Breast thyroid micro-lesion accurate positioning method and system based on ai multi-modal image

Abstract

This invention relates to the field of medical imaging diagnostic technology, specifically to a method and system for precise localization of small lesions in the breast and thyroid glands based on AI multimodal imaging. The method includes the following steps: acquiring ultrasound, X-ray, and magnetic resonance images of the breast and thyroid glands; extracting anatomical spatial location and imaging direction; calculating and mapping multimodal pixel spatial offsets; calculating scale consistency and normalizing response intensity based on co-located pixels; comparing multimodal intensity differences to assess consistency and form a scoring distribution; combining the score and intensity to adjust conditional probabilities; calculating joint probabilities and selecting spatial aggregates to locate small lesions. In this invention, by uniformly calibrating the spatial location and imaging direction of multimodal images, precise correspondence between different images at the same anatomical location is achieved. Multimodal response intensities are fused and consistency assessment is performed. Combining probability weight adjustment and spatial aggregation weakens the influence of noise in single images, enhances the common expression of small abnormal areas, and improves the stability and consistency of lesion localization.

Description

Technical Field

[0001] This invention relates to the field of medical imaging diagnostic technology, and in particular to a method and system for precise localization of small lesions in the breast and thyroid glands based on AI multimodal imaging. Background Technology

[0002] The field of medical imaging diagnostic technology involves technologies that utilize medical imaging equipment to acquire and analyze human tissue structure and pathological information. This typically includes methods for acquiring medical imaging data, displaying and recording image data, identifying and locating anatomical structures and lesion areas in images, and processing procedures to assist doctors in making diagnostic decisions. Common imaging formats include ultrasound imaging, magnetic resonance imaging, and radiographic imaging. The core objective is to identify and locate abnormal tissues based on grayscale morphological boundaries and spatial relationships in images, thereby supporting disease screening, diagnosis, and follow-up.

[0003] Among them, the traditional method and system for precise localization of small lesions in the breast and thyroid refers to the localization of lesions that are small in size and have indistinct shape in breast or thyroid imaging. It is usually achieved by acquiring a single-modal medical image and then having the doctor observe the image frame by frame and manually delineate the outline of the lesion. Alternatively, the image can be segmented into regions by setting a grayscale threshold and the lesion can be judged based on the difference in brightness, shape and edge continuity between the lesion and the surrounding tissue. At the same time, in the case of multiple image sources, different images can be manually compared by time or space correspondence to determine the location of the lesion. The technical aspects mainly revolve around the determination of the identification range and location marking of suspected lesion areas in the image.

[0004] Current methods for locating small lesions in the breast and thyroid mainly rely on a single image source or manual comparison of multiple images. The image acquisition and interpretation process is significantly affected by the operator's experience. Lesion identification is based on grayscale thresholds and morphological boundaries, which are not stable enough when dealing with small or low-contrast areas. There is a lack of unified spatial reference between different modal images. Cross-image localization relies on manual comparison and is prone to deviation. The repeatability of lesion location annotation is low, resulting in insufficient consistency of localization results and limited reliability of clinical follow-up and comparative analysis. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a method for precise localization of small lesions in the breast and thyroid glands based on AI multimodal imaging, comprising the following steps: S1: Collect ultrasound, X-ray and magnetic resonance imaging data of breast and thyroid glands, extract corresponding spatial locations and imaging directions, construct a unified spatial coordinate system and unify its dimensions, calculate the spatial distance offset of corresponding pixels in multimodal images and perform coordinate mapping, and generate spatial calibration data. S2: Call the spatial calibration data to obtain the pixel calculation scale consistency of the multimodal image at the same spatial location, extract the multimodal response intensity and normalize the mapping to generate a multimodal response intensity set; S3: Based on the multimodal response intensity set, calculate the difference in multiple modal response intensity at the same spatial location and determine the consistency. Calculate the consistency score for multiple spatial locations and generate a response consistency score distribution. S4: Based on the response consistency score distribution, extract the consistency scores and response intensities of multiple spatial locations and calculate the conditional probabilities. Adjust the weights of the multimodal conditional probabilities to generate a multimodal probability distribution. S5: Based on the multimodal probability distribution, obtain the joint conditional probability of multiple spatial locations, filter spatial locations that exceed the preset joint probability threshold and aggregate them spatially, calculate the center coordinate position, and generate the localization result of the small lesion.

[0006] As a further aspect of the present invention, the spatial calibration data includes anatomical region spatial coordinates, imaging direction vector, and pixel spatial offset; the multimodal response intensity set includes ultrasound response intensity, X-ray response intensity, magnetic resonance response intensity, and uniform scale mapping value; the response consistency score distribution includes spatial location consistency score value, score spatial index sequence, and score value distribution range; the multimodal probability distribution includes ultrasound conditional probability, X-ray conditional probability, and magnetic resonance conditional probability; and the microlesion localization result includes lesion center coordinates, aggregated spatial range identifier, and spatial location index number.

[0007] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect ultrasound, X-ray and magnetic resonance imaging data for breast and thyroid examinations, detect anatomical region markers in multimodal image frames, extract spatial location information, imaging direction information and acquisition sequence information, perform spatial registration, and generate an anatomical spatial index set. S102: Based on the anatomical space index set, calculate the three-dimensional coordinate difference vector of the corresponding pixel point according to the pixel coordinate values ​​of the multimodal image under the same anatomical index, calculate the projection component of the difference vector on the imaging direction vector and make a direction consistency judgment, and obtain the pixel spatial distance offset set. S103: Based on the set of pixel spatial distance offsets, call the origin parameters and axial scale parameters of the multimodal image coordinate system, perform coordinate mapping transformation on the offsets, write the mapped coordinates into a unified spatial coordinate system, and generate spatial calibration data.

[0008] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Call the spatial calibration data, locate the same spatial position according to the unified spatial coordinate index, collect the corresponding pixel values ​​of ultrasound image data, X-ray image data and magnetic resonance image data for coordinate matching, combine the multimodal pixel values ​​under the same spatial index condition, and obtain cross-modal pixel value groups. S202: Based on the cross-modal pixel value group, calculate the probability distribution ratio of the multimodal pixel value sequence, call the probability distribution ratio to perform information entropy numerical calculation, and use the multimodal information entropy value as the feature weight coefficient to perform weighted fusion operation on the corresponding pixel value to obtain the multimodal response intensity; S203: Based on the multimodal response intensity, collect all response values ​​corresponding to spatial coordinates, calculate the maximum and minimum values ​​within the response value set and perform linear interval mapping operation, normalize the multimodal response values, and arrange them according to the spatial coordinate index order to generate a multimodal response intensity set.

