[0053] The output image is compared with the standard quality image, and when the comparison result exceeds a preset range, a target loss value is calculated based on the output image and the standard quality image and by using the preset loss function; wherein a current loss value calculated by using the preset loss function based on a signal value of the output image and a signal value of the standard quality image is the comparison result;

[0054] When the comparison result is within the preset range, the target loss value is not calculated.

[0055] Optionally, back-propagating the target loss value, adjusting the neural network parameters based on the target loss value to obtain stable parameters, and determining the target neural network based on the stable parameters includes:

[0056] Back-propagating the target loss value to the neural network so as to adjust the neural network parameters based on the target loss value by stochastic gradient modification or adaptive moment estimation optimizer;

[0057] Calculating the preset quality image by adjusting the neural network parameters to obtain the corresponding output image, and re-comparing the output image with the standard quality image to obtain the corresponding comparison result;

[0058] When the comparison result is within a preset range, the adjusted neural network parameters are determined as stable parameters, and the target neural network is determined based on the stable parameters.

[0059] In a second aspect, the present application discloses a PET image processing device, comprising:

[0060] An image set acquisition module, used to acquire matched PET image data pairs through a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images;

[0061] An image preprocessing module, used to perform data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set;

[0062] Neural network building module, used to build neural networks through preset codes;

[0063] A loss function determination module, used to determine a preset loss function;

[0064] A loss value determination module, used for inputting the preset quality image into the neural network to obtain a corresponding output image, and calculating a target loss value through the preset loss function based on the output image and the standard quality image;

[0065] A parameter adjustment module, configured to back-propagate the target loss value, and adjust the neural network parameters based on the target loss value to obtain stable parameters, and determine the target neural network based on the stable parameters;

[0066] Wherein, the loss function determination module is specifically used for:

[0067] Determine any one function in the first loss function set as the first target loss function; the first loss function set includes an L1 loss function, an L2 loss function and a Bce loss function;

[0068] Determine any one or a combination of several functions in the second loss function set as the second target loss function; the second loss function set includes an SSIM loss function, a GAN loss function, a Branch loss function, and a Deepsupervision loss function;

[0069] Determine any one or a combination of several functions in the third loss function set as the third target loss function; the third loss function set includes a lesion attention loss function, a filtered lesion attention loss function, a weighted lesion attention loss function and a SUV feature loss function;

[0070] The first objective loss function, the second objective loss function and the third objective loss function are combined according to a preset weight ratio to obtain a preset loss function.

[0071] In a third aspect, the present application discloses an electronic device, comprising:

[0072] Memory, used to store computer programs;

[0073] The processor is used to execute the computer program to implement the steps of the PET image processing method disclosed above.

[0074] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the PET image processing method disclosed above is implemented.

[0075] It can be seen that the present application provides a PET image processing method, including: collecting matched PET image data pairs through a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images; performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set, and constructing a neural network through a preset code to determine a preset loss function; inputting the preset quality image into the neural network to obtain a corresponding output image, and based on the output image and the standard quality image, calculating a target loss value through the preset loss function; back-propagating the target loss value, and adjusting the neural network parameters based on the target loss value to obtain a stable parameter, and determining a target neural network based on the stable parameter. It can be seen that the present application obtains a corresponding output image by inputting the preset quality image in the target training set into the neural network for calculation. When there is a difference between the output image and the standard quality image, it indicates that the neural network parameters in the neural network need to be adjusted, and then the target loss value is calculated through a preset loss function, and the neural network parameters are adjusted based on the target loss value to obtain stable parameters and their corresponding target neural network. The image signal-to-noise ratio and image quality are improved, and artifacts are eliminated. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0077] Figure 1 This is a flowchart of the implementation of a PET image processing method embodiment 1 disclosed in this application;

[0078] Figure 2 This is a flowchart of the implementation of Embodiment 2 of a PET image processing method disclosed in this application;

[0079] Figure 3 This is a flowchart of the implementation of Embodiment 3 of a PET image processing method disclosed in this application;

[0080] Figure 4 A schematic diagram of the structure of the PET image processing device provided in this application;

[0081] Figure 5 A structural diagram of an electronic device provided for this application. DETAILED DESCRIPTION

[0082] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0083] At present, the imaging duration is shortened by shortening the signal acquisition time. This method has problems such as low signal-to-noise ratio and artifacts. The current imaging algorithms that can solve the above-mentioned low signal-to-noise ratio, artifacts and other problems and improve image quality can be divided into three categories: (1) iterative reconstruction algorithms, (2) image filtering and post-processing, and (3) machine learning. Traditional machine learning methods are data-driven, using methods such as mapping-based sparse representation, semi-supervised triple dictionary, and multi-polar canonical correlation analysis to complete the prediction of matching standard-duration PET images from fast PET images. In recent years, due to the great potential of deep learning in computer vision applications, this technology has been applied to image analysis, such as image segmentation and image registration. In the task of generating standard images from PET images with shorter data acquisition time, deep learning performs well. To this end, the present application provides a PET image processing method that can improve image signal-to-noise ratio and image quality and eliminate artifacts.

[0084] Embodiment 1 of the present invention discloses a PET image processing method, see Figure 1 As shown, the method includes:

[0085] Step S11: Acquire the matched PET image data pairs by a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images.

