Image annotation method and system applied to brain MRI (Magnetic Resonance Imaging) image segmentation

Abstract

The embodiment of the invention discloses an image annotation method and system applied to brain MRI image segmentation, and the method comprises the steps: obtaining a brain MRI image data set of a target object, and the brain MRI image data set comprises original image sequences of a plurality of scanning levels; performing multi-modal feature fusion processing on the original image sequence to generate an enhanced image feature set; calling a multi-layer cascade segmentation network to perform hierarchical feature extraction on the enhanced image feature set to obtain a multi-scale anatomical structure feature map; and performing region boundary optimization processing based on the multi-scale anatomical structure feature map, and generating a marked brain structure segmentation image. Therefore, the boundary of each structure of the brain can be accurately defined, the segmented image is more accurate and clearer, and the image segmentation and marking of the brain MRI image can be accurately and clearer realized.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more particularly to an image annotation method and system for brain MRI image segmentation. Background Art

[0002] Brain MRI image segmentation is crucial in the research and processing of medical images. By segmenting brain MRI images, different brain structures can be clearly distinguished. With the continuous advancement of medical imaging technology, the requirements for the accuracy and precision of brain MRI image segmentation are becoming increasingly stringent. Therefore, achieving more accurate and comprehensive brain MRI image segmentation has become a research hotspot in this field.

[0003] At present, some traditional brain MRI image segmentation technologies have difficulty in fully integrating the image features of multiple modalities when processing complex brain structures, resulting in incomplete image information and an inability to deeply explore the characteristics of brain structures. Another part of the technology lacks effective capture of features at different scales and levels during the feature extraction process, resulting in an insufficiently detailed presentation of brain anatomical structures and an inability to fully reflect the true situation of brain structures. Moreover, during image segmentation and annotation, the optimization capability of regional boundaries is insufficient, and the image boundaries after segmentation and annotation are not accurate and clear enough. Therefore, how to accurately and clearly implement image segmentation and annotation of brain MRI images is a technical problem that needs to be overcome. Summary of the Invention

[0004] The embodiments of the present invention provide an image annotation method and system for brain MRI image segmentation, which are used to accurately and clearly implement image segmentation and annotation of brain MRI images.

[0005] In a first aspect, an embodiment of the present invention provides an image annotation method for brain MRI image segmentation, which is applied to an image annotation system. The method includes: obtaining a brain MRI image data set of a target object, wherein the brain MRI image data set includes original image sequences of multiple scanning levels; performing multimodal feature fusion processing on the original image sequence to generate an enhanced image feature set; calling a multi-layer cascade segmentation network to perform hierarchical feature extraction on the enhanced image feature set to obtain a multi-scale anatomical structure feature map; performing regional boundary optimization processing based on the multi-scale anatomical structure feature map to generate a labeled brain structure segmentation image.

[0006] In a second aspect, an embodiment of the present invention provides an image annotation system, comprising: processor; a storage device having a computer program stored thereon, When the computer program is executed by the processor, the processor implements any of the image annotation methods for brain MRI image segmentation.

[0007] An embodiment of the present invention provides a readable storage medium storing a program or instruction. When the program or instruction is executed by a processor, the steps of the image annotation method applied to brain MRI image segmentation are implemented.

[0008] It can be seen that the embodiment of the present invention first obtains a brain MRI image data set containing multiple scanning level original image sequences, and then performs multimodal feature fusion processing to effectively integrate the features of multiple modalities, thereby generating an enhanced image feature set, which can make the information contained in the image more comprehensive and rich in depth, while mining features that are not easily perceived. Furthermore, a multi-layer cascade segmentation network is used to perform hierarchical feature extraction on the enhanced image feature set, which can capture brain anatomical structure features from different scales and levels, and obtain a multi-scale anatomical structure feature map; finally, based on the multi-scale anatomical structure feature map, regional boundary optimization processing is performed to generate a labeled brain structure segmentation image, which can accurately define the boundaries of each brain structure, making the segmented image more accurate and clear. In this way, image segmentation and annotation of brain MRI images can be accurately and clearly achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flowchart of an image annotation method for brain MRI image segmentation provided by an embodiment of the present invention.

[0010] Figure 2 A schematic diagram of the basic structure of an image annotation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0011] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0012] See also Figure 1 As shown in FIG, this figure is a flow chart of an image annotation method for brain MRI image segmentation provided by an embodiment of the present invention, which can be applied to an image annotation system. Figure 1 As shown, the method may include steps 110 to 140.

[0013] Step 110: Acquire a brain MRI image data set of the target object, wherein the brain MRI image data set includes original image sequences at multiple scanning levels.

[0014] In this embodiment, taking a patient suspected of having a brain disease as an example, the patient's brain can be scanned in all directions using existing MRI scanning equipment. During the scanning process, the equipment will acquire image data from multiple different scanning levels according to the preset layer spacing and scanning angle. This data is recorded in the form of raw image sequences. The raw image sequence of each scanning level contains detailed information about the brain tissue at that level. For example, the scanning equipment acquires image data from several scanning levels from the top to the bottom of the brain at different time points and angles. The image of each level is a two-dimensional grayscale image. The grayscale values ​​in these images represent the degree of response of different brain tissues to the MRI signal. For example, gray matter, white matter, cerebrospinal fluid and other tissues will show different grayscale characteristics in the image, thus forming a brain MRI image data set.

[0015] Step 120: Perform multimodal feature fusion processing on the original image sequence to generate an enhanced image feature set.

[0016] In this embodiment, multimodal feature fusion processing is performed on the original image sequence in the brain MRI image data set obtained above, aiming to integrate the image features under different modalities to obtain more comprehensive and accurate brain information, and then generate an enhanced image feature set.

[0017] In an optional embodiment, performing multimodal feature fusion processing on the original image sequence to generate an enhanced image feature set includes: Step 121: performing cross-modal feature alignment processing on each scanning level in the original image sequence to obtain a spatially aligned cross-modal image sequence.

[0018] In this embodiment, for the original image sequence of each scanning level, the images of different modalities may have differences in spatial position and scale. For example, the acquired image data includes different modalities such as T1-weighted images and T2-weighted images. Taking one of the scanning levels as an example, the brain tissue contours in the T1-weighted image and the brain tissue contours in the T2-weighted image may have slight deviations in position. In order to enable these images of different modalities to better perform feature fusion, cross-modal feature alignment processing needs to be performed. Through the image registration algorithm, with the image of one modality as the reference, the images of other modalities are translated, rotated, scaled, etc., so that they are aligned in space. For example, with the T1-weighted image as the reference, by calculating the position difference of the feature points of the same anatomical structure in the T2-weighted image and the T1-weighted image, the T2-weighted image is transformed accordingly using the image transformation algorithm, so that the brain structures in the two modal images are accurately aligned, thereby obtaining a spatially aligned cross-modal image sequence.

[0019] Step 122: Perform cross-modal feature conversion processing on the spatially aligned cross-modal image sequence to generate a conversion feature set with a target dimension mapping; wherein the cross-modal feature conversion processing includes performing local texture enhancement and global structure preservation operations on the original image sequence of each scanning level, so that each feature map in the conversion feature set contains cross-modal complementary anatomical information.

[0020] In this embodiment, the spatially aligned cross-modal image sequence obtained after the cross-modal feature alignment process is further subjected to cross-modal feature conversion processing. Local texture enhancement is performed on the original image sequence at each scanning level. For example, a local filtering algorithm is used to analyze each small area in the image. Taking a 3×3 pixel neighborhood as an example, the grayscale value changes of the pixels in the neighborhood are calculated and filtered using an adapted filter kernel to highlight the texture features of the area.