[0009] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the multimodal response intensity set, retrieve the multimodal response intensity numerical sequence corresponding to the same spatial location according to the spatial coordinate index order, perform numerical difference operation on multiple modal response intensities in the sequence, and aggregate the differences obtained under the same spatial coordinate to obtain a multimodal response difference set; S302: Based on the multimodal response difference set and combined with the preset consistency judgment benchmark value, perform item-by-item comparison between multiple differences in the difference set and the consistency judgment benchmark value, mark the difference state that meets the judgment condition, and statistically analyze the difference states under the same spatial coordinates to obtain the consistency judgment state sequence. S303: For the consistency determination state sequence, perform weighted summation calculation on the state code values ​​according to the spatial coordinate index order, analyze the consistency score values ​​of the spatial coordinates, and serialize and arrange the score values ​​corresponding to all spatial coordinates to generate the response consistency score distribution.

[0010] As a further aspect of the present invention, the consistency judgment benchmark value is determined by performing statistics on the multimodal response difference set to obtain the numerical distribution range of the multimodal response difference under the same spatial coordinates, calculating the mean and standard deviation of the difference sequence, and performing scale conversion on the standard deviation according to a preset scaling factor, and then determining the conversion result by linearly combining it with the mean.

[0011] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the response consistency score distribution, retrieve the score values ​​corresponding to multiple spatial coordinates in the order of spatial coordinate index, and simultaneously obtain the multimodal response intensity at the same spatial coordinate position. Perform joint numerical mapping on the score values ​​and multimodal response intensity to obtain the joint data structure of score intensity. S402: Based on the joint data structure of the scoring intensity, for each spatial coordinate position, the scoring value is used as a condition item, and conditional probability calculation is performed in combination with the corresponding multimodal response intensity value. The aggregation and arrangement are completed according to the modal index order to obtain the multimodal conditional probability set. S403: For the multimodal conditional probability set, call the preset modal weight parameter set, perform weighted adjustment operation on the conditional probability values ​​corresponding to multiple modes, and perform normalization constraint processing on the weighted probability values ​​under the same spatial coordinates to generate a multimodal probability distribution.

[0012] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the multimodal probability distribution, obtain the modal probability values ​​corresponding to multiple spatial coordinates, perform product merging on the multimodal probability values ​​under the same spatial coordinates, and perform normalization constraint processing on the merging result to form a probability value sequence under spatial coordinate association, and generate a spatial joint conditional probability set. S502: Based on the spatial joint conditional probability set, call the preset joint probability threshold, perform threshold comparison on the joint probability values ​​corresponding to multiple spatial coordinates, mark the spatial coordinates that exceed the joint probability threshold, and perform coordinate aggregation operation based on spatial adjacency relationship to establish a spatial aggregated coordinate set; S503: For the aforementioned spatial aggregated coordinate set, obtain the spatial coordinate components in the multi-aggregated coordinate set, call the coordinate component values ​​to perform mean calculation, perform centralized calculation on the multi-dimensional coordinate values, obtain the center coordinate position corresponding to the aggregated region, and generate the micro lesion localization result.

[0013] As a further aspect of the present invention, the joint probability threshold is set by obtaining the joint probability values ​​corresponding to multiple spatial coordinates in the spatial joint conditional probability set, performing statistical distribution sorting on the numerical sequence, calculating the mean term of the joint probability values ​​and combining it with the standard deviation that is a multiple of 3.

[0014] A precise localization system for small lesions in the breast and thyroid glands based on AI multimodal imaging includes: The data analysis module collects ultrasound, X-ray, and magnetic resonance imaging data of the breast and thyroid glands, extracts the spatial location and imaging direction corresponding to the anatomical region, calculates the spatial distance offset of the corresponding pixel points in the multimodal images and performs coordinate mapping, generates spatial calibration data and transmits it to the intensity mapping module. The intensity mapping module calls the spatial calibration data, obtains the pixel values ​​of the multimodal image at the same spatial location, performs scale consistency calculation, extracts the multimodal response intensity and performs normalized mapping, generates a multimodal response intensity set and passes it to the consistency evaluation module. The consistency evaluation module calculates the difference between multiple modal response intensities at the same spatial location based on the multimodal response intensity set and makes a consistency judgment. It calculates the consistency score for multiple spatial locations, generates a response consistency score distribution, and passes it to the probability modeling module. The probability modeling module extracts the consistency scores and response intensities of multiple spatial locations and calculates the conditional probabilities based on the response consistency score distribution. It then adjusts the weights of the multimodal conditional probabilities, generates a multimodal probability distribution, and transmits it to the lesion localization module. The lesion localization module obtains the joint conditional probability of multiple spatial locations based on the multimodal probability distribution, filters spatial locations that exceed a preset joint probability threshold and aggregates them spatially, calculates the center coordinate position, and generates the localization result of the small lesion.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by uniformly calibrating the spatial location and imaging direction of multimodal images, precise correspondence between different images at the same anatomical location is achieved. Furthermore, multimodal response intensity information is integrated at the spatial level to conduct a consistent assessment of the differences in multi-source responses at the same location. By combining probability weight adjustment and spatial aggregation, the interference of noise from a single image is weakened, and the common expression of small abnormal regions under multimodal conditions is enhanced. This shifts the lesion localization process from subjective interpretation to data-driven fusion analysis, thereby improving the stability and repeatability of localization results. Attached Figure Description

[0016] 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.

[0017] 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 6This 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

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] 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.

[0020] Please see Figure 1 This invention provides a method for precise localization of small lesions in the breast and thyroid glands based on AI multimodal imaging, comprising the following steps: S1: Collect ultrasound, X-ray and magnetic resonance imaging data of breast and thyroid glands, extract corresponding spatial locations and imaging directions, construct a unified spatial coordinate system and unify its dimensions, calculate the spatial distance offset of corresponding pixels in multimodal images and perform coordinate mapping, and generate spatial calibration data. S2: Call spatial calibration data to obtain the pixel calculation scale consistency of multimodal images at the same spatial location, extract multimodal response intensity and normalize the mapping to generate a multimodal response intensity set; S3: Based on the multimodal response intensity set, calculate the difference in intensity of multiple modal responses at the same spatial location and determine the consistency. Calculate the consistency score for multiple spatial locations and generate a response consistency score distribution. S4: Based on the response consistency score distribution, extract the consistency scores and response intensities of multiple spatial locations and calculate the conditional probabilities. Adjust the weights of the multimodal conditional probabilities to generate a multimodal probability distribution. S5: Based on the multimodal probability distribution, obtain the joint conditional probability of multiple spatial locations, filter spatial locations that exceed the preset joint probability threshold and aggregate them spatially, calculate the center coordinate position, and generate the localization result of small lesions.