[0086] In this embodiment, matched PET image data pairs are collected by a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images. It can be understood that the matched PET image data pairs refer to images whose real spatial positions corresponding to the pixel points at the same coordinates in the image matrix are consistent. It should be pointed out that the standard quality image is an image with a clarity greater than a first preset threshold, and the preset quality image is an image with a clarity less than a second preset threshold. In addition, the first preset threshold is greater than or equal to the second preset threshold.

[0087] In a specific embodiment, a preset acquisition mode for acquiring matched PET image data pairs is to reconstruct the standard quality image by using a first number of data and to reconstruct the preset quality image by using a second number of data in listmode data of the same PET scan. It should be noted that the first number is greater than the second number.

[0088] In another specific embodiment, a preset acquisition method for acquiring matched PET image data pairs is to perform two consecutive PET scans with different acquisition times on the same target body, and determine the image obtained by the scan with the longer acquisition time of the two scans as the standard quality image, and determine the image obtained by the scan with the shorter acquisition time of the two scans as the preset quality image.

[0089] Step S12: performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set, and constructing a neural network through a preset code to determine a preset loss function.

[0090] In this embodiment, the target image set is subjected to data preprocessing according to a preset preprocessing method to obtain a target training set, and a neural network is constructed through a preset code to determine a preset loss function. It should be noted that the preset preprocessing method includes methods such as image normalization, image slicing, and data amplification.

[0091] It should be pointed out that, in a specific embodiment, the image normalization method is to use the maximum value of all pixel values ​​of the PET image or a certain percentage of the larger value as the normalization maximum value max, and use the minimum value of all pixel values ​​of the PET image or a certain percentage of the smaller value as the normalization minimum value min, and normalize the pixel value x of the PET image to obtain the corresponding pixel value y. For example, the certain percentage ranges from 0 to 10%.

[0092] The normalization formula is:

[0093] Y=(x-min) / (max-min) or Y=(max-x) / (max-min).

[0094] In another specific embodiment, the image normalization method is to perform normalization based on the average value mean and standard deviation std of all pixel values ​​of the PET image, and normalize the pixel value x of the PET image to obtain the corresponding pixel value y.

[0095] The normalization formula is:

[0096] y = (x-mean) / std.

[0097] It can be understood that the neural network directly skips connections between different convolutional network modules, dense network modules, residual network modules, downsampling modules, upsampling modules and attention mechanism modules. Or, the modules of the neural network with multiple layers connected in series include convolutional network modules, dense network modules, residual network modules, downsampling modules, upsampling modules and attention mechanism modules.

[0098] It can be understood that the preset loss function is determined, specifically, any one function in the first loss function set is determined as the first target loss function; the first loss function set includes L1 loss function (L1 loss, minimizing absolute error), L2 loss function (L2 loss, minimizing square error) and Bce loss function; any one or a combination of several functions in the second loss function set is determined as the second target loss function; the second loss function set includes SSIM loss function (image quality loss function), GAN loss function, Branch loss function and Deep supervision loss function (deep supervision loss); any one or a combination of several functions in the third loss function set is determined as the third target loss function; the third loss function set includes lesion attention loss function, filtered lesion attention loss function, weighted lesion attention loss function and SUV feature loss function; the first target loss function, the second target loss function and the third target loss function are combined according to a preset weight ratio to obtain a preset loss function.

[0099] Step S13: inputting the preset quality image into the neural network to obtain a corresponding output image, and based on the output image and the standard quality image, calculating a target loss value through the preset loss function.

[0100] In this embodiment, the preset quality image is input into the neural network to obtain a corresponding output image, and based on the output image and the standard quality image, the target loss value is calculated by the preset loss function. It can be understood that the preset quality image is input into the neural network to obtain a corresponding output image. At this time, based on the difference between the output image and the standard quality image, it is judged whether the calculation parameters in the current neural network need to be adjusted. When the difference between the output image and the standard quality image is greater than the preset difference range, it indicates that the calculation parameters in the current neural network need to be adjusted. Then, based on the output image and the standard quality image, the target loss value is calculated by the preset loss function.

[0101] Step S14: back-propagate the target loss value, and adjust the neural network parameters based on the target loss value to obtain stable parameters, and determine the target neural network based on the stable parameters.

[0102] In this embodiment, after the target loss value is calculated by the preset loss function, the target loss value is back-propagated to the neural network, and then the neural network parameters are adjusted based on the target loss value to obtain stable parameters, and the target neural network is determined based on the stable parameters. It can be understood that the method of adjusting the neural network parameters includes a stochastic gradient modification method and an Adam (adaptive moment estimation) optimizer to iteratively train the neural network weights.

[0103] It can be seen that the present application provides a PET image processing method, including: collecting matched PET image data pairs through a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images; performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set, and constructing a neural network through a preset code to determine a preset loss function; inputting the preset quality image into the neural network to obtain a corresponding output image, and based on the output image and the standard quality image, calculating a target loss value through the preset loss function; back-propagating the target loss value, and adjusting the neural network parameters based on the target loss value to obtain a stable parameter, and determining a target neural network based on the stable parameter. It can be seen that the present application obtains a corresponding output image by inputting the preset quality image in the target training set into the neural network for calculation. When there is a difference between the output image and the standard quality image, it indicates that the neural network parameters in the neural network need to be adjusted, and then the target loss value is calculated through a preset loss function, and the neural network parameters are adjusted based on the target loss value to obtain stable parameters and their corresponding target neural network. The image signal-to-noise ratio and image quality are improved, and artifacts are eliminated.

