Bone tumor focus positioning method and device based on machine learning

Abstract

The invention relates to the technical field of focus positioning, and particularly discloses a bone tumor focus positioning method and device based on machine learning, and the method comprises the steps: carrying out the multi-level feature analysis of a bone tumor related CT medical image through employing an image processing method based on machine learning, so as to extract the deep features and shallow features of the medical image, and further, carrying out the positioning of a bone tumor focus. Attention joint perception based on semantic information guidance is carried out on deep features and shallow features of a medical image to realize comprehensive description of an anatomical structure and pathological features of the medical image, so that semantic segmentation and bone tumor focus positioning of a bone tumor related CT medical image are carried out on the basis. By means of the mode, the bone tumor focus positioning efficiency can be effectively improved, dependence on experience and professional knowledge of doctors is reduced, and therefore powerful support is provided for clinical diagnosis and treatment of bone tumors.

Description

Technical Field

[0001] The present application relates to the technical field of lesion localization, and more specifically, to a method and device for locating bone tumor lesions based on machine learning. Background Art

[0002] Bone tumors are tumors that occur in bones or their accessory tissues. In a broad sense, they include bone tumor-like lesions, benign bone tumors, and malignant bone tumors. Bone tumors have multiple histological subtypes, which can be classified into osseous, cartilaginous, fibrous, myogenic, adipose, vascular, and unspecified types according to whether the tumor originates from cells or matrix.

[0003] CT (Computed Tomography) images are reconstructed analog images that have been digitally converted. They are composed of a certain number of pixels with different grayscales from black to white arranged in an intrinsic matrix, where the grayscale of the pixels reflects the X-ray absorption coefficient of the corresponding voxels. CT images have high density resolution and can clearly display organs composed of soft tissues and bone structures, and accurately display the image of lesions on a good image background. In the diagnosis of bone tumors, CT examinations can detect small lesions earlier and more accurately display the scope of lesions, so they are widely used in clinical practice.

[0004] However, although CT images can provide rich diagnostic information, they still rely on the doctor's experience and expertise to identify and locate lesions. This method is not only time-consuming and labor-intensive, but also easily affected by subjective factors, which limits the accuracy and efficiency of the diagnostic results.

[0005] In recent years, with the advancement of medical imaging technology and computing power, the use of automation technology to assist in the localization of bone tumor lesions has become a research hotspot.

[0006] Therefore, a method and device for localizing bone tumor lesions based on machine learning are expected. Summary of the invention

[0007] In order to solve the above technical problems, this application is proposed. The embodiment of the present application provides a method and device for locating bone tumor lesions based on machine learning, which uses an image processing method based on machine learning to perform multi-level feature analysis on bone tumor-related CT medical images to extract deep features and shallow features of the medical images, and then, through the deep features and shallow features of the medical images, the attention joint perception based on semantic information guidance is performed to achieve a comprehensive description of the anatomical structure and pathological characteristics of the medical images, so as to perform semantic segmentation and bone tumor lesion localization of bone tumor-related CT medical images on this basis. In this way, the efficiency of bone tumor lesion localization can be effectively improved, and the dependence on doctors' experience and professional knowledge can be reduced, thereby providing strong support for the clinical diagnosis and treatment of bone tumors.

[0008] Accordingly, according to one aspect of the present application, a bone tumor lesion localization device based on machine learning is provided, comprising:

[0009] CT medical image acquisition module, used to acquire CT medical images related to bone tumors;

[0010] A standardization processing module, used for performing standardization processing on the bone tumor-related CT medical image to obtain a standardized bone tumor-related CT medical image;

[0011] A multi-level feature extraction module, used for performing multi-level feature extraction on the standardized bone tumor-related CT medical image to obtain a medical image shallow feature map and a medical image deep feature map;

[0012] A multi-level feature joint perception module, used for performing feature joint perception based on semantic information guidance on the shallow feature map of the medical image and the deep feature map of the medical image to obtain a deep and shallow joint perception coding feature map of the bone tumor-related medical image;

[0013] The bone tumor lesion positioning module is used to determine the position of the bone tumor lesion in the bone tumor related CT medical image based on the depth and shallowness joint perception coding feature map of the bone tumor related medical image.

[0014] According to another aspect of the present application, a method for locating bone tumor lesions based on machine learning is provided, comprising:

[0015] Obtain bone tumor-related CT medical images;

[0016] Performing standardization processing on the bone tumor-related CT medical image to obtain a standardized bone tumor-related CT medical image;

[0017] Performing multi-level feature extraction on the bone tumor-related CT medical image after the standardization process to obtain a medical image shallow feature map and a medical image deep feature map;

[0018] Performing feature joint perception based on semantic information guidance on the shallow feature map of the medical image and the deep feature map of the medical image to obtain a deep-shallow joint perception coding feature map of the medical image related to the bone tumor;

[0019] Based on the depth-shallowness joint perception coding feature map of the bone tumor-related medical image, the position of the bone tumor lesion in the bone tumor-related CT medical image is determined.