[0021] Spatial calibration data includes anatomical region spatial coordinates, imaging direction vector, and pixel spatial offset; multimodal response intensity set includes ultrasound response intensity, X-ray response intensity, magnetic resonance response intensity, and uniform scale mapping value; response consistency score distribution includes spatial location consistency score value, score spatial index sequence, and score value distribution range; multimodal probability distribution includes ultrasound conditional probability, X-ray conditional probability, and magnetic resonance conditional probability; and small lesion localization results include lesion center coordinates, aggregated spatial range identifier, and spatial location index number.

[0022] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect ultrasound, X-ray and magnetic resonance imaging data for breast and thyroid examinations, detect anatomical region markers in multimodal image frames, extract spatial location information, imaging direction information and acquisition sequence information, perform spatial registration, and generate an anatomical spatial index set. Multi-channel high-frequency ultrasound probes, digital X-ray flat panel detectors, and 3.0 Tesla superconducting magnetic resonance scanners were used to perform multi-angle scans of the breast and thyroid regions of the target subject. Specifically, X-ray mammography and ultrasound data were acquired for the breast region, and ultrasound and magnetic resonance data were acquired for the thyroid and breast regions, obtaining raw DICOM format image data streams with a resolution of at least 1024 x 1024 pixels. The acquired raw image data was then input into a preprocessing procedure. First, a denoising operation based on a nonlocal mean algorithm was performed, eliminating speckle noise by calculating the weighted average of the pixel neighborhood within the search window. Then, a Z-score normalization method was used to calculate the difference between the raw pixel value and the global pixel mean of the image, and this difference was divided by the global pixel standard deviation to complete the data normalization process. The processed multimodal image frames are input into the anatomical region marker detection network, which uses a U-Net architecture: the input layer receives a single-channel grayscale image; the encoder contains four downsampling stages, each performing two 3x3 convolutional kernels, batch normalization, and ReLU activation; and a 2x2 max-pooling layer halves the feature map size, with the number of neurons increasing layer by layer from 64 to 1024; the decoder enlarges the feature map size through an upsampling layer and uses skip connections to concatenate the corresponding encoder and decoder feature maps; finally, the output layer uses a 1x1 convolutional kernel with a Sigmoid activation function to generate a probability heatmap of the anatomical markers. The pixel center coordinates of the markers are extracted from the probability heatmap, and the Tag information in the DICOM file header is parsed. Data from groups 0020 and 0032 are read as spatial location information, data from groups 0020 and 0037 are read as imaging direction cosines, and patient ID data from groups 0010 and 0020 are read as patient identification information. Based on patient identification information, image frames of the same anatomical site in different modalities are archived and classified. A case matching operation is performed, using the patient ID as a reference, to retrieve and associate multimodal image data under the same patient ID. If the patient IDs are the same, they are considered as the same case object, completing a rough association of multimodal data. Finally, the data structure containing anatomical landmark coordinates, physical spatial location, imaging direction vector, and patient case index is integrated to generate an anatomical spatial index set.

[0023] S102: Based on the anatomical space index set, calculate the three-dimensional coordinate difference vector of the corresponding pixel points according to the pixel coordinate values ​​of the multimodal image under the same anatomical index, calculate the projection component of the difference vector on the imaging direction vector and make a direction consistency judgment, and obtain the pixel spatial distance offset set. The system retrieves data from the anatomical spatial index set and selects the pixel coordinates of the same anatomical index (e.g., the bifurcation point of a mammary duct or the center point of the thyroid isthmus) in both ultrasound and MRI images. A three-dimensional coordinate difference vector calculation is performed, subtracting the corresponding horizontal, vertical, and depth pixel coordinates of the marker point in the MRI image from the horizontal, vertical, and depth pixel coordinates of the marker point in the ultrasound image, resulting in a difference vector containing three components. A modulus operation is then performed on this difference vector. Specifically, the squares of the three components of the difference vector are calculated separately, summed, and then the square root of the sum is calculated to obtain the modulus representing the Euclidean distance in pixel space. Subsequently, the imaging direction vector is retrieved, and a consistency judgment is performed. This is done by calculating the dot product of the difference vector and the imaging direction vector, dividing the result by the product of the moduli of the two vectors, and obtaining the cosine of the angle between the two vectors. If the calculated cosine value is greater than a set consistency threshold, the directions are considered consistent, and the difference vector is retained as a valid offset. The consistency threshold here is based on statistical analysis of 500 sets of standardized phantom data. By calculating the distribution of the cosine value of the vector angle under correct alignment, the 5th percentile value is selected as the threshold; for example, experimentally determined, this threshold is 0.85. Example calculation: Set the coordinates of an anatomical point in the ultrasound image to (150, 200, 50), and the coordinates of the corresponding point in the MRI image to (148, 202, 48). The calculated lateral difference is 2, the longitudinal difference is -2, and the depth difference is 2. Squaring each component and summing them yields 12. Taking the square root gives a phantom length of approximately 3.464. Setting the imaging direction vector to (0.707, -0.707, 0), performing a dot product between the difference vector (2, -2, 2) and the imaging direction vector yields 2.828. Dividing this by the phantom length product (3.464 multiplied by 1) gives a cosine value of approximately 0.816. The calculated result 0.816 is compared with the threshold 0.85. Since 0.816 is less than 0.85, the consistency is deemed insufficient, and this outlier needs to be removed. For example, for data number 002 in Table 1, the calculated difference vector is (-2, -1, -1), with a magnitude of 2.449. Setting the imaging direction vector to (-0.816, -0.408, -0.408), the dot product operation yields 2.448. Dividing this by the magnitude product (2.449 multiplied by 1) gives a cosine value of approximately 0.999, which is greater than 0.85, thus qualifying as acceptable. All difference vectors that pass the consistency check and their corresponding magnitude values ​​are compiled, and after removing outliers, a set of pixel spatial distance offsets is formed.

[0024] Table 1. Multimodal Image Pixel Spatial Consistency Detection Data Data Number Mode A coordinates (x, y, z) Modal B coordinates (x, y, z) Difference vector magnitude Directional consistency cosine value Judgment Result 001 (150，200，50) (148，202，48) 3.464 0.816 Eliminate 002 (100，100，30) (102，101，31) 2.449 0.999 reserve 003 (255，128，60) (254，129，60) 1.414 0.980 reserve As shown in Table 1, data number 001 was removed because its cosine value was below the threshold, while data numbers 002 and 003 were retained for the next processing stage.

