A deep learning-based breast contrast-enhanced x-ray background enhancement automatic suppression method and related device

Abstract

The application discloses a breast contrast-enhanced X-ray background enhancement automatic suppression method based on deep learning and related equipment, and the application takes a subtraction image as a main input, introduces a low-energy image as an auxiliary input, so that a convolutional neural network model can accurately distinguish 'lesion enhancement' and 'background enhancement' according to tissue structure differences, and solves the problem that it is difficult to distinguish by simply relying on a subtraction image; by using a trained convolutional neural network model to perform image segmentation on the low-energy image and the subtraction image, interference of BPE, especially moderate and severe BPE, in the CEM imaging process of the low-energy image and the high-energy image can be suppressed, the lesion and the background can be automatically distinguished, directional removal of the background enhancement is realized, the contrast of the lesion and the background is significantly improved, and this is helpful for a clinician to find a tiny lesion hidden by a highlighted background and improve the sensitivity and specificity of diagnosis. The application is widely applied to the technical field of medical image processing.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and related equipment for automatic suppression of background enhancement in breast contrast-enhanced X-rays based on deep learning. Background Technology

[0002] Breast cancer is the most common malignant tumor among women worldwide, and early, accurate diagnosis is crucial for improving prognosis. Contrast-enhanced mammography (CEM) is a breast imaging technique that combines the high spatial resolution of traditional digital mammography (MG) with the functional imaging advantages of magnetic resonance imaging (MRI). CEM involves intravenously injecting an iodine contrast agent and using high- and low-energy X-rays for exposure. Subtraction techniques are then used to remove the background of breast tissue, thereby highlighting the tumor vascular bed information that has taken up the contrast agent.

[0003] However, background parenchymal enhancement (BPE) is a significant interfering factor in the clinical application of contrast imaging (CEM). BPE refers to the uptake of contrast agent by normal breast fibroglandular tissue, appearing as diffuse or focal high-density areas on subtraction images. In cases of moderate to severe BPE, the degree of background enhancement is often similar to the degree of lesion enhancement, causing suspicious lesions, especially small lesions or non-mass-like enhancing lesions, to be masked by the background signal. Thus, when visually examining CEM images, it is difficult for doctors to distinguish which parts of the image belong to lesions and which parts are simply background without lesions, leading to undesirable outcomes such as missed diagnoses and misdiagnoses. Furthermore, clinicians need to rely on subjective experience to eliminate background interference and search for lesion areas within large areas of bright background when interpreting images, resulting in poor diagnostic consistency among different doctors. Summary of the Invention

[0004] To address at least one of the aforementioned technical problems, the present invention aims to provide a method and related equipment for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning.

[0005] On one hand, embodiments of the present invention include a method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning, the method comprising the following steps: Establish a convolutional neural network model; the convolutional neural network model is a U-Net structure; Acquire low-energy images and subtraction images; Based on the low-energy image and the subtraction image, construct the first channel input data, the second channel input data, and the third channel input data; The first channel input data, the second channel input data, and the third channel input data are input into the convolutional neural network model for processing; Obtain the three-channel probability map output by the convolutional neural network model; the three-channel probability map represents the probability that a pixel in the subtraction image belongs to a region with no background enhancement, a region with substantial background enhancement, or a lesion region.

[0006] Furthermore, the acquisition of low-energy images and subtraction images includes: The subjects were intravenously injected with iodine contrast agent; The low-energy image is obtained by exposing the subject's breast glandular tissue with low-energy X-rays. High-energy X-rays are used to expose and image the breast glandular tissue of the subject to obtain high-energy images; The low-energy image and the high-energy image are subjected to subtraction processing to obtain the subtracted image.

[0007] Further, the step of constructing the first channel input data, the second channel input data, and the third channel input data based on the low-energy image and the subtraction image includes: The subtraction image is used as the input data for the first channel; The low-energy image is used as the input data for the second channel; The low-energy image is processed locally to obtain the third channel input data.