[0104] See also Figure 2 As shown, the second embodiment of the present invention discloses a PET image processing method. Compared with the previous embodiment, this embodiment further illustrates and optimizes the technical solution.

[0105] Step S21: Acquire standard quality images and preset quality images.

[0106] In this embodiment, standard quality images and preset quality images are collected. It can be understood that, in a specific embodiment, standard quality images and preset quality images generated based on different doses of preset drugs are collected; wherein, the image generated based on the standard dose preset drug is the standard quality image, and the image generated based on the preset dose preset drug is the preset quality image, and the preset dose is less than the standard dose. In another specific embodiment, standard quality images and preset quality images generated based on different acquisition times are collected; wherein, the image generated based on the standard acquisition time is the standard quality image, and the image generated based on the preset acquisition time is the preset quality image.

[0107] Step S22: determining the standard quality image and the preset quality image as a matched PET image data pair to obtain a target image set.

[0108] In this embodiment, after obtaining the preset quality image corresponding to the standard quality image, the standard quality image and the preset quality image are determined as a matched PET image data pair to obtain a target image set. It should be noted that when acquiring the matched PET image data pair, CT (Computer Tomography) images and MR (Magnetic Resonance Imaging) images are acquired at the same time. That is, the target image set includes the matched PET image data pair, the CT image, and the MR image.

[0109] Step S23: performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set.

[0110] In this embodiment, a data preprocessing operation is performed on the target image set according to an image normalization method to obtain a target training set.

[0111] In a specific implementation, the signal mean of the preset quality image is calculated, and the preset quality image and the signal mean are calculated using a first preset normalization method to obtain a normalized preset quality image; the standard quality image and the signal mean are calculated using a second preset normalization method to obtain a normalized standard quality image; the size of the normalized preset quality image and the normalized standard quality image is adjusted using a preset image adjustment algorithm to obtain an adjusted image, and a target training set is obtained based on the adjusted image.

[0112] It can be understood that the signal mean X_Mean of the preset quality image X is calculated, and then the preset quality image is normalized using the first preset normalization method to obtain the normalized preset quality image; the signal mean X_Mean of the preset quality image X and the standard quality image Y are calculated using the second preset normalization method to obtain the normalized standard quality image

[0113] The calculation formula of the first preset normalization method is:

[0114] X′=X / X_Mean;

[0115] The calculation formula of the second preset normalization method is:

[0116] Y′=Y / X_Mean;

[0117] Wherein, X is the preset quality image in the target image set, Y is the standard quality image in the target image set, X′ is the normalized preset quality image, Y′ is the normalized standard quality image, and X_Mean is the signal mean of the preset quality image X.

[0118] It can be understood that the calculation formula of the preset image adjustment algorithm is:

[0119] H1=ceil(h / pixel_unit)*pixel_unit;

[0120] H2=floor(h / pixel_unit)*pixel_unit;

[0121]

[0122] Wherein, ceil is a rounding function, floor is a rounding function, pixel_unit is a preset network structure characteristic value, H1 is the length of the adjusted image, H2 is the width of the adjusted image, h is the length of the normalized image, and H′ is the calculated optimal length.

[0123] In another specific embodiment, the image normalization method is to perform a normalization operation on the mean of all pixel points of the PET image or all pixel points of the PET foreground image in the target image set to obtain a target training set.

[0124] The normalization formula is:

[0125] Y=x / mean.

[0126] Step S24: construct a neural network through a preset code and determine a preset loss function.

[0127] In this embodiment, a neural network is constructed by a preset code. It can be understood that the input of the neural network is a single layer of the preset quality image or multiple layers of the preset quality image, and there is no bypass input; or, the input of the neural network is a single layer of the preset quality image or multiple layers of the preset quality image, and at the same time, any one or a combination of several items in the gradient map of the computer X-ray tomography image, the magnetic resonance imaging image, the computer X-ray tomography image after single-layer convolution processing, the computer X-ray tomography image after multi-layer convolution processing, the magnetic resonance imaging image after single-layer convolution processing, the magnetic resonance imaging image after multi-layer convolution processing, and the preset quality image is used as a bypass input. The output of the neural network may be a single layer or multiple layers of the standard quality image. It should be pointed out that the neural network includes a feature extraction network, an image optimization network, and an image discrimination network.

[0128] It can be understood that any one function in the first loss function set is determined as the first target loss function; the first loss function set includes L1 loss function, L2 loss function and Bce loss function; any one or a combination of several functions in the second loss function set is determined as the second target loss function; the second loss function set includes SSIM loss function, GAN loss function, Branch loss function and Deep supervision loss function; any one or a combination of several functions in the third loss function set is determined as the third target loss function; the third loss function set includes lesion attention loss function, filtered lesion attention loss function, weighted lesion attention loss function and SUV feature loss function; the first target loss function, the second target loss function and the third target loss function are combined according to a preset weight ratio to obtain a preset loss function.