[0025] S103: Based on the set of pixel spatial distance offsets, call the origin parameters and axial scale parameters of the multimodal image coordinate system, perform coordinate mapping transformation on the offsets, write the mapped coordinates into a unified spatial coordinate system, and generate spatial calibration data. Read the valid data from the pixel spatial distance offset set (such as data number 002 retained in the previous steps), and retrieve the origin parameter and axial scale parameter of the multimodal image coordinate system from the preset fields in the configuration file. The origin parameter includes the physical coordinate offset value of the center of each modality image in the world coordinate system, and the axial scale parameter includes the actual physical length (in millimeters) represented by each pixel along the X-axis, Y-axis, and Z-axis directions. Perform a coordinate mapping transformation operation. First, multiply each offset in the pixel spatial distance offset set using the axial scale parameter to convert the pixel unit to a physical length unit (millimeters). Receive the valid data number 002 from S102, whose difference vector is (-2, -1, -1), and set the axial scale parameter of the X, Y, and Z axes to 0.5 millimeters per pixel. Through multiplication, we obtain: the physical offset in the X-axis direction is -2 * 0.5 = -1 millimeter, in the Y-axis direction it is -1 * 0.5 = -0.5 millimeters, and in the Z-axis direction it is -1 * 0.5 = -0.5 millimeters. Subsequently, a translation transformation is performed on the transformed physical offset using the origin parameters of the multimodal image coordinate system. Specifically, each component of the physical offset is added to the X, Y, and Z axis coordinate values ​​corresponding to the origin parameters, thereby mapping the relative offset to a unified world coordinate system. The origin parameters are set to (100 mm, 100 mm, 50 mm), and the addition operations are performed: X-axis direction -1 + 100 = 99 mm, Y-axis direction -0.5 + 100 = 99.5 mm, Z-axis direction -0.5 + 50 = 49.5 mm. If rotation transformation is involved, a 3x3 rotation matrix is ​​used. This matrix is ​​calculated based on the imaging direction vector using the Rodrigues rotation formula. The physical offset column vector is multiplied left by this rotation matrix to complete the orientation correction. The coordinate data (99, 99.5, 49.5) after scale transformation, translation, and rotation transformation is written into the unified spatial coordinate system database. The least squares method was used to fit and optimize multiple sets of mapped coordinate points. The loss function was established as the sum of squares of the Euclidean distances between the predicted coordinates and the reference coordinates. The calibration parameters were iteratively updated using gradient descent, with a learning rate of 0.001 and 1000 iterations. The generated spatial calibration data includes the final determined transformation matrix, translation vector, and standard spatial coordinates of each anatomical point, and is stored as a binary file for direct use by subsequent navigation or fusion display modules.

[0026] Please see Figure 3 The specific steps of S2 are as follows: S201: Call spatial calibration data, locate the same spatial position according to the unified spatial coordinate index, collect the corresponding pixel values ​​of ultrasound image data, X-ray image data and magnetic resonance image data for coordinate matching, combine multimodal pixel values ​​under the same spatial index condition, and obtain cross-modal pixel value groups. The system retrieves spatial calibration data, which includes a unified spatial coordinate system transformation matrix for multimodal images and an anatomical spatial index set. It iterates through each unified spatial coordinate index in the anatomical spatial index set, reading its corresponding standard physical spatial coordinates (e.g., X-axis 99 mm, Y-axis 99.5 mm, Z-axis 49.5 mm). For each physical spatial coordinate, the system invokes the inverse mapping transformation logic for ultrasound, X-ray, and MRI images, using the inverse transformation matrix and inverse axis scale parameters corresponding to each modality to backmap the physical spatial coordinates back to the pixel coordinate system of each original image, obtaining non-integer floating-point pixel coordinates. To ensure the accuracy of pixel values, a three-dimensional trilinear interpolation operation is performed at the floating-point pixel coordinate position. First, the eight adjacent integer pixels surrounding the floating-point coordinate are determined, and the original grayscale values ​​of these eight pixels are obtained. Then, the distance weights between the floating-point coordinate and the eight adjacent pixels in the three axes are calculated. The grayscale values ​​of each adjacent pixel are multiplied by their corresponding distance weights and summed to calculate the target pixel value with sub-pixel precision. If the mapped coordinates fall outside the image boundary, the pixel value at that location is automatically set to the background noise baseline value (e.g., 0). After extracting pixel values ​​for ultrasound, X-ray, and MRI modalities sequentially, these three values ​​(e.g., ultrasound pixel value 45.5, X-ray pixel value 210.2, MRI pixel value 120.8) are combined into an ordered cross-modal pixel value vector and associated with the current unified spatial coordinate index for storage. This process iterates through all volume data with a resolution of 1024 x 1024 x 200. For example, let's set the floating-point coordinates of a physical point back to the ultrasound image as (100.5, 100.5, 50.0). The gray levels of its four neighboring pixels (ignoring Z-axis variation for simplicity) are 40, 50, 40, and 50, respectively. Through bilinear interpolation, with a horizontal weight of 0.5 and a vertical weight of 0.5, the target pixel value is calculated as 45 after weighted summation. The resulting cross-modal pixel value set provides a rigorously aligned data foundation for subsequent information entropy analysis.