[0020] Compared with the prior art, the bone tumor lesion localization method and device based on machine learning provided by the present application uses a machine learning-based image processing method to perform multi-level feature analysis on bone tumor-related CT medical images to extract deep and shallow features of the medical images, and then, by performing attention joint perception based on semantic information guidance on the deep and shallow features of the medical images, a comprehensive description of the anatomical structure and pathological characteristics of the medical images is achieved, and on this basis, semantic segmentation and bone tumor lesion localization of bone tumor-related CT medical images is performed. In this way, the efficiency of bone tumor lesion localization can be effectively improved, and the dependence on doctors' experience and professional knowledge can be reduced, thereby providing strong support for the clinical diagnosis and treatment of bone tumors. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 It is a block diagram of a bone tumor lesion localization device based on machine learning according to an embodiment of the present application.

[0023] Figure 2 Schematic diagram of data flow of a bone tumor lesion localization device based on machine learning according to an embodiment of the present application.

[0024] Figure 3 It is a block diagram of a multi-level feature joint perception module in a bone tumor lesion localization device based on machine learning according to an embodiment of the present application.

[0025] Figure 4 It is a block diagram of a bone tumor lesion localization module in a bone tumor lesion localization device based on machine learning according to an embodiment of the present application.

[0026] Figure 5 Flow chart of a method for locating bone tumor lesions based on machine learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0028] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0029] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0030] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0031] It is worth noting that in this application, all actions to obtain data are carried out in compliance with the relevant data protection laws and policies of the country where the data is located, and with the authorization given by the owner of the corresponding device.

[0032] In response to the technical problems described in the above background, this application proposes a bone tumor lesion localization device based on machine learning, which uses a machine learning-based image processing method to perform multi-level feature analysis on bone tumor-related CT medical images to extract deep and shallow features of the medical images, and then, through the joint perception of attention guided by semantic information on the deep and shallow features of the medical images, a comprehensive description of the anatomical structure and pathological characteristics of the medical images is achieved, and on this basis, semantic segmentation and bone tumor lesion localization of bone tumor-related CT medical images are performed. In this way, the efficiency of bone tumor lesion localization can be effectively improved, and the dependence on doctors' experience and professional knowledge can be reduced, thereby providing strong support for the clinical diagnosis and treatment of bone tumors.

[0033] Figure 1It is a block diagram of a bone tumor lesion localization device based on machine learning according to an embodiment of the present application. Figure 2 FIG. 1 is a data flow diagram of a bone tumor lesion localization device based on machine learning according to an embodiment of the present application. Figure 1 and Figure 2 As shown, the bone tumor lesion localization device 100 based on machine learning includes: a CT medical image acquisition module 110, which is used to acquire bone tumor-related CT medical images; a standardization processing module 120, which is used to perform standardization processing on the bone tumor-related CT medical images to obtain standardized bone tumor-related CT medical images; a multi-level feature extraction module 130, which is used to perform multi-level feature extraction on the standardized bone tumor-related CT medical images to obtain a medical image shallow feature map and a medical image deep feature map; a multi-level feature joint perception module 140, which is used to perform feature joint perception based on semantic information guidance on the medical image shallow feature map and the medical image deep feature map to obtain a bone tumor-related medical image deep-shallow joint perception coding feature map; a bone tumor lesion localization module 150, which is used to determine the position of the bone tumor lesion in the bone tumor-related CT medical image based on the bone tumor-related medical image deep-shallow joint perception coding feature map.

[0034] In the above-mentioned bone tumor lesion localization device based on machine learning, the CT medical image acquisition module 110 is used to acquire CT medical images related to bone tumors. It should be understood that CT (Computed Tomography) is an electronic computer tomography scan, which uses X-rays to perform tomography scans on the human body. The detector receives the X-ray attenuation signal after passing through the human body and converts it into digital information. After computer processing, a human body tomographic image is reconstructed, so that the morphology and structure of bones and surrounding tissues can be clearly displayed.

[0035] The core components of a CT scanner include an X-ray source, a detector array, and a computer system. The quality of the X-ray source directly affects the image contrast and noise level. New CT machines use high-energy X-ray sources, which can achieve better penetration at lower doses, thereby reducing radiation damage to patients. The detector array determines the spatial resolution of the image, and the latest detectors have higher sensitivity and faster data acquisition speeds, which means that clear images can be captured even on fast-moving parts of the body. The performance of the computer system is also critical, and advanced reconstruction algorithms can effectively reduce artifacts and improve image quality. For example, compared with traditional filtered back projection methods, iterative reconstruction technology can generate clearer images at lower doses, which is especially important for diseases such as bone tumors that require high-resolution imaging.