[0129] Among them, the calculation formula of the lesion attention loss function is:

[0130]

[0131] Where L1 is the mean absolute error, lesion mask is the label value indicating whether the current part is a lesion, and x i Represents the pixel value of the image output by the neural network, y i It is represented as the pixel value of the standard quality image, n is the number of pixels in the lesion, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, ∑ pixel i inlesion |x i -yi | is represented by the sum of the pixel value errors corresponding to all the pixel points in the lesion; L lesion Loss of attention for the lesion;

[0132] The calculation formula of the filtering lesion attention loss function is:

[0133]

[0134] Where, L1 is the mean absolute error, L lesion is the lesion attention loss, filter represents the weight of each lesion determined by the filter, lesion mask is the label value showing whether the current part is the lesion, x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, n is the number of pixels in the lesion, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, |x i -y i |*filter is the first processed error obtained after filtering the pixel value error, ∑ pixel i inlesion |x i -y i |*filter is the sum of the first processed errors corresponding to all pixels in the lesion;

[0135] The calculation formula of the weighted lesion attention loss function is:

[0136]

[0137] Where, L1 is the mean absolute error, L lesion is the lesion attention loss, lesion mask is the label value showing whether the current part is the lesion, filter is the weight of each part of the lesion determined by the filter, pixel is the pixel value of the standard quality image, and x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, y i ·|xi -y i |*filter is the weight of the pixel value error and the pixel value of the standard quality image, and then the second processed error is obtained by filtering the result, ∑ pixel i in lesion y i ·|x i -y i |*filter is the sum of the second processed errors corresponding to all pixels in the lesion;

[0138] The calculation formula of the SUV feature loss function is:

[0139] L SUVmax =|max(x i )-max(y i )| lesion SUVmap ;

[0140] Where, L SUVmax is the SUVmax error loss; SUVmax is the maximum value of the standard absorbed dose, x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, |max(x i )-max(y i )| is the maximum error between the maximum pixel value of the image output by the neural network in the lesion and the maximum pixel value of the standard quality image, |max(x i )-max(y i )| lesion SUVmap is the maximum error after conversion into SUV image.

[0141] It should be pointed out that the SUV feature loss includes the SUV mean feature loss and the SUV max feature loss. The SUV mean feature loss has the same definition as the lesion attention loss. The calculation formula of the SUV max feature loss is:

[0142] L SUVmax =|max(x i )-max(y i )| lesion SUVmap .

[0143] It can be understood that the lesion mask is a label value indicating whether the current part is the lesion. For example, the label is a (0, 1) label, the label value of the lesion part is 1, and the label value of the non-lesion part is 0.

[0144] Step S25: Input the preset quality image into the neural network to obtain a corresponding output image, and based on the output image and the standard quality image, calculate the target loss value through the preset loss function.

[0145] Step S26: back-propagate the target loss value, and adjust the neural network parameters based on the target loss value to obtain stable parameters, and determine the target neural network based on the stable parameters.

[0146] For the specific contents of the above steps S25 and S26, reference may be made to the corresponding contents disclosed in the above embodiments, which will not be repeated here.

[0147] It can be seen that the embodiment of the present application processes the standard quality image by using a preset processing method to obtain a preset quality image corresponding to the standard quality image, and simultaneously collects the CT image and the MR image to obtain a target image set. The target image set is subjected to data preprocessing operations according to the image normalization method to obtain a target training set, and then a neural network is constructed through a preset code, and a preset loss function is determined. The preset quality image is input into the neural network to obtain a corresponding output image, and based on the output image and the standard quality image, a target loss value is calculated through the preset loss function, and a target neural network is determined based on the target loss value. The image signal-to-noise ratio and image quality are improved, and artifacts are eliminated.

[0148] See also Figure 3 As shown, the third embodiment of the present invention discloses a PET image processing method. Compared with the previous embodiment, this embodiment further illustrates and optimizes the technical solution.

[0149] Step S31: Acquire the matched PET image data pairs by a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images.

[0150] Step S32: performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set, and constructing a neural network through a preset code to determine a preset loss function.

[0151] Step S33: inputting the preset quality image into the neural network to obtain a corresponding output image.

[0152] In this embodiment, a neural network is constructed by a preset code, and after determining a preset loss function, the preset quality image is input into the neural network, and the corresponding output image is obtained through calculation by the neural network. It can be understood that a single layer of the preset quality image or multiple layers of the preset quality image are input into the neural network without bypass input, or a single layer of the preset quality image or multiple layers of the preset quality image are input into the neural network, and at the same time, any one or a combination of several items of the computer X-ray tomography image, the magnetic resonance imaging image, the computer X-ray tomography image after single-layer convolution processing, the computer X-ray tomography image after multi-layer convolution processing, the magnetic resonance imaging image after single-layer convolution processing, the magnetic resonance imaging image after multi-layer convolution processing, and the gradient map of the preset quality image are bypassed and input into the neural network, and then the corresponding output image is obtained through calculation of parameters in the neural network. It should be pointed out that the output image can be a single layer or multiple layers of the standard quality image.

[0153] It can be understood that the preset quality image in the target training set is input into the feature extraction network to obtain the feature image and noise feature map of the preset quality image; the feature image and the noise feature map are input into the image discrimination network to obtain the medium type corresponding to the preset quality image; the feature extraction network and the image discrimination network are trained using the medium type to obtain the trained feature extraction network and the trained image discrimination network; the feature image, the noise feature map and the preset quality image are channel-superimposed to obtain the superimposed image, and the superimposed image is input into the image optimization network so that the image optimization network processes the superimposed image by the processing method corresponding to the preset network structure to obtain the corresponding output image. It should be pointed out that the preset network structure includes UNet with batch normalization layer, U2Net without batch normalization layer, residual network without batch normalization layer and RRDB network (Residual-in-Residual Dense Block).