[0008] Further, the step of performing local computational processing on the low-energy image to obtain the third channel input data includes: According to the formula

[0009] Perform calculations; among which, The coordinates in the low-energy image are... pixel values, This represents the radius of the local sliding window. =2 +1, This indicates the size of the local sliding window. and Represents the variable to be summed. The coordinates in the low-energy image are... The local mean of pixel values ​​within the local sliding window. Indicates local standard deviation; The local standard deviation is used as the input data for the third channel.

[0010] Further, the step of performing local computational processing on the low-energy image to obtain the third channel input data includes: According to the formula

[0011] Perform calculations; among which, The coordinates in the low-energy image are... pixel values, Represents the local gradient assignment plot; The local gradient assignment map is used as the input data for the third channel.

[0012] Furthermore, the deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement also includes: During the training phase, blind annotation is performed on the low-energy image and / or the subtraction image to obtain pixel-level semantic segmentation labels; the pixel-level semantic segmentation labels indicate that each pixel in the low-energy image and / or the subtraction image belongs to a non-enhanced background region, a substantially enhanced background region, or a lesion region. Calculate the joint loss function based on the three-channel probability map and the pixel-level semantic segmentation label; Based on the joint loss function, the model parameters of the convolutional neural network model are optimized using the backpropagation algorithm.

[0013] Further, the step of calculating the joint loss function based on the three-channel probability map and the pixel-level semantic segmentation label includes: Through formula

[0014] Perform calculations; among which, Indicates the channel number. Represents the category weight coefficient. This refers to the pixel-level semantic segmentation label. This represents the three-channel probability map. This represents the weighted cross-entropy loss; Through formula

[0015] Perform calculations; among which, Indicates the smoothing term. Indicates Dice loss; According to the formula

[0016] Perform calculations; among which, and Indicates weight, This represents the joint loss function.

[0017] Furthermore, the deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement also includes: In the reasoning stage, according to the formula

[0018] Perform calculations; among which, The probability in the three-channel probability map represents the probability that a pixel in the subtraction image belongs to the actual background enhancement region. To determine the threshold, The suppression coefficient, Background enhancement suppression mask; According to the formula

[0019] Perform calculations; among which, This refers to the subtraction image. This represents the enhanced image after suppressing the enhancement of the actual background.

[0020] On the other hand, embodiments of the present invention also include a computer device, including a memory and a processor, the memory for storing at least one program, and the processor for loading at least one program to execute the deep learning-based automatic suppression method for breast contrast enhancement X-ray background enhancement in the embodiments.

[0021] On the other hand, embodiments of the present invention also include a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement in the embodiments.

[0022] The beneficial effects of this invention are as follows: The deep learning-based automatic suppression method for background enhancement in breast contrast-enhanced X-rays, as described in the embodiments, uses subtraction images as the main input and introduces low-energy images as auxiliary inputs. This enables the convolutional neural network model to accurately distinguish between "lesion enhancement" and "background enhancement" based on differences in tissue structure, solving the problem of difficulty in distinguishing between them when relying solely on subtraction images. By using a trained convolutional neural network model to segment low-energy and subtraction images, interference from BPE, especially moderate to severe BPE, during CEM imaging of low-energy and high-energy images can be suppressed. This automatically distinguishes between lesions and background, achieving targeted removal of background enhancement, significantly improving the contrast between lesions and background. This helps clinicians discover tiny lesions obscured by bright backgrounds, enhancing the sensitivity and specificity of diagnosis. Attached Figure Description

[0023] Figure 1This is a schematic diagram illustrating the steps of the deep learning-based automatic suppression method for breast contrast enhancement X-ray background enhancement in the embodiment. Figure 2 This is a schematic diagram of the structure of the convolutional neural network model in the embodiment; Figure 3 This is a flowchart illustrating an application example of the deep learning-based automatic suppression method for breast contrast enhancement X-ray background enhancement in the embodiments. Detailed Implementation

[0024] Current image post-processing techniques mainly rely on simple window width and level adjustments or single threshold segmentation, which cannot effectively distinguish between "lesion enhancement" and "background enhancement," and are prone to losing lesion information while removing the background.