[0036] Specifically, before deciding to perform a CT scan, the doctor will consider the patient's medical history, clinical manifestations, and the results of previous examinations. In cases of suspected bone tumors, the doctor may recommend a CT scan to evaluate changes in bone structure in more detail. Depending on where the lesion may be located, the doctor will choose the most suitable scanning plan, including determining the scanning range, selecting appropriate slice thickness, slice interval, and other parameters. In order to improve diagnostic accuracy, other imaging techniques such as MRI or PET-CT are sometimes combined to obtain complementary information. The preparation stage also includes calibration and maintenance of the CT equipment. Before each use, technicians will check the status of the equipment according to the guidelines provided by the manufacturer to ensure that all components are working properly. In addition, regular quality control tests are also essential to verify that the image quality and radiation dose meet the standards. High-quality images are essential for accurate diagnosis, while precise control of radiation dose helps protect patient health. Modern CT machines are usually equipped with advanced software algorithms that can maintain or even improve image resolution while reducing radiation exposure. Low-dose scanning modes are becoming increasingly important, especially in children or adult patients who need frequent follow-up examinations.

[0037] After the preparation is completed, the patient will be guided into the scanning room and placed on the CT machine. Depending on the scanning area, the patient may need to be adjusted to supine, prone, or other positions. To ensure stability and comfort, technicians may use additional support pads or fixation belts. These auxiliary tools can not only help patients maintain the correct position, but also reduce artifacts caused by movement. For scanning of certain special parts, such as the spine or around joints, special body supports may be required to optimize the imaging angle.

[0038] During the CT scan, the machine rotates around the patient, and the radiation from the X-ray tube passes through different layers of the body and is ultimately received by the detector. The data collected by the detector is digitized and transmitted to a computer system, which reconstructs a cross-sectional image. The use of multi-slice spiral CT (MDCT) has greatly improved the speed and efficiency of scanning, allowing a large number of sections to be acquired in a short period of time. This rapid imaging capability is particularly suitable for emergency assessments in emergency situations and also makes it possible to dynamically observe organ movement. For cases that require contrast enhancement, technicians will administer iodinated contrast agents intravenously before the scan begins. This method can significantly improve the visualization of blood vessels and soft tissue structures, making the boundary between tumors and surrounding normal tissues clearer. However, the use of contrast agents is not without risks, so the indications must be carefully evaluated for each patient. In addition to iodine allergy, renal insufficiency is also an important consideration, because the kidneys are responsible for eliminating most of the contrast agent in the body. For patients with potential risks, doctors may order blood tests in advance to monitor renal function indicators and adjust medication strategies if necessary.

[0039] As the CT scan is completed, the generated image data will be immediately transmitted to the workstation of the radiology department. Here, radiologists can use various tools to post-process the images, such as changing the window width and window position settings, 3D reconstruction, etc., to better show the details of the area of ​​interest. In addition, in order to further improve the image quality, in some cases, other imaging techniques such as magnetic resonance imaging (MRI) or positron emission tomography (PET) can be combined to achieve multimodal imaging and provide more information for the diagnosis of bone tumors.

[0040] In the above-mentioned bone tumor lesion localization device based on machine learning, the standardization processing module 120 is used to perform standardization processing on the bone tumor-related CT medical images to obtain the bone tumor-related CT medical images after standardization processing. It should be understood that, considering the original CT medical images during the acquisition process, due to factors such as equipment differences, scanning parameter fluctuations and individual differences among patients, there may be problems such as inconsistent data distribution, insufficient resolution and noise interference, which in turn affects the training effect and accuracy of the subsequent machine learning model. Therefore, in order to improve the processing ability of the machine learning model for bone tumor-related CT medical images, it is necessary to perform standardization processing on the bone tumor-related CT medical images. In an embodiment of the present application, the standardization processing includes normalization, data enhancement and denoising.

[0041] Specifically, normalization is to map data to a specific interval to eliminate the influence of different dimensions between data. For CT images, the pixel value reflects the attenuation degree of human tissue to X-rays, and the pixel value range under different scanning conditions may vary greatly. Through normalization, the pixel values ​​of all images are uniformly mapped to the [0,1] interval, so that the model can treat each sample more fairly during training, thereby accelerating the convergence speed and improving the stability and accuracy of training. Data enhancement refers to enhancing the contrast, brightness, sharpness and other attributes of the bone tumor-related CT medical images through image processing technology to improve the quality of the image and make the details in the image clearer, which is helpful for subsequent lesion feature extraction and analysis. Denoising is to remove irrelevant interference information in the image, such as spots and stripes, to reduce the impact of noise on lesion identification. The noise in CT images mainly comes from the electronic noise of the equipment, the physiological movement of the patient, and quantum noise. The principle of denoising is to process the image using a filtering algorithm, and by designing a specific filter, retain the useful signals in the image (such as the edges and textures of bones and tumors) and remove the noise signal. In a specific example of the present application, a Gaussian filter algorithm is used for image denoising. Gaussian filtering is a commonly used linear smoothing filter algorithm, which is based on the characteristics of the Gaussian function and performs weighted averaging on each pixel in the image and its neighboring pixels to achieve effective noise suppression and smoothing. Since the Gaussian function takes the largest value at the center and the smaller the value is, the weight of the central pixel is the largest during the filtering process, and the weight of the neighboring pixel decreases with the increase of the distance, thereby achieving smoothing of the image, effectively removing noise, and keeping the edge and detail information of the image as much as possible. After standardization, the CT image data is unified in scale and distribution, and the noise is effectively suppressed, so that a clearer and purer image can be provided for the subsequent localization of bone tumor lesions, which helps to improve the accuracy of localization of bone tumor lesions.

