Artifact-targeted adversarial training based trustworthy prostate cancer identification method and device

Abstract

The application discloses a kind of based on artifact targetting countermeasure training credible prostate cancer identification method and device, comprising: collecting prostate data set, and pre-processing, obtain different dimensions of prostate original image;Different dimensions of prostate cancer identification model are constructed, and different dimensions of prostate original image are introduced into corresponding prostate cancer identification model and pre-training is carried out;Prostate original image is input into pre-trained prostate cancer identification model, and style sample is screened out;Based on style sample, in combination with style migration technology and countermeasure attack technology, construct special countermeasure sample generation model, generate special countermeasure sample;Prostate original image and special countermeasure sample are used to carry out countermeasure training to pre-trained prostate cancer identification model, and identification model parameter is optimized together, to obtain credible prostate cancer identification model.The application uses countermeasure training method, in combination with special sample and original image, improves the credibility of prostate cancer identification system.

Description

Technical Field

[0001] This invention belongs to the technical field of medical image processing, specifically relating to a reliable prostate cancer identification method and device based on artifact-targeted adversarial training. Background Technology

[0002] Currently, multiparameter magnetic resonance imaging (mpMRI) is widely used in the diagnosis of prostate cancer (PCa) due to its high sensitivity in detecting prostate lesions. However, traditional prostate mpMRI evaluation is based on subjective visual assessment, which leads to variability in evaluation results and poor performance in assessing lesions. Furthermore, manually interpreting mpMRI sequences requires significant experience and labor, limiting its clinical applicability. Therefore, efficient and accurate interpretation of mpMRI data to achieve satisfactory PCa diagnostic results in clinical practice is of paramount importance.

[0003] In recent years, deep learning technology has achieved significant success in medical image recognition tasks. Prostate magnetic resonance imaging (MRI) can provide anatomical and tissue function information of the prostate gland (PCa), which can be used for PCa detection, diagnosis, classification, and treatment response assessment. Currently, several MRI-based deep learning diagnostic methods for PCa exist. However, in clinical practice, MRI is prone to sensitivity artifacts, and existing deep learning-assisted methods typically do not consider the influence of rectal artifacts, thus their performance is easily affected by these artifacts. In particular, besides obvious abnormal textures (usually outside the lesion area), rectal artifacts have small and variable diffusion noise that can easily cover the lesion area, interfering with the performance of lesion-level PCa recognition models. Furthermore, the morphology, feature distribution, and size of rectal artifacts vary considerably depending on the degree of rectal dilation and intestinal contents. These issues lead to insufficient reliability of deep learning-assisted diagnosis. Therefore, how to enhance the reliability of the model has become a core issue in the clinical practice of intelligent diagnostic technology.

[0004] Therefore, it is necessary to provide a reliable PCa identification system that can resist artifact interference. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a reliable prostate cancer identification method and device based on artifact-targeted adversarial training. By utilizing adversarial training methods and combining proprietary samples with original images, the reliability of the prostate cancer identification system is improved, thereby enabling more accurate prostate cancer identification.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention provides a reliable prostate cancer identification method based on artifact-targeted adversarial training, comprising the following steps:

[0008] S1. Collect a prostate dataset and perform preprocessing operations on the prostate dataset to obtain raw prostate images of different dimensions;

[0009] S2. Construct prostate cancer recognition models of different dimensions, and import the original prostate images of different dimensions into the corresponding prostate cancer recognition models for pre-training to obtain pre-trained prostate cancer recognition models.

[0010] S3. Input the original prostate image as a reference set into the pre-trained prostate cancer recognition model to obtain predicted sample images. Select prostate sample images with rectal artifact attributes from the predicted sample images as style samples for style transfer.

[0011] S4. Based on the style samples, and combining style transfer technology and adversarial attack technology, construct a proprietary adversarial sample generation model, and optimize the proprietary adversarial sample generation model by minimizing the total loss to generate proprietary adversarial samples with rectal artifact attributes.

[0012] S5. Using the original prostate image and proprietary adversarial examples, perform adversarial training on the pre-trained prostate cancer recognition model to jointly optimize the parameters of the pre-trained prostate cancer recognition model and obtain a reliable prostate cancer recognition model.