[0025] Based on the above principles, this embodiment provides a deep learning-based method for automatically suppressing background enhancement in breast contrast-enhanced X-rays. (Refer to...) Figure 1 The deep learning-based method for automatic suppression of background enhancement in breast contrast-enhanced X-rays includes the following steps: S1. Establish a convolutional neural network model; S2. Acquire low-energy images and subtraction images; S3. Based on the low-energy image and the subtraction image, construct the first channel input data, the second channel input data, and the third channel input data; S4. Input the first channel input data, the second channel input data, and the third channel input data into the convolutional neural network model for processing; S5. Obtain the three-channel probability map output by the convolutional neural network model.

[0026] In step S1, establish Figure 2 The U-Net structure shown is a convolutional neural network model. (Refer to...) Figure 2This convolutional neural network model includes an input layer, an encoder, a decoder, and an output layer. The input layer has a size of 512×512 mm. The encoder includes four downsampling modules, each consisting of two 3×3 convolutional layers, a batch normalization layer, a ReLU activation function, and a 2×2 max-pooling layer. The number of convolutional kernels in these four downsampling modules are 64, 128, 256, and 512, respectively. Each downsampling module can be used to extract multi-scale deep features from the input data. A bottleneck layer is set in both the encoder and decoder, with 1024 convolutional kernels. The decoder includes four upsampling modules. The four upsampling modules are symmetrical to the four downsampling modules, with the number of convolutional kernels in the four upsampling modules being 512, 256, 128, and 64, respectively. Each upsampling module uses a deconvolutional layer to restore the feature map size. Skip connections are set between downsampling and upsampling modules of the same size. Through skip connections, each upsampling module can fuse the shallow features output by the encoder part when performing data upsampling processing, thereby preserving edge details. Finally, the output layer uses a 1×1 convolutional layer with a Softmax activation function to output a three-channel probability map with a size of 512×512×3.

[0027] In this embodiment, when performing step S2, which is to acquire the low-energy image and the subtraction image, the following steps can be specifically performed: S201. Intravenous injection of iodine contrast agent into the subject; S202. Low-energy X-rays are used to expose and image the breast glandular tissue of the subject to obtain low-energy images; S203. High-energy X-rays are used to expose and image the breast glandular tissue of the subject to obtain high-energy images; S204. Perform subtraction processing on the low-energy image and the high-energy image to obtain the subtracted image.

[0028] In step S202, the energy of the low-energy X-rays used can be in the range of 26-33 kVp. In step S203, the energy of the high-energy X-rays used can be in the range of 44-50 kVp.

[0029] The low-energy image obtained in step S202 and the high-energy image obtained in step S203 These images were obtained by directly performing X-ray imaging on subjects who had been injected with iodine contrast agent. During steps S202 and S203, the subject's pose and imaging angle remained constant, thus producing low-energy images. and high-energy images coordinates in The same point indicates the same location in the subject's breast glandular tissue.

[0030] In this embodiment, for the low-energy images obtained by actual measurement and high-energy images It can also perform preprocessing such as image registration, denoising, and grayscale normalization. Among these, image registration enables the processing of low-energy images. and high-energy images Pixels with the same coordinates in the image correspond to the same location in the subject's breast glandular tissue; denoising can reduce the low-energy image. and high-energy images In the noise, grayscale normalization can make low-energy images and high-energy images The pixel values ​​are mapped to the [0, 1] interval. For example, in this embodiment, this can be achieved using the formula...