[0154] Step S34: Compare the output image with the standard quality image. When the comparison result exceeds a preset range, a target loss value is calculated based on the output image and the standard quality image and through the preset loss function; wherein the current loss value calculated based on the signal value of the output image and the signal value of the standard quality image through the preset loss function is the comparison result.

[0155] In this embodiment, after the corresponding output image is obtained through calculation by the neural network, the output image is compared with the standard quality image. When the comparison result exceeds the preset range, the target loss value is calculated based on the output image and the standard quality image and through the preset loss function; wherein, the current loss value calculated by the preset loss function based on the signal value of the output image and the signal value of the standard quality image is the comparison result. It can be understood that after comparing the output image with the standard quality image, a corresponding comparison result is obtained, and then the comparison result is compared with the preset range. When the comparison result is within the preset range, it indicates that the parameters in the neural network do not need to be adjusted at this time, and the output image meets the expected setting, so it is not necessary to calculate the target loss value to adjust the neural network parameters. When the comparison result exceeds the preset range, it indicates that the output image does not meet the expected setting, so it is necessary to calculate the target loss value based on the output image and the standard quality image through the preset loss function. For example, when the target indicator error of the output image is within 5% compared with the standard quality image, it indicates that the output image meets the expected setting and it is not necessary to calculate the target loss value to adjust the neural network parameters. When the target indicator error of the output image exceeds 5% compared with the standard quality image, it indicates that the output image does not meet the expected setting, so the target loss value needs to be calculated. For example, the target indicator error can be an SUV indicator error.

[0156] Step S35: back-propagating the target loss value to the neural network so as to adjust the neural network parameters based on the target loss value through stochastic gradient modification or adaptive moment estimation optimizer.

[0157] In this embodiment, after the target loss value is calculated by the preset loss function, the target loss value is back-propagated to the neural network so that the neural network parameters are adjusted by stochastic gradient modification or adaptive moment estimation optimizer based on the target loss value. It can be understood that the neural network parameters are adjusted by stochastic gradient modification, for example, when the target indicator error of the output image exceeds 5% compared with the standard quality image, the target loss value is calculated. Since the target indicator error at this time exceeds 5% and the target indicator of the output image is greater than the target indicator of the standard quality image, it indicates that the neural network parameters need to be adjusted by gradient descent to reduce the target indicator error. Similarly, when the target indicator error exceeds 5% and the target indicator of the output image is less than the target indicator of the standard quality image, the neural network parameters need to be adjusted by gradient ascent to reduce the target indicator error.

[0158] Step S36: Calculate the preset quality image by adjusting the neural network parameters to obtain the corresponding output image, and re-compare the output image with the standard quality image to obtain the corresponding comparison result.

[0159] In this embodiment, the target loss value is back-propagated to the neural network, and after the neural network parameters are adjusted based on the target loss value, the preset quality image is calculated by the adjusted neural network parameters to obtain the corresponding output image, and the output image is re-compared with the standard quality image to obtain the corresponding comparison result. It can be understood that after the neural network parameters are adjusted based on the target loss value, the preset quality image is input into the neural network again, and the corresponding output image is obtained by calculating the adjusted neural network parameters in the neural network, and the output image is compared with the standard quality image again to obtain the corresponding comparison result.

[0160] Step S37: When the comparison result is within a preset range, the adjusted neural network parameters are determined as stable parameters, and the target neural network is determined based on the stable parameters.

[0161] In this embodiment, it is determined whether the comparison result is within a preset range. When the comparison result is within the preset range, it indicates that the adjusted neural network parameter is the stable parameter, and then the target neural network is determined based on the stable parameter. It can be understood that the target neural network is the trained neural network. For example, first collect FDG-PET (deoxyglucose positron emission tomography) scan data of 40 patients with lymphoma or lung tumors, each of which has a scan time of 1.5 minutes, and reconstruct the preset quality image by 1 / 3 listmode data, and then collect PSMA-PET (prostate cancer specific membrane antigen) scan data of 40 patients with prostate cancer, and reconstruct the standard quality image by all listmode data, and pre-process the preset quality images and the standard quality images of 60 groups of patients, slice the image, and perform data enhancement processing, and use the processed image as the input and output of the neural network, and then train the image discrimination network for judging the image medium, and then use L1 loss, SSIM loss, deep supervision loss, filter lesion attention loss and SUV feature loss with different weights as loss functions to train the image optimization network. The trained neural network can significantly improve the image quality of the generated images and ensure that the SUV features of the generated images are consistent with those of the standard quality images.

[0162] For the specific contents of the above steps S31 and S32, reference may be made to the corresponding contents disclosed in the above embodiments, which will not be repeated here.

[0163] It can be seen that the embodiment of the present application obtains a corresponding comparison result by comparing the output image with the standard quality image, and then compares the comparison result with the preset range. When the comparison result exceeds the preset range, the target loss value is calculated based on the output image and the standard quality image through the preset loss function, and then the target loss value is back-propagated to the neural network, and the neural network parameters are adjusted based on the target loss value. The preset quality image is calculated by adjusting the neural network parameters to obtain the corresponding output image, and the output image is re-compared with the standard quality image to obtain the corresponding comparison result. When the comparison result is within the preset range, the adjusted neural network parameters are determined as stable parameters, and the target neural network is determined based on the stable parameters. The image signal-to-noise ratio and image quality are improved, and artifacts are eliminated.