[0013] As a preferred technical solution, the preprocessing operations in step S1 include cropping, normalization, registration, and cutting;

[0014] S11. The cropping is to crop all images to the area around the prostate with a voxel size of a preset size and a preset spacing. The interpolation of all images is done using the third-order B-spline interpolation method.

[0015] S12. The normalization is performed independently on each channel of the image, so that the voxels in the image fall within the range of [0, 1]. The normalization formula is:

[0016]

[0017] Where x represents the normalized image; percentile(x init ,num) represents all the values ​​x of that image channel. init Sort the numbers in ascending order to form a sequence, and take the value at the num% position of the sequence. This operation avoids the interference of the maximum and minimum values ​​on the normalization.

[0018] S13. The registration is a non-rigid registration based on B-spline transformation to achieve consistency of all data in spatial location.

[0019] S14. The cutting is performed by generating a prostate segmentation mask using a prostate segmentation model, and simultaneously generating a lesion region mask by using a threshold level setting method, thereby cropping out the lesion region in each slice to form a lesion sample.

[0020] As a preferred technical solution, the different dimensions include lesion dimension, slice dimension, and patient dimension.

[0021] As a preferred technical solution, step S3 specifically includes:

[0022] Multiple images with obvious rectal artifacts are selected from the original prostate images as a reference set. This reference set is then input into a pre-trained recognition model to obtain predicted sample images. The predicted sample image with the lowest confidence is selected as the reference sample. A region is cropped from the reference sample as the style sample x for style transfer. s .

[0023] As a preferred technical solution, the specific process of constructing a proprietary adversarial sample generation model based on the style samples and combining style transfer technology and adversarial attack technology in step S4 is as follows:

[0024] Given a raw prostate image x with category label y and a prostate cancer recognition model h θ First, a strong first-order attack PGD is used to perform an adversarial attack on the original prostate image to obtain the adversarial sample x′, as shown in the formula:

[0025] x′=x+δ,h θ (x′)≠y,||xx||≤ε

[0026] Where δ represents the noise resistance, and ||·|| represents the noise resistance based on L ∞ The distance metric of the norm, ε, is a manually set perturbation budget. The adversarial example x′ is obtained by solving an optimization problem:

[0027]

[0028] Among them, y i Let p be the y-value corresponding to the i-th class. i To identify the model's predicted probability for class i, let C represent the number of classes; thus, the adversarial loss is derived.

[0029]

[0030] When conducting adversarial attacks, the optimal perturbation budget ε is selected as 8 / 255, and the attack step size is 2×10. -3 The attack iteration count is 10;

[0031] Using a given style sample x s The adversarial sample x′ is style-transferred to generate a proprietary adversarial sample with rectal artifact properties. The specific steps are as follows: using VggNet-16 as the feature extraction network E, the adversarial sample x′ and style sample x′ are style-transferred. s Feature extraction, and definition of adversarial sample x′ and style sample x s The style loss between them is:

[0032]

[0033] Where Ψ is the index set of style layers in E, including convolutional layers 2, 4, 7, and 10. ψ It is the ψth style layer in Ψ. The Gram matrix represents the deep features extracted from the style layer;

[0034] A semantic loss is constructed between the adversarial sample x′ and the original prostate image x to ensure that the semantic information of the original prostate image is not altered. The semantic loss is defined as:

[0035]

[0036] Where Φ is the set of indexes for semantic layers in E, including convolutional layers 10, 11, 12, and 13. φ It is the φth semantic layer in Φ.

[0037] As a preferred technical solution, step S4 involves optimizing the proprietary adversarial example generation model by minimizing the total loss, specifically as follows:

[0038] The final total loss function consists of three parts: an adversarial loss used to find the most powerful perturbation. Style loss used to generate proprietary adversarial examples of rectal artifacts and semantic loss used to prevent semantic information from being corrupted The final total loss function is:

[0039]

[0040] Here, α and β are hyperparameters used to balance the three terms. By making a series of changes to the hyperparameters, α and β are selected to achieve the optimal model performance. α and β are each set to 10. -3 and 5×10 -2 ;

[0041] The proprietary adversarial example generation model is optimized by minimizing the total loss of the model until the training meets the set termination condition, resulting in proprietary adversarial examples with rectal artifact attributes. The overall optimization objective is:

[0042]

[0043] Where N is the number of training data.

