Intelligent staging method and system for prostate cancer based on multi-modal image fusion

Abstract

The present application relates to the medical image diagnosis technical field, specifically to a prostate cancer intelligent staging method and system based on multi-modal image fusion, comprising the following steps: acquiring diffusion images and separating sequences, comparing signal changes of sequence differences, extracting response blocks to locate turning points, tracking boundaries to extract texture offsets, fusing edge matching signal types, and generating prostate cancer intelligent staging images. In the present application, the signal intensity contrast between multi-direction diffusion images is used to construct the change trend, the continuous image track is extracted by combining the difference area mapping and the path direction, the expression ability of the dynamic structure in the image is enhanced, the key points of change are located according to the signal turning characteristics, the structure extraction accuracy is strengthened by combining the boundary tracking and the texture offset identification, the area label division and filling are completed by fusing the signal arrangement and the image category, the coherence of the lesion recognition and the objective accuracy of the labeling are improved, and the positioning ability of the structure evolution process and the discrimination stability of the image characteristics are enhanced.

Description

Technical Field

[0001] This invention relates to the field of medical imaging diagnostic technology, and in particular to a method and system for intelligent staging of prostate cancer based on multimodal image fusion. Background Technology

[0002] The field of medical imaging diagnostics utilizes radiology, ultrasound, magnetic resonance imaging (MRI), computed tomography (CT), nuclear medicine imaging, and other medical imaging techniques to acquire image data of internal structures or lesions in the human body. These images are then analyzed manually or with computer assistance to identify, segment, and diagnose abnormalities, determining their location, shape, size, and the extent of lesion spread. This provides a basis for disease diagnosis, staging, and treatment planning. Traditional intelligent staging methods for prostate cancer involve using single-modal medical imaging (such as ultrasound or MRI) to acquire images of the prostate. Radiologists manually observe and identify abnormal areas, and the stage of the cancer is determined based on tumor size, invasion of surrounding structures, suspected metastasis, and clinical and pathological examination results. This method relies heavily on physician experience, visual image assessment, and comprehensive analysis of traditional pathological examinations and clinical indicators.

[0003] Existing technologies rely on single-modal imaging as a basis for judgment. The tissue boundaries in the images are not clear enough, and the signal levels are singular, making it difficult to accurately identify continuous changes in local structures. The differences between tumors and surrounding tissues in the images are unstable and easily interfered with by factors such as gray-scale distribution and structural overlap, resulting in blurred boundaries of response areas and one-sided extraction of morphological features. In the process of manual analysis, there is a lack of quantitative basis for image structure, and the division of regions relies on experience and judgment, which makes it difficult to unify standards. This leads to significant subjective influence on staging results, and there is a lack of ability to track the changing trends of potential lesions in the images. This limits the depth and breadth of identification and analysis of the evolution path of structural abnormalities, affecting the continuity and accuracy of diagnosis. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a smart staging method for prostate cancer based on multimodal image fusion, comprising the following steps: S1: Acquire magnetic resonance diffusion images, separate image sequences according to direction identifiers, perform sequence difference and trend comparison on signals at the same tissue location, merge image mapping difference regions, and generate diffusion direction superimposed images; S2: Compare the pixels of the superimposed image of the diffusion direction with those of the magnetic resonance diffusion image, extract the continuously distributed response blocks, delineate the change clusters along the axis and connect and identify them, draw the closed structure region, and generate the prostate fusion response structure image. S3: Call the path of the prostate fusion response structure image, locate the signal arrangement of the corresponding point, analyze the signal curve turning point in the differential image sequence, label the image category and source of the turning point, and generate the path modal signal breakpoint image; S4: Based on the turning points in the path modal signal breakpoint image, trace the closed boundary in the prostate fusion response structure image, extract the overlapping part of texture change and signal direction shift, and generate a lesion edge blurred region extraction image; S5: Combine the edge structure of the image extracted from the blurred area of ​​the lesion edge with the image category in the path mode signal breakpoint image, fill the closed area of ​​the entire image and match the signal type to generate a smart staging image for prostate cancer.

[0005] As a further aspect of the present invention, the diffusion direction overlay image includes anisotropic signal distribution, directional intensity comparison results, difference region fusion map, and continuous direction stitched image; the prostate fusion response structure image includes continuous response blocks, change cluster distribution, connected region boundaries, and closed structure graphics; the path modal signal breakpoint image includes signal turning points, path direction identifiers, signal trend categories, and image number labels; the lesion edge blurred region extraction image includes texture change boundaries, direction offset regions, boundary overlap structures, and artifact removal patches; and the prostate cancer intelligent staging image includes image category partitioning, closed structure filling, signal type matching results, and region label images.

[0006] As a further aspect of the present invention, the continuously distributed response blocks refer to a set of pixel regions in the diffusion direction superimposed image that have a consistent signal response and are spatially continuous.

