Embolism focus segmentation method and system based on medical image processing

Abstract

The invention relates to the technical field of medical image processing, in particular to an embolism focus segmentation method and system based on medical image processing. The system comprises a medical image feature registration module, an image frequency band fusion enhancement module, a deep network lesion segmentation module and a lesion segmentation image fusion optimization module, an embolism CT image and an embolism MRI medical image can be acquired, and feature point minimization registration and image frequency band fusion are performed to generate an embolism fusion image; performing histogram equalization processing on the embolism fusion image to generate an embolism comparison standard image; constructing a corresponding deep network lesion segmentation model, and performing network lesion segmentation processing, up-sampling and element-by-element splicing fusion to generate an embolism lesion feature fusion image; and performing morphological optimization operation on the lesion boundary corresponding to the embolism lesion feature fusion image to obtain an embolism lesion segmentation result. According to the invention, accurate segmentation of the embolism focus in the medical image can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for segmenting embolic lesions based on medical image processing. Background Art

[0002] Embolic lesions refer to pathological phenomena such as local ischemia and tissue necrosis caused by emboli (such as blood clots or fat masses) blocking blood vessels. Embolic lesions are commonly found in the vascular systems of vital organs such as the brain, lungs, and kidneys. In recent years, the application of deep learning, particularly convolutional neural networks (CNNs), to medical image segmentation has made significant progress. Deep learning can automatically learn features from large amounts of annotated data and effectively process complex image content, resulting in more accurate segmentation results. However, the appearance of embolic lesions varies significantly across different imaging modalities, and the lesions are complex in morphology and have fuzzy boundaries. This places high demands on existing deep learning segmentation methods. Currently, embolic lesion segmentation typically relies on medical imaging technologies such as CT (computed tomography), MRI (magnetic resonance imaging), and ultrasound. Traditional lesion segmentation methods often rely on manual annotation or simple threshold-based image processing techniques. These methods are not only limited in accuracy but also susceptible to factors such as noise and image quality, resulting in poor lesion segmentation results. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a method and system for segmenting embolic lesions based on medical image processing to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a medical image processing-based embolic lesion segmentation system includes the following modules:

[0005] a medical image feature registration module, configured to obtain an embolism CT image containing anatomical structure correspondences and an embolism MRI medical image containing soft tissue contrast correspondences, and perform feature point minimization registration on the embolism CT image and the embolism MRI medical image to generate an embolism CT registered image and an embolism MRI registered image;

[0006] The image band fusion enhancement module is used to perform image band fusion on the embolism CT registered image and the embolism MRI registered image to generate an embolism fusion image; and perform histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image;

[0007] The deep network lesion segmentation module is used to build a deep network lesion segmentation model corresponding to the self-attention module and the channel attention module. Based on the deep network lesion segmentation model, the embolic lesion area corresponding to the embolic contrast standard image is segmented. The cross entropy loss and Dice coefficient loss are used as the optimization targets of the deep network lesion segmentation model to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network.

[0008] The lesion segmentation map fusion optimization module is used to upsample and element-by-element concatenate the embolic lesion feature segmentation maps corresponding to each layer of the convolutional network to generate an embolic lesion feature fusion map; morphological optimization operations are performed on the lesion boundaries corresponding to the embolic lesion feature fusion map to obtain the embolic lesion segmentation results.

[0009] Furthermore, the medical image feature registration module includes the following functions:

[0010] Obtain embolic CT images containing corresponding anatomical structures;

[0011] Acquire embolic MRI medical images containing soft tissue contrast;

[0012] Performing same-scale space mapping on the embolism CT image and the embolism MRI medical image to generate corresponding embolism CT image and embolism MRI image in the same-scale space;

[0013] The corresponding embolism CT images and embolism MRI images in the same scale space are registered by minimizing feature points to generate embolism CT registered images and embolism MRI registered images.

[0014] Furthermore, performing feature point minimization registration on the corresponding embolism CT image and embolism MRI image in the same scale space includes:

[0015] Perform feature point matching and screening on the corresponding embolism CT images and embolism MRI images in the same scale space to obtain a set of embolism multimodal feature point pairs;

[0016] The corresponding feature point context descriptor is obtained for each pair of feature points corresponding to the CT image and the MRI image in the embolism multimodal feature point pair set, and the feature point similarity calculation is performed on each pair of feature points in the embolism multimodal feature point pair set based on the feature point context descriptor to obtain the embolism feature similarity corresponding to each pair of feature points;

[0017] Based on the embolism feature similarity corresponding to each pair of feature points, each pair of feature points in the embolism multimodal feature point pair set is screened for similarity and representativeness, so as to obtain a set of representative embolism feature point pairs;

[0018] Calculating the Euclidean distance between each pair of representative feature points in the set of representative feature points of embolism to obtain the Euclidean distance value between each pair of representative feature points of embolism;

[0019] Based on the minimization of the Euclidean distance between each pair of representative feature points of embolism, the corresponding embolism CT images and embolism MRI images in the same scale space are minimized and registered to generate embolism CT registered images and embolism MRI registered images.

[0020] Furthermore, the matching and screening of feature points of the corresponding embolism CT images and embolism MRI images in the same scale space includes:

[0021] Multimodal feature deep mining is performed on the corresponding embolic CT images and embolic MRI images in the same scale space. Edge detection is used to extract the edge contour features of the embolic area tissue in the CT image, including shape, size, and density. The texture features of the CT image are calculated using the gray-level co-occurrence matrix, including contrast, correlation, energy, and entropy. The phase difference, proton density, and relaxation time features corresponding to different embolic tissues in the MRI image are analyzed to obtain the embolic multimodal image feature set.

[0022] Perform feature correlation analysis between the CT image features and MRI image features at each position point in the same scale space within the embolism multimodal image feature set to obtain the feature correlation correspondence between the CT and MRI image features at each position point;

[0023] Based on the feature correlation correspondence between the CT and MRI image features at each position point, feature point matching and screening are performed between the CT image features and MRI image features at each position point in the same scale space in the embolism multimodal image feature set to obtain the embolism multimodal feature point pair set.

[0024] Furthermore, the image band fusion enhancement module includes the following functions:

[0025] Performing frequency domain conversion on the embolism CT registered image and the embolism MRI registered image to generate an embolism CT spectrum map and an embolism MRI spectrum map;

[0026] Perform wavelet frequency band decomposition on the embolism CT spectrum and the embolism MRI spectrum to obtain different characteristic frequency bands corresponding to the embolism CT and MRI spectrum;

[0027] Perform phase-consistent weighted fusion on the different characteristic frequency bands corresponding to the CT and MRI spectrograms of the embolism to calculate the phase similarity of the CT and MRI spectrograms in each frequency band, and perform weighted fusion of the frequency bands based on the phase similarity in each frequency band to generate an initial fused spectrogram of the embolism;

[0028] The fusion quality control of the initial fusion spectrum of embolism is optimized to obtain the peak signal-to-noise ratio (PSNR) of different monitoring fusion images during the fusion process, and the frequency band weights are reallocated according to the PSNR to perform fusion control optimization to generate the fusion spectrum of embolism bands. The spatial domain inverse operation is performed on the fusion spectrum of embolism bands to generate the fusion image of embolism.

[0029] The embolization fusion image was processed with histogram equalization to generate the embolization contrast standard image.

[0030] Furthermore, the different characteristic frequency bands corresponding to the embolism CT and MRI spectra include different frequency bands corresponding to differences in embolism density and structural characteristics in the embolism CT spectra and different frequency bands corresponding to differences in embolism morphology, size and position in the embolism MRI spectra.