[0044] As a preferred technical solution, in step S5, the optimization objective of optimizing the parameters of the pre-trained prostate cancer recognition model is:

[0045]

[0046] in, It is the original loss function of the prostate cancer identification model;

[0047] Backpropagation is performed and the optimization steps for the pre-trained prostate cancer recognition model parameters are repeated until the set termination condition is reached, thus achieving a reliable prostate cancer recognition model.

[0048] In another aspect, the present invention provides a reliable prostate cancer identification system based on artifact-targeted adversarial training, which is applied to the aforementioned reliable prostate cancer identification method based on artifact-targeted adversarial training, including a preprocessing module, an identification model module, a style sample selection module, an adversarial sample generation module, and an adversarial training module.

[0049] The preprocessing module is used to collect a prostate dataset and perform preprocessing operations on the prostate dataset to obtain raw prostate images of different dimensions.

[0050] The recognition model module is used to construct prostate cancer recognition models of different dimensions. The original prostate images of different dimensions are imported into the corresponding prostate cancer recognition models for pre-training to obtain pre-trained prostate cancer recognition models.

[0051] The style sample selection module is used to input the original prostate image as a reference set into the pre-trained prostate cancer recognition model to obtain predicted sample images, and to select prostate sample images with rectal artifact attributes from the predicted sample images as style samples for style transfer.

[0052] The adversarial sample generation module is used to construct a proprietary adversarial sample generation model based on the style sample, combined with style transfer technology and adversarial attack technology, and to optimize the proprietary adversarial sample generation model by minimizing the total loss to generate proprietary adversarial samples with rectal artifact attributes.

[0053] The adversarial training module is used to perform adversarial training on the pre-trained prostate cancer recognition model using the original prostate image and proprietary adversarial samples, jointly optimizing the parameters of the pre-trained prostate cancer recognition model to obtain a reliable prostate cancer recognition model.

[0054] In another aspect, the present invention provides an electronic device, characterized in that the electronic device comprises:

[0055] At least one processor; and,

[0056] A memory communicatively connected to the at least one processor; wherein,

[0057] The memory stores computer program instructions that can be executed by the at least one processor, which enables the at least one processor to execute the trusted prostate cancer identification method based on artifact-targeted adversarial training.

[0058] In another aspect, the present invention provides a computer-readable storage medium storing a program that, when executed by a processor, implements the aforementioned reliable prostate cancer identification method based on artifact-targeted adversarial training.

[0059] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0060] 1. Due to the special nature of medical imaging, generally only radiologists and people with medical clinical knowledge can detect prostate cancer based on images. This invention proposes a reliable prostate cancer identification method based on artifact-targeted adversarial training, which greatly reduces the complexity of doctors manually interpreting MRI images.

[0061] 2. Based on style transfer and adversarial attack techniques, this invention proposes a proprietary adversarial sample generation module that can perform noise modeling for rectal artifacts, thereby obtaining adversarial samples with rectal artifact attributes for adversarial training of prostate cancer recognition models.

[0062] 3. This invention utilizes an adversarial training method, combined with proprietary samples and original images, to propose a reliable prostate cancer identification system based on artifact-targeted adversarial training, thereby improving the reliability of the prostate cancer identification system and enabling more accurate prostate cancer identification. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of a reliable prostate cancer identification method based on artifact-targeted adversarial training according to an embodiment of the present invention;

[0065] Figure 2This is a schematic diagram of the overall structure of the reliable prostate cancer identification method based on artifact-targeted adversarial training according to an embodiment of the present invention.

[0066] Figure 3 (a) is the original T2 image of the embodiment of the present invention without rectal artifact attributes;

[0067] Figure 3 (b) is the original ADC image of the embodiment of the present invention that does not have rectal artifact properties;

[0068] Figure 3 (c) The original DWI image of the embodiment of the present invention does not have rectal artifact properties;

[0069] Figure 3 (d) is a style sample image with obvious rectal artifact attributes;

[0070] Figure 4 (a) is a prostate mask diagram according to an embodiment of the present invention;

[0071] Figure 4 (b) is a rectal artifact noise map after the restricted area in an embodiment of the present invention;

[0072] Figure 5 (a) A proprietary T2 image generated for an embodiment of the present invention;

[0073] Figure 5 (b) A proprietary ADC image generated for an embodiment of the present invention;

[0074] Figure 5 (c) A proprietary DWI image generated for an embodiment of the present invention;

[0075] Figure 6 This is a block diagram of a reliable prostate cancer recognition system based on artifact-targeted adversarial training, according to an embodiment of the present invention.

