Lung tumor CT image denoising method, system and device
Through deep self-coding technology, using deep autoencoder denoising network and optimization algorithm, the denoising problem of rare lung tumor CT images was solved, and the image details were maintained and the diagnostic accuracy was improved with low computational complexity.
Patent Information
- Application Number
- CN202510562583.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-09-19
AI Technical Summary
When processing CT images of rare lung tumors, existing technologies have problems such as large computational complexity, multiple parameters, and difficulty in preserving image details, resulting in limited diagnostic accuracy.
A deep autoencoder denoising network is adopted, combined with a structure preserver and an L1-norm sparse regularization term for quantization noise, and the proximal mapping operator and momentum acceleration mechanism are used to optimize the network to construct a lightweight denoising model that preserves image details and reduces computational complexity.
It achieves efficient denoising with limited hardware resources, maintains the structure and texture information of lung tumor CT images, and improves the accuracy and efficiency of diagnosis.
Smart Images
Figure CN120672601A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of medical information and intelligent medicine, and in particular to a method, system and device for denoising lung tumor CT images. Background Art
[0002] Lung tumor CT images play a crucial role in the diagnosis and treatment of lung tumors. They provide a wealth of valuable quantitative features and can deeply mine high-dimensional data within the tumor, providing powerful support for clinical decision-making. However, in actual clinical applications, the acquisition process inevitably generates noise, which can result in a high concentration of noise in lung tumor CT images. This can manifest as blurred image boundaries and overlapping artifacts. These issues can obscure critical details of subtle lesions, seriously compromising lung tumor diagnosis and surgery.
[0003] In recent years, the problem of denoising lung tumor CT images has attracted considerable attention. Traditional methods, such as Kalman filtering and Gaussian filtering, while simple to use, struggle to preserve texture detail. With the development of intelligent technology, deep learning has garnered widespread attention. While deep learning-based methods have overcome the limitations of traditional methods, training deep learning models requires numerous parameters and high computational complexity, rendering them ineffective in rare case scenarios with limited training data.
[0004] In the fields of medical information and intelligent healthcare, some rare cases require relatively small amounts of image data, making deep learning approaches unsuitable. Therefore, there is an urgent need for a lightweight lung tumor CT image denoising method that is computationally efficient, effective, and effective. This approach is crucial for the accuracy of clinical lung tumor diagnosis. Summary of the Invention
[0005] The present invention provides a lung tumor CT image denoising method, system and storage medium, aiming to preserve fine details of the image, meet the requirements of high efficiency and practicality, and have low computational complexity, short processing time and adaptability to general hardware resources.
[0006] The present invention specifically provides the following technical solutions:
[0007] In a first aspect, the present invention provides a method for denoising a lung tumor CT image, the method comprising the following steps:
[0008] Step S1: Deploy a decentralized image acquisition information transmission network to collect and transmit basic information and CT images of lung tumor patients, and store them in a "tree" structure to the main server;
[0009] Step S2: In the main server, a new channel for CT images of lung tumor patients is opened, basic information of lung tumor patients is digitized to form a coding array, and the resulting coding array is stored in the new channel for CT images of lung tumor patients, thereby integrating the basic information of lung tumor patients with the CT images, and dividing the images into a training set and a test set.
[0010] Step S3: Constructing a deep autoencoder denoising network for CT images of lung cancer patients. The deep autoencoder denoising network introduces a structure preserver and designs a target loss function for accurate noise quantization.
[0011] Step S4: using an optimization method that integrates a proximal mapping operator and a momentum acceleration mechanism to train the deep autoencoder denoising network, optimizing the number of network layers, and obtaining a trained deep autoencoder denoising network;
[0012] In step S5, a CT image of a patient with lung tumor is obtained, the CT image is processed according to step S2, and then the processed CT image is denoised using the trained deep autoencoder denoising network to finally obtain a denoised lung tumor CT image.