[0021] Some subtle textures in brain tissue, such as the direction of nerve fibers, can be more clearly displayed through this local texture enhancement operation. At the same time, a global structure preservation operation is performed, and some global feature extraction algorithms, such as principal component analysis, are used to extract the overall structural information of the image, ensuring that the global structure of the image is not destroyed while enhancing local textures. After these operations, the features of the different modal images are fused and transformed to generate a transformed feature set with a target dimension mapping. Each feature map in this set integrates the complementary anatomical information of different modal images. For example, a feature map contains both information about brain tissue density in T1-weighted images and information about tissue water content in T2-weighted images, making the feature map more comprehensive in reflecting the anatomical structure of the brain.

[0022] Step 123: Call the debugged attention weight allocation model to perform weight allocation on different modal features in the conversion feature set to generate a modal weight distribution map; perform weighted fusion processing on the conversion feature set according to the modal weight distribution map to generate the enhanced image feature set.

[0023] In this embodiment, a conversion feature set has been generated. Next, a previously debugged attention weight allocation model is invoked. This model, trained and optimized using extensive data, can weight the different modal features in the conversion feature set based on their importance in reflecting brain structural information. For example, for the analysis of certain brain diseases, features in T1-weighted images may be more important for identifying the boundaries of the lesion area, while features in T2-weighted images may be more critical for determining the nature of the lesion. Based on these characteristics, the attention weight allocation model assigns corresponding weights to each modal feature, generating a modal weight distribution map. In this modal weight distribution map, each modal feature has a corresponding weight value, and the size of the weight value reflects the importance of the modal feature. Then, based on this modal weight distribution map, a weighted fusion process is performed on the conversion feature set. For each feature map in the conversion feature set, a weighted calculation is performed according to the corresponding weight value. Finally, the weighted results are spliced ​​to generate an enhanced image feature set. This enhanced image feature set fully integrates the advantageous information of different modalities and can more accurately reflect the anatomical structure of the brain.

[0024] Step 130: Calling a multi-layer cascade segmentation network to perform hierarchical feature extraction on the enhanced image feature set to obtain a multi-scale anatomical structure feature map.

[0025] In this embodiment, the generated enhanced image feature set is further processed using a multi-layer cascade segmentation network, with the aim of extracting brain anatomical structure features at different scales, thereby obtaining a multi-scale anatomical structure feature map.

[0026] As an implementation method, calling a multi-layer cascade segmentation network to perform hierarchical feature extraction on the enhanced image feature set to obtain a multi-scale anatomical structure feature map includes: Step 131: Input the enhanced image feature set into the first encoder of the multi-layer cascade segmentation network to extract the first-level anatomical structure feature map; input the first-level anatomical structure feature map into the second encoder of the multi-layer cascade segmentation network to extract the second-level anatomical structure feature map; input the second-level anatomical structure feature map into the third encoder of the multi-layer cascade segmentation network to extract the third-level anatomical structure feature map.

[0027] In this embodiment, the enhanced image feature set is input into the first encoder of the multi-layer cascade segmentation network. The first encoder performs a series of operations such as convolution and pooling on the input feature set. For example, by performing convolution operations with feature maps in the enhanced image feature set using convolution kernels of different sizes, local features at different scales are extracted. For another example, the first encoder uses 3×3 and 5×5 convolution kernels. The 3×3 convolution kernel can extract smaller-scale detail features, such as some tiny textures in brain tissue; the 5×5 convolution kernel can extract relatively larger-scale structural features, such as the rough outlines of sulci and gyri. After these operations, the first-level anatomical structure feature map is obtained.

[0028] Next, the first-level anatomical structure feature map is fed into the second encoder, which extracts higher-level features based on the first-level feature map. It may use larger convolution kernels or more convolution layers to process the first-level feature map, extracting more macroscopic anatomical features, such as the approximate boundaries of different lobes of the brain, to generate the second-level anatomical structure feature map.

[0029] Then, the second-level anatomical structure feature map is input into the third encoder, which will continue to extract more abstract and advanced features, such as the overall structural features of the entire brain, to obtain the third-level anatomical structure feature map.

[0030] Step 132: Perform multi-scale feature fusion processing on the first-level anatomical structure feature map, the second-level anatomical structure feature map, and the third-level anatomical structure feature map to generate the multi-scale anatomical structure feature map; wherein the multi-scale feature fusion processing includes upsampling the first-level anatomical structure feature map, downsampling the third-level anatomical structure feature map, and adjusting the contribution weights of feature maps of different scales through a channel attention mechanism.

[0031] In this embodiment, after obtaining the first-level, second-level, and third-level anatomical structure feature maps, a multi-scale feature fusion process is performed. First, the first-level anatomical structure feature map is upsampled. Since the first-level feature map extracts smaller-scale detail features, through upsampling operations, such as bilinear interpolation, its size is enlarged to a size that matches other feature maps, so that it can be fused with other feature maps at the same scale. For the third-level anatomical structure feature map, a downsampling operation is performed, because the third-level feature map extracts larger-scale macro features, and downsampling can make its size consistent with other feature maps. Figure 1 Easy to integrate.

[0032] The contribution weights of feature maps at different scales are then adjusted through a channel-wise attention mechanism. This mechanism analyzes the importance of each feature map across different channels. For example, in some channels, first-level feature maps may be more important for revealing subtle brain structures, while in other channels, third-level feature maps may be more crucial for understanding the brain's overall layout. Based on these analysis results, a corresponding weight is assigned to each channel of the feature maps at different scales. The weighted feature maps are then concatenated to generate a multi-scale anatomical feature map. This map incorporates brain anatomical features at different scales, providing a more comprehensive representation of the brain's structural information.

[0033] Step 140: performing region boundary optimization processing based on the multi-scale anatomical structure feature map to generate a labeled brain structure segmentation image.

[0034] In this embodiment, based on the generated multi-scale anatomical structure feature map, regional boundary optimization processing is carried out to finally generate a labeled brain structure segmentation image to more clearly divide different structural regions of the brain.

[0035] In a preferred embodiment, the performing of region boundary optimization processing based on the multi-scale anatomical structure feature map to generate a labeled brain structure segmentation image includes: Step 141: Perform initial segmentation processing on the multi-scale anatomical structure feature map to generate an initial brain structure segmentation mask.

[0036] In this embodiment, an initial segmentation process is performed on the multi-scale anatomical structure feature map, aiming to preliminarily divide regions of different brain structures and generate an initial brain structure segmentation mask.

[0037] As a preferred technical solution, performing initial segmentation processing on the multi-scale anatomical structure feature map to generate an initial brain structure segmentation mask includes: Step 1410: Input the multi-scale anatomical structure feature map into the fully connected decoder of the multi-layer cascade segmentation network to generate an initial probability distribution map; perform threshold segmentation processing on the initial probability distribution map to obtain a binary segmentation result; perform connected domain analysis processing on the binary segmentation result to remove isolated noise areas to obtain a denoised binary segmentation result; perform topological correction processing on the denoised binary segmentation result based on a preset brain anatomical structure template to generate the initial brain structure segmentation mask.