[0031] Furthermore, the deep network lesion segmentation module includes the following functions:

[0032] Build a deep network lesion segmentation model that includes a self-attention module and a channel attention module;

[0033] Based on the deep network lesion segmentation model, network lesion segmentation training is performed on the corresponding embolic lesion area in the embolic contrast standard image to output the embolic lesion segmentation results corresponding to each layer of the convolutional network;

[0034] The cross entropy loss and Dice coefficient loss corresponding to the deep network lesion segmentation model are calculated based on the embolic lesion segmentation results corresponding to each layer of the convolutional network. The cross entropy loss is used to measure the classification error between the model prediction results and the actual lesion area label. The Dice coefficient loss is used to evaluate the degree of overlap corresponding to the lesion segmentation results. The cross entropy loss and Dice coefficient loss are used as the optimization targets corresponding to the deep network lesion segmentation model. The stochastic gradient descent algorithm is used to dynamically adjust the hyperparameters corresponding to the deep network lesion segmentation model to generate a deep lesion segmentation optimization model.

[0035] The embolism contrast standard image is re-input into the deep lesion segmentation optimization model for lesion optimization segmentation processing to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network.

[0036] Furthermore, the deep network lesion segmentation model is specifically composed of 5 layers of 5x5 convolutional layers and a 1x1 pooling layer connected to each convolutional layer, so that a self-attention module is connected between the corresponding convolutional layer and pooling layer of each layer to focus on the embolic lesion area in the fused image to capture the long-distance dependency between lesions, and a channel attention module is connected between the corresponding channels of each layer to automatically learn the importance between channel features of different layers.

[0037] Furthermore, the lesion segmentation map fusion optimization module includes the following functions:

[0038] Perform convolution and upsampling operations on the embolic lesion feature segmentation map corresponding to each convolutional network layer, so as to adjust the embolic lesion feature segmentation maps corresponding to different scales in each convolutional network layer to the same size, thereby obtaining the corresponding embolic lesion feature maps at the same size;

[0039] The corresponding embolic lesion feature maps of the same size are spliced ​​and fused element by element to generate an embolic lesion feature fusion map containing lesion detail feature information and corresponding embolic lesion semantic information;

[0040] A morphological optimization operation is performed on the lesion boundary corresponding to the fusion map of embolic lesion features to remove the corresponding noise area using dilation and erosion operations, and the corresponding lesion boundary is smoothed to obtain the embolic lesion segmentation result.

[0041] Furthermore, the present invention also provides a method for segmenting embolic lesions based on medical image processing, which is used to execute the above-mentioned system for segmenting embolic lesions based on medical image processing. The method for segmenting embolic lesions based on medical image processing includes:

[0042] Step S1: acquiring an embolism CT image containing anatomical structures and an embolism MRI medical image containing soft tissue contrast, and performing feature point minimization registration on the embolism CT image and the embolism MRI medical image to generate an embolism CT registered image and an embolism MRI registered image;

[0043] Step S2: performing image band fusion on the embolism CT registered image and the embolism MRI registered image to generate an embolism fusion image; performing histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image;

[0044] Step S3: Construct a deep network lesion segmentation model corresponding to the self-attention module and the channel attention module, and perform network lesion segmentation processing on the corresponding embolic lesion area in the embolic contrast standard image based on the deep network lesion segmentation model. Combine the cross entropy loss and Dice coefficient loss as the corresponding optimization targets of the deep network lesion segmentation model to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network;

[0045] Step S4: Upsample and element-by-element concatenation of the embolic lesion feature segmentation map corresponding to each layer of the convolutional network to generate an embolic lesion feature fusion map; perform morphological optimization operations on the lesion boundaries corresponding to the embolic lesion feature fusion map to obtain the embolic lesion segmentation result.

[0046] Beneficial effects of the present invention:

[0047] The embolic lesion segmentation system based on medical image processing proposed in the present invention is generally composed of a medical image feature registration module, an image band fusion enhancement module, a deep network lesion segmentation module, and a lesion segmentation map fusion optimization module. Compared with the existing technology, the beneficial effect of the present application is that by acquiring embolic CT images and embolic MRI images and performing feature point minimization registration, the two different modalities of medical images can be effectively spatially aligned to ensure their accurate comparison on the same anatomical structure. Embolic CT images usually have a high anatomical structure resolution and are suitable for detailed observation of the morphological structure of the embolic area; while embolic MRI images are superior in soft tissue contrast and can provide clearer soft tissue lesion information. Through precise image registration, the registration error caused by different imaging modalities, viewing angles and patient posture changes can be eliminated, so that the two image modalities can be analyzed in the same spatial coordinate system, providing accurate basic data for subsequent lesion detection, analysis and fusion, so that the accuracy and reliability of embolic lesion detection can be improved by comprehensive processing of multimodal images. Secondly, the embolic CT and embolic MRI registered images are processed using image band fusion technology, which effectively combines the advantages of the two images. Embolic CT images have high spatial resolution and can clearly display anatomical structures. Embolic MRI images, due to their high soft tissue contrast, have an advantage in displaying soft tissue-related features such as lesions and tumors. Through band fusion, the anatomical structure information in the embolic CT images and the soft tissue information in the embolic MRI images can be organically combined to create a fused image with more comprehensive anatomical and lesion information. This fused image provides a clearer and more comprehensive view for lesion detection and segmentation, effectively improving the accuracy of lesion area identification. Histogram equalization, as an image preprocessing technique, can further improve the contrast of the embolic fusion image, making the grayscale distribution of the image more balanced, effectively enhancing the recognizability of embolic lesions in the image, and making the embolic lesion area more clearly distinguishable from the surrounding normal tissue. This helps to improve the efficiency of subsequent automated lesion segmentation, thereby reducing the impact of image noise and quality factors.Then, by constructing a deep network lesion segmentation model including a self-attention module and a channel attention module, the model can automatically focus on the lesion area and ignore irrelevant background information when processing embolism contrast standard images. The self-attention mechanism can help the network to globally interact information between different regions in the image and can adaptively weight according to the important areas in the embolism image. The channel attention module can better capture the characteristic information of embolic lesions by adjusting the weights of different channels. By combining cross-entropy loss and Dice coefficient loss, the model training process is optimized, which not only ensures the pixel-level accuracy between the segmentation results and the true annotations, but also further improves the connectivity and consistency of the segmentation results through Dice coefficient loss, so that the network model can not only perform well in precision, but also achieve a high level of recall rate, further improving the lesion segmentation effect. Finally, by upsampling and element-by-element concatenation of the embolic lesion feature segmentation map in each layer of the convolutional network, this process utilizes feature information at different levels and combines low-level features (such as edges, textures, etc.) with high-level features (such as overall structural information) to perform more accurate lesion segmentation. The upsampling operation enables the low-resolution feature map to be restored to a higher resolution, thereby enhancing the segmentation accuracy and further improving the expressiveness of the feature map, which can better capture complex lesion morphology and boundary information. In addition, the morphological optimization operation can refine the boundaries of the embolic lesions, eliminate segmentation errors caused by noise or false positives, ensure that the segmentation results are smoother and more coherent, reduce unnatural breaks or over-smoothing of the boundaries, and make the final embolic lesion segmentation results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0049] Figure 1 Schematic diagram of the modules of the embolic lesion segmentation system based on medical image processing of the present invention;

[0050] Figure 2 for Figure 1 Schematic diagram of the functional flow of the Traditional Chinese Medicine image feature registration module;

[0051] Figure 3 for Figure 1 Schematic diagram of the functional flow of the mid-image frequency band fusion enhancement module. DETAILED DESCRIPTION

[0052] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0053] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0054] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0055] To achieve this, please refer to Figures 1 to 3 The present invention provides an embolic lesion segmentation system based on medical image processing, the system comprising the following modules:

[0056] a medical image feature registration module, configured to obtain an embolism CT image containing anatomical structure correspondences and an embolism MRI medical image containing soft tissue contrast correspondences, and perform feature point minimization registration on the embolism CT image and the embolism MRI medical image to generate an embolism CT registered image and an embolism MRI registered image;

[0057] The image band fusion enhancement module is used to perform image band fusion on the embolism CT registered image and the embolism MRI registered image to generate an embolism fusion image; and perform histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image;

[0058] The deep network lesion segmentation module is used to build a deep network lesion segmentation model corresponding to the self-attention module and the channel attention module. Based on the deep network lesion segmentation model, the embolic lesion area corresponding to the embolic contrast standard image is segmented. The cross entropy loss and Dice coefficient loss are used as the optimization targets of the deep network lesion segmentation model to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network.