[0007] As a further aspect of the present invention, the analysis of signal curve inflection points in the differential image sequence refers to identifying the locations of changes and trend reversals by comparing the signal intensity curves at the same location in images with different diffusion directions.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Acquire the diffused images collected by the magnetic resonance imaging device, and divide the sequence into subsets according to the orientation markers in the images. The images within the subsets are classified and sorted according to orientation to generate an orientation-grouped image sequence. S102: Based on the directional grouped image sequence, extract the signal intensity value of the same tissue location, construct a smooth function model in the directional domain using cubic spline interpolation, calculate the signal intensity value in the unsampled direction, obtain the trend of signal change with direction, and generate a tissue signal change matrix; S103: Based on the trend of difference changes in the tissue signal change matrix, extract the region that exceeds the directional difference threshold, normalize and enhance the intensity value of the signal change region in the directional image to obtain a directional enhancement image of the same size, and then generate a diffusion direction superimposed image based on the weighted fusion result of the directional image.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Obtain the superimposed image of the diffusion direction and the magnetic resonance diffusion image, perform pixel-by-pixel registration and comparison between the two, extract the regions in the superimposed image where the pixel intensity of the magnetic resonance diffusion image changes, and extract the pixel block regions that are continuously distributed according to the continuity of image coordinates to generate a set of continuous response blocks. S202: Based on the continuous response block set, retrieve the position sequence of the response blocks along the image axis, perform clustering processing on the pixel distance and axis projection angle between adjacent response blocks, filter the region combinations that form a changing trend in the axial direction, and generate a changing cluster position index group. S203: Based on the changed cluster location index group, the connectivity of the pixel structure within the location area is judged, areas where the pixel connectivity does not meet the contiguous standard are filtered out, and the remaining area boundaries are closed and drawn and then encoded into an image frame format to generate a prostate fusion response structure image.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Call the path direction of the closed region in the prostate fusion response structure image, retrieve the image coordinates on the path line, and extract the signal value of the corresponding position in the differential image sequence to generate a path signal sequence group; S302: Based on the path signal sequence group, calculate the amplitude difference between adjacent signal points, determine whether it exceeds the amplitude difference reference value, filter the points where the curve changes inflection point, and map the corresponding image category according to the preset interval where the amplitude difference is located to obtain the signal inflection index set. S303: Based on the signal turning index set, the turning point position is bound to the image category information, and an image annotation data frame is constructed. After superimposing the path coordinates, the image is encoded to generate a path modal signal breakpoint image.

[0011] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the turning point position in the path modal signal breakpoint image, call the closed boundary in the prostate fusion response structure image, continuously track the boundary line segment of the turning point, extract the boundary fragments associated with signal changes in all closed paths, and generate a signal offset boundary set. S402: Based on the signal offset boundary set, retrieve the corresponding image texture change region in the original image, extract pixel blocks from the overlapping part of the boundary and texture, construct an independent image block sequence, and aggregate them according to the block number and path position to generate a texture boundary overlapping block group. S403: Based on the texture boundary overlay block group, identify the attached label information, perform artifact removal operations in sequence, and output image frames after re-encoding the region boundaries to generate an image of the blurred lesion edge region.

[0012] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: The image is called to extract the edge structure of the stripped area in the blurred area of ​​the lesion edge, and the boundary line is pixel tracked by the image contour extraction algorithm to extract the set of closed structure edges and record the coordinate mapping relationship to generate the image edge structure set. S502: Based on the image edge structure set and the image category corresponding to the turning point in the path mode signal breakpoint image, the image category information is used as the classification basis for the region division line. The edge-enclosed regions are grouped by category, and region closure processing is performed to generate an image category closed partition map. S503: Call the number information of the region in the closed partition map of the image category, match the corresponding signal curve type inside the region, perform a filling mark operation on the partition, and output image frame data to generate a prostate cancer intelligent staging image.

[0013] A prostate cancer intelligent staging system based on multimodal image fusion includes: The directional fusion image generation module is used to achieve S1: acquiring magnetic resonance diffusion images, separating image sequences according to directional identifiers, performing sequence difference and trend comparison on signals at the same tissue location, merging image mapping difference regions, and generating diffusion direction superimposed images; The response structure extraction module is used to implement S2: compare the pixels of the diffusion direction superimposed image with the magnetic resonance diffusion image, extract the continuously distributed response blocks, delineate the change clusters along the axis and connect and identify them, draw the closed structure region, and generate the prostate fusion response structure image; The path breakpoint identification module is used to implement S3: call the path direction of the prostate fusion response structure image, locate the corresponding signal arrangement, analyze the signal curve turning point in the differential image sequence, label the image category and source of the turning point, and generate the path modal signal breakpoint image; The blurred region stripping module is used to implement S4: based on the turning point in the path modal signal breakpoint image, trace the closed boundary in the prostate fusion response structure image, extract the overlapping part of texture change and signal direction offset, and generate a blurred region extraction image of the lesion edge; The intelligent staging image generation module is used to implement S5: combining the edge structure of the image extracted from the blurred area of ​​the lesion edge with the image category in the path mode signal breakpoint image, filling the closed area of ​​the entire image and matching the signal type to generate an intelligent staging image for prostate cancer.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, a trend of change is constructed by comparing the signal intensity between multi-directional diffusion images. By combining differential region mapping and path direction extraction, the expressive ability of dynamic structures in the image is enhanced. Key points of change are located based on signal turning features. The accuracy of structure extraction is enhanced by combining boundary tracking and texture offset recognition. Region labeling and filling are completed by fusing signal arrangement and image category. This improves the coherence of lesion identification and the objective accuracy of labeling, and enhances the ability to locate the structural evolution process and the discrimination stability of image features. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides an intelligent staging method for prostate cancer based on multimodal image fusion, comprising the following steps: S1: Acquire the diffusion image acquired by the magnetic resonance imaging device, separate the image sequence according to the direction identifier, perform sequence difference processing on the signal intensity of the same tissue location, compare the change trend according to the direction, reconstruct the image of the difference region by image merging, stitch the direction to form a continuous image, and generate a diffusion direction superimposed image. S2: Compare the image content in the diffusion direction superimposed image with the magnetic resonance diffusion image pixel by pixel, extract the response blocks that are continuously distributed in the image, delineate the change clusters along the image axis, perform connectivity recognition processing on the regions in the change clusters, and uniformly draw them as closed structure regions to generate a prostate fusion response structure image. S3: Call the path direction of the closed region in the prostate fusion response structure image, locate the signal arrangement of the corresponding points on the path line, compare the trend changes of the signal curve in the differential image sequence, extract the image category where the curve turns, number the points where the turning point appears and label the source, and generate the path modal signal breakpoint image. S4: Based on the turning points in the path modal signal breakpoint image, call the closed boundary in the prostate fusion response structure image, perform continuous tracking processing on the boundary line segments, extract the overlapping part of the image texture change area and the boundary where the signal direction is offset into independent blocks, perform label removal operation on the block area, and generate the lesion edge blurred area extraction image. S5: Extract the image edge structure of the stripped area in the image by calling the blurred area of ​​the lesion edge, combine it with the image category already labeled in the path mode signal breakpoint image, use the image category as the dividing line, fill the closed area of ​​the whole image and match the signal type, output the image label map after the region division, and generate a smart staging image for prostate cancer.