[0031] Perform grayscale normalization. Specifically, for low-energy images... When performing grayscale normalization, For low-energy images , For low-energy images The minimum value among all pixel values ​​(grayscale values), For low-energy images The maximum value among all pixel values ​​(grayscale values), For low-energy images The grayscale normalization result; when dealing with high-energy images When performing grayscale normalization, For high-energy images , For high-energy images The minimum value among all pixel values ​​(grayscale values), For high-energy images The maximum value among all pixel values ​​(grayscale values), For high-energy images The result of grayscale normalization.

[0032] In this embodiment, since grayscale normalization only changes the individual values ​​of each pixel but does not affect the feature information of each pixel, it can still be used. and This represents low-energy and high-energy images after grayscale normalization.

[0033] In step S204, the low-energy images can be processed separately. and high-energy images This is converted into corresponding absorbance information. (Using low-energy images...) For example, it can record low-energy images. The intensity of incident X-rays used for imaging Low-energy images Represents pixels The corresponding pixel value is the measured X-ray intensity, according to the formula. =

[0034] The absorbance information of the low-energy image was calculated. The low-energy images of the above process. Replace with high-energy image By processing, the absorbance information of high-energy images can be calculated. Set the calibration weighting factors (related to the characteristics of the X-ray imaging equipment, such as tube voltage, filter plates, and detectors). and According to the formula

[0035] Calculated subtraction image .

[0036] In step S3, based on the low-energy image and subtraction image The first channel input data, the second channel input data, and the third channel input data are constructed.

[0037] In this embodiment, a (grayscale normalized) subtraction image can be used. As the first input data channel, it reflects the distribution of iodine contrast agent in the subject's breast glandular tissue, and can provide information on blood flow enhancement in the subject's breast glandular tissue; using (grayscale normalized) low-energy images. As a second-channel input data, it can provide information on the breast glandular structure (e.g., breast glandular density and anatomical structure) and lesions of the subject's breast tissue; for low-energy images The result of local computation processing is used as the input data for the third channel.

[0038] In this embodiment, two methods can be used to process low-energy images. Local calculations are performed. Specifically, in the first method, according to the formula...

[0039] Perform the calculations. Among them, Indicates the coordinates in the low-energy image as pixel values, This represents the radius of the local sliding window. =2 +1, Indicates the size of the local sliding window. and The variable being summed, in the above formula, is equivalent to setting a variable of size . Local sliding window for low-energy images Perform sampling. Indicates the coordinates in the low-energy image as The local mean of pixel values ​​within a local sliding window is used to calculate the low-energy image. Local standard deviation It can be used as a low-energy image. The texture feature map is then used as the third channel input data for processing in the convolutional neural network model. This third channel input data can help identify normal gland texture.

[0040] In this embodiment, in the second method, according to the formula

[0041] Perform calculations to obtain low-energy images. Local gradient assignment plot This data serves as the third channel input to the convolutional neural network model for processing. This third channel input highlights the low-energy image. The edge and texture details in the image help convolutional neural network models distinguish between the edges of lesions and the edges of glands.

[0042] In step S4, the first channel input data (subtraction image) is... Second channel input data (low-energy image) ) and third channel input data (local standard deviation) Or a local gradient assignment graph The input is fed into a convolutional neural network model for processing.

[0043] In step S5, the processing results of the first channel input data, the second channel input data, and the third channel input data output by the convolutional neural network model are obtained, i.e., the three-channel probability map. .in, The subtraction image represented by the convolutional neural network model. Pixels The probability of belonging to an unenhanced background region. The subtraction image represented by the convolutional neural network model. Pixels The probability of belonging to a region of substantial background enhancement. The subtraction image represented by the convolutional neural network model. Pixels The probability of belonging to the lesion area satisfies + + =1.