[0059] The lesion segmentation map fusion optimization module is used to upsample and element-by-element concatenate the embolic lesion feature segmentation maps corresponding to each layer of the convolutional network to generate an embolic lesion feature fusion map; morphological optimization operations are performed on the lesion boundaries corresponding to the embolic lesion feature fusion map to obtain the embolic lesion segmentation results.

[0060] In the embodiment of the present invention, please refer to Figure 1 FIG. 1 is a schematic diagram of the modules of the embolic lesion segmentation system based on medical image processing according to the present invention. In this example, the embolic lesion segmentation system based on medical image processing includes the following modules:

[0061] S1: A medical image feature registration module, which is used to obtain an embolism CT image containing anatomical structures and an embolism MRI medical image containing soft tissue contrast, and perform feature point minimization registration on the embolism CT image and the embolism MRI medical image to generate an embolism CT registered image and an embolism MRI registered image;

[0062] In an embodiment of the present invention, a set of embolic CT images containing anatomical structures and embolic MRI medical images containing soft tissue contrast are obtained from a database or clinical data source. The embolic CT images typically provide high-contrast bone and vascular structures, while the embolic MRI images can present detailed information about soft tissue, such as lesions in the brain or other organs. After obtaining these original medical images, image registration technology is used to register the embolic CT images with the embolic MRI images. The purpose is to align the anatomical structures in the two image spaces, thereby combining the advantages of both for subsequent processing. The registration method uses feature point minimization registration, employing a mutual information-based registration algorithm to calculate the transformation matrix between the CT and MRI images. An iterative optimization algorithm is then used to spatially align the anatomical landmarks in the CT and MRI images. The registration results in the generation of two registered images, ultimately obtaining an embolic CT registered image and an embolic MRI registered image, which are spatially consistent.

[0063] S2: Image band fusion enhancement module, used to perform image band fusion on the embolism CT registered image and the embolism MRI registered image to generate an embolism fusion image; perform histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image;

[0064] In an embodiment of the present invention, after registration, the embolization CT and MRI registered images are subjected to image band fusion. Image band fusion is a technique that combines image frequency domain information (such as low-frequency and high-frequency components) by decomposing them. Specifically, the embolization CT and MRI images are Fourier transformed to obtain frequency domain representations. Subsequently, by selectively retaining low-frequency information (primarily from anatomical structures) and high-frequency information (primarily from lesion details) from each image, the frequency band information of the two is fused together to generate a fused embolization image. After fusion, the fused embolization image incorporates the advantages of both imaging modalities, including the anatomical details of the CT image and the soft tissue contrast of the MRI image. The fused embolization image is then processed using histogram equalization. By adjusting the image's grayscale distribution, the image contrast is enhanced, making the lesion area more prominent. Histogram equalization automatically adjusts the image's grayscale levels to achieve a more uniform brightness, highlighting the important lesion area, and ultimately generating a standard embolization contrast image.

[0065] S3: Deep network lesion segmentation module, which is used to build a deep network lesion segmentation model corresponding to the self-attention module and the channel attention module. Based on the deep network lesion segmentation model, the embolic lesion area corresponding to the embolic contrast standard image is segmented. The cross entropy loss and Dice coefficient loss are used as the optimization targets of the deep network lesion segmentation model to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network.

[0066] In an embodiment of the present invention, a deep learning model suitable for embolic lesion segmentation is constructed. The model adopts a combination of a self-attention module and a channel attention module to enhance the model's ability to focus on the lesion area. The self-attention module is used to capture the relationship between distant pixels in the image, enabling the network to model global information; while the channel attention module can adaptively adjust the importance of different feature channels to enhance the recognition of lesion area features. Specifically, a convolutional neural network (CNN) architecture is designed, combining self-attention and channel attention mechanisms, and an attention module is added after each convolution layer to adjust the weight of the feature map. The deep network lesion segmentation model uses embolic contrast standard images as input, extracts features through a multi-layer convolutional network, and generates a predicted segmentation map of the embolic lesion. During the training process, cross entropy loss and Dice coefficient loss are used as optimization targets. Cross entropy loss is used to measure the difference between the model output and the true label, while Dice coefficient loss focuses on improving the segmentation accuracy of the model in the lesion area. The network weights are optimized through the backpropagation algorithm, and finally a feature segmentation map of the embolic lesion is generated for each layer of the convolutional network.

[0067] S4: Lesion segmentation map fusion optimization module, which is used to upsample and element-by-element concatenate the embolic lesion feature segmentation maps corresponding to each layer of the convolutional network to generate an embolic lesion feature fusion map; perform morphological optimization operations on the lesion boundaries corresponding to the embolic lesion feature fusion map to obtain the embolic lesion segmentation results.

[0068] In an embodiment of the present invention, an upsampling operation is performed on the embolic lesion feature segmentation map output by each layer of the convolutional network. Upsampling is to increase the resolution of the image through interpolation method, so that the low-resolution feature map can be aligned with the higher-resolution input image, thereby better locating the lesion area. After upsampling, the feature maps of each layer of the convolutional network are spliced ​​into a high-resolution embolic lesion feature fusion map through element-by-element splicing and fusion operation. At this time, the feature map contains information from each layer of the network, which helps to more accurately locate the shape and position of the lesion. Then, a morphological optimization operation is performed on the embolic lesion feature fusion map. Morphological optimization mainly includes corrosion, expansion, opening operation and closing operation and other technologies to remove noise, fill holes and refine the lesion boundary. Through these morphological operations, the connectivity of the lesion area can be further enhanced, ensuring that the segmentation result of the lesion is more accurate and smooth. After morphological optimization, the final embolic lesion segmentation result is finally generated.

[0069] Furthermore, the medical image feature registration module includes the following functions:

[0070] Obtain embolic CT images containing corresponding anatomical structures;

[0071] Acquire embolic MRI medical images containing soft tissue contrast;

[0072] Performing same-scale space mapping on the embolism CT image and the embolism MRI medical image to generate corresponding embolism CT image and embolism MRI image in the same-scale space;

[0073] The corresponding embolism CT images and embolism MRI images in the same scale space are registered by minimizing feature points to generate embolism CT registered images and embolism MRI registered images.

[0074] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Schematic diagram of the functional flow of the traditional Chinese medicine image feature registration module. In this embodiment, the medical image feature registration module includes the following functions:

[0075] S11: Acquire embolic CT images corresponding to anatomical structures;

[0076] In an embodiment of the present invention, embolism CT images are typically acquired using a CT scanning device. Specifically, a patient is first placed in a CT scanner, and the patient's body is scanned using an X-ray source. During the scanning process, the CT device collects image data of anatomical structures based on different slice thicknesses and scanning angles. An appropriate reconstruction algorithm (such as filtered back projection or iterative reconstruction) can be selected to post-process the image data to generate high-resolution embolism CT images, ensuring that the image quality can clearly reflect the details of the anatomical structure. During image acquisition, the scanning range should be precisely adjusted based on the area where the embolic lesion is located to ensure that the image of the embolic area is presented as completely as possible.