[0013] Preferably, step S1 specifically includes:
[0014] Clinical CT image acquisition sensors are deployed in cancer hospitals and affiliated branches to form a decentralized image acquisition information transmission network. Basic information of lung cancer patients (such as name, gender, age, examination date, etc.) and CT images are uploaded to the cloud computing storage system via the network. The cloud computing storage system transmits them to the hospital information server in a multi-threaded parallel manner. Within the server, a "tree" storage structure is formed with the hospital number as the "root" node and the basic information of lung cancer patients as the "leaf" node, and the corresponding CT images are stored in the "leaf" nodes.
[0015] Preferably, step S2 specifically includes:
[0016] Using word embedding technology, the basic information of lung cancer patients (such as name, gender, age, examination date, etc.) is digitally converted to form a coding array. At the same time, a hash coding algorithm based on matrix decomposition technology is used to encode the basic structural information of the coding array as "0" and "1". Then, the collected CT images are rotated and scaled for preprocessing, and new channels are expanded. The "0" and "1" coding numbers are attached to the new channels to achieve the fusion of the basic information of lung cancer patients and CT images.
[0017] Preferably, step S3 specifically includes:
[0018] A deep autoencoder denoising network for CT images of lung tumor patients includes an encoder, an encoding structure preserver, a decoder, and a decoding structure preserver, wherein the encoder includes an encoding layer, the decoder includes a decoding layer, the encoding structure preserver includes an encoding structure preserving layer, and the decoding structure preserver includes a decoding structure preserving layer;
[0019] The number of layers of the encoder and the encoding structure preserver is set to l1, and the number of layers of the decoder and the decoder structure preserver is set to l2. The L1-norm sparse regularization term of the quantization noise ΔX is introduced into the target loss function. The target loss function expression is as follows:
[0020] in, X i is the i-th layer of the CT image of a patient with lung tumor; H i is the intermediate variable of the i-th layer; ΔX is the noise term; W 1p is the p-th encoding matrix; Z 1p Maintaining matrix for the p-th encoding structure; W 2q is the qth decoding matrix; Z 2q Maintain the matrix for the qth decoding structure; ||·|| 2 represents the square of the F-norm; f(·) is the nonlinear mapping function; n is the total number of image channels; λ is the regularization coefficient; ||·||1 is the L1-norm.
[0021] Preferably, in step S4, the optimization method of fusing the proximal mapping operator and the momentum acceleration mechanism is used to train the deep autoencoder denoising network and optimize the number of network layers, specifically including:
[0022] For W 1p ,(p∈{1,2,...,l1-1})have:
[0023]
[0024] for have:
[0025]
[0026] For W 21 have:
[0027]
[0028] For W 2q ,(q∈{2,3,...,l2})have:
[0029]
[0030] For Z 1p ,(p∈{1,2,...,l1-1})have:
[0031]
[0032] for have:
[0033]
[0034] For Z 21 have:
[0035]
[0036] For Z 2q ,(q∈{2,3,...,l2})have:
[0037]
[0038] Add the approximation term ΔS of the quantization noise ΔX and use the proximal mapping operator to approximate ΔX:
[0039]
[0040] The variable update formula of ΔS is formed using the momentum acceleration mechanism:
[0041]
[0042] in, is the p-th encoding matrix of the k-th iteration; Keep the matrix for the pth encoding of the kth iteration; is the qth decoding matrix of the kth iteration; is the qth decoding holding matrix of the kth iteration; η is the learning rate; * T represents the transpose operation of the matrix *; ΔS is the proximal term of the noise term ΔX; ΔX k is the kth iteration of ΔX; ΔS k is the kth iteration of ΔS; a is the weight term; I represents the identity matrix with the same diagonal as ΔX and the same dimensions as ΔX; is a transition variable; By the formula establish;
[0043] Network layer optimization: Set the network layer adjustment interval [b, t], keep the regularization coefficient λ, learning rate η, and weight term a fixed, increase the number of network layers from b layer by layer, record the change of image signal-to-noise ratio with the number of layers, and select the number of layers c corresponding to the maximum signal-to-noise ratio, which is the optimal number of network layers.