[0044] In this embodiment, steps S1-S5 can be applied to either the training or inference phase. During the training phase, the convolutional neural network model is not yet fully trained and therefore does not accurately recognize subtraction images. The performance of the non-enhanced background region, the substantially enhanced background region, and the lesion region in the model can be evaluated. Therefore, the following steps can be performed to train the convolutional neural network model: S6A. Perform blind annotation on low-energy images and / or subtraction images to obtain pixel-level semantic segmentation labels; S7A. Calculate the joint loss function based on the three-channel probability map and pixel-level semantic segmentation labels; S8A. Based on the joint loss function, optimize the model parameters of the convolutional neural network model using the backpropagation algorithm.

[0045] In step S6A, a "double-blind annotation + arbitration" strategy can be used to construct pixel-level semantic segmentation labels. Specifically, due to the limitations of low-energy images... and high-energy images Image registration was performed, so low-energy images can be selected. Or subtraction image One of them is used for blind annotation. Using subtraction images... For example, two attending physicians (Physician A and Physician B) with more than 5 years of experience in breast imaging diagnosis can independently analyze the subtraction images. Initial annotation is performed; for example, physician A and physician B each draw the subtraction image. The image is divided into lesion areas, areas with enhanced solid background, and areas without enhanced background. The system then automatically compares the annotation results of two physicians. If the intersection-over-union (IoU) ratio of the two (e.g., the lesion area drawn by physician A and the lesion area drawn by physician B) is greater than 0.8, they are considered consistent. For disputed areas, another chief physician with over 15 years of experience conducts final review and correction to determine the final pixel-level semantic segmentation label. Pixel-level semantic segmentation labels represent subtraction images. Pixels It can be classified as a non-enhancing background area, a solid background enhancement area, or a lesion area. Among them, a lesion area refers to the area of ​​the lesion site in the corresponding subject's breast glandular tissue, a solid background enhancement area refers to the non-lesion site in the corresponding subject's breast glandular tissue where BPE has occurred, and a non-enhancing background area refers to the non-lesion site in the corresponding subject's breast glandular tissue where BPE has not occurred.

[0046] The above process is based on a subtraction image This explanation uses annotation as an example. Multiple subtraction images can be obtained. For each subtraction image By performing appropriate annotations, each subtraction image can be obtained separately. The corresponding pixel-level semantic segmentation tags.

[0047] In step S7A, using the formula

[0048] Calculate the weighted cross-entropy loss .in, Indicates the channel number. This indicates the weight corresponding to this channel, which can balance the problem of having too few lesion pixels. For example... When =1, This represents the subtraction image labeled with pixel-level semantic segmentation tags. A certain pixel on Tags belonging to channel 1 (unenhanced background area) =0 indicates that this pixel does not belong to the unenhanced background area. =1 indicates that this pixel belongs to an unenhanced background area. A three-channel probability map representing the output of a convolutional neural network model. The same pixel corresponding to the first channel probability value The above process can calculate the weighted cross-entropy loss corresponding to a single pixel, traversing the subtraction image. The subtractive image is obtained by averaging the weighted cross-entropy loss of all pixels on the image. Weighted cross-entropy loss .

[0049] Next, through the formula

[0050] Calculate Dice loss .in, This represents the smoothing term, which is a constant. It is used in calculating the weighted cross-entropy loss. The process is similar; by calculating the Dice loss corresponding to a single pixel, the subtraction image is traversed. The average Dice loss of all pixels on the image is calculated to obtain the subtractive image. Dice loss .

[0051] Finally, set the weights. and According to the formula

[0052] Calculated subtraction image joint loss function .

[0053] In step S8A, the joint loss function calculated in step S7A is used. By optimizing the model parameters of the convolutional neural network model through backpropagation, it is possible to use subtraction images. Train the convolutional neural network model.

[0054] The examples in steps S7A-S8A above illustrate the use of a single subtraction image. The convolutional neural network model was trained. Multiple subtraction images were obtained. Given their pixel-level semantic segmentation labels, repeat the above steps to achieve the use of multiple subtraction images. The convolutional neural network model is trained. The training phase ends when the number of rounds of steps S7A-S8A is reached, or when the model parameters of the optimized convolutional neural network model converge, resulting in a well-trained convolutional neural network model.