[0042] In the above-mentioned bone tumor lesion localization device based on machine learning, the multi-level feature extraction module 130 is used to perform multi-level feature extraction on the standardized bone tumor-related CT medical image to obtain a shallow feature map of the medical image and a deep feature map of the medical image. In a specific example of the present application, the multi-level feature extraction module 130 is used to: input the standardized bone tumor-related CT medical image into a multi-scale medical image feature scanner based on a hollow pyramid network model to obtain the shallow feature map of the medical image and the deep feature map of the medical image. Specifically, since the characteristics of bone tumors have significant multi-scale characteristics, tumors of different sizes and shapes and their surrounding tissues present feature information of different scales in CT images. For example, tiny tumor nodules may require high-resolution detail features to identify, while the overall relationship and macroscopic structure of the tumor and surrounding tissues require more abstract, large-scale features to describe, and the traditional single-scale feature extraction method is difficult to fully capture the rich feature information in the CT image. Therefore, the present application adopts a hollow pyramid network model to construct a multi-scale medical image feature scanner to extract the multi-scale feature information of the standardized bone tumor-related CT medical image. It should be understood that the dilated convolution is based on the standard convolution kernel. By inserting holes in the convolution kernel (i.e. skipping some pixels), the convolution kernel can cover a larger area when performing the convolution operation without increasing parameters and calculations. In the dilated pyramid network, convolution layers with different dilation rates are arranged according to a specific structure to form a pyramid-like structure. The convolution layer with a smaller dilation rate can capture the local detail information of the image, such as the edge, texture and fine structure of the tumor, so as to generate a shallow feature map of the medical image, which helps to accurately depict the boundary and morphology of the tumor; while the convolution layer with a larger dilation rate can obtain a wider range of contextual information, such as the overall structure of the tissue, the category characteristics of the tumor and the spatial relationship with the surrounding tissue, so as to generate a deep feature map of the medical image, which helps to understand the nature of the tumor and the relationship with the surrounding tissue from a macroscopic level. Through this multi-scale feature extraction method, the feature information of the bone tumor-related CT medical image at different scales after the standardized processing can be fully captured, which helps to improve the accuracy and reliability of bone tumor lesion localization.

[0043] In the above-mentioned bone tumor lesion localization device based on machine learning, the multi-level feature joint perception module 140 is used to perform feature joint perception based on semantic information guidance on the shallow feature map of the medical image and the deep feature map of the medical image to obtain a deep and shallow joint perception encoding feature map of the bone tumor-related medical image. It should be understood that the shallow feature map of the medical image and the deep feature map of the medical image respectively reflect the local detail information and the overall structural information of the CT image. Due to the different receptive fields of feature extraction, the two have differences in feature dimensions and spatial positions. Directly performing simple feature splicing and fusion may lead to information loss and confusion. In this regard, the present application proposes a feature joint perception method based on semantic information guidance, which first uses the high-resolution detail information in the shallow feature map of the medical image to guide the low-resolution overall structural information in the deep feature map of the medical image for feature alignment, so that each feature point in the deep feature map can be associated with the corresponding detail information in the shallow feature map, thereby enhancing the deep feature's perception of local details, guiding the information migration and fusion between deep and shallow features, so that the fused feature map retains the local detail information of the image and integrates the overall structural information, realizing the effective joint perception of deep and shallow features. Among them, Figure 3 FIG. 1 is a block diagram of a multi-level feature joint perception module in a bone tumor lesion localization device based on machine learning according to an embodiment of the present application. Figure 3 As shown, the multi-level feature joint perception module 140 includes: a field-guided modulation unit 141, which is used to perform field-guided modulation on the deep feature map of the medical image based on the shallow feature map of the medical image to obtain a field-modulated deep feature map of the medical image; and a deep-shallow feature joint perception unit 142, which is used to fuse the shallow feature map of the medical image and the field-modulated deep feature map of the medical image to obtain the deep-shallow joint perception coding feature map of the bone tumor-related medical image.