[0044] Preferably, the structure of the deep self-editing denoising network is: starting from the input layer, after passing through one encoding layer, a layer of encoding structure maintenance layer with the same dimension is connected, and the encoding layer and the encoding structure maintenance layer are cyclically alternated until the number of layers reaches a preset upper limit; then, after passing through one decoding layer, a layer of decoding structure maintenance layer with the same dimension is connected, and the decoding layer and the decoding structure maintenance layer are cyclically alternated until the number of layers reaches a preset upper limit, and finally output through the output layer.
[0045] In a second aspect, the present invention provides a lung tumor CT image denoising system for implementing the lung tumor CT image denoising method, the system comprising:
[0046] Data acquisition module: used to collect basic information and CT images of lung tumor patients, and transmit the collected basic information and CT images of lung tumor patients from the cloud computing storage system to the main server;
[0047] Data processing module: In the main server, a new channel for CT images of lung tumor patients is opened, basic information of lung tumor patients is digitized to form a coding array, and the resulting coding array is stored in the new channel for CT images of lung tumor patients, thus realizing the fusion of basic information of lung tumor patients with CT images;
[0048] Image denoising module: Utilizes a deep autoencoder denoising network of CT images of lung cancer patients to denoise the fused image of the basic information of lung cancer patients and CT images;
[0049] Visualization module: used to display denoised lung tumor CT images.
[0050] In a third aspect, the present invention provides a lung tumor CT image device for implementing the lung tumor CT image denoising method, comprising a memory, a processor, a controller, a display, an input unit and an output unit; the memory is used to store the image obtained by fusion of the basic information of the lung tumor patient and the CT image; the processor is used to run a computer program, the program ensuring that the basic information of the lung tumor patient correctly matches the CT image and the fused image is denoised; the controller is used to coordinate the data transmission and processing flow between the memory and the processor; the display displays the denoised lung tumor CT image; the input unit is used to input the lung tumor CT image and the basic information of the lung tumor patient; the output unit is used to output the denoised lung tumor CT image.
[0051] Compared with existing autoencoder networks, this invention makes significant innovations and improvements, with the following significant features:
[0052] 1. Design a new autoencoder network: By establishing a new autoencoder network, we can explore the deep features of CT images and achieve good denoising effects by utilizing implicit feature information.
[0053] 2. Construct a specialized objective function: By introducing the quantization noise ΔX term and the L1-norm regularization term, the noise of lung tumor CT images can be accurately calculated.
[0054] 3. Fusion optimization algorithm: By integrating the proximal mapping operator and momentum acceleration mechanism, the denoising network can be quickly trained, reducing training time and computing resource consumption.
[0055] 4. Introduction of structure preserver: By introducing the structure preserver, a basis and guarantee are provided for maintaining the structure and texture information of lung tumor CT images.
[0056] 5. Low-complexity design: By optimizing the denoising network structure, the best denoising effect is achieved, and the storage requirements are adapted to general hardware resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 This is a technical flow chart of the present invention;
[0058] Figure 2 This is the structure diagram of the deep autoencoder denoising network;
[0059] Figure 3 This is the change graph of peak signal-to-noise ratio (PSNR) corresponding to the number of autoencoder layers;
[0060] Figure 4 This is a graph showing the peak signal-to-noise ratio (PSNR) changes with the number of iterations corresponding to different autoencoder layers;
[0061] Figure 5 is the original noisy lung tumor CT image acquired;
[0062] Figure 6 This is a clear denoised lung tumor CT image processed by a deep autoencoder denoising network. DETAILED DESCRIPTION
[0063] The following, in conjunction with the accompanying drawings, clearly and completely describes the technical solutions of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts should fall within the scope of protection of the present invention.
[0064] A method, system and device for denoising lung tumor CT images. The denoising method is designed to integrate four parts: data acquisition, data processing, image denoising and network optimization, so as to retain subtle details of the image and meet the requirements of efficiency and practicality.
[0065] like Figure 1As shown, a lung tumor CT image denoising method includes the following steps:
[0066] Step 1: Deploy a decentralized image acquisition and information transmission network to collect and transmit basic information and CT images of lung cancer patients, and store them in a "tree" structure to the main server.