[0023] The diffusion direction overlay image includes heterogeneous signal distribution, directional intensity contrast results, difference region fusion map, and continuous direction stitched image; the prostate fusion response structure image includes continuous response blocks, change cluster distribution, connected region boundaries, and closed structure graphics; the path modal signal breakpoint image includes signal turning points, path direction indicators, signal trend categories, and image number labels; the lesion edge blurred region extraction image includes texture change boundaries, directional offset regions, boundary overlap structures, and artifact removal patches; and the prostate cancer intelligent staging image includes image category partitioning, closed structure filling, signal type matching results, and region label images.

[0024] Please see Figure 2 The specific steps of S1 are as follows: S101: Acquire the diffused images collected by the magnetic resonance imaging device, and divide the sequence into subsets according to the orientation markers in the images. The images within the subsets are classified and sorted according to orientation to generate an orientation-grouped image sequence. Images of this type are imported from scanned examinations through an image archiving and communication system. Typically, image data is stored in DICOM format, containing the image itself and its acquisition parameters. First, the b-value corresponding to the image needs to be identified to confirm it is a diffusion image. Then, the diffusion direction encoding information is read as the grouping basis. This direction encoding information is generally stored in the image's metadata field, representing the spatial direction of the diffusion gradient applied to each image acquisition in vector form. Subsequently, all images are categorized according to their direction vectors. This is done by extracting the direction vector information of each image one by one, calculating the angle between this direction vector and the already categorized direction. If the angle is within a preset tolerance range, such as less than 15 degrees, the directions are considered consistent, and the images are grouped accordingly. The image is incorporated into an existing directional classification. If the classification is not satisfied, a new classification is created and the current image is included. After classification, each category of images constitutes a directional subset. Each subset represents an image sequence in a certain diffusion direction. Then, the images within each subset are sorted according to the acquisition time. The acquisition time can be extracted from the timestamp information of the image. To ensure accurate sorting, the shooting time field is used for ascending order. For example, if a group of images were acquired at 10:01:20, 10:01:21, and 10:01:23, they will be sorted in the order from earliest to latest to form an ordered image sequence. Finally, all image subsets that have completed directional classification and sequence sorting are organized and output to form a multi-directional grouped image sequence.

[0025] S102: Based on the directional grouping image sequence, extract the signal intensity value of the same tissue location, use the cubic spline interpolation method to construct a smooth function model in the directional domain, calculate the signal intensity value under the unsampled direction, obtain the trend of signal change with direction, and generate the tissue signal change matrix; To obtain the signal characteristics of tissue under different diffusion directions, it is necessary to extract the signal intensity value of each location point in images of each direction. In practice, a fixed pixel coordinate position is selected in the image, for example, a location coordinate of (100, 150) in an image. This point has a corresponding pixel value in each direction image. By extracting the signal intensity at the same coordinates in these images one by one, a set of directional signal datasets is constructed. Simultaneously, the corresponding direction vector information is recorded for subsequent analysis of signal change trends. Since the actual acquisition directions may not cover all theoretical directions, it is necessary to interpolate the signal variation with direction. A cubic spline interpolation method is used to construct a smooth function model, which is derived from the discrete directional signals. To measure the signal intensity in the unsampled direction, the target direction distribution needs to be determined before interpolation. A spherical coordinate system can be used to convert the direction vector into angular form. Then, the angular step size of the interpolation points can be set, for example, one interpolation point every 10 degrees, to ensure coverage of the entire spatial direction domain. After that, for each interpolation direction, the cubic spline interpolation method is used to construct a smooth function model in the direction domain, calculate the signal intensity value in the unsampled direction, generate the fitted signal value, and then summarize the interpolation results of all pixels in the corresponding directions and store them in the tissue signal change matrix. This matrix uses image coordinates as an index to record the signal trend information of each point in multiple directions, and fully expresses the spatial characteristics of the signal change with direction in the tissue region.