[0038] In this embodiment, the multi-scale anatomical structure feature map is input into the fully connected decoder of the multi-layer cascade segmentation network. The fully connected decoder performs a series of deconvolution, full connection and other operations on the input feature map to convert the feature map into an initial probability distribution map. Each pixel value in the initial probability distribution map represents the probability that the pixel belongs to a different brain structure. For example, in a probability distribution map, the probability value of a pixel is 0.8, indicating that there is an 80% probability that the pixel belongs to the gray matter area. Then, the initial probability distribution map is thresholded and a suitable threshold value, such as 0.5, is set. Pixels with a probability value greater than 0.5 are set to 1, and pixels with a probability value less than 0.5 are set to 0, thereby obtaining a binary segmentation result. In the binary image, white pixels represent areas that may belong to the target brain structure, and black pixels represent other areas.

[0039] However, some isolated noise points may be generated during the binarization process, which may interfere with subsequent analysis. Therefore, the binary segmentation results are subjected to connected domain analysis. By using the connected domain marking algorithm, the connected regions in the image are found, and the isolated regions with smaller areas are treated as noise and removed to obtain the denoised binary segmentation results. Finally, the denoised binary segmentation results are topologically corrected based on the preset brain anatomical structure template. The preset brain anatomical structure template contains the topological information of the normal brain structure, such as the relative position and shape of different brain regions. The denoised binary segmentation results are compared and adjusted with the template, and the areas that do not conform to the template topological structure are corrected to generate the initial brain structure segmentation mask.

[0040] Step 142: performing boundary confidence analysis on the initial brain structure segmentation mask to determine a confidence distribution map of the boundary fuzzy area.

[0041] In this embodiment, in order to further optimize the boundary of brain structure segmentation, a boundary confidence analysis is performed on the generated initial brain structure segmentation mask to determine a confidence distribution map of the boundary fuzzy area.

[0042] In one implementation, performing boundary confidence analysis on the initial brain structure segmentation mask to determine a confidence distribution map of a fuzzy boundary region includes: Step 1420: performing boundary gradient calculation on the initial brain structure segmentation mask to generate a boundary gradient intensity map; performing feature uncertainty assessment on the multi-scale anatomical structure feature map to generate a feature uncertainty distribution map; performing weighted fusion processing on the boundary gradient intensity map and the feature uncertainty distribution map to generate an initial confidence map; performing regional connectivity analysis on the initial confidence map to screen out continuous regions below a preset confidence threshold as the boundary fuzzy regions; and generating the confidence distribution map based on the spatial distribution of the boundary fuzzy regions.

[0043] In this embodiment, the boundary gradient calculation is first performed on the initial brain structure segmentation mask. Using a gradient calculation algorithm, such as the Sobel operator, the gradient value of each pixel in the mask is calculated to generate a boundary gradient intensity map. In this boundary gradient intensity map, areas with larger gradient values ​​indicate areas where boundary changes are obvious, which may be the boundaries of different brain structures. For example, at the boundary between gray matter and white matter, the grayscale value of the pixel changes greatly, and the gradient value will also be correspondingly larger. Next, the feature uncertainty assessment is performed on the multi-scale anatomical structure feature map. Using some uncertainty assessment algorithms, the stability and reliability of each feature in the feature map are analyzed to generate a feature uncertainty distribution map. The values ​​in the feature uncertainty distribution map reflect the degree of uncertainty of each position feature.

[0044] The boundary gradient intensity map and the feature uncertainty distribution map are then weighted and fused. Weights are assigned to the two maps based on their importance, for example, a weight of 0.6 for the boundary gradient intensity map and a weight of 0.4 for the feature uncertainty distribution map. The values ​​of the corresponding pixels in the two maps are weighted and calculated to generate an initial confidence map. In this initial confidence map, the value of each pixel represents the confidence level that the location is a boundary. Regional connectivity analysis is performed on the initial confidence map. Using the connectivity analysis algorithm, continuous regions in the map with values ​​below a preset confidence threshold (e.g., 0.3) are identified. These regions are the fuzzy boundary regions.

[0045] Finally, based on the spatial distribution of these fuzzy boundary areas, a confidence distribution map is generated. In this confidence distribution map, each fuzzy boundary area has a corresponding confidence value, which can help to optimize the boundary more accurately in the future.

[0046] Step 143: performing an iterative morphological optimization operation on the initial brain structure segmentation mask based on the confidence distribution map to generate an optimized brain structure segmentation mask; wherein the iterative morphological optimization operation includes performing a dilation-erosion sequence processing on the boundary fuzzy area and adjusting the size of the dilation kernel based on the confidence distribution map.

[0047] In this embodiment, an iterative morphological optimization operation is performed on the initial brain structure segmentation mask based on the generated confidence distribution map to further optimize the boundaries of the brain structure segmentation and generate an optimized brain structure segmentation mask.

[0048] In an optional embodiment, performing an iterative morphological optimization operation on the initial brain structure segmentation mask based on the confidence distribution map to generate an optimized brain structure segmentation mask includes: Step 1430: Determine the corresponding morphological operation intensity parameter according to the confidence value of each area in the confidence distribution map; perform a morphological closing operation with a fixed kernel size on the target confidence area in the initial brain structure segmentation mask; perform a morphological opening operation with an adaptive kernel size on the boundary fuzzy area, wherein the kernel size is negatively correlated with the confidence value; repeatedly perform the morphological closing operation and the morphological opening operation until the confidence values ​​of all areas in the confidence distribution map reach a preset optimization threshold; and use the morphological processing result obtained from the last iteration as the optimized brain structure segmentation mask.

[0049] In this embodiment, the corresponding morphological operation strength parameter is first determined based on the confidence values ​​of each region in the confidence distribution map. For target confidence regions with high confidence, whose boundaries are relatively clear, a morphological closing operation with a fixed kernel size is performed. This morphological closing operation can fill small holes and make the boundaries more continuous. For example, a 3×3 structuring element is used to close these regions. For regions with blurred boundaries, due to their lower confidence, a morphological opening operation with an adaptive kernel size is performed. The kernel size is negatively correlated with the confidence value; the lower the confidence, the larger the kernel size. For example, for regions with blurred boundaries with a confidence value of 0.2, a 5×5 structuring element is used for the opening operation; for regions with a confidence value of 0.1, a 7×7 structuring element is used. The opening operation can remove some noise and small protrusions, making the boundaries smoother. The morphological closing and opening operations are then repeated. After each iteration, the confidence values ​​of each region in the confidence distribution map are reassessed until the confidence values ​​of all regions reach a preset optimization threshold (e.g., 0.6). Finally, the morphological processing result obtained in the last iteration is used as the optimized brain structure segmentation mask.

[0050] Step 144: superimpose and fuse the optimized brain structure segmentation mask with the original image sequence to generate the labeled brain structure segmentation image.

[0051] In this embodiment, the optimized brain structure segmentation mask is superimposed and fused with the original image sequence to generate a labeled brain structure segmentation image, making the segmentation result more intuitive and accurate.