[0067] In this step, a decentralized image acquisition and information transmission network is deployed to collect and transmit basic information and CT images of lung tumor patients, and store them in a "tree" structure to the main server. The steps include:
[0068] Clinical CT image acquisition sensors are deployed throughout the cancer hospital and its affiliated branches, forming a decentralized image acquisition and information transmission network. Basic information about lung cancer patients (such as name, gender, age, and examination date) and lung CT images are uploaded to a cloud computing storage system via the network. The cloud computing storage system then transmits these to the hospital's central information server in a multi-threaded, parallel manner. Within the central server, a "tree"-like information storage structure is formed, with the hospital number as the "root" node and basic patient information as the "leaf" nodes. The corresponding CT images are stored within the leaf nodes.
[0069] Step 2: Fuse the basic information of lung tumor patients (such as name, gender, age, examination date, etc.) with the CT image.
[0070] In this step, word embedding technology is used to digitally convert the basic information of lung cancer patients (such as name, gender, age, examination date, etc.) into a coded array. At the same time, a hash coding algorithm based on matrix decomposition is used to implement the "0" and "1" encoding of the basic structural information. The original CT images are then preprocessed by rotation and scaling, and new image channels are expanded. The "0" and "1" coded numbers are attached to the new channels to achieve the fusion of the patient's basic information and the CT images, which are then divided into training and test sets.
[0071] Step 3: Construct a deep autoencoder denoising network for CT images of lung cancer patients. The deep autoencoder denoising network introduces a structure preserver and designs a target loss function that can accurately quantify the noise.
[0072] In this step, the deep autoencoder denoising network for CT images of lung tumor patients is mainly composed of an encoder, an encoding structure retainer, a decoder and a decoding structure retainer. The encoder includes an encoding layer, the decoder includes a decoding neural layer, the encoding structure retainer includes an encoding structure retaining layer, and the decoding structure retainer includes a decoding structure retaining layer. Specifically, the structure of the deep autoencoder denoising network is as follows: starting from the input layer, after passing through one encoding layer, a layer of encoding structure retaining layer with the same dimension is connected, and the encoding layer and the encoding structure retaining layer are cyclically alternated until the number of layers reaches a preset upper limit; then, after passing through one decoding layer, a layer of decoding structure retaining layer with the same dimension is connected, and the decoding layer and the decoding structure retaining layer are cyclically alternated until the number of layers reaches a preset upper limit, and finally output through the output layer.
[0073] The number of layers of the encoder and the encoder structure preserver is set to l1, and the number of layers of the decoder and the decoder structure preserver is set to l2. In the designed network objective function, the L1-norm sparse regularization term of the quantization noise ΔX is introduced to form the following objective loss function:
[0074]
[0075] Among them, X i is the i-th layer of the CT image of a patient with lung tumor; H i is the intermediate variable of the i-th layer; ΔX is the noise term; W 1p is the p-th encoding matrix; Z 1p Maintaining matrix for the p-th encoding structure; W 2q is the qth decoding matrix; Z 2q Maintain the matrix for the qth decoding structure; ||·|| 2 represents the square of the F-norm; f(·) is the nonlinear mapping function; n is the total number of image channels; λ is the regularization coefficient; ||·||1 is the L1-norm.
[0076] Step 4: Use the optimization method of fusing the proximal mapping operator and the momentum acceleration mechanism to implement deep autoencoder denoising network training, optimize the number of network layers, and obtain the trained deep autoencoder denoising network.
[0077] In this step, the method of fusing the proximal mapping operator and momentum acceleration is used to form the deep autoencoder denoising network training rules for CT images of patients with lung tumors.