[0076] Figure 7 This is a structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0077] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0078] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0079] In clinical practice, medical images are prone to rectal magnetic susceptibility artifacts, which can interfere with the performance of prostate cancer recognition models. Therefore, this invention studies a reliable prostate cancer recognition method based on artifact-targeted adversarial training to improve the anti-interference ability of prostate cancer recognition models against rectal artifacts in medical images and enhance the reliability of the model.

[0080] Please see Figure 1 , Figure 2 As shown, in one embodiment of this application, a reliable prostate cancer identification method based on artifact-targeted adversarial training is provided, including the following steps:

[0081] S1. Collect a prostate dataset and preprocess the dataset to obtain raw prostate images of different dimensions; such as... Figure 3 (a) Figure 3 (b) and Figure 3 As shown in (c).

[0082] Furthermore, the preprocessing operations include cropping, normalization, registration, and cutting; specifically:

[0083] S11. The cropping is to crop all images to the prostate area with a voxel size of 160×160×24 and a spacing of (0.5,0.5,3)mm. The interpolation of all images is performed using the third-order B-spline interpolation method.

[0084] S12. The normalization is performed independently on each channel of the image, so that most voxels in the image fall within the range of [0,1]. The normalization formula is as follows:

[0085]

[0086] Where x represents the normalized image; percentile(x init ,num) represents all the values ​​x of that image channel. init Sort the numbers in ascending order to form a sequence, and take the value at the num% position of the sequence; this operation can avoid the interference of the maximum and minimum values ​​on the normalization.

[0087] S13. The registration is a non-rigid registration based on B-spline transformation to achieve consistency of all data in spatial location.

[0088] S14. The cutting process uses a prostate segmentation model to generate a prostate segmentation mask, and at the same time, a lesion area mask is generated by using a threshold level setting method, thereby cropping out the lesion area in each slice to form a lesion sample.

[0089] Furthermore, the different dimensions include lesion dimension, slice dimension, and patient dimension.

[0090] S2. Construct prostate cancer recognition models with different dimensions, and import the original prostate images of different dimensions into the corresponding prostate cancer recognition models for pre-training to obtain pre-trained prostate cancer recognition models.

[0091] Furthermore, step S2 involves pre-training a prostate cancer recognition model using three different dimensions of the pre-processed original prostate image: lesion dimension, slice dimension, and patient dimension.

[0092] S3. Input the original prostate image as a reference set into the pre-trained prostate cancer recognition model to obtain predicted sample images. Select the prostate sample image with the most rectal artifact attributes from the predicted sample images as style samples for style transfer.

[0093] Furthermore, based on style selection, five images with obvious rectal artifacts were chosen from the original data by several radiologists with many years of experience through voting, serving as a reference set. This reference set was then input into a pre-trained recognition model to obtain predicted sample images. The predicted sample image with the lowest confidence was selected as the reference sample. Following the radiologists' guidance, a region was cropped from the reference sample as the style sample x for style transfer. s Please see Figure 3 (d)

[0094] S4. Based on the style samples, and combining style transfer technology and adversarial attack technology, construct a proprietary adversarial sample generation model, and optimize the proprietary adversarial sample generation model by minimizing the total loss to generate proprietary adversarial samples with rectal artifact attributes.

[0095] S41. Given a raw prostate image x with category label y and a prostate cancer identification model h. θ First, a strong first-order attack PGD is used to perform an adversarial attack on the original prostate image to obtain the adversarial sample x′, as shown in the formula:

[0096] x′=x+δ,h θ (x′)≠y,||xx||≤ε

[0097] Where δ represents the noise resistance, and ||·|| represents the noise resistance based on L ∞ The distance metric for the norm, ε is an artificially set perturbation budget; see [link to relevant documentation]. Figure 4 (a) Figure 4 As shown in (b), noise with rectal artifact properties is generated using a prostate mask as a restricted region, which is then used to generate proprietary adversarial samples.