[0026] S103: Based on the trend of difference changes in the tissue signal change matrix, extract the region that exceeds the directional difference threshold, normalize and enhance the intensity value of the signal change region in the directional image to obtain a directional enhancement image of the same size, and then generate a diffusion direction superimposed image based on the weighted fusion result of the directional image. Further exploration of regions exhibiting significant differences in the directional domain involves analyzing the signal value sequence of each pixel along the directional dimension. First, the signal difference between adjacent directions in this sequence is calculated. For example, the signal intensity values ​​at direction indices k and k+1 are taken, and the magnitude of the difference is determined. This difference is then compared to a preset threshold. The threshold can be set based on the overall signal intensity range; for example, when the image signal intensity range is 0 to 200, the threshold can be set to 20. If the difference exceeds this threshold, the direction is considered to have a significant change, and this direction is marked as a direction of significant change. To further filter for true difference points, the percentage of pixels classified as significant differences across all directions can be statistically analyzed. The number of significantly different directions is determined. If this number exceeds a set threshold, such as 3 directions, the pixel is identified as a significant directional response region. When processing these regions, the signal intensity in the corresponding directional image needs to be extracted and normalized. For example, the original intensity is linearly mapped to the range of 0 to 255. Then, the normalized pixel values ​​in these directional images are superimposed. The intensity values ​​of the signal change region in the directional image are normalized and enhanced to obtain a directional enhancement image of the same size. Subsequently, based on the weighted fusion result of the directional images, a complete diffusion direction superimposed image is formed, realizing the aggregate representation of significant signal regions in different directions.

[0027] Please see Figure 3 The specific steps of S2 are as follows: S201: Obtain the superimposed image of the diffusion direction and the magnetic resonance diffusion image, perform pixel-by-pixel registration and comparison between the two, extract the regions in the superimposed image where the pixel intensity of the magnetic resonance diffusion image changes, and extract the pixel block regions that are coherently distributed based on the continuity of image coordinates to generate a set of continuous response blocks. The two images are aligned in the same spatial coordinate system through registration. Magnetic resonance diffusion images are typically T2-weighted images, characterized by clear anatomical structures. The overlay image is derived from the orientation intensity response image generated in the previous step. To ensure the accuracy of pixel-by-pixel comparison, the two images are first resized to check for consistency. If they are inconsistent, a resampling operation is performed to unify the resolution to 0.5 mm / pixel. Then, image registration is performed, and after unifying the resolution to 0.5 mm / pixel, positional registration is only performed between the base image and the diffusion orientation overlay image in the spatial coordinate plane (intensity comparison is not performed). This ensures that spatial alignment facilitates the mapping of structural regions and contour comparison analysis. A pixel-by-pixel comparison operation is performed on the two images, i.e., traversing each pixel position (x, y), and comparing pixels at the same coordinate position... The intensity values ​​are calculated by difference. If the difference between corresponding pixels in two images is greater than the set change threshold V_th, then the point is considered to have a response change. The threshold V_th is set to 30, that is, if |I1(x,y)-I2(x,y)|≥30, it is determined to be a response pixel. After obtaining all response pixels, connected component extraction is required to statistically analyze the spatial distribution of response pixels. If multiple pixels are arranged continuously on the image coordinates and their interval is within 3 pixels, they are classified as the same region. Connectivity detection is performed using the 4-neighborhood or 8-neighborhood method to extract all pixel block regions that conform to the continuous distribution. If the number of pixels in a certain region is ≥10, the region is considered to constitute a valid response block. The coordinates of its upper left and lower right corners are recorded as the bounding rectangle of the region. Finally, the set of these regions is sorted and output to generate a set of continuous response blocks.

[0028] S202: Based on a continuous set of response blocks, retrieve the position sequence of response blocks along the image axis, perform clustering processing on the pixel distance and axis projection angle between adjacent response blocks, filter the region combinations that form a changing trend in the axial direction, and generate a change cluster position index group. First, the geometric center coordinates of each response block are extracted and projected along the image's axis. The projected coordinates of the response block center along the axis are calculated. Then, these center points are sorted according to their position along the axis. For example, along the horizontal axis, the center coordinates of multiple response blocks are arranged from left to right. The pixel distance D_ij between the centers of two adjacent response blocks and the angle θ_ij between the connecting line and the axis are calculated. Clustering criteria are set: when the distance D_ij between two blocks is less than 20 pixels and the angle θ_ij is less than 10 degrees, the two blocks are considered continuous and are grouped into the same cluster. In the clustering combination, the entire retrieval process uses a traversal method to check whether all block pairs meet the condition. If they do, they are added to the current cluster combination; otherwise, a new combination is created. The minimum cluster combination size is set to 3, meaning that if a combination contains fewer than 3 blocks, it is removed. In addition, distance and angle thresholds can be set based on image resolution. For example, if the image resolution is 0.5mm / pixel, then 20 pixels correspond to 10mm, which is within the acceptable error range for clinical lesions. Finally, all cluster combinations that meet the conditions are numbered, and the index of the response blocks they contain is recorded to form a variable cluster location index group.