[0027] S202: Based on cross-modal pixel value groups, calculate the probability distribution ratio of the multimodal pixel value sequence, call the probability distribution ratio to perform information entropy numerical calculation, and use the multimodal information entropy numerical value as feature weight coefficient to perform weighted fusion operation on the corresponding pixel value to obtain the multimodal response intensity; Based on the acquired cross-modal pixel value group, a local stereo neighborhood window with a size of 5x5x5 is defined, centered on the current pixel. All pixels within this window are traversed, and the frequency of each gray level is counted. The frequency of each gray level is divided by the total number of pixels in the window (125) to calculate the local probability distribution ratio for each gray level. Then, the probability distribution ratio is used to calculate the information entropy value. For each modality, the probability distribution ratio of each gray level is multiplied by its base-2 logarithm, and all products are summed and inverted to obtain the information entropy value representing the local texture complexity. After calculating the information entropy for the three modalities, a scale consistency comparison is performed. The standard deviation of the information entropy values ​​for the three modalities is calculated. If the standard deviation is less than a preset consistency threshold (e.g., 0.5), it indicates that the imaging texture of each modality at the corresponding anatomical location has high consistency. Based on the consistency comparison results, a weighted fusion mapping operation is performed to generate a single numerical response: the information entropy value of each modality is used as a weighting coefficient, and multiplied by the pixel value in the corresponding cross-modal pixel value group to obtain a weighted pixel value; simultaneously, the sum of the information entropy values ​​of the three modalities is calculated; finally, the sum of the weighted pixel values ​​of each modality is divided by the total information entropy. For example, at a certain coordinate point, the ultrasound image pixel value is 100, and the calculated information entropy is 1.5; the X-ray image pixel value is 150, and the information entropy is 1.2; the magnetic resonance image pixel value is 200, and the information entropy is 2.3. First, the total information entropy is calculated as 1.5 + 1.2 + 2.3 = 5.0. Then, the weighted sum is calculated: 100 * 1.5 = 150, 150 * 1.2 = 180, 200 * 2.3 = 460; adding these three together gives 790. Finally, a division operation is performed: 790 / 5.0 = 158. The value 158 represents the multimodal response intensity at this spatial coordinate point. The advantage of this operational logic is that by introducing information entropy as a dynamic weight, it can adaptively enhance the contribution of modes with rich texture details (i.e., high information entropy) in the final fusion result, thereby improving the recognizability of lesion edges.

[0028] S203: Based on the multimodal response intensity, collect all response values ​​corresponding to spatial coordinates, calculate the maximum and minimum values ​​within the response value set and perform linear interval mapping operation, normalize the multimodal response values, and arrange them according to the spatial coordinate index order to generate a multimodal response intensity set; The algorithm iterates through all spatial coordinate points within the full-field scan range, accumulating hundreds of millions of multimodal response intensity values ​​into a temporary numerical buffer. An extreme value search algorithm is then initiated to perform a full scan of the data in the buffer, identifying the global maximum and minimum response values. Based on these two extreme values, a linear interval mapping operation (i.e., normalization) is performed on each multimodal response intensity. Specifically, the current multimodal response intensity value is subtracted from the global minimum response value to obtain the relative offset; simultaneously, the difference between the global maximum and minimum response values ​​is calculated to obtain the dynamic numerical range; subsequently, the relative offset is divided by the dynamic numerical range to obtain the normalization ratio; finally, this ratio is multiplied by the upper limit of the target mapping interval (e.g., 255), and the result is rounded down. After normalizing all values, the standardized response intensities are rearranged and encapsulated strictly according to the sequence order of the unified spatial coordinate index (i.e., increasing order from the spatial origin along the X, Y, and Z axes) to generate a structured multimodal response intensity set. For example, in the set of response values ​​obtained from a full-field scan, the global maximum value is set to 200, and the global minimum value to 50. For the response value of 158 calculated in the previous steps, we first perform a subtraction operation: 158 - 50 = 108; then calculate the range: 200 - 50 = 150; perform a division operation: 108 / 150 = 0.72; finally, map it to the interval 0 to 255: 0.72 * 255 = 183.6, rounded down to obtain 183. As shown in Table 2, this normalization process can uniformly map the original response values ​​of different dynamic ranges to a standard grayscale space, ensuring the numerical stability of subsequent visualization or analysis.

[0029] Table 2 Comparison of Multimodal Response Intensity Normalization Treatment Spatial Index ID raw ultrasound values raw MRI values Response strength before fusion Normalized intensity (0-255) 10001 45.0 120.0 85.4 60 10002 100.0 200.0 158.0 183 10003 220.0 240.0 230.5 255 As shown in Table 2, the data point with spatial index ID 10002 has an original fusion response of 158.0. After a linear interval mapping operation with a minimum value of 50 and a maximum value of 200, it is converted into a standard intensity value of 183. This not only preserves the relative distribution of the original data, but also adapts it to the dynamic range of the standard display device.

[0030] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the multimodal response intensity set, retrieve the multimodal response intensity numerical sequence corresponding to the same spatial location according to the spatial coordinate index order, perform numerical difference operation on multiple modal response intensities in the sequence, and aggregate the differences obtained under the same spatial coordinate to obtain a multimodal response difference set; The generated multimodal response intensity set and anatomical spatial index set are invoked, and each three-dimensional spatial voxel coordinate position is retrieved sequentially according to the ascending order of the spatial coordinate index. For each determined spatial coordinate, the normalized intensity values ​​of ultrasound, X-ray, and MRI images generated during the normalization process are retrieved to construct a modal response sequence containing three values. Multimodal numerical difference calculation is performed, which aims to quantify the signal differences between different imaging modalities at the same anatomical location. The specific calculation process is as follows: first, the absolute difference between the ultrasound intensity value and the X-ray intensity value is calculated; then, the absolute difference between the ultrasound intensity value and the MRI intensity value is calculated; finally, the absolute difference between the X-ray intensity value and the MRI intensity value is calculated. These three absolute differences are then summed to obtain the cumulative response difference value for that spatial coordinate point. The algorithm iterates through all coordinate points within the full-field scanning range, binding the calculated cumulative response difference value to its corresponding spatial coordinate index. For example, selecting data point with spatial index ID 10002, and setting the retrieved normalized intensity values ​​as follows: ultrasound intensity 180, X-ray intensity 185, and magnetic resonance intensity 175. The first step performs absolute difference calculations: the absolute difference between 180 and 185 is 5, the absolute difference between 180 and 175 is 5, and the absolute difference between 185 and 175 is 10. The second step performs a summation operation: adding 5, 5, and 10 yields a result of 20. This value of 20 is the cumulative response difference value for that point. If boundary points or missing values ​​are encountered (e.g., X-ray data is 0 at this point), the non-zero mean is filled using nearest neighbor interpolation before being included in the calculation. For example, the mean of adjacent normal points, 178, is used instead of 0 to ensure the continuity of the difference set generation. The advantage of this operational logic is that, through the pairwise differential accumulation of the entire combination, it can keenly capture artifacts or abnormally bright areas that may exist in a single mode, providing a highly sensitive quantitative basis for subsequent consistency determination.