[0098] The adversarial example x′ is obtained by solving the following optimization problem:

[0099]

[0100] Among them, y i Let p be the y-value corresponding to the i-th class. i To identify the model's predicted probability for class i; C represents the number of classes; thus, the adversarial loss can be obtained.

[0101]

[0102] When conducting adversarial attacks, the optimal perturbation budget ε is selected as 8 / 255, and the attack step size is 2×10. -3 The attack iteration count is 10;

[0103] S42. Using the given style sample x s The adversarial sample x′ is style-transferred to generate a proprietary adversarial sample with rectal artifact properties. The specific steps are as follows: using VggNet-16 as the feature extraction network E, the adversarial sample x′ and style sample x′ are style-transferred. s Feature extraction, defining adversarial sample x′ and style sample x s The style loss between them is:

[0104]

[0105] Where Ψ is the index set of style layers in E, including convolutional layers 2, 4, 7, and 10. ψ It is the ψth style layer in Ψ. The Gram matrix represents the deep features extracted from the style layer;

[0106] A semantic loss is constructed between the adversarial sample x′ and the original prostate image x to ensure that the semantic information of the original prostate image is not altered. The semantic loss is defined as:

[0107]

[0108] Where Φ is the set of indexes for semantic layers in E, including convolutional layers 10, 11, 12, and 13.φ It is the φth semantic layer in Φ.

[0109] S43. Furthermore, the optimization of the proprietary adversarial example generation model by minimizing the total loss specifically involves: the final loss function consisting of three parts: an adversarial loss used to find the most powerful perturbation. Style loss used to generate rectal artifact styles and semantic loss used to prevent semantic information from being corrupted The final loss function is as follows:

[0110]

[0111] Here, α and β are hyperparameters used to balance the three terms. By making a series of changes to the hyperparameters, α and β are selected to achieve the optimal model performance. α and β are each set to 10. -3 and 5×10 -2 ;

[0112] The generative model is optimized by minimizing its total loss until the training meets a predetermined termination condition, resulting in proprietary adversarial examples with rectal artifact attributes. The overall optimization objective is as follows:

[0113]

[0114] Where N is the number of training data.

[0115] S5. Using the original prostate image and proprietary adversarial examples, perform adversarial training on the pre-trained prostate cancer recognition model to jointly optimize the parameters of the pre-trained prostate cancer recognition model and obtain a reliable prostate cancer recognition model.

[0116] Furthermore, the optimization objective for the parameters of the pre-trained prostate cancer recognition model is:

[0117]

[0118] in It is the original loss function of the prostate cancer identification model;

[0119] The process involves backpropagation and repeating the optimization steps for the pre-trained prostate cancer recognition model parameters until a predetermined termination condition is reached, thus achieving a reliable prostate cancer recognition model. This invention uses a wide range of metrics to evaluate the prostate cancer recognition model, including accuracy, sensitivity, true positive rate, specificity, true negative rate, false positive rate, false negative rate, and the area under the ROC curve.

[0120] Finally, as Figure 5 (a) Figure 5 (b) Figure 5 As shown in (c), a proprietary image is generated using a reliable prostate cancer identification model.

[0121] Please see Figure 3 (a) Figure 3 (b) Figure 3 (c) Figure 3 (d) and Figure 5 (a) Figure 5 (b) Figure 5 As shown in (c), using style samples as the basis for style transfer, the aforementioned method for reliable prostate cancer identification based on artifact-targeting adversarial training is applied. Figure 3 The original image data in (a)-(c) are processed to obtain the following: Figure 5 Generated images from (a)-(c). From Figure 5 As can be seen from (a)-(c), Figure 5 The images in (a)-(c) are well preserved. Figure 3 Visual features of the original prostate images in (a)-(c). However, since these images are obtained by processing real images, they exhibit obvious rectal artifacts and are completely different from images of real data.

[0122] In summary, this invention develops a reliable prostate cancer identification method and device based on artifact-targeted adversarial training. Using style samples as a foundation, style transfer is performed on adversarial samples generated by the adversarial sample generation model through the construction of style loss and semantic loss to generate proprietary adversarial samples with rectal artifact attributes. These proprietary adversarial samples are then combined with the original image for further adversarial training of the prostate cancer identification model, thereby improving the model's robustness and reliability against rectal artifacts. Through this method, the generation of adversarial samples with rectal artifact attributes is effectively achieved, while simultaneously realizing a reliable prostate cancer identification model.