[0164] See also Figure 4 As shown, the embodiment of the present application also discloses a PET image processing device, including:

[0165] An image set acquisition module 11 is used to acquire matched PET image data pairs through a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images;

[0166] An image preprocessing module 12 is used to perform data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set;

[0167] A neural network construction module 13, used to construct a neural network through a preset code;

[0168] A loss function determination module 14, used to determine a preset loss function;

[0169] A loss value determination module 15, configured to input the preset quality image into the neural network to obtain a corresponding output image, and obtain a target loss value by calculating the preset loss function based on the output image and the standard quality image;

[0170] A parameter adjustment module 16, configured to back-propagate the target loss value, and adjust the neural network parameters based on the target loss value to obtain stable parameters, and determine the target neural network based on the stable parameters;

[0171] The loss function determination module 14 is specifically used for:

[0172] Determine any one function in the first loss function set as the first target loss function; the first loss function set includes an L1 loss function, an L2 loss function and a Bce loss function;

[0173] Determine any one or a combination of several functions in the second loss function set as the second target loss function; the second loss function set includes an SSIM loss function, a GAN loss function, a Branch loss function, and a Deepsupervision loss function;

[0174] Determine any one or a combination of several functions in the third loss function set as the third target loss function; the third loss function set includes a lesion attention loss function, a filtered lesion attention loss function, a weighted lesion attention loss function and a SUV feature loss function;

[0175] The first objective loss function, the second objective loss function and the third objective loss function are combined according to a preset weight ratio to obtain a preset loss function.

[0176] It can be seen that the present application includes: collecting matched PET image data pairs through a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images; performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set, and constructing a neural network through a preset code to determine a preset loss function; inputting the preset quality image into the neural network to obtain a corresponding output image, and calculating a target loss value based on the output image and the standard quality image through the preset loss function; back-propagating the target loss value, and adjusting the neural network parameters based on the target loss value to obtain a stable parameter, and determining the target neural network based on the stable parameter. It can be seen that the present application calculates by inputting the preset quality image in the target training set into the neural network to obtain a corresponding output image, and when there is a difference between the output image and the standard quality image, it indicates that the neural network parameters in the neural network need to be adjusted, and then the target loss value is calculated through a preset loss function, and the neural network parameters are adjusted based on the target loss value to obtain stable parameters and their corresponding target neural network. The image signal-to-noise ratio and image quality are improved, and artifacts are eliminated.

[0177] In some specific embodiments, the image set acquisition module 11 specifically includes:

[0178] A standard quality image acquisition unit, used for acquiring standard quality images;

[0179] A preset quality image acquisition unit, used to process the standard quality image using a preset processing method to obtain a preset quality image corresponding to the standard quality image;

[0180] a data pair determination unit, configured to determine the standard quality image and the preset quality image as a matched PET image data pair;

[0181] The image acquisition unit is used to acquire computer tomography images and magnetic resonance imaging images.

[0182] In some specific embodiments, the loss value determination module 15 specifically includes:

[0183] An output image acquisition unit, used for inputting the preset quality image into the neural network, and obtaining the corresponding output image through calculation by the neural network;

[0184] An image comparison unit, used for comparing the output image with the standard quality image to obtain a corresponding comparison result;

[0185] The first result judgment unit is used to judge whether the comparison result is within a preset range. When the comparison result exceeds the preset range, a target loss value is calculated based on the output image and the standard quality image by using the preset loss function. When the comparison result is within the preset range, the target loss value is not calculated.

[0186] In some specific embodiments, the parameter adjustment module 16 specifically includes:

[0187] A parameter adjustment unit, configured to back-propagate the target loss value to the neural network so as to adjust the neural network parameters by stochastic gradient modification or adaptive moment estimation optimizer based on the target loss value;

[0188] A calculation unit, configured to calculate the preset quality image by adjusting the neural network parameters to obtain the corresponding output image, and re-compare the output image with the standard quality image to obtain the corresponding comparison result;

[0189] The second result judgment unit is used to judge whether the comparison result is within a preset range. When the comparison result is within the preset range, the adjusted neural network parameters are determined as stable parameters, and the target neural network is determined based on the stable parameters.

[0190] Furthermore, an embodiment of the present application also provides an electronic device. Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram cannot be considered as any limitation on the scope of use of the present application.

[0191] Figure 5 The present invention provides a schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the PET image processing method disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0192] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0193] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0194] The operating system 221 is used to manage and control the hardware devices and computer programs 222 on the electronic device 20, and may be Windows Server, Netware, Unix, Linux, etc. In addition to computer programs that can be used to complete the PET image processing method performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 may further include computer programs that can be used to complete other specific tasks.

[0195] Furthermore, an embodiment of the present application also discloses a storage medium, in which a computer program is stored. When the computer program is loaded and executed by a processor, the steps of the PET image processing method disclosed in any of the aforementioned embodiments are implemented.