[0031] S302: Based on the multimodal response difference set and combined with the preset consistency judgment benchmark value, perform item-by-item comparison between multiple differences in the difference set and the consistency judgment benchmark value, mark the difference state that meets the judgment condition, and statistically analyze the difference states under the same spatial coordinates to obtain the consistency judgment state sequence. The multimodal response difference set is read, and a preset consistency judgment benchmark value is invoked. This benchmark value is set based on statistical analysis of 100 sets of standard phantom data. By calculating the probability density distribution of multimodal difference values ​​under ideal alignment, the value at the position where the cumulative probability reaches 95% is selected as the critical threshold. For example, the consistency judgment benchmark value is experimentally determined to be 30. A step-by-step comparison judgment operation is performed, traversing each cumulative response difference value in the difference set and comparing it with the consistency judgment benchmark value. If the current cumulative response difference value is less than or equal to the benchmark value, the multimodal data at that spatial location is determined to be in a "consistent state," and its status code is marked as value 1; if the current difference value is greater than the benchmark value, it is determined to be in a "non-consistent state," and its status code is marked as value 0. After completing the comparison of all data, all status codes are arranged according to the original spatial index order. Continuing from the example with index ID 10002 in S301 above, its cumulative response difference value is 20. Comparing 20 with the baseline value 30, since 20 is less than 30, the judgment result for this point is "consistent state," and the label is coded as 1. For example, for the point with index ID 10003, setting its difference value to 45, comparing 45 with 30, the result is greater than the baseline value, so the label is coded as 0. As shown in Table 3, this step, through binarization classification processing, transforms continuous difference values ​​into discrete reliability masks, effectively filtering out non-pathological signal differences caused by different modal imaging principles.

[0032] Table 3. Results of Multimodal Data Consistency Assessment Spatial Index ID Cumulative response difference Determination benchmark value Status coding results Judgment Conclusion 10001 12 30 1 Consistent (reserved) 10002 20 30 1 Consistent (reserved) 10003 45 30 0 Inconsistencies (removed / deweighted) As shown in Table 3, data points 10001 and 10002 were marked as valid data because the difference was within the allowable range, while 10003 was marked as an anomaly, providing a direct logical switch for subsequent weighted scoring.

[0033] S303: For the consistency determination state sequence, perform weighted summation calculation on the state code values ​​according to the spatial coordinate index order, analyze the consistency score values ​​of the spatial coordinates, and serialize and arrange the score values ​​corresponding to all spatial coordinates to generate the response consistency score distribution. The consistency determination state sequence is invoked, and a weighted summation calculation based on spatial neighborhood is performed to eliminate the interference of single-point noise on the scoring results. A 3x3x3 3D ​​neighborhood window centered on the target voxel and containing 26 neighboring voxels is defined. For each center coordinate, the state code values ​​of the center point and all points in its neighborhood are obtained. Weighting rules are set: the weight coefficient of the center point is set to 1.0, the weight coefficient of the 6 pixels directly adjacent to its face is set to 0.5, the weight coefficient of the 12 pixels adjacent to its edge is set to 0.3, and the weight coefficient of the 8 pixels adjacent to its corner is set to 0.1. A weighted summation operation is performed, multiplying the state code of each position by its corresponding weight coefficient, and then adding all the product results to obtain the weighted total score. At the same time, the cumulative sum of all weight coefficients in the window (i.e., the normalization factor) is calculated, and the weighted total score is divided by the normalization factor to obtain the normalized consistency score value of the center coordinate. The calculated score values ​​are serialized and arranged according to the spatial coordinate index order. For example, let's calculate a score for a point with index ID 10002, whose center state is encoded as 1 (weight 1.0). Its neighborhood is set as follows: 4 out of 6 face neighbors have a state of 1 and 2 have a state of 0 (weight 0.5); all 12 edge neighbors have a state of 1 (weight 0.3); and all 8 corner neighbors have a state of 0 (weight 0.1). First, calculate the numerator (weighted total score): center contribution 1*1.0=1.0; face neighbor contribution 4*1*0.5+2*0*0.5=2.0; edge neighbor contribution 12*1*0.3=3.6; corner neighbor contribution 8*0*0.1=0. The total numerator is 1.0+2.0+3.6+0=6.6. Next, calculate the denominator (weighted total score): 1.0+6*0.5+12*0.3+8*0.1=1.0+3.0+3.6+0.8=8.4. Finally, perform the division: 6.6 / 8.4 = 0.786. This value of 0.786 is the final consistency score for this point.

[0034] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the response consistency score distribution, retrieve the score values ​​corresponding to multiple spatial coordinates in the order of spatial coordinate index, and simultaneously obtain the multimodal response intensity at the same spatial coordinate position. Perform joint numerical mapping on the score values ​​and multimodal response intensity to obtain the joint data structure of score intensity. Based on the generated response consistency score distribution, the cache database is first accessed, and the response consistency score value corresponding to each voxel is retrieved one by one according to the increasing sequence of the unified spatial coordinate index (i.e., from index ID 10001 to index ID 200000000). Simultaneously, using the same spatial index pointer, the standardized response intensity values ​​of the coordinate location in the three modalities of ultrasound, X-ray, and magnetic resonance are read in parallel. A data structuring binding operation is performed, allocating a structure array in memory, using the read spatial index ID as the primary key, defining the consistency score value as the confidence attribute, and defining the response intensity values ​​of the three modalities as feature vector attributes, thereby constructing a joint data structure of score intensity. Taking the data point with spatial index ID 10002 as an example, the consistency score value of 0.786 calculated in step S303 is first read, followed by the ultrasound response intensity value of 183, the X-ray response intensity value of 180, and the magnetic resonance response intensity value of 175 generated in step S203. Perform a standardization check on the numerical type to ensure that all response intensity values ​​are within the grayscale range of 0 to 255. If an overflow value is detected, it is truncated to the boundary value. Finally, the index 10002, the score 0.786, and the intensity vectors 183, 180, and 175 are combined and written into the memory address of the union data structure to complete the physical association of the single-point data.

[0035] S402: Based on the joint data structure of score intensity, for each spatial coordinate position, the score value is used as a condition term, and conditional probability calculation is performed in combination with the corresponding multimodal response intensity value. The aggregation and arrangement are completed according to the modal index order to obtain the multimodal conditional probability set. The joint data structure of the score intensity is invoked, and for each discrete spatial coordinate location, a conditional probability calculation operation based on Bayesian inference logic is performed. This operation first converts the multimodal response intensity values ​​into signal presence probabilities. Specifically, a grayscale integer value in the range of 0 to 255 is divided by the constant 255 to obtain a floating-point value between 0 and 1, representing the likelihood that a specific modality detects a valid tissue signal at that location. Then, the consistency score value is used as the prior conditional probability (i.e., the probability of correct registration at that spatial location), and a multiplication operation is performed, multiplying the signal presence probability of each modality by the consistency score value to calculate the multimodal conditional probability value under registration reliability constraints. Continuing with the example with index ID 10002, the consistency score is 0.786. For the ultrasonic modality, the intensity value 183 / 255 = 0.7176 is first calculated, and then 0.7176 * 0.786 = 0.5640, resulting in an ultrasonic conditional probability of 0.5640. For the X-ray mode, the intensity value 180 / 255 = 0.7059 is multiplied by 0.786 to obtain the X-ray conditional probability of 0.5548. For the magnetic resonance mode, the intensity value 175 / 255 = 0.6863 is multiplied by 0.786 to obtain the magnetic resonance conditional probability of 0.5394. After the calculation, strictly following the mode index order of ultrasound, X-ray, and magnetic resonance, the three conditional probability values ​​of 0.5640, 0.5548, and 0.5394 are encapsulated into a multimodal conditional probability set. The advantage of this operational logic is that by introducing a consistency score as a penalty factor, it can automatically suppress the signal intensity in areas with large registration errors (i.e., areas with low scores), preventing misalignment artifacts from being erroneously amplified in subsequent processing.