[0123] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously.

[0124] Based on the same idea as the trusted prostate cancer identification method based on artifact-targeted adversarial training in the above embodiments, the present invention also provides a trusted prostate cancer identification system based on artifact-targeted adversarial training. This system can be used to execute the aforementioned trusted prostate cancer identification method based on artifact-targeted adversarial training. For ease of explanation, the structural diagram of the embodiment of the trusted prostate cancer identification system based on artifact-targeted adversarial training only shows the parts related to the embodiments of the present invention. Those skilled in the art will understand that the illustrated structure does not constitute a limitation on the device, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0125] Please see Figure 6 In another embodiment of this application, a reliable prostate cancer identification system 100 based on artifact-targeted adversarial training is provided. The system includes a preprocessing module 101, an identification model module 102, a style sample selection module 103, an adversarial sample generation module 104, and an adversarial training module 105.

[0126] The preprocessing module 101 is used to collect a prostate dataset and perform preprocessing operations on the prostate dataset to obtain original prostate images of different dimensions.

[0127] The recognition model module 102 is used to construct prostate cancer recognition models of different dimensions, and to import the original prostate images of different dimensions into the corresponding prostate cancer recognition models for pre-training to obtain pre-trained prostate cancer recognition models.

[0128] The style sample selection module 103 is used to input the original prostate image as a reference set into the pre-trained prostate cancer recognition model to obtain predicted sample images, and to select the prostate sample image with the most rectal artifact attributes from the predicted sample images as the style sample for style transfer.

[0129] The adversarial sample generation module 104 is used to construct a proprietary adversarial sample generation model based on the style sample, combined with style transfer technology and adversarial attack technology, and optimize the proprietary adversarial sample generation model by minimizing the total loss to generate proprietary adversarial samples with rectal artifact attributes.

[0130] The adversarial training module 105 is used to perform adversarial training on the pre-trained prostate cancer recognition model using the original prostate image and proprietary adversarial samples, jointly optimizing the parameters of the pre-trained prostate cancer recognition model to obtain a reliable prostate cancer recognition model.

[0131] It should be noted that the trusted prostate cancer identification system based on artifact-targeted adversarial training of the present invention corresponds one-to-one with the trusted prostate cancer identification method based on artifact-targeted adversarial training of the present invention. The technical features and beneficial effects described in the embodiments of the trusted prostate cancer identification method based on artifact-targeted adversarial training described above are applicable to the embodiments of the trusted prostate cancer identification system based on artifact-targeted adversarial training. For details, please refer to the description in the embodiments of the method of the present invention, which will not be repeated here.

[0132] Furthermore, in the embodiments of the trusted prostate cancer recognition system based on artifact targeting adversarial training described above, the logical division of each program module is merely illustrative. In practical applications, the above functions can be assigned to different program modules as needed, for example, for the sake of corresponding hardware configuration requirements or the convenience of software implementation. That is, the internal structure of the trusted prostate cancer recognition system based on artifact targeting adversarial training can be divided into different program modules to complete all or part of the functions described above.

[0133] Please see Figure 7 In one embodiment, an electronic device is provided for implementing a trusted prostate cancer identification method based on artifact-targeted adversarial training. The electronic device 200 may include a first processor 201, a first memory 202, and a bus, and may also include a computer program stored in the first memory 202 and executable on the first processor 201, such as a trusted prostate cancer identification program 203 based on artifact-targeted adversarial training.

[0134] The first memory 202 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the first memory 202 can be an internal storage unit of the electronic device 200, such as the portable hard drive of the electronic device 200. In other embodiments, the first memory 202 can also be an external storage device of the electronic device 200, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 200. Furthermore, the first memory 202 can include both internal and external storage units of the electronic device 200. The first memory 202 can be used not only to store application software and various types of data installed on the electronic device 200, such as the code of a trusted prostate cancer identification program 203 based on artifact-targeting adversarial training, but also to temporarily store data that has been output or will be output.

[0135] In some embodiments, the first processor 201 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The first processor 201 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the first memory 202 and calls data stored in the first memory 202 to perform various functions of the electronic device 200 and process data.