[0077] S12: Acquire an MRI medical image of embolism including soft tissue contrast;

[0078] In an embodiment of the present invention, the acquisition of embolism MRI medical images is completed by MRI scanning equipment, usually in a magnetic resonance imaging device. The patient is arranged to enter the MRI scanner, and the body tissue is excited by a strong magnetic field and radio frequency pulses to obtain magnetic resonance images of the soft tissue. Since MRI has a high soft tissue contrast and can clearly distinguish soft tissue structures from surrounding blood vessels and other tissues, it is particularly suitable for observing embolic lesions. To ensure accurate presentation of the embolic area, it is necessary to adjust the scanning parameters according to the patient's physical signs and the location of the embolism, including the scanning sequence (such as T1-weighted, T2-weighted, enhanced imaging, etc.) and the imaging resolution. In particular, in the lesion area, an enhanced contrast agent (such as a gadolinium-based contrast agent) can be used to further improve the visualization of the lesion. Through this process, the obtained embolism MRI image can clearly show the soft tissue structure surrounding the lesion.

[0079] S13: performing same-scale space mapping on the embolism CT image and the embolism MRI medical image to generate corresponding embolism CT image and embolism MRI image in the same-scale space;

[0080] In embodiments of the present invention, after acquiring embolic CT and embolic MRI images, a unified mapping of the image scale space is typically performed. The key to this process is to scale-normalize the two different modal medical images so that they are at the same physical spatial scale. Specifically, interpolation processing (e.g., nearest neighbor interpolation, bilinear interpolation, cubic interpolation, etc.) is performed on the two images to unify their resolutions and ensure that their spatial dimensions (e.g., voxel size) are the same. An affine transformation algorithm can be employed to adjust the images using scaling techniques based on the original dimensions and coordinate systems of the two images. During the image mapping process, the embolic regions should be aligned in the two images to minimize distortion caused by interpolation. Furthermore, an image registration algorithm (e.g., mutual information method or region-based registration method) is employed to further optimize the spatial alignment of the two images, ensuring their precise alignment in the same scale space. This facilitates subsequent analysis and ultimately generates corresponding embolic CT and embolic MRI images at the same scale space.

[0081] S14: performing feature point minimization registration on the corresponding embolism CT image and embolism MRI image in the same scale space to generate an embolism CT registered image and an embolism MRI registered image.

[0082] In an embodiment of the present invention, an embolization CT image and an embolization MRI image are precisely registered to ensure optimal overlap of the two images in anatomical space. This process utilizes a feature point minimization registration method. First, significant feature points (such as corners, edges, or textures) are extracted from the two images. During feature extraction, algorithms such as image gradients, SIFT (Scale Invariant Feature Transform), or SURF (Speeded Up Robust Features) can be used to identify key feature points in the images. Then, a registration algorithm (such as one based on minimizing mutual information or least squares) is used to calculate the transformation matrix between the images and optimize the alignment between the two images. The minimization process minimizes the error between the feature points in the embolization CT image and the embolization MRI image by adjusting translation, rotation, and scaling parameters, thereby generating a registered CT embolization image and a registered MRI embolization image. This step precisely aligns the two images, providing consistent structural information in the same coordinate system, ensuring more accurate subsequent embolic lesion analysis. Ultimately, a registered CT embolization image and a registered MRI embolization image are generated.

[0083] Furthermore, performing feature point minimization registration on the corresponding embolism CT image and embolism MRI image in the same scale space includes:

[0084] Perform feature point matching and screening on the corresponding embolism CT images and embolism MRI images in the same scale space to obtain a set of embolism multimodal feature point pairs;

[0085] In an embodiment of the present invention, feature point matching and screening are performed on embolism CT images and embolism MRI images in the same scale space, so as to extract key feature points from the CT images and MRI images respectively by applying image feature extraction algorithms, such as SIFT (Scale Invariant Feature Transform) or SURF (Speeded Up Robust Features). To ensure the accuracy of feature point matching, a descriptor-based matching method is adopted, and each feature point in the CT image and the MRI image is compared through its context descriptor. The similarity between the feature points of the two images is calculated, and feature point pairs with high similarity are screened out by setting a threshold, and those feature point pairs with low matching quality are eliminated to obtain a set of embolism multimodal feature point pairs. In this process, the nearest neighbor search (such as KD tree) can be used to optimize the matching efficiency of the feature points to ensure high efficiency and matching accuracy.

[0086] Preferably, a corresponding feature point context descriptor is obtained for each pair of feature points corresponding to the CT image and the MRI image in the embolism multimodal feature point pair set, and a feature point similarity calculation is performed on each pair of feature points in the embolism multimodal feature point pair set based on the feature point context descriptor to obtain the embolism feature similarity corresponding to each pair of feature points;

[0087] In an embodiment of the present invention, by extracting the corresponding feature point context descriptors from the embolism CT image and the embolism MRI image based on each pair of feature points in the embolism multimodal feature point pair set, the context descriptors can be generated by describing the structural information of the feature point neighborhood, such as the gradient histogram based on the local image region or the texture features around the feature points. For each pair of feature points, the similarity between their context descriptors is calculated, usually using measurement methods such as Euclidean distance, cosine similarity or Hamming distance. The embolism feature similarity of each pair of feature points can be calculated based on the similarity between the feature point context descriptors. If the similarity is high, it means that the expression of this pair of feature points in different modal images is relatively consistent, representing that they correspond to the same anatomical structure or lesion area, and finally the embolism feature similarity corresponding to each pair of feature points is obtained.

[0088] Preferably, based on the embolism feature similarity corresponding to each pair of feature points, each pair of feature points in the embolism multimodal feature point pair set is screened for similarity and representativeness, so as to obtain a set of representative embolism feature point pairs;

[0089] In an embodiment of the present invention, the set of embolic multimodal feature point pairs is screened for similarity and representativeness based on the embolic feature similarity of each pair of feature points. The purpose of the screening is to select those feature point pairs that are representative and stable in the embolic lesion area from the matched feature point pairs. This process is based on a preset threshold and sorts the feature points according to the embolic feature similarity of each pair of feature points, and screens out feature point pairs with higher similarity as representative feature points. In order to further improve the quality of the representative feature points, an iterative optimization method can also be used to eliminate noise in the matching and retain those key point pairs that can reflect the characteristics of the embolic lesion. The screened feature point pairs constitute the set of representative embolic feature point pairs, which becomes an important basis for the subsequent registration process.

[0090] Preferably, the Euclidean distance between each pair of representative feature points in the set of representative feature points of embolism is calculated to obtain the Euclidean distance value between each pair of representative feature points of embolism;

[0091] In the embodiment of the present invention, the Euclidean distance value of each pair of feature points in the set of representative feature points of embolism is calculated by calculating the Euclidean distance between each pair of feature points. The Euclidean distance is used to quantify the degree of spatial difference between two feature points and provide a quantitative basis for subsequent image registration. In the specific operation, the Euclidean distance of each pair of feature points is first calculated based on the pixel coordinates of the feature points in the image. The formula is: Among them, x1, y1 and x2, y2 are the coordinates of the two feature points respectively. By calculating these Euclidean distances, the spatial consistency between the feature points can be evaluated, and finally the Euclidean distance value between each pair of representative feature points of embolism can be obtained.

[0092] Preferably, the corresponding embolism CT images and embolism MRI images in the same scale space are minimized and registered based on minimization of the Euclidean distance value between each pair of representative feature points of embolism as the optimization target to generate embolism CT registered images and embolism MRI registered images.