[0055] After the training phase is complete, the convolutional neural network model can accurately identify subtraction images. The performance of the non-enhanced background region, the solid background enhanced region, and the lesion region in the image is considered. Therefore, the following steps can be performed to use a trained convolutional neural network model to process the subtraction image. Segmentation process: S6B. According to the formula

[0056] Calculations are performed to obtain the background enhancement suppression mask. .in, Three-channel probability plot The value in the image represents the probability that a pixel in the subtraction image belongs to a region of substantial background enhancement. To determine the threshold, you can set... =0.5; The suppression coefficient can be set to 0 ≤ <1, where =0 indicates that the subtraction image is completely removed. The pixel values ​​in >0 indicates that the subtraction image is retained. In this embodiment, the background outline can be set as a portion of the image. =0.1. Then, according to the formula...

[0057] Perform calculations to make the subtraction image Each pixel value in the image is multiplied by the background enhancement suppression mask. The values ​​at the same positions are used to obtain the enhanced image after suppressing the enhancement of the actual background. .

[0058] By executing step S6B, the subtraction image was processed. The probability that the middle part belongs to the actual background enhancement region Larger (greater than the judgment threshold) The pixel value of the pixel with a value of 0.5 is multiplied by the suppression coefficient. Thus, the proportion of these pixels is retained. (In this embodiment) =0.1 (i.e., 10%) of the background signal is used to maintain the anatomical reference, and these pixels are removed by 1- (In this embodiment, it is 90%) of the actual background enhancement; while for other pixels (the probability of belonging to the actual background enhancement region) If the probability of a pixel belonging to a region of substantial background enhancement is relatively small, then the pixel value of these pixels is multiplied by 1 to keep the original pixel value unchanged. The above method applies different processing to pixels with different probabilities of belonging to the substantial background enhancement region, thereby achieving the effect of processing the subtraction image. The suppression of background enhancement of breast parenchyma in BPE, i.e., the obtained enhanced image. This is an image after BPE (broadcast emulsion) was suppressed to enhance the background of the breast parenchyma.

[0059] The deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement in this embodiment has the following advantages: 1. By using subtraction images (first channel input data) as the main input, and introducing low-energy images (second channel input data) and texture features of low-energy images (third channel input data) as auxiliary inputs, the deep learning model (convolutional neural network model) can accurately distinguish between "lesion enhancement" and "background enhancement" based on differences in tissue structure, thus solving the problem that it is difficult to distinguish by simply relying on subtraction images; 2. No manual drawing or threshold adjustment is required; one-click automatic identification and suppression of BPE greatly simplifies the doctor's workflow. 3. By using a trained convolutional neural network model to segment low-energy and subtraction images, interference from BPE, especially moderate to severe BPE, during CEM imaging of low-energy and high-energy images can be suppressed. This significantly improves the contrast between lesions and background, helping clinicians to detect tiny lesions obscured by bright backgrounds and enhancing the sensitivity and specificity of diagnosis.

[0060] The following is an application example of a deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement (steps S1-S5). The flowchart of this application example is as follows: Figure 3 As shown. (Refer to...) Figure 3 This application example includes the following steps: 1. Data Preparation CEM (Cyclic Emission Modeling) data of a breast cancer patient were exported from the hospital's PACS system. This case included various breast density subtypes and showed enhancement against a solid background. Low-energy and subtraction images were collected from the patient in cephalothorax, medial and lateral oblique views. To accommodate network input, the image resolution was uniformly scaled to 512×512 pixels, and grayscale normalization was performed to the [0,1] range.

[0061] 2. Feature fusion input Construct three-channel input data: Channel 1 is taken from... This reflects the distribution of the iodine contrast agent; Channel 2 is taken from... This reflects the density and anatomical structure of mammary gland tissue; Channel 3 calculation The local gradient magnitude plot is shown in the following formula:

[0062] This channel highlights edge and texture details in low-energy images, helping the model distinguish lesion edges from gland edges.