[0052] As a preferred embodiment, the step of superimposing and fusing the optimized brain structure segmentation mask with the original image sequence to generate the labeled brain structure segmentation image includes: Step 1440: performing edge smoothing on the optimized brain structure segmentation mask to generate a smooth segmentation contour; performing transparency blending on the smooth segmentation contour and the grayscale values ​​of the original image sequence to generate a preliminary annotated image; performing color coding on the boundary regions of different anatomical structures in the preliminary annotated image to generate a distinguishable pseudo-color annotation layer; spatially aligning and superimposing the pseudo-color annotation layer with the original image sequence to generate the annotated brain structure segmentation image; In this embodiment, the optimized brain structure segmentation mask is first subjected to edge smoothing. The edges of the mask are processed by a smoothing algorithm, such as a Gaussian filter, to make the edges smoother and generate a smooth segmentation contour. For example, a Gaussian filter with a standard deviation of 1 is used to filter the mask edge. Then, the smooth segmentation contour is blended with the grayscale value of the original image sequence for transparency. A transparency parameter is set, such as 0.5, and the smooth segmentation contour is blended with the grayscale value of the original image sequence according to the transparency to generate a preliminary annotated image. In this preliminary annotated image, both the information of the original image and the approximate position of the segmentation contour can be clearly identified. Next, color coding is performed on the boundary areas of different anatomical structures in the preliminary annotated image.

[0053] As another preferred embodiment, performing color coding on boundary regions of different anatomical structures in the preliminary annotated image to generate a pseudo-color annotation layer with distinguishability includes: Step 14401: Assign unique color labels to different anatomical regions according to preset brain anatomical structure classification rules.

[0054] In this embodiment, the preset brain anatomical structure classification rules are based on medical anatomical knowledge. For example, the brain is divided into different lobes such as the frontal lobe, parietal lobe, temporal lobe, and occipital lobe, as well as different brain tissues such as the thalamus and hippocampus. Each anatomical region is assigned a unique color identifier, such as red for the frontal lobe, blue for the parietal lobe, green for the temporal lobe, yellow for the occipital lobe, purple for the thalamus, and orange for the hippocampus. These color identifiers are carefully selected to ensure sufficient differentiation in subsequent displays, making it easier for doctors or researchers to observe and analyze brain structures.

[0055] Step 14402: Perform category matching processing on each connected region in the optimized brain structure segmentation mask to determine the corresponding target color identifier.

[0056] In this embodiment, each connected region in the optimized brain structure segmentation mask is matched against preset brain anatomical structure classification rules based on its location, shape, and other characteristics within the brain. For example, if the shape and location of a connected region are analyzed and found to match the characteristics of the frontal lobe, the target color corresponding to that connected region is determined to be red. This process uses relevant common image analysis algorithms and medical knowledge models to accurately classify each connected region, thereby determining the correct target color.

[0057] Step 14403: performing gradient color filling processing on the boundary pixels of the connected area based on the target color identifier to generate the pseudo color annotation layer; wherein the color saturation of the gradient color filling processing is positively correlated with the distance from the boundary pixel to the center of the segmentation contour.

[0058] In this embodiment, for each connected area with a target color identifier, a gradient color filling process is performed on its boundary pixels. Taking a certain connected area as an example, its target color identifier is red. Starting from the center of the segmentation contour, gradient color filling is performed toward the boundary pixels. The farther the boundary pixels are from the center of the segmentation contour, the higher the color saturation. For example, at the boundary pixels closer to the center, the saturation of red may be 0.3. As the distance increases, the saturation gradually increases, and at the boundary pixels farthest from the center, the saturation reaches 0.8. Through this gradient color filling process, a pseudo-color annotation layer with distinguishability is generated, making the boundaries of different anatomical structures clearer and easier to observe and distinguish.

[0059] In an alternative technical solution, spatially aligning and superimposing the pseudo-color annotation layer with the original image sequence to generate the annotated brain structure segmentation image includes: Step 14404: Perform normalized coordinate transformation processing on the pixel coordinates of the pseudo-color annotation layer to generate a standardized coordinate mapping table aligned with the physical coordinate system of the original image sequence; wherein the normalized coordinate transformation processing includes mapping the pixel positions of the pseudo-color annotation layer to the scanning inter-layer resolution ratio space of the original image sequence, and compensating for inter-layer displacement errors through a bilinear interpolation algorithm.

[0060] In this embodiment, since the pseudo-color annotation layer and the original image sequence may differ in terms of coordinate system and resolution, a normalized coordinate conversion process is required. First, the pixel position of the pseudo-color annotation layer is mapped to the scanning inter-layer resolution ratio space of the original image sequence. For example, the original image sequence has a certain proportional relationship in inter-layer resolution, and the inter-layer spacing is 1mm, and the pixel position of the pseudo-color annotation layer needs to be repositioned according to the inter-layer resolution. The inter-layer displacement error is compensated by the bilinear interpolation algorithm. For the pixels located between the two layers in the pseudo-color annotation layer, their exact position in the coordinate system of the original image sequence is calculated by bilinear interpolation based on their relative position between the two layers. After these operations, a standardized coordinate mapping table aligned with the physical coordinate system of the original image sequence is generated to ensure that the pseudo-color annotation layer and the original image sequence can accurately correspond in spatial position.

[0061] Step 14405: Perform spatial resampling processing on the color channels of the pseudo-color annotation layer according to the standardized coordinate mapping table to generate a resampled pseudo-color layer that completely matches the pixel spatial distribution of the original image sequence; wherein the spatial resampling processing adopts an isotropic interpolation kernel function, and the kernel width parameter of the isotropic interpolation kernel function is unified with the dimension of the pixel spacing within the layer of the original image sequence.

[0062] In this embodiment, the color channels of the pseudo-color annotation layer are spatially resampled based on the generated standardized coordinate mapping table. An isotropic interpolation kernel function is used, which can perform uniform interpolation in all directions. The kernel width parameter is dimensionally consistent with the pixel spacing within the layer of the original image sequence. For example, if the pixel spacing within the layer of the original image sequence is 0.5 mm, the kernel width parameter will also be set according to this spacing. Through this spatial resampling process, the color channels of the pseudo-color annotation layer are resampled so that the resampled pseudo-color layer completely matches the original image sequence in terms of pixel spatial distribution. In this way, in the subsequent overlay process, the accurate fusion of color information and the original image can be ensured.

[0063] Step 14406: Input the resampled pseudo-color layer into the transparency blending channel, perform multi-channel pixel-level overlay operation processing, and generate a mixed annotated image; wherein, the overlay operation processing includes performing a linear brightness-preserving transformation on the grayscale values ​​of the original image sequence, and performing dynamic range compression on the RGB color components of the resampled pseudo-color layer, so that the pixel value after overlay does not exceed a preset maximum display dynamic range.

[0064] In this embodiment, the resampled pseudo-color layer is input into the transparency blending channel for multi-channel pixel-level overlay processing. A linear brightness-preserving transformation is performed on the grayscale values ​​of the original image sequence. This is intended to preserve the brightness information of the original image during the overlay process, ensuring that the overall brightness distribution of the brain does not change significantly. For example, a linear transformation function is used to appropriately scale and offset the grayscale values ​​of the original image to ensure that the brightness is within a reasonable range. Simultaneously, dynamic range compression is performed on the RGB color components of the resampled pseudo-color layer. Because the value range of the RGB color components can be large, compression is required to prevent the pixel values ​​after overlay from exceeding a preset maximum display dynamic range (e.g., 0-255). For example, a compression function is used to map the color component values ​​to an appropriate range. After these operations, a mixed annotated image is generated, which contains both the information of the original image and the color information of the pseudo-color annotation layer.

[0065] Step 14407: Perform edge sharpening and anti-aliasing processing on the mixed annotated image to generate an intermediate annotated image with enhanced boundaries; wherein the edge sharpening processing adopts an anisotropic diffusion filtering algorithm, and the gradient threshold parameter of the anisotropic diffusion filtering algorithm matches the grayscale gradient distribution statistic of the original image sequence.