[0078] For W 1p ,(p∈{1,2,...,l1-1})have:
[0079]
[0080] for have:
[0081]
[0082] For W 21 have:
[0083]
[0084] For W 2q ,(q∈{2,3,...,l2})have:
[0085]
[0086] For Z 1p ,(p∈{1,2,...,l1-1})have:
[0087]
[0088] for have:
[0089]
[0090] For Z 21 have:
[0091]
[0092] For Z 2q ,(q∈{2,3,...,l2})have:
[0093]
[0094] Add the approximation term ΔS of the quantization noise ΔX and use the proximal mapping operator to approximate ΔX:
[0095]
[0096] The variable update formula of ΔS is formed using the momentum acceleration mechanism:
[0097]
[0098] in, is the p-th encoding matrix of the k-th iteration; Keep the matrix for the pth encoding of the kth iteration; is the qth decoding matrix of the kth iteration; is the qth decoding holding matrix of the kth iteration; η is the learning rate; * T represents the transpose operation of the matrix *; ΔS is the proximal term of the noise term ΔX. ΔX k is the kth iteration of ΔX; ΔS kis the kth iteration of ΔS. a is a weight term; I represents the identity matrix with the same dimensions as ΔX and a diagonal of 1; is a transition variable; By the formula establish.
[0099] Optimize the number of network layers: Set the network layer adjustment interval [b, t], keep other parameters (such as regularization coefficient λ, learning rate η, weight term a, etc.) unchanged, increase the number of network layers from b layer by layer, record the image signal-to-noise ratio changes with the number of layers, and select the number of layers c corresponding to the maximum signal-to-noise ratio, which is the optimal number of network layers.
[0100] Step 5: Obtain a CT image of a lung tumor patient, process the CT image in step 2, and then use the trained deep autoencoder denoising network to denoise the processed CT image, ultimately obtaining a denoised lung tumor CT image.
[0101] Based on the above methods and statements, the present invention provides a lung tumor CT image denoising system, comprising:
[0102] Data acquisition module: used to collect basic information and CT images of lung tumor patients, and transmit the collected basic information and CT images of lung tumor patients from the cloud computing storage system to the main server;
[0103] Data processing module: In the main server, a new channel for CT images of lung tumor patients is opened, basic information of lung tumor patients is digitized to form a coding array, and the resulting coding array is stored in the new channel for CT images of lung tumor patients, thus realizing the fusion of basic information of lung tumor patients with CT images;
[0104] Image denoising module: Utilizes a deep autoencoder denoising network of CT images of lung cancer patients to denoise the image that is a fusion of basic information of lung cancer patients and CT images;
[0105] Visualization module: used to display denoised lung tumor CT images.
[0106] A lung tumor CT image device comprises a memory, a processor, a controller, a display, an input unit, and an output unit; the memory is used to store an image formed by fusing basic information of a lung tumor patient with a CT image; the processor is used to run a computer program, wherein the program ensures that the basic information of the lung tumor patient is correctly matched with the CT image and that the fused image is denoised; the controller is used to coordinate data transmission and processing flows between the memory and the processor; the display displays a denoised lung tumor CT image; the input unit is used to input a lung tumor CT image and basic information of a lung tumor patient; and the output unit is used to output a denoised lung tumor CT image.
[0107] The following is a preferred description of the concept, specific algorithm structure and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features and effects of the present invention.
[0108] According to the implementation process, the application steps of the present invention in a practical scenario are as follows:
[0109] 1. Shanghai First People's Hospital was selected as the actual application site. By deploying clinical CT image acquisition sensors at Lianyungang First People's Hospital and its affiliated branches, a decentralized image acquisition information transmission network was formed. Basic information of lung cancer patients (such as name, gender, age, examination date, etc.) and lung CT images were uploaded to the cloud computing storage system via the network. The data was then transmitted from the cloud to the Lianyungang First People's Hospital information server in a multi-threaded parallel manner. Within the server, a "tree"-like information storage structure was formed, with the hospital number as the "root" node and basic patient information as the "leaf" node. The corresponding CT images were stored in the "leaf" nodes.
[0110] 2. In the information server of Lianyungang First People's Hospital, word embedding technology is used to digitize the basic information of lung cancer patients (such as name, gender, age, examination date, etc.) to form a coding array. At the same time, a hash coding algorithm based on matrix decomposition is used to implement the "0" and "1" encoding of the basic structural information. The original CT images collected are then rotated, scaled, and other preprocessing operations, and new image channels are expanded. The "0" and "1" coding numbers are attached to the new channels to achieve the fusion of the patient's basic information and the CT images.