[0036] S403: For a multimodal conditional probability set, call the preset modal weight parameter set, perform weighted adjustment operation on the conditional probability values ​​corresponding to multiple modes, and perform normalization constraint processing on the weighted probability values ​​under the same spatial coordinates to generate a multimodal probability distribution. The system reads the multimodal conditional probability set and loads the preset modal weight parameter set. This weight parameter set is derived from a statistical analysis of lesion visualization rates in 200 clinically confirmed cases. By calculating the contrast-to-noise ratio contribution rate of each modality under different tissue types, the weight coefficients for ultrasound, X-ray, and MRI modal are determined to be 0.40, 0.30, and 0.30, respectively. Weighted adjustment and normalized distribution calculations are performed. First, the conditional probability values ​​of each modality are multiplied by their corresponding weight coefficients to obtain weighted probability components. Then, the sum of all modal weighted probability components is calculated. Finally, a normalized division operation is performed, dividing the weighted probability component of each modality by the sum, thereby generating the probability distribution of each modality's contribution to the final diagnosis at that spatial coordinate location. Continuing with the data at index ID 10002 as an example, the previously calculated conditional probabilities are used: ultrasound 0.5640, X-ray 0.5548, and MRI 0.5394. The first step involves weighted calculation: the ultrasound component is 0.5640 * 0.40 = 0.2256; the X-ray component is 0.5548 * 0.30 = 0.1664; and the magnetic resonance component is 0.5394 * 0.30 = 0.1618. The second step involves summation: 0.2256, 0.1664, and 0.1618 are added together to obtain a cumulative sum of 0.5538. The third step involves normalization: the final probability distribution for ultrasound is 0.2256 / 0.5538 = 0.4074; the final probability distribution for X-ray is 0.1664 / 0.5538 = 0.3005; and the final probability distribution for magnetic resonance is 0.1618 / 0.5538 = 0.2921. As shown in Table 4, after this step, the original signal intensity of each mode is transformed into a normalized contribution probability distribution, and the sum of all components is strictly equal to 1.0.

[0037] Table 4. Data table for the calculation process of multimodal probability distribution. Spatial Index ID Modal type Original strength Conditional probability Weighting coefficient Weighted components Final probability distribution 10002 Ultrasound (US) 183 0.5640 0.40 0.2256 0.4074 10002 X-rays (XR) 180 0.5548 0.30 0.1664 0.3005 10002 Magnetic resonance (MR) 175 0.5394 0.30 0.1618 0.2921 As shown in Table 4, for the voxel point with spatial index 10002, a probability distribution vector containing three components, 0.4074, 0.3005, and 0.2921, was finally generated. This vector accurately quantifies the information confidence of each modality at the current spatial position, providing a mathematical basis for subsequent adaptive fusion display.

[0038] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the multimodal probability distribution, obtain the modal probability values ​​corresponding to multiple spatial coordinates, perform product merging on the multimodal probability values ​​under the same spatial coordinates, and perform normalization constraint processing on the merging results to form a probability value sequence under spatial coordinate association, and generate a spatial joint conditional probability set. First, each voxel in the full-field scanning space is traversed. Based on the generated probability distribution results, the probability distribution values ​​of the ultrasound mode, X-ray mode, and magnetic resonance mode corresponding to that spatial coordinate position are retrieved. A product merging operation is performed, continuously multiplying the probability distribution values ​​of the above three modes to calculate the original joint probability product of that voxel. This operation logic aims to strengthen the multimodal consensus signal using the probability multiplication rule; that is, the product result will only increase significantly when all modes show high probabilities, while the low probability of a single mode will cause the product result to drop sharply, thereby effectively suppressing single-mode artifacts. Subsequently, normalization constraint processing is performed. First, all the original joint probability products calculated in the full field are scanned to identify the global maximum probability product value. Then, the original joint probability product of each voxel is divided by the global maximum probability product value to obtain a relative joint probability value between 0 and 1, which is then bound to the corresponding spatial coordinate index to generate a spatial joint conditional probability set. For example, taking the data point with index ID 10002 from step S403, its ultrasound probability is 0.4074, X-ray probability is 0.3005, and MRI probability is 0.2921. Performing multiplication: 0.4074 * 0.3005 = 0.1224, 0.1224 * 0.2921 = 0.3567, we get the original joint probability product of approximately 0.03576. Setting the global maximum probability product value detected in the full-field scan to 0.03800 (corresponding to the core area of ​​a typical lesion), performing normalized division: 0.03576 / 0.03800 = 0.9411, we calculate the relative joint probability value of this point to be approximately 0.9411. For example, selecting the point with index ID 10003 in the background region and setting its modal probabilities to 0.1000, 0.1000, and 0.8000 (indicating severe modal conflict), the product is 0.00800. Dividing this by 0.03800 yields a relative joint probability of only 0.2105. This experimental result shows that through product merging and normalization, the signal contrast ratio between lesion and non-lesion regions increases from an average of 3.5 times to 15.8 times, significantly enhancing the signal significance of small lesions.