[0029] S203: Based on the change cluster location index group, the connectivity of the pixel structure within the location area is judged, the areas where the pixel connectivity does not meet the contiguous standard are filtered out, and the remaining area boundaries are closed and drawn and then encoded into image frame format to generate a prostate fusion response structure image. Further connectivity judgment is needed for the pixel structure within the location area covered by each combination. After extracting all response pixels contained in the combination, a set of region pixels is constructed. Connectivity analysis is performed on all points in the set. The judgment criterion is that the minimum number of pixels in a single connected region is not less than 15 pixels. If multiple response blocks in a combination meet the distance and orientation clustering, but the actual pixels do not form a continuous structure, the combination does not meet the connectivity requirements and is removed. For regions that pass the connectivity judgment, their edge pixels are extracted, and closed boundary curves are constructed. By detecting the boundary contours of pixels with a response value of 1 in the image, all boundary points are connected to form a complete closed loop, and the edge map of the region is drawn. The drawn closed image structure is encoded in the form of image frames. During the encoding process, the image data is serialized into image format data by row and column, and the resolution and pixel format settings are unified. For example, a grayscale image format is used. Each pixel value indicates whether it is in the response structure. If it is, it is set to 255, and if it is not, it is set to 0. Finally, a prostate fusion response structure image is formed.

[0030] Please see Figure 4 The specific steps of S3 are as follows: S301: Call the path direction of the closed region in the prostate fusion response structure image, retrieve the image coordinates on the path line, and extract the signal value of the corresponding position in the differential image sequence to generate a path signal sequence group; First, the boundary information of the closed region is obtained from the image. The boundary usually exists in the form of contour lines. After extracting the contour lines, path fitting is performed according to the image coordinate order. During the path fitting process, the continuity of the contour lines must be maintained to avoid breakpoints or duplicate points affecting subsequent path retrieval. After fitting, the image coordinates of all path points are recorded in sequence. For example, a contour on the path may be composed of image coordinate points (85, 60), (86, 61), (87, 62), etc., forming a continuous sequence of coordinate points. This path coordinate sequence serves as the index basis for subsequent extraction of signal values. Then, based on the acquired differential image sequence, The signal intensity value corresponding to the coordinates of the path point is retrieved in each image. Each position has a different signal value in different images. For example, a certain coordinate point has a value of 135 in the T2-weighted image, 90 in the ADC image, and 110 in the diffusion direction image. These values ​​are extracted and arranged in the order of the images to form a signal vector based on a single coordinate point. The signal vectors of all path coordinate points are collected and organized to obtain a signal sequence containing the entire path. The information of the category of the source image is retained in each signal sequence. Finally, a path signal sequence group is formed for subsequent analysis of the signal fluctuation characteristics on the path.

[0031] S302: Based on the path signal sequence group, calculate the amplitude difference between adjacent signal points, determine whether it exceeds the amplitude difference reference value, filter the points where the curve changes inflection point, and map the corresponding image category according to the preset interval where the amplitude difference is located to obtain the signal inflection index set. For each path point, a point-by-point comparison of signal values ​​in different image modalities is performed. During processing, two adjacent points on the path are selected, namely the current point and its next point. The amplitude difference between the signal values ​​of these two points in the same image category is calculated. The magnitude of this difference is then compared with a predefined amplitude difference benchmark to determine whether an inflection point has been reached. The benchmark value should consider the signal-to-noise ratio of the image and the actual range of tissue signal fluctuations. For example, a benchmark value of 30 is set for T2-weighted images with a signal range of 0 to 200, and a benchmark value of 20 is set for diffusion-related images. This process is repeated to determine the signal values ​​of each path point in different image modalities. If the difference exceeds the range, mark the point as a signal inflection point and record the path index number and corresponding image modality category of the point. For cases with long path lengths, there may be multiple significant signal jump locations between every 10 to 15 points. Therefore, it is necessary to scan the entire path and record all inflection point information to form a complete signal inflection index set. Each element contains the path point number and its change attributes under a certain image modality. For example, recording (point 25, ADC image) indicates that the signal at this point undergoes a sudden change in the ADC image, forming a clear set of identifiers that can be used for subsequent localization analysis.

[0032] S303: Based on the signal turning index set, bind the turning point position with the image category information, construct the image annotation data frame, overlay the path coordinates and perform image encoding to generate the path modal signal breakpoint image; Further steps are needed to bind the image coordinates of each inflection point to its corresponding image category, construct a standardized data structure to record the specific information of the inflection event, including the coordinate position of the path point in the image, the image mode in which the point is located, the signal values ​​before and after the point, and the difference value, etc., and save them in a unified format data frame, with each line representing an inflection point event. After organizing each item in the data frame, image overlay processing is performed. In the original path image, a mark is drawn according to the coordinate position of the inflection point, for example, a small circle or cross mark is drawn at the image coordinates (120, 85), and the corresponding image mode category is identified by color coding, for example, red represents the signal transition point of the T2 image, and green represents the transition point of the ADC image. After marking, the original image and the labeled image are merged and encoded into an image frame, and the encoding format is kept consistent with the original image. An 8-bit or 16-bit grayscale image overlay RGB label layer can be selected. At the same time, the metadata of the image, such as image number, path number, number of labels, etc., is recorded. Finally, a path mode signal breakpoint image containing all signal breakpoint markers is output.