[0044] Specifically, in a specific example of the present application, the field guided modulation unit 141 is used to: upsample the medical image deep feature map to obtain an upsampled medical image deep feature map, wherein the upsampled medical image deep feature map has the same size as the medical image shallow feature map, which is expressed by the formula:

[0045] F h =Upsampling(F h )

[0046] Among them, F h represents the deep feature map of the medical image, Upsampling(·) represents the upsampling operation, F h Represents the upsampled medical image deep feature map.

[0047] Specifically, during the downsampling process of the network, the deep feature map of the medical image obtains more abstract and high-level semantic information, but at the same time, the spatial resolution is reduced and the feature size becomes smaller. Therefore, in order to enable deep and shallow features to interact with each other at the same spatial scale, first, the deep feature map of the medical image is upsampled so that it has the same feature size as the shallow feature map of the medical image, so as to restore the spatial structure of the deep features and obtain the upsampled deep feature map of the medical image. Here, the upsampling process can use algorithms such as bilinear interpolation and transposed convolution.

[0048] In a specific example of the present application, the field guided modulation unit 141 is further used to: perform feature connection on the upsampled medical image deep feature map and the medical image shallow feature map and input them into a semantic information field predictor based on gated convolution to obtain a medical image deep and shallow feature semantic information field, which is expressed as:

[0049] Ω=Conv 1×1 {Conv 3×3 [cat(F l , F h ')]}

[0050] Among them, F l represents the shallow feature map of medical images, cat(·,·) represents feature cascade, Conv 3×3 Represents convolution processing based on 3×3 convolution kernel, Conv 1×1 Represents the convolution processing based on 1×1 convolution kernel, and Ω represents the semantic information field of the depth and shallowness features of medical images.

[0051] That is, the upsampled medical image deep feature map and the medical image shallow feature map are further feature-connected so that features at different levels can interact directly and are input into a semantic information field predictor based on gated convolution. The activation degree of the convolution output is adjusted by introducing additional learning parameters, thereby dynamically adjusting the weight distribution of the deep features according to the contextual association information in the deep and shallow connection features of the medical images, and generating a semantic information field of the deep and shallow features of the medical images.

[0052] In a specific example of the present application, the field guided modulation unit 141 is further used to: map the upsampled medical image deep feature map to the medical image deep and shallow feature semantic information field to obtain the field modulated medical image deep feature map, which is expressed by the formula:

[0053] F t =F h ⊙Ω

[0054] Among them, ⊙ represents the point product by position, F tRepresents the deep feature map of field-modulated medical images.

[0055] That is, by mapping the upsampled medical image deep feature map to the medical image deep and shallow feature semantic information field, each feature point in the deep feature map can be finely adjusted according to the guidance of the semantic information field to make it more matched with the detail information in the shallow feature map, thereby enhancing the perception accuracy of the deep feature for local details and obtaining the field modulated medical image deep feature map. It should be understood that this process is similar to superimposing a semantic guide guided by shallow features on the deep features, so that the deep features can more accurately locate the key details in the image, such as the boundaries of the tumor, texture changes, etc., under the guidance of semantic information, while retaining the overall structural information, thereby optimizing the information flow of the deep and shallow features at the semantic level.

[0056] Specifically, in a specific example of the present application, the depth-shallowness joint perception unit 142 is used to: input the shallow feature map of the medical image and the deep feature map of the field-modulated medical image into a feature joint perception module based on a multiple attention structure to obtain the depth-shallowness joint perception encoding feature map of the bone tumor-related medical image, which is expressed by the formula:

[0057] L(X)=σ(BN(PWconv1(F t )))

[0058] G(X) = σ(BN(PWConv1(GAP(F l ))))

[0059]

[0060] Among them, PWConv1(·) represents point-by-point convolution, σ represents the Sigmoid activation function, BN represents batch normalization, L(X) represents the deep attention modulation coding feature map of medical images, G(X) represents the shallow attention modulation coding feature map of medical images, represents point addition, and F(X) represents the depth-shallowness joint perceptual coding feature map of bone tumor-related medical images.

[0061] Specifically, in order to make full use of the information complementary advantages of shallow and deep features, the present application further sends the shallow feature map of the medical image and the deep feature map of the field modulated medical image to a feature joint perception module based on a multiple attention structure for attention fusion processing. The multiple attention structure is used to enhance the sensitivity to key feature information, dynamically adjust the feature attention, so that the features that are more relevant to the location of bone tumor lesions receive greater attention, while the secondary features are relatively weakened, so as to achieve effective focusing and enhanced fusion of the deep features and shallow features related to bone tumor lesions, and then form a multi-level feature representation system, namely, the deep and shallow joint perception encoding feature map of bone tumor-related medical images, thereby providing more comprehensive and accurate feature information support for subsequent bone tumor lesion identification.