[0093] In an embodiment of the present invention, the registration of the embolism CT image and the embolism MRI image is performed by minimizing the objective function based on the Euclidean distance value between the representative feature points of the embolism. The objective function achieves registration optimization by minimizing the Euclidean distance between each pair of representative feature points. An optimization algorithm such as gradient descent or Gauss-Newton method is often used for minimization calculation. The optimization goal is to minimize the difference between the feature points in the spatial coordinate system, thereby achieving accurate alignment of the CT image and the MRI image in the same coordinate system. The minimization process continues until a predetermined convergence condition is reached to ensure the best registration effect of the CT and MRI images. The finally generated embolism CT registered image and embolism MRI registered image will have spatial consistency.

[0094] Furthermore, the matching and screening of feature points of the corresponding embolism CT images and embolism MRI images in the same scale space includes:

[0095] Multimodal feature deep mining is performed on the corresponding embolic CT images and embolic MRI images in the same scale space. Edge detection is used to extract the edge contour features of the embolic area tissue in the CT image, including shape, size, and density. The texture features of the CT image are calculated using the gray-level co-occurrence matrix, including contrast, correlation, energy, and entropy. The phase difference, proton density, and relaxation time features corresponding to different embolic tissues in the MRI image are analyzed to obtain the embolic multimodal image feature set.

[0096] In an embodiment of the present invention, multimodal feature deep mining is performed on embolism CT images and embolism MRI images. For CT images, edge contour features of the embolism area in the image are extracted by edge detection technology. Commonly used edge detection methods include Canny edge detection and Sobel operator, etc. These methods can effectively extract the shape, size and boundary information of the embolism area. After obtaining the edge image, the density features of the embolism area can be further calculated. In order to extract the texture features, the gray level co-occurrence matrix (GLCM) is used to analyze the texture features of the image. The specific operation is to first calculate the co-occurrence matrix of the image gray value, and then further calculate the contrast of the embolism area based on this matrix. Texture features such as intensity, correlation, energy, and entropy help describe the tissue structure and distribution of the embolic area in CT images. For MRI images, the appearance of embolic tissue in the image is usually different from that of other tissues. Therefore, by analyzing the phase difference, proton density, and relaxation time of embolic tissue in MRI images, the embolic area can be effectively distinguished from the surrounding tissue. The phase difference is mainly caused by the difference in magnetic resonance characteristics of different tissues and can be extracted through phase imaging technology. The proton density reflects the difference in proton content in the tissue, and the relaxation time mainly reflects the reaction speed of the tissue to the magnetic field. Combining this information, a multimodal image feature set of embolism is finally constructed.

[0097] Preferably, feature correlation analysis is performed between the CT image features and the MRI image features at each position point in the same scale space within the embolism multimodal image feature set to obtain a feature correlation correspondence between the CT and MRI image features at each position point;

[0098] In an embodiment of the present invention, a correlation analysis is performed between the CT image features and the MRI image features at each position point in the same scale space within the embolism multimodal image feature set. The key to this step is to achieve the correspondence between the CT image and MRI image features, which can be achieved by calculating the feature similarity of the two images at the same position point. Mutual information (MI) is usually used to measure the correlation between images of different modalities. Mutual information is an indicator of the amount of information shared by two images. For each position point, the feature vectors of the CT image and the MRI image are calculated, and their mutual information values ​​are evaluated to obtain the correlation strength between the features. Through this analysis, the correspondence between the embolism area in the two images can be determined, and finally the feature correlation correspondence between the CT and MRI image features at each position point is obtained.

[0099] Preferably, based on the feature association correspondence between the CT and MRI image features at each position point, feature point matching screening is performed between the CT image features and the MRI image features at each position point in the same scale space in the embolism multimodal image feature set to obtain a set of embolism multimodal feature point pairs.

[0100] In an embodiment of the present invention, feature point matching and screening are performed on the CT image features and MRI image features at each position point in the same scale space in the embolism multimodal image feature set based on the previously obtained feature association correspondence. The specific operation is to pair and match the CT and MRI image features of each position point on the basis of feature association analysis. The feature vectors can be matched by metrics such as Euclidean distance or cosine similarity to ensure that the similarity between the matching pairs meets the predetermined standard. At the same time, in order to improve the matching accuracy, feature point descriptors (such as SIFT or SURF algorithms) can be used to further extract key feature points in the image and perform matching screening. The screening process can exclude matching pairs that do not meet the requirements through methods such as threshold setting or iterative optimization. Finally, all qualified CT and MRI image feature point pairs are collected to form an embolism multimodal feature point pair set, which is used as input data for embolic lesion segmentation, providing high-quality feature point support for the next step of lesion identification and segmentation.

[0101] Furthermore, the image band fusion enhancement module includes the following functions:

[0102] Performing frequency domain conversion on the embolism CT registered image and the embolism MRI registered image to generate an embolism CT spectrum map and an embolism MRI spectrum map;

[0103] Perform wavelet frequency band decomposition on the embolism CT spectrum and the embolism MRI spectrum to obtain different characteristic frequency bands corresponding to the embolism CT and MRI spectrum;

[0104] Perform phase-consistent weighted fusion on the different characteristic frequency bands corresponding to the CT and MRI spectrograms of the embolism to calculate the phase similarity of the CT and MRI spectrograms in each frequency band, and perform weighted fusion of the frequency bands based on the phase similarity in each frequency band to generate an initial fused spectrogram of the embolism;

[0105] The fusion quality control of the initial fusion spectrum of embolism is optimized to obtain the peak signal-to-noise ratio (PSNR) of different monitoring fusion images during the fusion process, and the frequency band weights are reallocated according to the PSNR to perform fusion control optimization to generate the fusion spectrum of embolism bands. The spatial domain inverse operation is performed on the fusion spectrum of embolism bands to generate the fusion image of embolism.

[0106] The embolization fusion image was processed with histogram equalization to generate the embolization contrast standard image.

[0107] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 Schematic diagram of the functional flow of the image band fusion enhancement module in this embodiment. The image band fusion enhancement module includes the following functions:

[0108] S21: performing image frequency domain conversion on the embolism CT registered image and the embolism MRI registered image to generate an embolism CT spectrum map and an embolism MRI spectrum map;

[0109] In an embodiment of the present invention, registered image data of embolism CT and embolism MRI are obtained. For the embolism CT image, the image is first converted into a grayscale image to facilitate frequency domain analysis. Then, the embolism CT image is subjected to frequency domain conversion using a fast Fourier transform (FFT). The FFT converts the image from the spatial domain to the frequency domain to obtain an embolism CT spectrum diagram. Each point in the spectrum diagram represents different frequency information of the image. A similar frequency domain conversion is also performed on the embolism MRI image. The same FFT method is used to perform frequency domain conversion on the MRI image to obtain an embolism MRI spectrum diagram. This process can effectively extract frequency features in the image, providing a basis for subsequent frequency band decomposition and fusion operations.

[0110] S22: performing wavelet frequency band decomposition on the embolism CT spectrum and the embolism MRI spectrum to obtain different characteristic frequency bands corresponding to the embolism CT and MRI spectrum;

[0111] In an embodiment of the present invention, wavelet transform is performed on the embolism CT spectrum map and the embolism MRI spectrum map. In the specific operation, a wavelet transform method such as discrete wavelet transform (DWT) is used to decompose the spectrum map. This method can decompose the spectrum map into multiple frequency bands, each frequency band representing a different frequency range. In the embolism CT image, wavelet transform can extract frequency bands related to the embolism density and structural characteristics, such as low-frequency bands are used to capture the global structural information of the image, and high-frequency bands can capture detailed features and changes. In the embolism MRI image, wavelet transform helps to extract frequency band information related to the embolism morphology, size and position. Different frequency bands correspond to different features, and can effectively separate useful visual features from CT and MRI images, and finally obtain different characteristic frequency bands corresponding to the embolism CT and MRI spectrum maps.