[0111] 3. By constructing a deep autoencoder denoising network for CT images of lung cancer patients, introducing a structure preserver, and designing an objective function that can accurately quantify noise
[0112] 4. Using the optimization method of integrating the proximal mapping operator and the momentum acceleration mechanism, deep autoencoder network training is implemented, the network layer adjustment range is set [1,9] and the network layer number is optimized layer by layer. Figure 3 and Figure 4 According to the data results, the signal-to-noise ratio is optimal when the network layer is 5, and the signal-to-noise ratio is also the best at different numbers of iterations, indicating that the denoising performance is best when the network layer is 5. The image corresponding to the 5th layer of the optimal network layer is selected, and the trained deep autoencoder denoising network is used to denoise the image after the fusion of the patient's basic information and the CT image, which is the final denoised lung tumor CT image. Figure 5 For the collected CT images, Figure 6 It can be seen that the image processed by the 5-layer autoencoder network with the optimal network layer number of the present invention is more informative than the original image, and the structure and texture information of the lung tumor CT image is maintained.
[0113] The above embodiments merely illustrate the implementation methods of the present invention. Although the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and such modifications and improvements are all within the scope of protection of the present invention.
Claims
1. A lung tumor CT image denoising method, characterized by: The method comprises the following steps: Step S1: Deploy a decentralized image acquisition and information transmission network to collect and transmit basic information and CT images of lung tumor patients, and store them in a "tree" structure to the main server; Step S2: In the main server, a new channel for CT images of lung tumor patients is opened, basic information of lung tumor patients is digitized to form a coding array, and the resulting coding array is stored in the new channel for CT images of lung tumor patients, thereby integrating the basic information of lung tumor patients with the CT images, and dividing the images into a training set and a test set. Step S3: Constructing a deep autoencoder denoising network for CT images of lung cancer patients. The deep autoencoder denoising network introduces a structure preserver and designs a target loss function for accurate noise quantization. Step S4: using an optimization method that integrates a proximal mapping operator and a momentum acceleration mechanism to train the deep autoencoder denoising network, optimizing the number of network layers, and obtaining a trained deep autoencoder denoising network; In step S5, a CT image of a patient with lung tumor is obtained, the CT image is processed according to step S2, and then the processed CT image is denoised using the trained deep autoencoder denoising network to finally obtain a denoised lung tumor CT image.
2. The lung tumor CT image denoising method according to claim 1, characterized in that: Step S1 specifically includes: Clinical CT image acquisition sensors are deployed in the cancer hospital and its affiliated branches to form a decentralized image acquisition information transmission network. Basic information and CT images of lung cancer patients are uploaded to the cloud computing storage system via the network. The cloud computing storage system then transmits them to the hospital's main information server in a multi-threaded parallel manner. Within the main server, a "tree" storage structure is formed with the hospital number as the "root" node and the basic information of lung cancer patients as the "leaf" node, and the corresponding CT images are stored in the "leaf" nodes.
3. The lung tumor CT image denoising method according to claim 1, characterized in that: Step S2 specifically includes: Using word embedding technology, the basic information of lung cancer patients is digitally converted to form a coding array. At the same time, a hash coding algorithm based on matrix decomposition technology is used to encode the basic structural information of the coding array as "0" and "1". Then, the collected CT images are rotated and scaled for preprocessing, and new channels are expanded. The "0" and "1" coding numbers are attached to the new channels to achieve the fusion of the basic information of lung cancer patients and CT images.
4. The lung tumor CT image denoising method according to claim 1, characterized in that: Step S3 specifically includes: A deep autoencoder denoising network for CT images of lung tumor patients includes an encoder, an encoding structure preserver, a decoder, and a decoding structure preserver, wherein the encoder includes an encoding layer, the decoder includes a decoding layer, the encoding structure preserver includes an encoding structure preserving layer, and the decoding structure preserver includes a decoding structure preserving layer; The number of layers of the encoder and the encoding structure preserver is set to l1, and the number of layers of the decoder and the decoder structure preserver is set to l2. The L1-norm sparse regularization term of the quantization noise ΔX is introduced into the target loss function. The target loss function expression is as follows: Among them, X i is the i-th layer of the CT image of a patient with lung tumor; H i is the intermediate variable of the i-th layer; ΔX is the noise term; W 1p is the p-th encoding matrix; Z 1p Maintain the matrix for the p-th encoding structure; W 2q is the qth decoding matrix; Z 2q Maintain the matrix for the qth decoding structure; ||·|| 2 represents the square of the F-norm; f(·) is the nonlinear mapping function; n is the total number of image channels; λ is the regularization coefficient; ||·||1 is the L1-norm.