[0062] In the above-mentioned bone tumor lesion localization device based on machine learning, the bone tumor lesion localization module 150 is used to determine the position of the bone tumor lesion in the bone tumor-related CT medical image based on the depth-shallowness joint perception coding feature map of the bone tumor-related medical image. Figure 4 FIG. 1 is a block diagram of a bone tumor lesion localization module in a bone tumor lesion localization device based on machine learning according to an embodiment of the present application. Figure 4 As shown, the bone tumor lesion localization module 150 includes: a semantic segmentation unit 151, which is used to perform semantic segmentation on the depth-shallowness joint perceptual coding feature map of the bone tumor-related medical image to obtain a medical image semantic segmentation result; and a lesion localization unit 152, which is used to determine the position of the bone tumor lesion in the bone tumor-related CT medical image based on the medical image semantic segmentation result.

[0063] Specifically, in a specific example of the present application, the semantic segmentation unit 151 is used to: input the joint perception coding feature map of the bone tumor-related medical image depth into a segmentation model based on U-Net to obtain the semantic segmentation result of the medical image. It should be understood that semantic segmentation is the task of classifying each pixel in the image into different semantic categories, and by classifying each area in the medical image pixel by pixel, the accurate identification and positioning of bone tumor lesions can be achieved. U-Net is a classic convolutional neural network model specially designed for medical image segmentation. Its unique encoder-decoder structure and jump connection design make it perform well in the field of medical image segmentation. The present application inputs the joint perception coding feature map of the bone tumor-related medical image depth into the U-Net model, which can make full use of its powerful segmentation ability to accurately classify different tissues and structures in the medical image, thereby clearly segmenting the bone tumor area and providing accurate regional division for subsequent lesion positioning. Specifically, during the decoding process, the U-Net model gradually restores the spatial resolution of the feature map through upsampling operations, and introduces jump connections to splice the bone tumor-related medical image depth-shallow joint perception encoding feature map with the feature map of the decoding stage, integrating feature information at different levels to enhance the model's perception of local details. Finally, at the output of the model, each pixel is classified by the softmax function to obtain the semantic segmentation result of the medical image, that is, the probability distribution map of each pixel being classified as a bone tumor lesion area or other normal tissue area. Through this semantic segmentation result, the location, morphology, and size of the bone tumor lesion can be intuitively observed.

[0064] In a preferred example of the present application, inputting the bone tumor-related medical image depth-shallowness joint perceptual coding feature map into a U-Net-based segmentation model to obtain a medical image semantic segmentation result includes:

[0065] First, the bone tumor-related medical image depth-shallowness joint perception coding feature map is expanded into a bone tumor-related medical image depth-shallowness joint perception coding feature vector, and the L1 distance and L2 distance between the i-th eigenvalue and the j-th eigenvalue of the bone tumor-related medical image depth-shallowness joint perception coding feature vector are calculated, thereby obtaining the first bone tumor-related medical image depth-shallowness joint perception description distance matrix and the second bone tumor-related medical image depth-shallowness joint perception description distance matrix, which are expressed as follows:

[0066] D (1) i，j =d L1 (v i , v j )

[0067] D (2) i，j =d L2(v i , v j )

[0068] Among them, d L1 (·,·) and d L2 (·,·) represent the L1 distance and L2 distance metric functions respectively, v i and v j denote the i-th eigenvalue and j-th eigenvalue of the joint perception encoding feature vector of the bone tumor-related medical image depth and shallowness, respectively, and D (1) i，j represents the (i, j)th eigenvalue of the depth-shallowness joint perception description distance matrix of the first bone tumor-related medical image, D (2) i，j represents the (i, j)th eigenvalue of the depth-shallowness joint perceptual description distance matrix of the second bone tumor-related medical image;

[0069] Secondly, the weighted sum of the first bone tumor-related medical image depth-shallow joint perception description distance matrix and the second bone tumor-related medical image depth-shallow joint perception description distance matrix is ​​calculated, and each eigenvalue λ1 to λ of the distance weighted sum matrix is ​​determined. n , so as to arrange the eigenvalues ​​to obtain the distance eigenvector of the joint perception description of the depth of the bone tumor-related medical image, which can be expressed as follows:

[0070] V λ =(λ1,λ2,…,λ n )

[0071] Among them, λ1, λ2 and λ n Respectively represent the first, second, and nth eigenvalues ​​of the distance weighted sum matrix, V λ Represents the distance eigenvector of the joint perception description of the depth and shallowness of bone tumor related medical images;

[0072] Then, the depth-shallowness joint perception description distance eigenvector V of the bone tumor-related medical image is λ Performing interpolation to obtain an interpolated eigenvector of a bone tumor-related medical image depth-shallowness joint perception description having the same length as the bone tumor-related medical image depth-shallowness joint perception encoding eigenvector;