[0136] Figure 7 Only electronic devices with components are shown; those skilled in the art will understand that... Figure 7 The structure shown does not constitute a limitation on the electronic device 200, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0137] The trusted prostate cancer recognition program 203, trained on artifact targeting and adversarial methods and stored in the first memory 202 of the electronic device 200, is a combination of multiple instructions. When run in the first processor 201, it can achieve the following:

[0138] S1. Collect a prostate dataset and preprocess the prostate dataset to obtain raw prostate images of different dimensions;

[0139] S2. Construct prostate cancer recognition models of different dimensions, and import the original prostate images of different dimensions into the corresponding prostate cancer recognition models for pre-training to obtain pre-trained prostate cancer recognition models.

[0140] S3. Input the original prostate image as a reference set into the pre-trained prostate cancer recognition model to obtain predicted sample images. Select the prostate sample image with the most rectal artifact attributes from the predicted sample images as style samples for style transfer.

[0141] S4. Based on the style samples, and combining style transfer technology and adversarial attack technology, construct a proprietary adversarial sample generation model, and optimize the proprietary adversarial sample generation model by minimizing the total loss to generate proprietary adversarial samples with rectal artifact attributes.

[0142] S5. Using the original prostate image and proprietary adversarial examples, the pre-trained prostate cancer recognition model is adversarially trained to jointly optimize the parameters of the pre-trained prostate cancer recognition model, thereby obtaining a reliable prostate cancer recognition model. Further, if the modules / units integrated in the electronic device 200 are implemented as software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0143] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0144] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A reliable prostate cancer identification method based on artifact-targeted adversarial training, characterized in that, Includes the following steps: S1. Collect a prostate dataset and perform preprocessing operations on the prostate dataset to obtain raw prostate images of different dimensions; S2. Construct prostate cancer recognition models of different dimensions, and import the original prostate images of different dimensions into the corresponding prostate cancer recognition models for pre-training to obtain pre-trained prostate cancer recognition models. S3. Input the original prostate image as a reference set into the pre-trained prostate cancer recognition model to obtain predicted sample images. Select prostate sample images with rectal artifact attributes from the predicted sample images as style samples for style transfer. S4. Based on the style samples, and combining style transfer technology and adversarial attack technology, a proprietary adversarial sample generation model is constructed, and the proprietary adversarial sample generation model is optimized by minimizing the total loss to generate proprietary adversarial samples with rectal artifact attributes. S5. Using the original prostate image and proprietary adversarial examples, perform adversarial training on the pre-trained prostate cancer recognition model to jointly optimize the parameters of the pre-trained prostate cancer recognition model and obtain a reliable prostate cancer recognition model. The specific process of constructing a proprietary adversarial sample generation model based on the style samples in S4, combined with style transfer and adversarial attack techniques, is as follows: Given a raw prostate image x with category label y and a prostate cancer recognition model First, adversarial examples are obtained by performing a strong first-order attack, PGD, on the original prostate image. The formula is: ， ， in, It is to combat noise. Indicates based on The distance metric of norm, It is an artificially set perturbation budget, adversarial sample is obtained by solving an optimization problem: in, Let y be the value corresponding to the i-th class. To identify the model's predicted probability for the i-th class, This indicates the number of categories; from this, we obtain the adversarial loss. : ； When conducting adversarial attacks, select the optimal perturbation budget. The value is 8 / 255, and the attack step size is... The attack iteration count is 10; Using a given style sample adversarial examples Style transfer is performed to generate proprietary adversarial examples with rectal artifact attributes. The specific steps are as follows: using VggNet-16 as the feature extraction network E, adversarial examples are generated. and style samples Feature extraction and definition of adversarial examples and style samples The style loss between them is: in This is the index set of style layers in E, including convolutional layers 2, 4, 7, and 10. yes The Middle A style layer, The Gram matrix represents the deep features extracted from the style layer; Constructing adversarial examples The semantic loss between the original prostate image x and the original prostate image is used to ensure that the semantic information of the original prostate image is not altered. The semantic loss is defined as: in It is the index set of semantic layers in E, including convolutional layers 10, 11, 12, and 13. yes The Middle A semantic layer.