[0033] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the turning point position in the path modal signal breakpoint image, call the closed boundary in the prostate fusion response structure image, continuously track the boundary line segment of the turning point, extract the boundary fragments associated with signal changes in all closed paths, and generate a signal offset boundary set. The closed boundary information stored in the prostate fusion response structure image needs to be called one by one. The closed boundary is a closed path composed of multiple adjacent pixels. First, by reading the image coordinates of the turning point, it is found in the fusion response structure image whether the coordinates are located on any closed boundary path. If a match is found, the position of the turning point in the closed boundary path is taken as the starting point, and line segment tracing is performed along the path. The tracing process adopts the method of sequentially connecting adjacent boundary points. Each time, the adjacent point of the current point is taken as the next tracing node to ensure the connectivity and closure of the path. When the path travels continuously for more than a certain number of pixels, such as a length threshold of 15 pixels, the tracing stops, and the line segment between the starting point and the ending point is cut into a boundary segment. Then, the segment is bound to the corresponding signal change point, and the start and end coordinates of the segment and the closed path number to which it belongs are recorded. At the same time, it is checked whether there are multiple turning point influence areas in the boundary segment. If so, it is classified as a signal change related segment. All boundary segments that meet the conditions are sorted and numbered. Each boundary segment needs to contain attribute fields such as path ID, start and end coordinates, and number of influence signal points. A set structure is constructed through these fields, and finally a signal offset boundary set is formed.

[0034] S402: Based on the signal offset boundary set, retrieve the corresponding image texture change region in the original image, extract pixel blocks from the overlapping part of the boundary and texture, construct an independent image block sequence, and aggregate them according to the block number and path position to generate a texture boundary overlapping block group; The texture features corresponding to each boundary segment in the original image need to be extracted. First, the original image data is read, and the specific pixel range of the region is retrieved in the image space by referring to the coordinate information of the boundary segment. To ensure that the texture information is not truncated, the boundary is extended by 5 pixels inward and outward along the normal direction during extraction, forming a pixel block containing the boundary and the surrounding area. Then, the texture statistics of the pixel block are calculated, including gray mean, variance, texture direction gradient, etc., for subsequent analysis of the actual changes in the image structure at the boundary. After that, the pixel block is saved as an independent image block and assigned a unique number. Then, the image block is bound to its corresponding position on the path, and its path point index, image modality source, boundary segment number and other information are recorded. An index table structure is established, and all image blocks are grouped and sorted according to the path position to form a complete image block sequence. Each block group corresponds to a boundary texture set region on a specific path, and its number information and texture feature attributes are retained. Finally, the combination and summary of texture boundary regions are completed to generate texture boundary superimposed block groups.

[0035] S403: Based on the texture boundary overlay block group, identify the attached label information, perform artifact removal operations in sequence, and output image frames after re-encoding the region boundary to generate the lesion edge blurred region extraction image; It is necessary to identify whether each image block contains attached label information. The label information is usually present in the form of color or binary mask. For example, in medical images, suspected areas are often marked with white or red blocks. First, the image content is read block by block and the pixel value is scanned. The gray level between 240 and 255 is set as the label identifier. If there is a continuous area with a pixel value of 250 in an image block, it is judged as a label area. For the image blocks judged to contain labels, a pixel value replacement operation is performed to reset the pixel value of the label area to the average value of the surrounding background pixels, or directly set it to zero value for stripping. Then, the boundary information of the image block is re-extracted. The edges of the discontinuous areas caused by artifact removal are re-closed. During closure, the broken boundaries are connected point by point according to the edge detection results to ensure the integrity of the contour. After closure, the new boundary area is re-encoded into an image frame. The image frame encoding retains the original image size and pixel structure to ensure that the output image can be used for subsequent processing. Finally, an image of the blurred lesion edge area with clear structural boundaries and no label interference is generated.

[0036] Please see Figure 6 The specific steps of S5 are as follows: S501: Extract the edge structure of the stripped area in the image by calling the blurred area of ​​the lesion edge, perform pixel tracking processing on the boundary line by combining the image contour extraction algorithm, extract the set of closed structure edges, record the coordinate mapping relationship, and generate the image edge structure set; First, the hollow contours formed by the peeled-off label regions in the image are extracted. In the initial stage, the image pixel matrix is ​​read. By scanning the regions with pixel values ​​of 0 or background values, it is determined whether they form a boundary with the surrounding non-background regions. If a continuous pixel block is surrounded by non-zero pixels, it is marked as a boundary structure to be extracted. Then, the boundary pixels of the region are tracked using contour extraction. The tracking rule is to start from a boundary starting point and expand to adjacent pixels point by point according to the image coordinate direction. Points with boundary pixel values ​​in the eight adjacent directions are selected as the next tracking targets. A complete closed contour is obtained by constructing a boundary pixel sequence. During the tracking process, the image coordinates of each pixel are recorded and stored in an array. At the same time, the region number information corresponding to the boundary line segment is bound to its contour sequence and stored. After the boundary tracking is completed, all closed regions are integrated to form an edge structure set. Each edge structure unit contains fields such as contour point sequence, region number, area size, start and end coordinates, etc. At the same time, these edge structures are mapped to the original image position before label peeling and a coordinate transformation table or matrix is ​​established for the subsequent steps to associate the spatial position relationship between the original image and the processed image, and finally, an image edge structure set is generated.