[0073] Next, the bone tumor-related medical image depth and shallow joint perception coding feature vector is matrix-multiplied with the first bone tumor-related medical image depth and shallow joint perception description distance matrix to obtain the bone tumor-related medical image depth and shallow joint perception description intermediate feature vector, and the second bone tumor-related medical image depth and shallow joint perception description distance matrix is ​​multiplied with the self-correlation matrix of the bone tumor-related medical image depth and shallow joint perception coding feature vector, that is, Perform matrix multiplication to obtain the intermediate feature matrix of the joint perception description of the depth and shallowness of bone tumor-related medical images;

[0074] Then, after matrix multiplication of the intermediate feature vector of the bone tumor-related medical image depth-shallowness joint perception description and the intermediate feature matrix of the bone tumor-related medical image depth-shallowness joint perception description, further dot multiplication is performed with the interpolation eigenvector of the bone tumor-related medical image depth-shallowness joint perception description to obtain an optimized bone tumor-related medical image depth-shallowness joint perception description encoding vector;

[0075] Finally, the optimized bone tumor-related medical image depth-shallowness joint perception encoding feature vector is input into a U-Net-based segmentation model to obtain the medical image semantic segmentation result.

[0076] Here, when the shallow feature map of the medical image and the deep feature map of the medical image respectively represent the shallow image semantic features and deep image semantic features of the bone tumor-related CT medical image after standardization processing, when performing feature joint perception, the difference in image feature depth and order will cause sparse joint perception of the shallow and deep joint perception encoding feature map of the bone tumor-related medical image, thereby reducing the accuracy of the medical image semantic segmentation result obtained by inputting the U-Net-based segmentation model due to the lack of semantic segmentation inference degree.

[0077] Therefore, the first distance matrix and the second distance matrix of the bone tumor-related medical image depth and shallow joint perceptual coding feature vector after the bone tumor-related medical image depth and shallow joint perceptual coding feature map is expanded are used as the fine-grained metric association cluster representation of the bone tumor-related medical image depth and shallow joint perceptual coding feature vector, and the dynamic planning of the relationship between association clusters of different association clusters is performed on the bone tumor-related medical image depth and shallow joint perceptual coding feature vector and the self-association representation of the bone tumor-related medical image depth and shallow joint perceptual coding feature vector, respectively, to simulate the sparse activation of the association system based on neuron clusters, and the fine-grained predictable sparsity of the bone tumor-related medical image depth and shallow joint perceptual coding feature vector is coordinated with the intrinsic representation of the metric association cluster of the first distance matrix and the second distance matrix of the bone tumor-related medical image depth and shallow joint perceptual coding feature vector, so as to avoid the lack of association caused by sparsity affecting the loss of semantic segmentation reasoning, and improve the accuracy of the medical image semantic segmentation result obtained by inputting the bone tumor-related medical image depth and shallow joint perceptual coding feature map into the U-Net-based segmentation model.

[0078] Specifically, the lesion localization unit 152 is used to determine the position of the bone tumor lesion in the bone tumor-related CT medical image based on the semantic segmentation result of the medical image. That is, according to the pixel distribution of the tumor area in the semantic segmentation result, the position of the tumor can be determined using a geometric calculation method to provide key information for subsequent treatment plans. For example, the centroid coordinates of the tumor area can be calculated as the approximate position of the tumor.

[0079] In summary, the bone tumor lesion localization device based on machine learning according to the embodiment of the present application is explained, which uses a machine learning-based image processing method to perform multi-level feature analysis on bone tumor-related CT medical images to extract deep and shallow features of the medical images, and then, through the deep and shallow features of the medical images, the attention joint perception based on semantic information guidance is performed to achieve a comprehensive description of the anatomical structure and pathological characteristics of the medical images, so as to perform semantic segmentation and bone tumor lesion localization of bone tumor-related CT medical images on this basis. In this way, the efficiency of bone tumor lesion localization can be effectively improved, and the dependence on doctors' experience and professional knowledge can be reduced, thereby providing strong support for the clinical diagnosis and treatment of bone tumors.

[0080] Furthermore, a method for locating bone tumor lesions based on machine learning is also provided.

[0081] Figure 5 FIG. 1 is a flow chart of a method for locating bone tumor lesions based on machine learning according to an embodiment of the present application. Figure 5 As shown, the bone tumor lesion localization method based on machine learning includes the following steps: S1, acquiring bone tumor related CT medical images; S2, performing standardization processing on the bone tumor related CT medical images to obtain the standardized bone tumor related CT medical images; S3, performing multi-level feature extraction on the standardized bone tumor related CT medical images to obtain a medical image shallow feature map and a medical image deep feature map; S4, performing feature joint perception based on semantic information guidance on the medical image shallow feature map and the medical image deep feature map to obtain a bone tumor related medical image deep and shallow joint perception coding feature map; S5, determining the position of the bone tumor lesion in the bone tumor related CT medical image based on the bone tumor related medical image deep and shallow joint perception coding feature map.