2. The reliable prostate cancer identification method based on artifact-targeted adversarial training according to claim 1, characterized in that, The preprocessing operations in S1 include cropping, normalization, registration, and cutting; S11. The cropping is to crop all images to the area around the prostate with a voxel size of a preset size and a preset spacing. The interpolation of all images is done using the third-order B-spline interpolation method. S12. The normalization is performed independently on each channel of the image, so that the voxels in the image fall within the range of [0,1]. The normalization formula is: in, This represents the normalized image; This represents all the values ​​of that image channel. Sort the numbers in ascending order to form a sequence, and then take the sequence. The value at that point is used to avoid interference from the maximum and minimum values ​​on the normalization; S13. The registration is a non-rigid registration based on B-spline transformation to achieve consistency of all data in spatial location. S14. The cutting is performed by generating a prostate segmentation mask using a prostate segmentation model, and simultaneously generating a lesion region mask by using a threshold level setting method, thereby cropping out the lesion region in each slice to form a lesion sample.

3. The reliable prostate cancer identification method based on artifact-targeted adversarial training according to claim 1, characterized in that, The different dimensions include lesion dimension, slice dimension, and patient dimension.

4. The reliable prostate cancer identification method based on artifact-targeted adversarial training according to claim 1, characterized in that, Specifically, in S3: Multiple images with obvious rectal artifacts were selected from the original prostate images as a reference set. This reference set was then input into a pre-trained recognition model to obtain predicted sample images. The predicted sample image with the lowest confidence was selected as the reference sample. A region was then cropped from the reference sample as the style sample for style transfer. .

5. The reliable prostate cancer identification method based on artifact-targeted adversarial training according to claim 1, characterized in that, In step S4, the proprietary adversarial example generation model is optimized by minimizing the total loss, specifically as follows: The final total loss function consists of three parts: an adversarial loss used to find the most powerful perturbation. Style loss used to generate proprietary adversarial examples of rectal artifacts and semantic loss used to prevent semantic information from being corrupted The final total loss function is: in and These are hyperparameters used to balance the three terms. By making a series of changes to the hyperparameters, the optimal model performance is selected. and , and Set to respectively and ; The proprietary adversarial example generation model is optimized by minimizing the total loss of the model until the training meets the set termination condition, resulting in proprietary adversarial examples with rectal artifact attributes. The overall optimization objective is: Where N is the number of training data.

6. The reliable prostate cancer identification method based on artifact-targeted adversarial training according to claim 1, characterized in that, In step S5, the optimization objective for the parameters of the pre-trained prostate cancer recognition model is: in, It is the original loss function of the prostate cancer identification model; Backpropagation is performed and the optimization steps for the pre-trained prostate cancer recognition model parameters are repeated until the set termination condition is reached, thus achieving a reliable prostate cancer recognition model.

7. A reliable prostate cancer identification system based on artifact-targeted adversarial training, characterized in that, The reliable prostate cancer identification method based on artifact-targeted adversarial training, as described in any one of claims 1-6, includes a preprocessing module, an identification model module, a style sample selection module, an adversarial sample generation module, and an adversarial training module. The preprocessing module is used to collect a prostate dataset and perform preprocessing operations on the prostate dataset to obtain raw prostate images of different dimensions. The recognition model module is used to construct prostate cancer recognition models of different dimensions. The original prostate images of different dimensions are imported into the corresponding prostate cancer recognition models for pre-training to obtain pre-trained prostate cancer recognition models. The style sample selection module is used to input the original prostate image as a reference set into the pre-trained prostate cancer recognition model to obtain predicted sample images, and to select prostate sample images with rectal artifact attributes from the predicted sample images as style samples for style transfer. The adversarial sample generation module is used to construct a proprietary adversarial sample generation model based on the style sample, combined with style transfer technology and adversarial attack technology, and to optimize the proprietary adversarial sample generation model by minimizing the total loss to generate proprietary adversarial samples with rectal artifact attributes. The adversarial training module is used to perform adversarial training on the pre-trained prostate cancer recognition model using the original prostate image and proprietary adversarial samples, jointly optimizing the parameters of the pre-trained prostate cancer recognition model to obtain a reliable prostate cancer recognition model.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores computer program instructions that can be executed by the at least one processor to enable the at least one processor to perform the reliable prostate cancer identification method based on artifact-targeted adversarial training as described in any one of claims 1-6.

9. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the reliable prostate cancer identification method based on artifact-targeted adversarial training as described in any one of claims 1-6.

Images