[0112] S23: performing phase-consistent weighted fusion on different characteristic frequency bands corresponding to the CT and MRI spectrograms of the embolism to calculate the phase similarity of the CT and MRI spectrograms in each frequency band, and performing weighted fusion of the frequency bands according to the phase similarity in each frequency band to generate an initial fused spectrogram of the embolism;

[0113] In an embodiment of the present invention, phase congruence weighted fusion is performed on different frequency bands of embolism CT and embolism MRI images. To perform this operation, the phase information of the CT and MRI images in each frequency band must be calculated. The phase information of each frequency band is extracted through Fourier transform, and phase similarity is calculated. Phase congruence reflects the degree of structural alignment of the CT and MRI images in that frequency band. Higher phase congruence indicates more consistent structures of the CT and MRI images in that frequency band. Based on the calculated phase congruence results, each frequency band is weightedly fused. When assigning weights, frequency bands with higher phase congruence are given greater weights, so that the fused image can fully retain the characteristic information of the CT and MRI images. After weighted fusion, an initial fused spectrum of embolism is ultimately generated.

[0114] S24: Optimizing fusion quality control of the initial fusion spectrum of embolism, optimizing the peak signal-to-noise ratios corresponding to different monitoring fusion images during the fusion process, and reallocating frequency band weights according to the peak signal-to-noise ratios to optimize fusion control, thereby generating an embolism frequency band fusion spectrum; performing spatial domain inverse operation on the embolism frequency band fusion spectrum to generate an embolism fusion image;

[0115] In an embodiment of the present invention, fusion quality control optimization is performed on the initial fusion spectrum of embolism. This process evaluates the fusion quality by calculating the peak signal-to-noise ratio (PSNR) of the initial fusion spectrum of embolism. The peak signal-to-noise ratio is an important indicator for measuring image quality. A higher PSNR indicates a higher image quality. By analyzing the PSNR of each frequency band, frequency bands with poor quality are identified during the fusion process. To further optimize the fusion effect, frequency band weights are reallocated according to the PSNR values ​​of each frequency band, and frequency bands with high PSNRs are given greater weights, thereby optimizing the quality of the fusion spectrum of the embolism frequency bands. After completing the frequency band weighting, the fusion spectrum of the embolism frequency bands is converted from the frequency domain back to the spatial domain through an inverse Fourier transform, ultimately generating an embolism fusion image. This image integrates the advantageous information of CT and MRI images and optimizes image quality.

[0116] S25: performing histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image.

[0117] In an embodiment of the present invention, histogram equalization is performed on the embolization fusion image. Histogram equalization is a commonly used image enhancement technology used to adjust the contrast of an image to make the brightness distribution of the image more uniform. In this step, the grayscale histogram of the embolization fusion image is redistributed, and the detail contrast in the image is improved by uniformizing the grayscale distribution. After the histogram equalization process, an embolization contrast standard image is generated. This image has a higher visual contrast, making the embolic lesion more obvious in the image, and finally, the embolization contrast standard image is generated.

[0118] Furthermore, the different characteristic frequency bands corresponding to the embolism CT and MRI spectra include different frequency bands corresponding to differences in embolism density and structural characteristics in the embolism CT spectra and different frequency bands corresponding to differences in embolism morphology, size and position in the embolism MRI spectra.

[0119] Furthermore, the deep network lesion segmentation module includes the following functions:

[0120] Build a deep network lesion segmentation model that includes a self-attention module and a channel attention module;

[0121] In an embodiment of the present invention, a deep network lesion segmentation model including a self-attention module and a channel attention module is constructed. The model consists of a five-layer convolutional neural network (CNN). Each convolution layer uses a 5x5 convolution kernel, and after each convolution operation, a 1x1 pooling layer is followed. In the specific implementation, the convolution layer adopts a standard convolution operation, extracts local features through a 5x5 convolution kernel, and compresses the spatial dimension through a 1x1 pooling layer. The role of the pooling layer is to retain important spatial information, reduce the amount of calculation, and reduce the risk of overfitting. A self-attention module is added between each layer of convolution and pooling. The role of this module is to enhance the expression ability of the lesion area by calculating the relationship between each pixel in the image, especially when processing embolic lesion areas. It can capture the long-distance dependencies between lesions. The self-attention mechanism assigns an attention weight to each pixel, so that the relevant pixels in the lesion area receive more prominent attention. This mechanism is implemented by calculating the similarity matrix of the input feature map, generating attention weights, and then performing weighted operations on the input image. At the same time, after each layer of convolution, it is also necessary to connect the channel attention module, which is responsible for performing channel-level weighted processing on the feature map output by each layer of convolution to learn the relative importance of each channel feature. This module generates a channel attention map and weights different feature channels, emphasizing the channel features that are helpful for lesion segmentation. The overall network structure can effectively improve the accurate segmentation of embolic lesion areas through this hierarchical feature weighting mechanism.

[0122] Preferably, network lesion segmentation training is performed on the corresponding embolic lesion area in the embolic contrast standard image based on the deep network lesion segmentation model to output the embolic lesion segmentation result corresponding to each layer of the convolutional network;

[0123] In an embodiment of the present invention, the embolism contrast standard image is trained by utilizing the constructed deep network lesion segmentation model. The embolism contrast standard image used in the training process is the image data with the lesion area labeled. The goal is to train with the standard image so that the model can learn how to identify and segment the lesion area from the new embolism image. During training, the standard image is first input into the constructed deep neural network. After feature extraction of five layers of convolution and pooling layers, the model will gradually generate embolic lesion segmentation results corresponding to each layer of convolution network. The segmentation results of these output layers are achieved through the collaborative work of the convolutional neural network, the self-attention module and the channel attention module. The output of each convolution layer will be used as the feature map of the next layer network input to obtain the lesion segmentation map of the embolic area. These segmentation results need to be compared with the lesion area labeled in the standard image and used for subsequent loss calculation and model optimization, and finally the embolic lesion segmentation results corresponding to each layer of convolution network are output.

[0124] Preferably, the cross entropy loss and Dice coefficient loss corresponding to the deep network lesion segmentation model are calculated according to the embolic lesion segmentation results corresponding to each layer of the convolutional network, so as to use the cross entropy loss to measure the classification error between the model prediction result and the actual lesion area label, and the Dice coefficient loss to evaluate the degree of overlap corresponding to the lesion segmentation results. The cross entropy loss and the Dice coefficient loss are used as the optimization targets corresponding to the deep network lesion segmentation model, and the stochastic gradient descent algorithm is used to dynamically adjust the hyperparameters corresponding to the deep network lesion segmentation model to generate a deep lesion segmentation optimization model;

[0125] In an embodiment of the present invention, the embolic lesion segmentation results corresponding to each layer of the convolutional network are evaluated, and the accuracy of the model segmentation is quantified by calculating the cross entropy loss and the Dice coefficient loss. The cross entropy loss function is used to measure the classification error between the model prediction result and the true label. The formula is the logarithmic difference between the probability distribution of the model output category and the label. By calculating the cross entropy loss of all pixels, an overall error value can be obtained to guide network optimization. The Dice coefficient loss is used to evaluate the degree of overlap of the lesion segmentation results. The Dice coefficient is calculated by calculating the intersection of the predicted segmentation area and the true segmentation area. The ratio of the union can better reflect the accuracy of lesion segmentation. The higher the value of the loss function, the higher the overlap between the segmentation result and the actual lesion area, and the better the performance of the model. The cross entropy loss and the Dice coefficient loss are used together as optimization targets, and the stochastic gradient descent algorithm (SGD) is used to dynamically adjust the hyperparameters of the model. Specifically, the SGD algorithm calculates the gradient of the loss function, updates the model parameters, and finally obtains a trained deep lesion segmentation optimization model. During the optimization process, hyperparameters such as the learning rate and regularization coefficient will be adjusted according to the feedback of the loss function to ensure that the model gradually converges to the optimal state.