[0066] In this embodiment, edge sharpening and anti-aliasing processing are performed on the mixed annotated image. An anisotropic diffusion filter algorithm is used for edge sharpening, and the algorithm can perform targeted diffusion processing based on the gradient information in different directions in the image. The gradient threshold parameter matches the grayscale gradient distribution statistics of the original image sequence. For example, the appropriate gradient threshold is determined by performing statistical analysis on the grayscale gradient of the original image sequence. When the gradient value in the image is greater than the threshold, the algorithm performs a smaller diffusion processing to highlight the edge; when the gradient value is less than the threshold, a larger diffusion processing is performed to smooth the image. At the same time, anti-aliasing processing is performed, and the jagged edges in the image are smoothed by an anti-aliasing algorithm, such as an interpolation-based algorithm, to generate an intermediate annotated image with enhanced boundaries, so that the boundaries of the brain structure are clearer and smoother.

[0067] Step 14408: Perform adaptive contrast adjustment processing based on the boundary intensity distribution of the intermediate annotated image to generate a brain structure segmentation image for final display; wherein the adaptive contrast adjustment processing includes performing a histogram equalization operation on the low contrast area and performing nonlinear gamma correction on the high gradient area, and the correction coefficient of the nonlinear gamma correction is negatively correlated with the local gradient amplitude.

[0068] In this embodiment, adaptive contrast adjustment processing is performed based on the boundary intensity distribution of the intermediate annotated image. For low-contrast areas in the image, a histogram equalization operation is performed. By statistically analyzing the grayscale histogram of the low-contrast area, the grayscale values ​​are redistributed so that the contrast of the area is improved. For example, the grayscale value of the low-contrast area is expanded from a narrower range to a wider range, thereby increasing the visual clarity of the image. At the same time, nonlinear gamma correction is performed on the high-gradient area. The correction coefficient of the nonlinear gamma correction is negatively correlated with the local gradient amplitude, that is, the larger the gradient amplitude, the smaller the correction coefficient, thereby avoiding excessive contrast enhancement in the high-gradient area, resulting in the image appearing too bright or too dark. After these operations, the final displayed brain structure segmentation image is generated, the image contrast is more appropriate, and the segmentation of the brain structure is clearer and more discernible.

[0069] Step 14409: Perform color space verification processing on the final displayed brain structure segmentation image so that the chromaticity channel of the pseudo-color annotation layer and the luminance channel of the original image sequence have no overlapping conflicts within the display color gamut; wherein, the color space verification processing includes converting the RGB color space into the YUV space and then detecting the orthogonality index of the luminance component and the chromaticity component. If an overlapping area is detected, the chromaticity coordinates are remapped to a preset safe palette range.

[0070] In this embodiment, the final displayed brain structure segmentation image is subjected to color space verification processing. First, the RGB color space is converted to the YUV space. This conversion can separate the color information into the luminance component (Y) and the chrominance component (U, V). Then, the orthogonality index of the luminance component and the chrominance component is detected, and the correlation between the luminance component and the chrominance component is calculated. If an overlapping area is detected, that is, there is mutual interference between the luminance component and the chrominance component, it is necessary to remap the chromaticity coordinates to a preset safe palette interval. The preset safe palette interval is determined based on the color gamut range of the display device and the requirements for medical image display. By mapping the chromaticity coordinates to this interval, it is ensured that the chromaticity channel of the pseudo-color annotation layer and the luminance channel of the original image sequence have no overlapping conflicts within the display color gamut, thereby ensuring the display quality and accuracy of the image.

[0071] In an independent embodiment, after generating the labeled brain structure segmentation image, the method further includes: performing cross-level 3D voxel fusion processing on the labeled brain structure segmentation image to generate a 3D brain anatomical structure model; wherein the cross-level 3D voxel fusion processing includes mapping the annotation results of each scanning level to a 3D spatial coordinate system according to the inter-layer resolution ratio, and compensating for inter-layer displacement errors through an isotropic interpolation algorithm; performing surface mesh optimization processing on the 3D brain anatomical structure model to generate a smooth and continuous 3D visualization model; wherein the surface mesh optimization processing includes eliminating voxel splicing noise based on a Laplace smoothing algorithm and adjusting the mesh density using an adaptive triangulation algorithm, wherein the mesh density is negatively correlated with the local curvature radius; spatially aligning the 3D visualization model with a standard brain anatomical atlas and calculating the volume difference coefficient of each anatomical structure; wherein the spatial registration includes alignment based on rigid transformation of feature points and optimizing non-rigid deformation parameters through a mutual information maximization algorithm; generating a brain structure abnormality detection report based on the volume difference coefficient, and associating the 3D visualization model with the abnormality detection report and storing it in a medical imaging database.

[0072] In this embodiment, cross-level 3D voxel fusion processing is performed on the generated, annotated brain structure segmentation image. First, the annotation results for each scan level are mapped to a 3D coordinate system based on the inter-slice resolution ratio. For example, given an inter-slice resolution of 1 mm, the position of each voxel in the 2D segmentation image at each scan level is determined in 3D space according to this resolution. Inter-slice displacement errors are compensated using an isotropic interpolation algorithm. For voxels in the inter-slice transition region, the exact position and attributes are calculated using an interpolation algorithm, thereby generating a 3D brain anatomical model.

[0073] Next, the surface mesh of the 3D brain anatomical model is optimized. Voxel patchwork noise is eliminated using a Laplace smoothing algorithm, which adjusts the positions of mesh vertices to smooth the mesh surface. An adaptive triangulation algorithm is also used to adjust the mesh density, determining it based on the local curvature radius of the model surface. Regions with smaller local curvature radii, such as brain wrinkles, receive a higher mesh density to more accurately represent these complex structures; regions with larger local curvature radii receive a lower mesh density. After these processes, a smooth and continuous 3D visualization model is generated.

[0074] The 3D visualization model is then spatially registered with a standard brain anatomical atlas. First, based on the rigid transformation alignment of the feature points, some iconic feature points in the model and atlas are found, and these feature points are aligned through rigid transformations such as translation and rotation. The non-rigid deformation parameters are then optimized using a mutual information maximization algorithm, further adjusting the shape of the model to better match the standard atlas. During the registration process, the volume difference coefficient of each anatomical structure is calculated. For example, the difference between the volume of an anatomical structure in the model and the volume of the corresponding anatomical structure in the standard atlas is calculated and normalized to obtain the volume difference coefficient.

[0075] Finally, a brain structural abnormality detection report is generated based on the volume difference coefficient, analyzing which anatomical structures have volume abnormalities. The 3D visualization model and the abnormality detection report are linked and stored in a medical imaging database for subsequent review and analysis by doctors and researchers.

[0076] In an independent embodiment, after the annotated brain structure segmentation image is generated, the method further includes: performing segmentation quality assessment on the annotated brain structure segmentation image to generate a regional confidence heat map; wherein the segmentation quality assessment includes calculating the segmentation probability variance of each pixel and normalizing the variance value to a preset interval as a confidence indicator; according to the coordinates of the region below the preset threshold in the regional confidence heat map, extracting the corresponding local image blocks from the original image sequence to generate a set of image blocks to be verified; inputting the set of image blocks to be verified into a debugged segmentation error classification network to output an error type label and a correction vector; wherein the segmentation error classification network adopts a multi-scale convolution structure and classifies boundary blur, topological fracture and modal confusion errors based on a cross entropy loss function; performing pixel-level geometric deformation correction on the annotated brain structure segmentation image according to the correction vector to generate an optimized segmentation image; wherein the geometric deformation correction includes applying an elastic transformation field to the error region, and the transformation field parameters are controlled by the amplitude and direction components of the correction vector.