5. The lung tumor CT image denoising method according to claim 4, characterized in that: In step S4, the optimization method of integrating the proximal mapping operator and the momentum acceleration mechanism is used to train the deep autoencoder denoising network and optimize the number of network layers, specifically including: For W 1p ,(p∈{1,2,...,l1-1})have: for have: For W 21 have: For W 2q ,(q∈{2,3,...,l2})have: For Z 1p ,(p∈{1,2,...,l1-1})have: for have: For Z 21 have: For Z 2q ,(q∈{2,3,...,l2})have: Add the approximation term ΔS of the quantization noise ΔX and use the proximal mapping operator to approximate ΔX: The variable update formula of ΔS is formed using the momentum acceleration mechanism: in, is the p-th encoding matrix of the k-th iteration; Keep the matrix for the pth encoding of the kth iteration; is the qth decoding matrix of the kth iteration; is the qth decoding holding matrix of the kth iteration; η is the learning rate; * T represents the transpose operation of the matrix *; ΔS is the proximal term of the noise term ΔX; ΔX k is the kth iteration of ΔX; ΔS k is the kth iteration of ΔS; a is the weight term; I represents the identity matrix with the same diagonal as ΔX and the same dimensions as ΔX; is a transition variable; By the formula establish; Network layer optimization: Set the network layer adjustment interval [b, t], keep the regularization coefficient λ, learning rate η, and weight term a fixed, increase the number of network layers from b layer by layer, record the change of image signal-to-noise ratio with the number of layers, and select the number of layers c corresponding to the maximum signal-to-noise ratio, which is the optimal number of network layers.
6. The lung tumor CT image denoising method according to claim 5, characterized in that: The structure of the deep self-editing denoising network is as follows: starting from the input layer, after passing through one encoding layer, a layer of encoding structure preservation layer with the same dimension is connected, and the encoding layer and the encoding structure preservation layer are cyclically alternated until the number of layers reaches a preset upper limit; Then, after passing through a decoding layer, a decoding structure preservation layer with the same dimension is connected, and the decoding layer and the decoding structure preservation layer are cyclically alternated until the number of layers reaches the preset upper limit, and finally output through the output layer.
7. A lung tumor CT image denoising system, for implementing the lung tumor CT image denoising method according to any one of claims 1 to 6, characterized in that: The system comprises: Data acquisition module: used to collect basic information and CT images of lung tumor patients, and transmit the collected basic information and CT images of lung tumor patients from the cloud computing storage system to the main server; Data processing module: In the main server, a new channel for CT images of lung tumor patients is opened, basic information of lung tumor patients is digitized to form a coding array, and the resulting coding array is stored in the new channel for CT images of lung tumor patients, thus realizing the fusion of basic information of lung tumor patients with CT images; Image denoising module: Utilizes a deep autoencoder denoising network of CT images of lung cancer patients to denoise the fused image of the basic information of lung cancer patients and CT images; Visualization module: used to display denoised lung tumor CT images.
8. A lung tumor CT image device, used to implement the lung tumor CT image denoising method according to any one of claims 1 to 6, characterized in that: The system comprises a memory, a processor, a controller, a display, an input unit and an output unit; the memory is used to store an image formed by fusing basic information of a lung tumor patient with a CT image; the processor is used to run a computer program, which ensures that the basic information of the lung tumor patient correctly matches the CT image and denoises the fused image; the controller is used to coordinate the data transmission and processing flow between the memory and the processor; the display displays the denoised lung tumor CT image; the input unit is used to input the lung tumor CT image and the basic information of the lung tumor patient; and the output unit is used to output the denoised lung tumor CT image.