[0126] Preferably, the embolism contrast standard image is re-input into the deep lesion segmentation optimization model for lesion optimization segmentation processing to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network.

[0127] In an embodiment of the present invention, the optimized deep lesion segmentation model is reapplied to the segmentation task of embolism contrast standard images, and the optimized model is input into the standard image. The model will output embolic lesion feature segmentation maps for each layer of convolutional network based on the weights and parameters obtained in the previous training process. These segmentation maps contain the lesion feature information extracted by each layer of convolutional network, and after weighting by the self-attention and channel attention mechanisms, they can accurately reflect the characteristics of the embolic lesion area. The final embolic lesion feature segmentation map will be used as the segmentation result of the lesion area and provided to subsequent medical analysis for further processing. In the process of generating these feature segmentation maps, the deep network can maintain high segmentation accuracy and stability when processing complex embolic images through efficient feature learning and optimization processes, and finally generate the corresponding embolic lesion feature segmentation maps in each layer of convolutional network.

[0128] Furthermore, the deep network lesion segmentation model is specifically composed of 5 layers of 5x5 convolutional layers and a 1x1 pooling layer connected to each convolutional layer, so that a self-attention module is connected between the corresponding convolutional layer and pooling layer of each layer to focus on the embolic lesion area in the fused image to capture the long-distance dependency between lesions, and a channel attention module is connected between the corresponding channels of each layer to automatically learn the importance between channel features of different layers.

[0129] Furthermore, the lesion segmentation map fusion optimization module includes the following functions:

[0130] Perform convolution and upsampling operations on the embolic lesion feature segmentation map corresponding to each convolutional network layer, so as to adjust the embolic lesion feature segmentation maps corresponding to different scales in each convolutional network layer to the same size, thereby obtaining the corresponding embolic lesion feature maps at the same size;

[0131] In an embodiment of the present invention, a convolution operation is performed on the embolic lesion feature segmentation map by using a convolutional neural network (CNN). Each layer of the convolutional network extracts feature information of the corresponding lesion area through a specific convolution kernel. Each layer of the convolutional network will produce embolic lesion feature maps of different sizes at different scales. Therefore, in order to ensure the uniformity of multi-layer features, it is necessary to perform an upsampling operation on the embolic lesion feature maps of different sizes. The upsampling method can be implemented through deconvolution or interpolation operations (such as bilinear interpolation) so that the embolic lesion feature maps output by each layer are the same in size. In this way, the sizes of the embolic lesion feature maps extracted at different scales are adjusted to be consistent, thereby preparing for the fusion operation in subsequent steps, ensuring that the feature maps of each layer can be effectively processed at the same spatial size, and finally obtaining the corresponding embolic lesion feature maps at the same size.

[0132] Preferably, the corresponding embolic lesion feature maps of the same size are spliced ​​and fused element by element, so as to splice and fuse the lesion detail feature information and the corresponding embolic lesion semantic information to generate an embolic lesion feature fusion map;

[0133] In an embodiment of the present invention, an element-by-element splicing operation is performed on the embolic lesion feature map of the same size obtained by each layer of the convolutional network. This process utilizes the semantic association and detail information between the feature maps. By splicing the feature maps of different levels element-by-element, the multi-scale embolic lesion feature information extracted by different convolutional layers can be fused. In specific implementation, the embolic lesion feature map obtained by the upsampling operation of each layer is spliced ​​in the channel dimension to obtain a feature map containing rich detail information. This splicing and fusion method can ensure that the details of embolic lesions of different scales are retained, while enhancing the semantic information of the lesions, thereby improving the segmentation accuracy and robustness. The spliced ​​embolic lesion feature map fuses the advantageous features of each layer, and finally generates an embolic lesion feature fusion map.

[0134] Preferably, a morphological optimization operation is performed on the lesion boundary corresponding to the fusion map of embolic lesion features, so as to remove the corresponding noise area by using dilation and corrosion operations, and smooth the corresponding lesion boundary to obtain the embolic lesion segmentation result.

[0135] In an embodiment of the present invention, the segmentation effect is further improved through morphological optimization operations. Specifically, the boundary of the embolic lesion is expanded through the dilation operation, making the boundary of the lesion area more obvious, thereby reducing the missed detection phenomenon in the segmentation. The dilation operation uses structural elements to expand the lesion area, which can effectively enhance the continuity of the lesion area. Subsequently, the corrosion operation is used to shrink the expanded area to remove the noise area and irrelevant background parts, making the segmentation result more accurate. Through these two morphological operations, the boundary of the embolic lesion can be smoothed, and small errors can be removed to ensure that the final embolic lesion segmentation result is both complete and accurate. The morphological optimization operation helps to improve the quality of the segmentation result, making the lesion area more consistent with the performance of the actual medical image, and finally obtaining the embolic lesion segmentation result.

[0136] Furthermore, the present invention also provides a method for segmenting embolic lesions based on medical image processing, which is used to execute the above-mentioned system for segmenting embolic lesions based on medical image processing. The method for segmenting embolic lesions based on medical image processing includes:

[0137] Step S1: acquiring an embolism CT image containing anatomical structures and an embolism MRI medical image containing soft tissue contrast, and performing feature point minimization registration on the embolism CT image and the embolism MRI medical image to generate an embolism CT registered image and an embolism MRI registered image;

[0138] Step S2: performing image band fusion on the embolism CT registered image and the embolism MRI registered image to generate an embolism fusion image; performing histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image;

[0139] Step S3: Construct a deep network lesion segmentation model corresponding to the self-attention module and the channel attention module, and perform network lesion segmentation processing on the corresponding embolic lesion area in the embolic contrast standard image based on the deep network lesion segmentation model. Combine the cross entropy loss and Dice coefficient loss as the corresponding optimization targets of the deep network lesion segmentation model to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network;

[0140] Step S4: Upsample and element-by-element concatenation of the embolic lesion feature segmentation map corresponding to each layer of the convolutional network to generate an embolic lesion feature fusion map; perform morphological optimization operations on the lesion boundaries corresponding to the embolic lesion feature fusion map to obtain the embolic lesion segmentation result.

[0141] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0142] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. An embolic lesion segmentation system based on medical image processing, characterized in that: Includes the following modules: a medical image feature registration module, configured to obtain an embolism CT image containing anatomical structure correspondences and an embolism MRI medical image containing soft tissue contrast correspondences, and perform feature point minimization registration on the embolism CT image and the embolism MRI medical image to generate an embolism CT registered image and an embolism MRI registered image; The image band fusion enhancement module is used to perform image band fusion on the embolism CT registered image and the embolism MRI registered image to generate an embolism fusion image; and perform histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image; The deep network lesion segmentation module is used to build a deep network lesion segmentation model corresponding to the self-attention module and the channel attention module. Based on the deep network lesion segmentation model, the embolic lesion area corresponding to the embolic contrast standard image is segmented. The cross entropy loss and Dice coefficient loss are used as the optimization targets of the deep network lesion segmentation model to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network. The lesion segmentation map fusion optimization module is used to upsample and element-by-element concatenate the embolic lesion feature segmentation maps corresponding to each layer of the convolutional network to generate an embolic lesion feature fusion map; morphological optimization operations are performed on the lesion boundaries corresponding to the embolic lesion feature fusion map to obtain the embolic lesion segmentation results.

2. The embolic lesion segmentation system based on medical image processing according to claim 1, characterized in that: The medical image feature registration module includes the following functions: Obtain embolic CT images containing corresponding anatomical structures; Acquire embolic MRI medical images containing soft tissue contrast; Performing same-scale space mapping on the embolism CT image and the embolism MRI medical image to generate corresponding embolism CT image and embolism MRI image in the same-scale space; The corresponding embolism CT images and embolism MRI images in the same scale space are registered by minimizing feature points to generate embolism CT registered images and embolism MRI registered images.