[0077] In this embodiment, the segmentation quality of the labeled brain structure segmentation image is evaluated. First, the segmentation probability variance of each pixel is calculated. For example, for each pixel, the variance value is calculated based on its probability distribution in different segmentation results. The variance value is normalized to a preset interval as a confidence index. The preset interval is [0, 1]. Through normalization, the variance value is mapped to the interval. The larger the value, the lower the segmentation confidence of the pixel. A regional confidence heat map is generated based on the calculation results. In the heat map, different colors represent different confidence levels. Then, based on the coordinates of the regions below the preset threshold (such as 0.3) in the regional confidence heat map, the corresponding local image blocks are extracted from the original image sequence. These image blocks may have segmentation errors and are combined into a set of image blocks to be verified.

[0078] Next, the set of image patches to be verified is fed into a debugged segmentation error classification network. This network employs a multi-scale convolutional architecture, extracting features from the image patches using convolution kernels of varying scales. It then classifies boundary blur, topological breakage, and modal aliasing errors using a cross-entropy loss function, outputting error type labels and correction vectors.

[0079] Finally, pixel-level geometric deformation correction is performed on the annotated brain structure segmentation image based on the correction vectors. For regions with errors, an elastic transformation field is applied, with the transformation field parameters controlled by the magnitude and directional components of the correction vector. For example, a large magnitude of the correction vector indicates that a larger deformation is required to correct the error, and the directional component determines the direction of the deformation. After these operations, an optimized segmentation image is generated, improving the accuracy and quality of the segmentation.

[0080] In another implementation of an embodiment of the present invention, obtaining a brain MRI image data set of a target object is the beginning of the entire process. A full-scale scan is performed on a patient suspected of having a brain disease using existing MRI scanning equipment. According to preset layer spacing and scanning angles, a plurality of original image sequences of different scanning levels are obtained. The grayscale value of each two-dimensional grayscale image in these sequences reflects the degree of response of different brain tissues to the MRI signal, thereby forming an image data set containing rich brain information.

[0081] Multimodal feature fusion processing of raw image sequences is a key step in improving image quality. Images of different modalities, such as T1-weighted and T2-weighted images, may differ in spatial position and scale. First, cross-modal feature alignment is performed. Using an image registration algorithm, images from one modality are spatially aligned by performing operations such as translation, rotation, and scaling on the images from the other modalities, generating a spatially aligned cross-modal image sequence. Next, cross-modal feature transformation is performed on the raw image sequence at each scan level, performing local texture enhancement and global structure preservation. Local filtering algorithms are used to highlight subtle textures, such as the orientation of neural fibers, while global feature extraction algorithms, such as principal component analysis, are employed to preserve the overall image structure. These operations generate a transformed feature set with a target dimensionality mapping, where each feature map incorporates complementary anatomical information from different modalities. Finally, an attention weight allocation model is used to assign weights to the different modal features in the transformed feature set, generating a modality weight distribution map. This weighted fusion is then performed to produce an enhanced image feature set that more accurately reflects brain anatomy.

[0082] A multi-layer cascaded segmentation network is used to perform hierarchical feature extraction on the enhanced image feature set to obtain multi-scale anatomical structure feature maps. The enhanced image feature set is sequentially input into each encoder level of the multi-layer cascaded segmentation network. Through operations such as convolution and pooling, anatomical structure features at different scales are extracted to obtain anatomical structure feature maps at each level. These feature maps are then subjected to multi-scale feature fusion processing. The first-level feature map is upsampled and the third-level feature map is downsampled to match the scale. The contribution weights of the feature maps at different scales are adjusted through a channel attention mechanism to generate a multi-scale anatomical structure feature map that integrates brain anatomical structure features at different scales.

[0083] Based on the multi-scale anatomical feature map, regional boundary optimization is performed to generate an annotated brain structure segmentation image. Initial segmentation is first performed on the feature map. After inputting into a fully connected decoder to generate an initial probability distribution map, threshold segmentation is performed, connected domain analysis denoising is performed, and topology correction based on a preset brain anatomical template is performed to generate an initial brain structure segmentation mask. Next, boundary confidence analysis is performed on the mask. By calculating boundary gradients, assessing feature uncertainty, weighted fusion is performed to generate an initial confidence map, and regional connectivity analysis is performed to determine the confidence distribution map for fuzzy boundary regions. Based on this distribution map, an iterative morphological optimization operation is performed on the initial mask. Morphological closing and opening operations with different kernel sizes are performed on regions with different confidence levels until the confidence value reaches a preset threshold. The optimized brain structure segmentation mask is then superimposed and fused with the original image sequence. A series of operations, including edge smoothing, transparency blending, color encoding, spatial alignment overlay, edge sharpening and anti-aliasing, adaptive contrast adjustment, and color space validation, are performed to generate an annotated brain structure segmentation image.

[0084] After generating an annotated brain structure segmentation image, further processing and analysis are required. Cross-level 3D voxel fusion can be performed, mapping the annotation results of each scan level to a 3D spatial coordinate system. Interpolation algorithms are used to compensate for inter-slice displacement errors, generating a 3D brain anatomical model. This model undergoes surface mesh optimization to eliminate voxel splicing noise and adjust the mesh density to produce a smooth and continuous 3D visualization model. This model is then spatially registered with a standard brain anatomical atlas, and the volumetric difference coefficients of each anatomical structure are calculated. A brain structural abnormality detection report is generated and stored in a medical imaging database. Segmentation quality assessment can also be performed by calculating the variance of the segmentation probability for each pixel and normalizing it as a confidence metric, generating a regional confidence heat map. Based on this heat map, a set of image blocks to be verified is extracted and fed into a segmentation error classification network. The resulting error type label and correction vector are then used to perform pixel-level geometric deformation correction on the segmented image, generating an optimized segmented image.

[0085] Furthermore, deep learning can be used to generate high-resolution structural MRI images, enabling virtual multimodal fusion of single-modality PET data. An automated segmentation algorithm based on generative MRI can reduce errors in gray-white matter boundary image recognition. For image segmentation model debugging, a cascaded 3D U-Net segmentation framework was constructed based on real MRI and generated synthetic MRI, with real MRI segmentation labels used as the gold standard for segmentation data generation. A loss function that minimizes the multiscale Dice loss (weight 0.6) and the boundary Hausdorff distance loss (weight 0.4) was used, combined with an Adam optimizer (initial lr = 5×10^-4), and debugged until convergence. 3D morphological closing (3×3×3 spherical kernel) was applied to segmentation edges to eliminate voxel-level aliasing artifacts. Fusion of SUVR heatmaps with synthetic MRI anatomy also generates interactive annotated images of Aβ deposition distribution.

[0086] In summary, the embodiment of the present invention first obtains a brain MRI image data set containing multiple scanning level original image sequences, and then performs multimodal feature fusion processing to effectively integrate the features of multiple modalities, thereby generating an enhanced image feature set, which can make the information contained in the image more comprehensive and in-depth, while mining features that are not easily perceived. Furthermore, a multi-layer cascade segmentation network is used to perform hierarchical feature extraction on the enhanced image feature set, which can capture brain anatomical structure features from different scales and levels, and obtain a multi-scale anatomical structure feature map; finally, based on the multi-scale anatomical structure feature map, regional boundary optimization processing is performed to generate a labeled brain structure segmentation image, which can accurately define the boundaries of each brain structure, making the segmented image more accurate and clear. In this way, image segmentation and annotation of brain MRI images can be accurately and clearly achieved.