[0196] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0197] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0198] The PET image processing method, device, equipment and storage medium provided by the present invention are introduced in detail above. The principle and implementation mode of the present invention are explained by using specific examples in this article. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation mode and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A PET image processing method, characterized in that: include: Acquire the matched PET image data pairs by a preset acquisition method to obtain a target image set; The PET image data pair includes a standard quality image and a preset quality image; Performing data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set, and constructing a neural network through a preset code to determine a preset loss function; Inputting the preset quality image into the neural network to obtain a corresponding output image, and calculating a target loss value through the preset loss function based on the output image and the standard quality image; Back-propagating the target loss value, and adjusting the neural network parameters based on the target loss value to obtain stable parameters, and determining the target neural network based on the stable parameters; Wherein, determining the preset loss function includes: Determine any one function in the first loss function set as the first target loss function; the first loss function set includes an L1 loss function, an L2 loss function and a Bce loss function; Determine any one or a combination of several functions in the second loss function set as the second target loss function; the second loss function set includes an SSIM loss function, a GAN loss function, a Branch loss function, and a Deep supervision loss function; Determine any one or a combination of several functions in the third loss function set as the third target loss function; the third loss function set includes a lesion attention loss function, a filtered lesion attention loss function, a weighted lesion attention loss function and a SUV feature loss function; Combining the first objective loss function, the second objective loss function and the third objective loss function according to a preset weight ratio to obtain a preset loss function; The matched PET image data pairs are acquired by a preset acquisition method to obtain a target image set; the PET image data pairs include standard quality images and preset quality images, including: Acquire standard quality images and preset quality images generated based on different doses of preset medicines; wherein the image generated based on the standard dose of preset medicine is the standard quality image, and the image generated based on the preset dose of preset medicine is the preset quality image, and the preset dose is less than the standard dose; Determining the standard quality image and the preset quality image as a matched PET image data pair to obtain a target image set; The neural network is constructed by using preset codes, including: A neural network including a feature extraction network, an image optimization network and an image discrimination network is constructed by a preset code; the input of the neural network is a single layer of the preset quality image or multiple layers of the preset quality image; The step of inputting the preset quality image into the neural network to obtain a corresponding output image, and calculating a target loss value based on the output image and the standard quality image by using the preset loss function, comprises: Inputting the preset quality image in the target training set into the feature extraction network to obtain a feature image and a noise feature map of the preset quality image; Inputting the characteristic image and the noise characteristic image into the image discrimination network to obtain the medium type corresponding to the preset quality image; Using the medium type to train the feature extraction network and the image discrimination network to obtain a trained feature extraction network and a trained image discrimination network; Channel-superimpose the feature image, the noise feature map and the preset quality image to obtain a superimposed image, and input the superimposed image into the image optimization network so that the image optimization network processes the superimposed image by a processing method corresponding to a preset network structure to obtain the corresponding output image; The output image is compared with the standard quality image, and when the comparison result exceeds a preset range, a target loss value is calculated based on the output image and the standard quality image and by using the preset loss function; wherein a current loss value calculated by using the preset loss function based on a signal value of the output image and a signal value of the standard quality image is the comparison result; When the comparison result is within the preset range, the target loss value is not calculated.

2. The PET image processing method according to claim 1, characterized in that: The performing data preprocessing operation on the target image set according to a preset preprocessing method to obtain a target training set includes: Calculating a signal mean of the preset quality image, and calculating the preset quality image and the signal mean in the target image set by using a first preset normalization method to obtain a normalized preset quality image; Calculating the standard quality image and the signal mean in the target image set by using a second preset normalization method to obtain a normalized standard quality image; Using a preset image adjustment algorithm, the sizes of the normalized preset quality image and the normalized standard quality image are adjusted to obtain an adjusted image, and a target training set is obtained based on the adjusted image; The calculation formula of the first preset normalization method is: X′=X / X_Mean; The calculation formula of the second preset normalization method is: Y′=Y / X_Mean; Wherein, X is the preset quality image in the target image set, Y is the standard quality image in the target image set, X′ is the normalized preset quality image, Y′ is the normalized standard quality image, and X_Mean is the signal mean of the preset quality image X; The calculation formula of the preset image adjustment algorithm is: H1=ceil(h / pixel_unit)*pixel_unit; H2=floor(h / pixel_unit)*pixel_unit; Wherein, ceil is a rounding function, floor is a rounding function, pixel_unit is a preset network structure characteristic value, H1 is the length of the adjusted image, H2 is the width of the adjusted image, h is the length of the normalized image, and H′ is the calculated optimal length.