[0063] 3. Deep learning model inference The three-channel data mentioned above are input into a pre-trained U-Net model. The specific model parameters are set as follows: the input layer is 512 × 512 × 3. The encoder has 4 layers with convolutional kernels of 64, 128, 256, and 512 respectively. Each convolutional layer is followed by Batch Normalization and ReLU activation. The bottleneck layer has 1024 convolutional kernels. The decoder has 4 layers, symmetrical to the encoder, with the number of convolutional kernels halved layer by layer. The output layer has an output size of 512 × 512 × 3, corresponding to the probability maps for the background, BPE, and lesion categories.

[0064] 4. Background Suppression and Output Based on the model output The image generates a suppression mask. A threshold of T = 0.5 is set, and the suppression coefficient is... 10% of the background signal is retained to maintain anatomical reference, while 90% of the enhancement intensity is removed. For all For pixels with a grayscale value >0.5, multiply their original grayscale value by 0.1; for... For pixels with a grayscale value greater than 0.5, the original grayscale value is preserved. Finally, the processed image is remapped back to its original resolution and displayed on the monitor.

[0065] A computer program for automatically suppressing breast contrast-enhanced X-ray background enhancement based on deep learning, as described in this embodiment, can be written into a computer device or storage medium. When the computer program is read and run, the method for automatically suppressing breast contrast-enhanced X-ray background enhancement based on deep learning, as described in this embodiment, can be executed, thereby achieving the same technical effect as the method for automatically suppressing breast contrast-enhanced X-ray background enhancement based on deep learning in this embodiment.

[0066] It should be noted that, unless otherwise specified, when a feature is referred to as "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. Furthermore, the descriptions of "upper," "lower," "left," and "right" used in this disclosure are only relative to the relative positional relationships of the components of this disclosure in the accompanying drawings. The singular forms "a," "an," and "the" used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. Moreover, unless otherwise defined, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this embodiment specification is only for describing particular embodiments and is not intended to limit the invention. The term "and / or" as used in this embodiment includes any combination of one or more of the associated listed items.

[0067] It should be understood that although various elements may be described in this disclosure using terms such as "second," "third," etc., these elements should not be limited to these terms. These terms are used only to distinguish elements of the same type from one another. For example, an element may also be referred to as a second element without departing from the scope of this disclosure, and similarly, a second element may also be referred to as an element. The use of any and all instances or exemplary language ("e.g.," "such as," etc.) provided in this embodiment is intended only to better illustrate embodiments of the invention and, unless otherwise required, does not impose a limitation on the scope of the invention.

[0068] It should be recognized that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a specific and predefined manner—according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0069] Furthermore, the procedures described in this embodiment can be performed in any suitable order unless otherwise indicated by this embodiment or otherwise obviously contradict the context. The procedures (or variations and / or combinations thereof) described in this embodiment can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. A computer program includes a plurality of instructions executable by one or more processors.

[0070] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices, etc. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention of this embodiment includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps above in conjunction with a microprocessor or other data processor. When programmed according to the methods and techniques of the invention, the invention also includes the computer itself.

[0071] A computer program can be applied to input data to perform the functions of this embodiment, thereby transforming the input data to generate output data stored in non-volatile memory. The output information can also be applied to one or more output devices, such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including specific visual depictions of physical and tangible objects generated on the display.

[0072] The above are merely preferred embodiments of the present invention. The present invention is not limited to the above-described embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention, as long as they achieve the technical effects of the present invention by the same means, should be included within the scope of protection of the present invention. Within the scope of protection of the present invention, the technical solutions and / or implementation methods can have various modifications and variations.