[0087] It is worth mentioning that the core of the embodiment of the present invention is to improve the analysis efficiency and accuracy of medical images through computer image processing technology, and does not directly involve the implementation of disease diagnosis or treatment methods. The embodiment of the present invention focuses on establishing a complete set of brain MRI image processing processes, specifically including quantifiable technical modules such as multimodal feature fusion, hierarchical feature extraction, and boundary optimization algorithms. Its innovation is reflected in the mathematical modeling and computer automation processing methods of image data. For example, the spatial registration problem of different image sequences is solved by a cross-modal feature alignment algorithm, the multi-scale feature fusion process is optimized by the channel attention mechanism, and the segmentation boundary accuracy is improved by combining morphological operations and confidence analysis. These technical means are essentially technical improvements in the field of computer vision and image processing, and have clear technical implementation paths and industrial application scenarios.

[0088] In addition, the implementation effect of the embodiment of the present invention is entirely dependent on the automatic processing capabilities of the computer system, and does not need to rely on the subjective judgment or medical diagnosis behavior of the doctor. Although the processing object is brain medical images, the innovation of the technical solution lies in the mathematical transformation and feature optimization of the image data, such as achieving image space alignment through coordinate normalization processing, improving edge clarity by using anisotropic diffusion filtering, and correcting segmentation errors based on elastic transformation fields. The improvement direction of these technical means always revolves around technical indicators such as image processing accuracy, computational efficiency, and feature expression capabilities. The results are manifested as optimized segmented images or three-dimensional visualization models, which are computer-generated data products rather than medical diagnostic conclusions. Even though "suspected patients" are mentioned in the example, the technical solution itself is not limited to disease diagnosis scenarios. Its method is also applicable to brain image analysis of healthy people and has universal technical characteristics.

[0089] Furthermore, the technical contribution of the embodiments of the present invention is reflected in the improvement and optimization of the functions of medical image processing equipment, which belongs to the improvement in the field of medical equipment technology. For example, the efficiency of feature extraction is improved through a multi-layer cascade segmentation network architecture, the accuracy of anatomical structure reconstruction is enhanced by a three-dimensional voxel fusion algorithm, and the visualization effect of human-computer interaction is improved with the help of pseudo-color annotation layer generation technology. These technical achievements can be directly applied to the development and performance improvement of medical image analysis systems. The implementation of the entire technical solution does not involve the imposition of medical measures or intervention behaviors on the human body. Its core value lies in providing higher-quality image analysis tools for medical diagnosis, rather than replacing or affecting the doctor's diagnostic decision-making process. Therefore, the embodiments of the present invention meet the requirements of the object attributes of the technical solution, and are inventions and creations that use technical means to solve specific technical problems. They are essentially different from disease diagnosis and treatment methods.

[0090] It is worth noting that in actual applications, cross-modal feature alignment can be performed based on multimodal registration methods in existing medical image processing technologies (such as open source tools such as ANTs and Elastix), and inter-layer displacement compensation can be achieved through the B-spline elastic registration algorithm.

[0091] For local texture enhancement in cross-modal feature conversion, a direction-controlled Gabor filter bank can be used to extract multi-scale texture features, combined with non-local mean filtering to achieve global structure preservation.

[0092] The attention weight allocation model can build a channel attention mechanism based on the Squeeze-and-Exception network architecture and capture the global statistical information of multimodal features through the adaptive pooling layer.

[0093] The multi-layer cascade segmentation network can be built with reference to the 3D U-Net framework. The residual connection module is integrated in the encoding path to prevent gradient disappearance, and the decoding path uses transposed convolution to restore the feature map resolution.

[0094] The adaptive kernel size calculation in the morphological optimization operation can be achieved by establishing a linear mapping function between the confidence value and the size of the structural element, where the upper limit of the kernel size needs to be physically constrained according to the resolution between image layers.

[0095] The three-dimensional voxel fusion process requires establishing a spatial transformation matrix based on the pixel spacing and layer thickness parameters in the DICOM standard, and using a trilinear interpolation algorithm to achieve cross-layer data fusion.

[0096] Surface mesh optimization can implement Laplace smoothing based on the MeshLab open source toolkit, and dynamically adjust the density of triangle faces by combining local curvature radius calculation.

[0097] The probability variance calculation used in segmentation quality assessment can be calculated using Monte Carlo Dropout to obtain pixel-level prediction distributions, and a sliding window strategy can be used to extract image blocks for verification. The segmentation error classification network can be constructed using cascaded Inception modules to capture multi-scale features, and a Focal Loss function can be used to mitigate class imbalance.

[0098] In addition, operations involving spatial transformation can ensure dimensional uniformity through DICOM coordinate system conversion, and the kernel size of morphological operations needs to be converted to millimeter units according to the physical resolution of the image to avoid scale confusion.

[0099] With the above-mentioned technologies, those skilled in the art can clearly and completely implement the embodiments of the present invention.

[0100] See also Figure 2 As shown in FIG. 2 , this figure is a schematic diagram of the basic structure of an image annotation system 200 provided by an embodiment of the present invention. The image annotation system 200 includes: Processor 201; a storage device 202 having a computer program 2020 stored thereon; When the computer program 2020 is executed by the processor 201 , the processor 201 implements any of the image annotation methods for brain MRI image segmentation.

[0101] Based on the above, a readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0102] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

Claims

1. An image annotation method for brain MRI image segmentation, characterized in that: include: Acquiring a brain MRI image data set of a target object, wherein the brain MRI image data set includes original image sequences at multiple scanning levels; Performing multimodal feature fusion processing on the original image sequence to generate an enhanced image feature set; Calling a multi-layer cascade segmentation network to perform hierarchical feature extraction on the enhanced image feature set to obtain a multi-scale anatomical structure feature map; Based on the multi-scale anatomical structure feature map, regional boundary optimization processing is performed to generate a labeled brain structure segmentation image.

2. The method according to claim 1, characterized in that The performing multimodal feature fusion processing on the original image sequence to generate an enhanced image feature set includes: Performing cross-modal feature alignment processing on each scanning level in the original image sequence to obtain a spatially aligned cross-modal image sequence; performing a cross-modal feature conversion process on the spatially aligned cross-modal image sequence to generate a conversion feature set having a target dimensional mapping; wherein the cross-modal feature conversion process includes performing local texture enhancement and global structure preservation operations on the original image sequence at each scanning level, so that each feature map in the conversion feature set contains cross-modal complementary anatomical information; Calling the debugged attention weight allocation model to perform weight allocation on different modal features in the conversion feature set to generate a modal weight distribution graph; The conversion feature set is weightedly fused according to the modality weight distribution map to generate the enhanced image feature set.