3. The embolic lesion segmentation system based on medical image processing according to claim 2, characterized in that: The performing feature point minimization registration on the corresponding embolism CT image and embolism MRI image in the same scale space includes: Perform feature point matching and screening on the corresponding embolism CT images and embolism MRI images in the same scale space to obtain a set of embolism multimodal feature point pairs; The corresponding feature point context descriptor is obtained for each pair of feature points corresponding to the CT image and the MRI image in the embolism multimodal feature point pair set, and the feature point similarity calculation is performed on each pair of feature points in the embolism multimodal feature point pair set based on the feature point context descriptor to obtain the embolism feature similarity corresponding to each pair of feature points; Based on the embolism feature similarity corresponding to each pair of feature points, each pair of feature points in the embolism multimodal feature point pair set is screened for similarity and representativeness, so as to obtain a set of representative embolism feature point pairs; Calculating the Euclidean distance between each pair of representative feature points in the set of representative feature points of embolism to obtain the Euclidean distance value between each pair of representative feature points of embolism; Based on the minimization of the Euclidean distance between each pair of representative feature points of embolism, the corresponding embolism CT images and embolism MRI images in the same scale space are minimized and registered to generate embolism CT registered images and embolism MRI registered images.

4. The embolic lesion segmentation system based on medical image processing according to claim 3, characterized in that: The matching and screening of feature points of the corresponding embolism CT images and embolism MRI images in the same scale space includes: Multimodal feature deep mining is performed on the corresponding embolic CT images and embolic MRI images in the same scale space. Edge detection is used to extract the edge contour features of the embolic area tissue in the CT image, including shape, size, and density. The texture features of the CT image are calculated using the gray-level co-occurrence matrix, including contrast, correlation, energy, and entropy. The phase difference, proton density, and relaxation time features corresponding to different embolic tissues in the MRI image are analyzed to obtain the embolic multimodal image feature set. Perform feature correlation analysis between the CT image features and MRI image features at each position point in the same scale space within the embolism multimodal image feature set to obtain the feature correlation correspondence between the CT and MRI image features at each position point; Based on the feature correlation correspondence between the CT and MRI image features at each position point, feature point matching and screening are performed between the CT image features and MRI image features at each position point in the same scale space in the embolism multimodal image feature set to obtain the embolism multimodal feature point pair set.

5. The embolic lesion segmentation system based on medical image processing according to claim 1, characterized in that: The image band fusion enhancement module includes the following functions: Performing frequency domain conversion on the embolism CT registered image and the embolism MRI registered image to generate an embolism CT spectrum map and an embolism MRI spectrum map; Perform wavelet frequency band decomposition on the embolism CT spectrum and the embolism MRI spectrum to obtain different characteristic frequency bands corresponding to the embolism CT and MRI spectrum; Perform phase-consistent weighted fusion on the different characteristic frequency bands corresponding to the CT and MRI spectrograms of the embolism to calculate the phase similarity of the CT and MRI spectrograms in each frequency band, and perform weighted fusion of the frequency bands based on the phase similarity in each frequency band to generate an initial fused spectrogram of the embolism; The fusion quality control of the initial fusion spectrum of embolism is optimized to obtain the peak signal-to-noise ratio of different monitoring fusion images during the fusion process, and the frequency band weights are reallocated according to the peak signal-to-noise ratio to perform fusion control optimization to generate the fusion spectrum of embolism frequency bands. Performing spatial domain inverse operation on the embolism band fusion spectrum to generate an embolism fusion image; The embolization fusion image was processed with histogram equalization to generate the embolization contrast standard image.

6. The method for segmenting embolic lesions based on medical image processing according to claim 5, characterized in that: The different characteristic frequency bands corresponding to the embolism CT and MRI spectrograms include different frequency bands corresponding to differences in embolism density and structural characteristics in the embolism CT spectrogram and different frequency bands corresponding to differences in embolism morphology, size and position in the embolism MRI spectrogram.

7. The embolic lesion segmentation system based on medical image processing according to claim 1, characterized in that: The deep network lesion segmentation module includes the following functions: Build a deep network lesion segmentation model that includes a self-attention module and a channel attention module; Based on the deep network lesion segmentation model, network lesion segmentation training is performed on the corresponding embolic lesion area in the embolic contrast standard image to output the embolic lesion segmentation results corresponding to each layer of the convolutional network; The cross entropy loss and Dice coefficient loss corresponding to the deep network lesion segmentation model are calculated based on the embolic lesion segmentation results corresponding to each layer of the convolutional network. The cross entropy loss is used to measure the classification error between the model prediction results and the actual lesion area label. The Dice coefficient loss is used to evaluate the degree of overlap corresponding to the lesion segmentation results. The cross entropy loss and Dice coefficient loss are used as the optimization targets corresponding to the deep network lesion segmentation model. The stochastic gradient descent algorithm is used to dynamically adjust the hyperparameters corresponding to the deep network lesion segmentation model to generate a deep lesion segmentation optimization model. The embolism contrast standard image is re-input into the deep lesion segmentation optimization model for lesion optimization segmentation processing to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network.

8. The embolic lesion segmentation system based on medical image processing according to claim 7, characterized in that: The deep network lesion segmentation model is specifically composed of 5 layers of 5x5 convolutional layers and a 1x1 pooling layer connected to each convolutional layer. A self-attention module is connected between the corresponding convolutional layer and pooling layer in each layer to focus on the embolic lesion area in the fused image to capture the long-distance dependency between lesions, and a channel attention module is connected between the corresponding channels in each layer to automatically learn the importance between channel features in different layers.

9. The embolic lesion segmentation system based on medical image processing according to claim 1, characterized in that: The lesion segmentation map fusion optimization module includes the following functions: Perform convolution and upsampling operations on the embolic lesion feature segmentation map corresponding to each convolutional network layer, so as to adjust the embolic lesion feature segmentation maps corresponding to different scales in each convolutional network layer to the same size, thereby obtaining the corresponding embolic lesion feature maps at the same size; The corresponding embolic lesion feature maps of the same size are spliced ​​and fused element by element to generate an embolic lesion feature fusion map containing lesion detail feature information and corresponding embolic lesion semantic information; A morphological optimization operation is performed on the lesion boundary corresponding to the fusion map of embolic lesion features to remove the corresponding noise area using dilation and erosion operations, and the corresponding lesion boundary is smoothed to obtain the embolic lesion segmentation result.

10. A method for segmenting embolic lesions based on medical image processing, characterized in that: The system for segmenting embolic lesions based on medical image processing according to claim 1, wherein the method for segmenting embolic lesions based on medical image processing comprises: Step S1: acquiring an embolism CT image containing anatomical structures and an embolism MRI medical image containing soft tissue contrast, and performing feature point minimization registration on the embolism CT image and the embolism MRI medical image to generate an embolism CT registered image and an embolism MRI registered image; Step S2: performing image band fusion on the embolism CT registered image and the embolism MRI registered image to generate an embolism fusion image; performing histogram equalization processing on the embolism fusion image to generate an embolism contrast standard image; Step S3: Construct a deep network lesion segmentation model corresponding to the self-attention module and the channel attention module, and perform network lesion segmentation processing on the corresponding embolic lesion area in the embolic contrast standard image based on the deep network lesion segmentation model. Combine the cross entropy loss and Dice coefficient loss as the corresponding optimization targets of the deep network lesion segmentation model to generate the corresponding embolic lesion feature segmentation map in each layer of the convolutional network; Step S4: Upsample and element-by-element concatenation of the embolic lesion feature segmentation map corresponding to each layer of the convolutional network to generate an embolic lesion feature fusion map; perform morphological optimization operations on the lesion boundaries corresponding to the embolic lesion feature fusion map to obtain the embolic lesion segmentation result.

Images