Claims

1. A method for automatic suppression of background enhancement in breast contrast-enhanced X-rays based on deep learning, characterized in that, The deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement includes: Establish a convolutional neural network model; the convolutional neural network model is a U-Net structure; Acquire low-energy images and subtraction images; Based on the low-energy image and the subtraction image, construct the first channel input data, the second channel input data, and the third channel input data; The first channel input data, the second channel input data, and the third channel input data are input into the convolutional neural network model for processing; Obtain the three-channel probability map output by the convolutional neural network model; the three-channel probability map represents the probability that a pixel in the subtraction image belongs to a region with no background enhancement, a region with substantial background enhancement, or a lesion region.

2. The method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning according to claim 1, characterized in that, The acquisition of low-energy images and subtraction images includes: The subjects were intravenously injected with iodine contrast agent; The low-energy image is obtained by exposing the subject's breast glandular tissue with low-energy X-rays. High-energy X-rays are used to expose and image the breast glandular tissue of the subject to obtain high-energy images; The low-energy image and the high-energy image are subjected to subtraction processing to obtain the subtracted image.

3. The method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning according to claim 1, characterized in that, The step of constructing first-channel input data, second-channel input data, and third-channel input data based on the low-energy image and the subtraction image includes: The subtraction image is used as the input data for the first channel; The low-energy image is used as the input data for the second channel; The low-energy image is processed locally to obtain the third channel input data.

4. The method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning according to claim 3, characterized in that, The step of performing local computational processing on the low-energy image to obtain the third channel input data includes: According to the formula Perform calculations; among which, The coordinates in the low-energy image are... pixel values, This represents the radius of the local sliding window. =2 +1, This indicates the size of the local sliding window. and Represents the variable to be summed. The coordinates in the low-energy image are... The local mean of pixel values ​​within the local sliding window. Indicates local standard deviation; The local standard deviation is used as the input data for the third channel.

5. The method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning according to claim 3, characterized in that, The step of performing local computational processing on the low-energy image to obtain the third channel input data includes: According to the formula Perform calculations; among which, The coordinates in the low-energy image are... pixel values, Represents the local gradient assignment plot; The local gradient assignment map is used as the input data for the third channel.

6. The method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning according to claim 2, characterized in that, The deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement also includes: During the training phase, blind annotation is performed on the low-energy image and / or the subtraction image to obtain pixel-level semantic segmentation labels; the pixel-level semantic segmentation labels indicate that each pixel in the low-energy image and / or the subtraction image belongs to a non-enhanced background region, a substantially enhanced background region, or a lesion region. Calculate the joint loss function based on the three-channel probability map and the pixel-level semantic segmentation label; Based on the joint loss function, the model parameters of the convolutional neural network model are optimized using the backpropagation algorithm.

7. The method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning according to claim 6, characterized in that, The step of calculating the joint loss function based on the three-channel probability map and the pixel-level semantic segmentation label includes: Through formula Perform calculations; among which, Indicates the channel number. Represents the category weight coefficient. This refers to the pixel-level semantic segmentation label. This represents the three-channel probability map. This represents the weighted cross-entropy loss; Through formula Perform calculations; among which, Indicates the smoothing term. Indicates Dice loss; According to the formula Perform calculations; among which, and Indicates weight, This represents the joint loss function.

8. The method for automatic suppression of breast contrast-enhanced X-ray background enhancement based on deep learning according to any one of claims 1-7, characterized in that, The deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement also includes: In the reasoning stage, according to the formula Perform calculations; among which, The probability in the three-channel probability map represents the probability that a pixel in the subtraction image belongs to the actual background enhancement region. To determine the threshold, The suppression coefficient, Background enhancement suppression mask; According to the formula Perform calculations; among which, This refers to the subtraction image. This represents the enhanced image after suppressing the enhancement of the actual background.

9. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory is used to store at least one program, and the processor is used to load at least one program to execute the deep learning-based method for automatic suppression of breast contrast-enhanced X-ray background enhancement as described in any one of claims 1-8.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the deep learning-based automatic suppression method for breast contrast-enhanced X-ray background enhancement as described in any one of claims 1-8.

Images