3. The method according to claim 2, characterized in that The calling of a multi-layer cascade segmentation network to perform hierarchical feature extraction on the enhanced image feature set to obtain a multi-scale anatomical structure feature map includes: Inputting the enhanced image feature set into the first encoder of the multi-layer cascade segmentation network to extract a first-level anatomical structure feature map; Inputting the first-level anatomical structure feature map into a second encoder of the multi-layer cascade segmentation network to extract a second-level anatomical structure feature map; Inputting the second-level anatomical structure feature map into a third encoder of the multi-layer cascade segmentation network to extract a third-level anatomical structure feature map; Multi-scale feature fusion processing is performed on the first-level anatomical structure feature map, the second-level anatomical structure feature map, and the third-level anatomical structure feature map to generate the multi-scale anatomical structure feature map; wherein the multi-scale feature fusion processing includes upsampling the first-level anatomical structure feature map, downsampling the third-level anatomical structure feature map, and adjusting the contribution weights of feature maps of different scales through a channel attention mechanism.

4. The method according to claim 1, wherein The performing of regional boundary optimization processing based on the multi-scale anatomical structure feature map to generate a labeled brain structure segmentation image includes: Performing initial segmentation processing on the multi-scale anatomical structure feature map to generate an initial brain structure segmentation mask; Performing boundary confidence analysis on the initial brain structure segmentation mask to determine a confidence distribution map of the boundary fuzzy area; performing an iterative morphological optimization operation on the initial brain structure segmentation mask based on the confidence distribution map to generate an optimized brain structure segmentation mask; wherein the iterative morphological optimization operation includes performing a dilation-erosion sequence on the boundary fuzzy region and adjusting the size of the dilation kernel based on the confidence distribution map; The optimized brain structure segmentation mask is superimposed and fused with the original image sequence to generate the labeled brain structure segmentation image.

5. The method according to claim 4, characterized in that The performing initial segmentation processing on the multi-scale anatomical structure feature map to generate an initial brain structure segmentation mask includes: Inputting the multi-scale anatomical structure feature map into the fully connected decoder of the multi-layer cascade segmentation network to generate an initial probability distribution map; Performing threshold segmentation processing on the initial probability distribution map to obtain a binary segmentation result; Performing connected domain analysis on the binary segmentation result to remove isolated noise areas and obtain a denoised binary segmentation result; The denoised binary segmentation result is subjected to topological correction based on a preset brain anatomical structure template to generate the initial brain structure segmentation mask.

6. The method according to claim 4, characterized in that The performing boundary confidence analysis on the initial brain structure segmentation mask to determine a confidence distribution map of the boundary fuzzy area includes: performing boundary gradient calculation on the initial brain structure segmentation mask to generate a boundary gradient intensity map; performing feature uncertainty assessment on the multi-scale anatomical structure feature map to generate a feature uncertainty distribution map; Performing weighted fusion processing on the boundary gradient intensity map and the feature uncertainty distribution map to generate an initial confidence map; Performing regional connectivity analysis on the initial confidence map, and screening out continuous regions with confidence levels below a preset threshold as the fuzzy boundary regions; The confidence distribution map is generated based on the spatial distribution of the boundary fuzzy area.

7. The method according to claim 4, characterized in that The performing an iterative morphological optimization operation on the initial brain structure segmentation mask based on the confidence distribution map to generate an optimized brain structure segmentation mask includes: Determining corresponding morphological operation intensity parameters according to the confidence values ​​of each region in the confidence distribution map; performing a morphological closing operation with a fixed kernel size on the target confidence region in the initial brain structure segmentation mask; Performing a morphological opening operation with an adaptive kernel size on the fuzzy boundary area, wherein the kernel size is negatively correlated with the confidence value; Repeating the morphological closing operation and the morphological opening operation until the confidence values ​​of all regions in the confidence distribution map reach a preset optimization threshold; The morphological processing result obtained in the last iteration is used as the optimized brain structure segmentation mask.

8. The method according to claim 4, characterized in that The step of superimposing and fusing the optimized brain structure segmentation mask with the original image sequence to generate the labeled brain structure segmentation image comprises: performing edge smoothing on the optimized brain structure segmentation mask to generate a smooth segmentation contour; Performing transparency blending processing on the smooth segmentation contour and the grayscale value of the original image sequence to generate a preliminary annotated image; performing color coding processing on boundary regions of different anatomical structures in the preliminary annotated image to generate a pseudo-color annotation layer with distinguishability; spatially aligning and superimposing the pseudo-color annotation layer with the original image sequence to generate the annotated brain structure segmentation image; The step of performing color coding on boundary regions of different anatomical structures in the preliminary annotated image to generate a pseudo-color annotation layer with distinguishability includes: Assign unique color labels to different anatomical regions based on preset brain anatomical structure classification rules; Performing category matching processing on each connected region in the optimized brain structure segmentation mask to determine the corresponding target color identification; The boundary pixels of the connected area are subjected to gradient color filling processing based on the target color identifier to generate the pseudo color annotation layer; wherein the color saturation of the gradient color filling processing is positively correlated with the distance from the boundary pixel to the center of the segmentation contour.

9. The method according to claim 8, characterized in that The step of spatially aligning and superimposing the pseudo-color annotation layer with the original image sequence to generate the annotated brain structure segmentation image includes: Performing normalized coordinate transformation processing on the pixel coordinates of the pseudo-color annotated layer to generate a standardized coordinate mapping table aligned with the physical coordinate system of the original image sequence; wherein the normalized coordinate transformation processing includes mapping the pixel positions of the pseudo-color annotated layer to the inter-slice resolution ratio space of the scanning layer of the original image sequence and compensating for inter-slice displacement errors using a bilinear interpolation algorithm; Performing spatial resampling processing on the color channels of the pseudo-color annotated layer according to the standardized coordinate mapping table to generate a resampled pseudo-color layer that completely matches the pixel spatial distribution of the original image sequence; wherein the spatial resampling processing uses an isotropic interpolation kernel function, and the kernel width parameter of the isotropic interpolation kernel function is dimensionally consistent with the pixel spacing within the layer of the original image sequence; Inputting the resampled pseudo-color layer into a transparency blending channel, performing a multi-channel pixel-level overlay operation to generate a mixed annotated image; wherein the overlay operation includes performing a linear brightness-preserving transformation on the grayscale values ​​of the original image sequence and performing dynamic range compression on the RGB color components of the resampled pseudo-color layer so that the pixel values ​​after overlay do not exceed a preset maximum display dynamic range; Performing edge sharpening and anti-aliasing processing on the mixed annotated image to generate an intermediate annotated image with enhanced boundaries; wherein the edge sharpening processing uses an anisotropic diffusion filter algorithm, and the gradient threshold parameter of the anisotropic diffusion filter algorithm matches the grayscale gradient distribution statistic of the original image sequence; performing adaptive contrast adjustment processing based on the boundary intensity distribution of the intermediate annotated image to generate a final displayed brain structure segmentation image; wherein the adaptive contrast adjustment processing includes performing a histogram equalization operation on low-contrast areas and performing nonlinear gamma correction on high-gradient areas, wherein the correction coefficient of the nonlinear gamma correction is negatively correlated with the local gradient amplitude; Color space verification processing is performed on the final displayed brain structure segmentation image to ensure that the chromaticity channel of the pseudo-color annotation layer and the luminance channel of the original image sequence have no overlapping conflicts within the display color gamut. The color space verification processing includes converting the RGB color space to YUV space and then detecting the orthogonality index of the luminance component and the chromaticity component. If an overlapping area is detected, the chromaticity coordinates are remapped to a preset safe palette range.

10. An image annotation system, characterized in that: include: processor; A storage device having a computer program stored thereon, wherein when the computer program is executed by the processor, the processor implements the image annotation method for brain MRI image segmentation according to any one of claims 1 to 9.

Images