[0037] S502: Based on the image edge structure set and the image category corresponding to the turning point in the path mode signal breakpoint image, the image category information is used as the classification basis for the region division line. The edge-enclosed regions are grouped by category, and region closure processing is performed to generate an image category closed partition map. By combining the image category information corresponding to the inflection points in the path mode signal breakpoint image, the classification of the region corresponding to each edge structure is determined. The process begins by extracting the enclosing region of each edge structure, that is, the pixel block enclosed by the closed contour is considered a complete region. Then, the number of signal inflection points contained within this region is counted, and the image category corresponding to these inflection points is obtained. When multiple image categories appear in the same region, the image category with the highest frequency is prioritized as the classification label for that region. For example, if five inflection points are detected in a region, and three of them are of image category T2, then the region is classified as T2. 2. Image partitioning: After establishing a one-to-one correspondence between the classification results and edge structures, a set merging operation is performed on all edge structures of the same label to form a set of closed regions grouped by image category in the image space. To ensure that each partition has a closed boundary, closure completion processing is performed on all edge structures with non-closed contours. The completion method is to interpolate and fill in the contour between the start and end points or directly connect the start and end points to avoid region recognition failure due to contour discontinuity. After completing all region classification and closure processing, the partition number, category label, and boundary pixel sequence are saved to the structure data, and the final output is the image category closed partition map.

[0038] S503: Call the number information of the region in the closed partition map of the image category, match the corresponding signal curve type inside the region, perform a filling mark operation on the partition, and output image frame data to generate a smart staging image for prostate cancer. Each region is processed individually, its corresponding image coordinate set is read, and the pixel range covered by the region is located in the original image data. The path signal sequence within the region is extracted, and then a curve type matching operation is performed on the signal sequence. Curve type matching is completed by analyzing the trend of the signal changing with the path position. The judgment is based on the length and amplitude of the signal rising, falling, or fluctuating interval. For example, if the path signal in the region shows a monotonically rising trend and the amplitude is greater than the set benchmark value of 30, it is matched as an enhanced curve. If there is a clear inflection point, it is classified as a mutation curve. After the matching is completed, each region is assigned a corresponding curve type label, and then a filling operation is performed. All coordinate points of the region are marked with specific pixel values ​​in the partition map. The filling value is assigned a unique identifier code according to the different curve types. For example, enhanced is 128, mutation is 200, and stable is 64. After the region filling is completed, the image is frame encoded, and the encoding format retains the original image resolution. A grayscale image is output according to the filling value of each pixel, and finally, a smart staging image for prostate cancer is generated.

[0039] Please see Figure 7 A prostate cancer intelligent staging system based on multimodal image fusion includes: The directional fusion image generation module is used to achieve S1: acquiring magnetic resonance diffusion images, separating image sequences according to directional identifiers, performing sequence difference and trend comparison on signals at the same tissue location, merging image mapping difference regions, and generating diffusion direction superimposed images; The response structure extraction module is used to implement S2: compare the pixels of the diffusion direction superimposed image with the magnetic resonance diffusion image, extract the continuously distributed response blocks, delineate the change clusters along the axis and connect and identify them, draw the closed structure region, and generate the prostate fusion response structure image; The path breakpoint identification module is used to implement S3: call the path direction of the prostate fusion response structure image, locate the signal arrangement of the corresponding point, analyze the signal curve turning point in the differential image sequence, label the image category and source of the turning point, and generate path modal signal breakpoint images; The blurred region stripping module is used to implement S4: based on the turning points in the path modal signal breakpoint image, trace the closed boundary in the prostate fusion response structure image, extract the overlapping part of texture change and signal direction offset, and generate a blurred region extraction image of the lesion edge; The intelligent staging image generation module is used to implement S5: combining the edge structure of the image with the path modal signal breakpoint image extracted from the blurred area of ​​the lesion edge, the image category is filled in the closed area of ​​the whole image and the signal type is matched to generate an intelligent staging image for prostate cancer.

[0040] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A smart staging method for prostate cancer based on multimodal image fusion, characterized in that, Includes the following steps: S1: Acquire magnetic resonance diffusion images, separate image sequences according to direction identifiers, perform sequence difference and trend comparison on signals at the same tissue location, merge image mapping difference regions, and generate diffusion direction superimposed images; S2: Compare the pixels of the superimposed image of the diffusion direction with those of the magnetic resonance diffusion image, extract the continuously distributed response blocks, delineate the change clusters along the axis and connect and identify them, draw the closed structure region, and generate the prostate fusion response structure image. S3: Call the path of the prostate fusion response structure image, locate the signal arrangement of the corresponding point, analyze the signal curve turning point in the differential image sequence, label the image category and source of the turning point, and generate the path modal signal breakpoint image; S4: Based on the turning points in the path modal signal breakpoint image, trace the closed boundary in the prostate fusion response structure image, extract the overlapping part of texture change and signal direction shift, and generate a lesion edge blurred region extraction image; S5: Combine the edge structure of the image extracted from the blurred area of ​​the lesion edge with the image category in the path mode signal breakpoint image, fill the closed area of ​​the entire image and match the signal type to generate a smart staging image for prostate cancer.

2. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The diffusion direction overlay image includes anisotropic signal distribution, directional intensity comparison results, difference region fusion map, and continuous direction stitched image. The prostate fusion response structure image includes continuous response blocks, change cluster distribution, connected region boundaries, and closed structure graphics. The path modal signal breakpoint image includes signal turning points, path direction indicators, signal trend categories, and image number labels. The lesion edge blurred region extraction image includes texture change boundaries, direction offset regions, boundary overlap structures, and artifact removal patches. The prostate cancer intelligent staging image includes image category partitioning, closed structure filling, signal type matching results, and region label images.

3. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The continuously distributed response blocks refer to a set of pixel regions in a diffusion-direction superimposed image that have a consistent signal response and are spatially continuous.

4. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The analysis of signal curve inflection points in the differential image sequence refers to identifying the locations of changes and trend reversals by comparing the signal intensity curves at the same location in images with different diffusion directions.

5. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Acquire the diffused images collected by the magnetic resonance imaging device, and divide the sequence into subsets according to the orientation markers in the images. The images within the subsets are classified and sorted according to orientation to generate an orientation-grouped image sequence. S102: Based on the directional grouped image sequence, extract the signal intensity value of the same tissue location, construct a smooth function model in the directional domain using cubic spline interpolation, calculate the signal intensity value in the unsampled direction, obtain the trend of signal change with direction, and generate a tissue signal change matrix; S103: Based on the trend of difference changes in the tissue signal change matrix, extract the region that exceeds the set directional difference threshold, normalize and enhance the intensity value of the signal change region in the directional image to obtain a directional enhancement image of the same size, and then generate a diffusion direction superimposed image based on the weighted fusion result of the directional image.

6. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Obtain the superimposed image of the diffusion direction and the magnetic resonance diffusion image, perform pixel-by-pixel registration and comparison between the two, extract the regions in the superimposed image where the pixel intensity of the magnetic resonance diffusion image changes, and extract the pixel block regions that are continuously distributed according to the continuity of image coordinates to generate a set of continuous response blocks. S202: Based on the continuous response block set, retrieve the position sequence of the response blocks along the image axis, perform clustering processing on the pixel distance and axis projection angle between adjacent response blocks, filter the region combinations that form a changing trend in the axial direction, and generate a changing cluster position index group. S203: Based on the changed cluster location index group, the connectivity of the pixel structure within the location area is judged, areas where the pixel connectivity does not meet the contiguous standard are filtered out, and the remaining area boundaries are closed and drawn and then encoded into an image frame format to generate a prostate fusion response structure image.

7. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Call the path direction of the closed region in the prostate fusion response structure image, retrieve the image coordinates on the path line, and extract the signal value of the corresponding position in the differential image sequence to generate a path signal sequence group; S302: Based on the path signal sequence group, calculate the amplitude difference between adjacent signal points, determine whether it exceeds the amplitude difference reference value, filter the points where the curve changes inflection point, and map the corresponding image category according to the preset interval where the amplitude difference is located to obtain the signal inflection index set. S303: Based on the signal turning index set, the turning point position is bound to the image category information, and an image annotation data frame is constructed. After superimposing the path coordinates, the image is encoded to generate a path modal signal breakpoint image.

8. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the turning point position in the path modal signal breakpoint image, call the closed boundary in the prostate fusion response structure image, continuously track the boundary line segment of the turning point, extract the boundary fragments associated with signal changes in all closed paths, and generate a signal offset boundary set. S402: Based on the signal offset boundary set, retrieve the corresponding image texture change region in the original image, extract pixel blocks from the overlapping part of the boundary and texture, construct an independent image block sequence, and aggregate them according to the block number and path position to generate a texture boundary overlapping block group. S403: Based on the texture boundary overlay block group, identify the attached label information, perform artifact removal operations in sequence, and output image frames after re-encoding the region boundaries to generate an image of the blurred lesion edge region.

9. The intelligent staging method for prostate cancer based on multimodal image fusion according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: The image is called to extract the edge structure of the stripped area in the blurred area of ​​the lesion edge, and the boundary line is pixel tracked by the image contour extraction algorithm to extract the set of closed structure edges and record the coordinate mapping relationship to generate the image edge structure set. S502: Based on the image edge structure set and the image category corresponding to the turning point in the path mode signal breakpoint image, the image category information is used as the classification basis for the region division line. The edge-enclosed regions are grouped by category, and region closure processing is performed to generate an image category closed partition map. S503: Call the number information of the region in the closed partition map of the image category, match the corresponding signal curve type inside the region, perform a filling mark operation on the partition, and output image frame data to generate a prostate cancer intelligent staging image.

10. A prostate cancer intelligent staging system based on multimodal image fusion, characterized in that, The system is used to implement the intelligent staging method for prostate cancer based on multimodal image fusion as described in any one of claims 1-9, and the system comprises: The directional fusion image generation module is used to achieve S1: acquiring magnetic resonance diffusion images, separating image sequences according to directional identifiers, performing sequence difference and trend comparison on signals at the same tissue location, merging image mapping difference regions, and generating diffusion direction superimposed images; The response structure extraction module is used to implement S2: compare the pixels of the diffusion direction superimposed image with the magnetic resonance diffusion image, extract the continuously distributed response blocks, delineate the change clusters along the axis and connect and identify them, draw the closed structure region, and generate the prostate fusion response structure image; The path breakpoint identification module is used to implement S3: call the path direction of the prostate fusion response structure image, locate the corresponding signal arrangement, analyze the signal curve turning point in the differential image sequence, label the image category and source of the turning point, and generate the path modal signal breakpoint image; The blurred region stripping module is used to implement S4: based on the turning point in the path modal signal breakpoint image, trace the closed boundary in the prostate fusion response structure image, extract the overlapping part of texture change and signal direction offset, and generate a blurred region extraction image of the lesion edge; The intelligent staging image generation module is used to implement S5: combining the edge structure of the image extracted from the blurred area of ​​the lesion edge with the image category in the path mode signal breakpoint image, filling the closed area of ​​the entire image and matching the signal type to generate an intelligent staging image for prostate cancer.

Images