3. The PET image processing method according to any one of claims 1 to 2, characterized in that: The third target loss function is determined as any one or a combination of several functions in the third loss function set; the third loss function set includes a lesion attention loss function, a filtered lesion attention loss function, a weighted lesion attention loss function and a SUV feature loss function, including: The calculation formula of the lesion attention loss function is: Where L1 is the mean absolute error, lesion mask is the label value indicating whether the current part is a lesion, and x i Represents the pixel value of the image output by the neural network, y i It is represented as the pixel value of the standard quality image, n is the number of pixels in the lesion, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, ∑ pixeliinlesion |x i -y i | is represented by the sum of the pixel value errors corresponding to all the pixel points in the lesion; L lesion Table shows focal attention loss; The calculation formula of the filtering lesion attention loss function is: Where, L1 is the mean absolute error, L lesion is the lesion attention loss, filter represents the weight of each lesion determined by the filter, lesion mask is the label value showing whether the current part is the lesion, x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, n is the number of pixels in the lesion, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, |x i -y i |*filter is the first processed error obtained after filtering the pixel value error, ∑ pixeliinlesion |x i -y i |*filter is the sum of the first processed errors corresponding to all pixels in the lesion; The calculation formula of the weighted lesion attention loss function is: Where, L1 is the mean absolute error, L lesion is the lesion attention loss, lesion mask is the label value showing whether the current part is the lesion, filter is the weight of each part of the lesion determined by the filter, pixel is the pixel value of the standard quality image, and x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, |x i -y i | is the pixel value error between the image output by the neural network and the standard quality image, y i ·|x i -y i |*filter is the weight of the pixel value error and the pixel value of the standard quality image, and then the second processed error is obtained by filtering the result, ∑ pixeliinlesion y i ·|x i -y i |*filter is the sum of the second processed errors corresponding to all pixels in the lesion; The calculation formula of the SUV feature loss function is: L SUVmax =|max(x i )-max(y i )| lesionSUVmap ; Where, L SUVmax is the SUVmax error loss; SUVmax is the maximum value of the standard absorbed dose, x i is the pixel value of the image output by the neural network, y i is the pixel value of the standard quality image, |max(x i )-max(y i )| is the maximum error between the maximum pixel value of the image output by the neural network in the lesion and the maximum pixel value of the standard quality image, |max(x i )-max(y i )| lesionSUVmap is the maximum error after conversion into SUV image.

4. The PET image processing method according to claim 1, characterized in that: The step of back-propagating the target loss value and adjusting the neural network parameters based on the target loss value to obtain stable parameters, and determining the target neural network based on the stable parameters comprises: Back-propagating the target loss value to the neural network so as to adjust the neural network parameters based on the target loss value by stochastic gradient modification or adaptive moment estimation optimizer; Calculating the preset quality image by adjusting the neural network parameters to obtain the corresponding output image, and re-comparing the output image with the standard quality image to obtain the corresponding comparison result; When the comparison result is within a preset range, the adjusted neural network parameters are determined as stable parameters, and the target neural network is determined based on the stable parameters.

5. A PET image processing device, characterized in that: include: An image set acquisition module, used to acquire matched PET image data pairs through a preset acquisition method to obtain a target image set; The PET image data pair includes a standard quality image and a preset quality image; An image preprocessing module, used to perform data preprocessing operations on the target image set according to a preset preprocessing method to obtain a target training set; Neural network building module, used to build neural networks through preset codes; A loss function determination module, used to determine a preset loss function; A loss value determination module, used for inputting the preset quality image into the neural network to obtain a corresponding output image, and calculating a target loss value through the preset loss function based on the output image and the standard quality image; A parameter adjustment module, configured to back-propagate the target loss value, and adjust the neural network parameters based on the target loss value to obtain stable parameters, and determine the target neural network based on the stable parameters; Wherein, the loss function determination module is specifically used for: Determine any one function in the first loss function set as the first target loss function; the first loss function set includes an L1 loss function, an L2 loss function and a Bce loss function; Determine any one or a combination of several functions in the second loss function set as the second target loss function; the second loss function set includes an SSIM loss function, a GAN loss function, a Branch loss function, and a Deep supervision loss function; Determine any one or a combination of several functions in the third loss function set as the third target loss function; the third loss function set includes a lesion attention loss function, a filtered lesion attention loss function, a weighted lesion attention loss function and a SUV feature loss function; Combining the first objective loss function, the second objective loss function and the third objective loss function according to a preset weight ratio to obtain a preset loss function; The image set acquisition module is specifically used to collect standard quality images and preset quality images generated based on different doses of preset medicines; wherein, the image generated based on the standard dose of preset medicine is the standard quality image, and the image generated based on the preset dose of preset medicine is the preset quality image, and the preset dose is less than the standard dose; Determining the standard quality image and the preset quality image as a matched PET image data pair to obtain a target image set; The neural network construction module is specifically used to construct a neural network including a feature extraction network, an image optimization network and an image discrimination network through a preset code; the input of the neural network is a single layer of the preset quality image or multiple layers of the preset quality image; The loss value determination module is specifically used to input the preset quality image in the target training set into the feature extraction network to obtain a feature image and a noise feature map of the preset quality image; Inputting the characteristic image and the noise characteristic image into the image discrimination network to obtain the medium type corresponding to the preset quality image; Using the medium type to train the feature extraction network and the image discrimination network to obtain a trained feature extraction network and a trained image discrimination network; Channel-superimpose the feature image, the noise feature map and the preset quality image to obtain a superimposed image, and input the superimposed image into the image optimization network so that the image optimization network processes the superimposed image by a processing method corresponding to a preset network structure to obtain the corresponding output image; The output image is compared with the standard quality image, and when the comparison result exceeds a preset range, a target loss value is calculated based on the output image and the standard quality image and by using the preset loss function; wherein a current loss value calculated by using the preset loss function based on a signal value of the output image and a signal value of the standard quality image is the comparison result; When the comparison result is within the preset range, the target loss value is not calculated.

6. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the PET image processing method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that: Used to store computer programs; wherein, when the computer program is executed by a processor, the PET image processing method according to any one of claims 1 to 4 is implemented.

Images