[0082] Here, those skilled in the art will appreciate that the specific operations of each step in the above-mentioned method for locating bone tumor lesions based on machine learning have been described in the above reference. Figures 1 to 4 The description of the machine learning-based bone tumor lesion localization device has been introduced in detail, and therefore, its repeated description will be omitted.

[0083] The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to being implemented by adopting the above specific details.

[0084] In the above embodiments, the description of each embodiment has its own emphasis. For the parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiment described above is only schematic. For example, the unit division is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0085] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference to a figure in a claim should not be considered as limiting the claim to which it relates.

[0086] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0087] Finally, it should be noted that the above description has been given for the purpose of illustration and description. In addition, the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A bone tumor lesion localization device based on machine learning, characterized in that: include: CT medical image acquisition module, used to acquire CT medical images related to bone tumors; A standardization processing module, used for performing standardization processing on the bone tumor-related CT medical image to obtain a standardized bone tumor-related CT medical image; A multi-level feature extraction module, used for performing multi-level feature extraction on the standardized bone tumor-related CT medical image to obtain a medical image shallow feature map and a medical image deep feature map; A multi-level feature joint perception module, used for performing feature joint perception based on semantic information guidance on the shallow feature map of the medical image and the deep feature map of the medical image to obtain a deep and shallow joint perception coding feature map of the bone tumor-related medical image; The bone tumor lesion localization module is used to determine the position of the bone tumor lesion in the bone tumor related CT medical image based on the depth and shallowness joint perception coding feature map of the bone tumor related medical image.

2. The bone tumor lesion localization device based on machine learning according to claim 1, characterized in that: The standardization process includes normalization, data enhancement and denoising.

3. The bone tumor lesion localization device based on machine learning according to claim 2, characterized in that: The multi-level feature extraction module is used to: The standardized bone tumor-related CT medical image is input into a multi-scale medical image feature scanner based on a hollow pyramid network model to obtain a shallow feature map of the medical image and a deep feature map of the medical image.

4. The bone tumor lesion localization device based on machine learning according to claim 3, characterized in that: The multi-level feature joint perception module includes: A field guided modulation unit, configured to perform field guided modulation on the medical image deep feature map based on the medical image shallow feature map to obtain a field modulated medical image deep feature map; The deep and shallow feature joint perception unit is used to fuse the shallow feature map of the medical image and the deep feature map of the field modulated medical image to obtain the deep and shallow joint perception coding feature map of the bone tumor related medical image.

5. The bone tumor lesion localization device based on machine learning according to claim 4, characterized in that: The field guidance modulation unit is used for: Upsampling the medical image deep feature map to obtain an upsampled medical image deep feature map, wherein the upsampled medical image deep feature map has the same size as the medical image shallow feature map; The upsampled medical image deep feature map and the medical image shallow feature map are feature-connected and then input into a semantic information field predictor based on gated convolution to obtain a medical image deep and shallow feature semantic information field; The upsampled medical image deep feature map is mapped to the medical image deep and shallow feature semantic information field to obtain the field-modulated medical image deep feature map.

6. The bone tumor lesion localization device based on machine learning according to claim 5, characterized in that: The depth-shallowness feature joint perception unit is used to: The shallow feature map of the medical image and the deep feature map of the field modulated medical image are input into a feature joint perception module based on a multiple attention structure to obtain the shallow and deep joint perception coding feature map of the bone tumor-related medical image.

7. The bone tumor lesion localization device based on machine learning according to claim 6, characterized in that: The bone tumor lesion positioning module comprises: A semantic segmentation unit, used for performing semantic segmentation on the bone tumor-related medical image depth-shallowness joint perception coding feature map to obtain a medical image semantic segmentation result; A lesion localization unit is used to determine the position of the bone tumor lesion in the bone tumor related CT medical image based on the semantic segmentation result of the medical image.

8. The bone tumor lesion localization device based on machine learning according to claim 7, characterized in that: The semantic segmentation unit is used to: The depth-shallowness joint perceptual coding feature map of the bone tumor-related medical image is input into a U-Net-based segmentation model to obtain the semantic segmentation result of the medical image.

9. A method for locating bone tumor lesions based on machine learning, characterized in that: include: Obtain bone tumor-related CT medical images; Performing standardization processing on the bone tumor-related CT medical image to obtain a standardized bone tumor-related CT medical image; Performing multi-level feature extraction on the bone tumor-related CT medical image after the standardization process to obtain a medical image shallow feature map and a medical image deep feature map; Performing feature joint perception based on semantic information guidance on the shallow feature map of the medical image and the deep feature map of the medical image to obtain a deep-shallow joint perception coding feature map of the medical image related to the bone tumor; Based on the depth-shallowness joint perception coding feature map of the bone tumor-related medical image, the position of the bone tumor lesion in the bone tumor-related CT medical image is determined.

Images