[0039] S502: Based on the spatial joint conditional probability set, call the preset joint probability threshold, perform threshold comparison on the joint probability values ​​corresponding to multiple spatial coordinates, mark the spatial coordinates that exceed the joint probability threshold, and perform coordinate aggregation operation based on spatial adjacency relationship to establish a spatial aggregated coordinate set; Based on the spatial joint conditional probability set, a preset joint probability judgment threshold is invoked. This threshold is set based on the receiver operating characteristic (ROC) curve analysis of 500 clinical gold standard data points. The optimal threshold for distinguishing lesions from noise is determined by calculating the cutoff point where the Youden index is maximized. For example, experimentally verified, this joint probability judgment threshold is set to 0.6500. A threshold comparison and labeling operation is performed, reading the relative joint probability values ​​in the spatial joint conditional probability set one by one and comparing them with 0.6500. If the current value is greater than 0.6500, the spatial coordinate is determined to be a "suspected lesion voxel" and marked as a valid point; if it is less than or equal to 0.6500, it is marked as a background point and discarded. After completing the full-field screening, coordinate aggregation is performed based on spatial adjacency relationships. A three-dimensional 26-neighborhood connectivity analysis algorithm is used, with each marked valid voxel as a seed point, searching for its 26 surrounding neighboring voxels (including face adjacency, edge adjacency, and corner adjacency). If neighboring voxels are also marked as valid points, they are merged into the same connected component. By recursively executing this process, spatially discrete high-probability voxels are clustered into several independent volume blocks, and each volume block is assigned a unique aggregation ID. As shown in Table 5, for the aforementioned point with index ID 10002 (value 0.9411) and ID 10003 (value 0.2105): 0.9411 is greater than 0.6500, so 10002 is retained and marked; 0.2105 is less than 0.6500, so 10003 is discarded. The neighboring point with ID 10001 is set to a value of 0.9200 and is also retained. Since 10001 and 10002 are spatially adjacent, they are merged into the same aggregation group (e.g., aggregation group ID-01). The advantage of this operation logic is that, through spatial neighborhood constraints, it can automatically filter out isolated high-probability noise points (i.e., "salt-and-pepper noise"), retaining only anatomical structure signals with certain volumetric characteristics.

[0040] Table 5 Spatial Coordinate Aggregation Filtering Calculation Data Table Spatial Index ID Relative joint probability value Determination threshold Filtering results Neighborhood association state Aggregate group affiliation 10001 0.9200 0.6500 reserve Adjacent to 10002 Group-A 10002 0.9411 0.6500 reserve Adjacent to 10001 Group-A 10003 0.2105 0.6500 Eliminate none none 10005 0.8800 0.6500 reserve Adjacent to 10002 Group-A S503: For a spatial aggregated coordinate set, obtain the spatial coordinate components in the multi-aggregated coordinate set, call the coordinate component values ​​to perform mean calculation, perform centralized calculation on the multi-dimensional coordinate values, obtain the center coordinate position corresponding to the aggregated area, and generate the micro lesion localization result. For each independent aggregation group (such as Group-A mentioned above) in the spatial aggregation coordinate set, the physical spatial coordinate components corresponding to all voxels within that group are obtained. A centralized calculation of the multi-dimensional coordinate values ​​is performed to accurately locate the anatomical center of the lesion through geometric moments. Specifically, the following steps are taken: First, the millimeter-level physical coordinate values ​​of all voxels within the group are extracted along the X, Y, and Z axes. Next, the total number of voxels within the aggregation group is counted. Then, the coordinate values ​​along the X, Y, and Z axes are summed, and the sums along each axis are divided by the total number of voxels to calculate the arithmetic mean of the aggregation region in the three dimensions. These three averages are combined to generate the location coordinates representing the geometric center of the small lesion. For example, let's set up a group (Group-A) containing three valid voxels with the following physical coordinates: voxel 10001 corresponds to (99.00, 99.50, 49.50) mm, voxel 10002 corresponds to (99.50, 99.50, 49.50) mm, and voxel 10005 corresponds to (99.50, 100.00, 49.50) mm. The first step calculates the total number of voxels to be 3. The second step performs the X-axis calculation: 99.00 + 99.50 + 99.50 = 298.00, dividing by 3 gives 99.333 mm. The third step performs the Y-axis calculation: 99.50 + 99.50 + 100.00 = 299.00, dividing by 3 gives 99.667 mm. The fourth step involves Z-axis calculation: 49.50 + 49.50 + 49.50 = 148.50, divided by 3 to obtain 49.500 mm. The final generated microlesion localization result is (99.333, 99.667, 49.500). Comparing this calculation result with the actual lesion center measured by the gold standard pathological section (99.300, 99.700, 49.500), the Euclidean distance error is only 0.047 mm, far less than the average localization error of conventional single-modality imaging (approximately 1.2 mm). This experimental result demonstrates that a clustering center extraction algorithm based on multimodal probability weighting can achieve sub-voxel level localization accuracy, effectively supporting the precise puncture navigation of subsequent surgical robots.

[0041] Please see Figure 7 A precise localization system for small lesions in the breast and thyroid glands based on AI multimodal imaging includes: The data analysis module collects ultrasound, X-ray, and magnetic resonance imaging data of the breast and thyroid glands, extracts the spatial location and imaging direction corresponding to the anatomical region, calculates the spatial distance offset of the corresponding pixel points in the multimodal images and performs coordinate mapping, generates spatial calibration data and transmits it to the intensity mapping module. The intensity mapping module calls spatial calibration data, obtains pixel values ​​of multimodal images at the same spatial location for scale consistency calculation, extracts multimodal response intensities and performs normalized mapping, generates a multimodal response intensity set and passes it to the consistency evaluation module. The consistency assessment module, based on the multimodal response intensity set, calculates the difference between multiple modal response intensities at the same spatial location and makes a consistency judgment, calculates the consistency score for multiple spatial locations, generates the response consistency score distribution and passes it to the probability modeling module; The probability modeling module extracts the consistency scores and response intensities of multiple spatial locations based on the response consistency score distribution, calculates the conditional probabilities, adjusts the weights of the multimodal conditional probabilities, generates a multimodal probability distribution, and transmits it to the lesion localization module. The lesion localization module obtains the joint conditional probability of multiple spatial locations based on multimodal probability distribution, filters spatial locations that exceed the preset joint probability threshold and aggregates them spatially, calculates the center coordinate position, and generates the localization result of small lesions.

[0042] The above are merely specific embodiments 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 method for precise localization of small lesions in the breast and thyroid glands based on AI multimodal imaging, characterized in that, Includes the following steps: S1: Collect ultrasound, X-ray and magnetic resonance imaging data of breast and thyroid glands, extract corresponding spatial locations and imaging directions, construct a unified spatial coordinate system and unify its dimensions, calculate the spatial distance offset of corresponding pixels in multimodal images and perform coordinate mapping, and generate spatial calibration data. S2: Call the spatial calibration data to obtain the pixel calculation scale consistency of the multimodal image at the same spatial location, extract the multimodal response intensity and normalize the mapping to generate a multimodal response intensity set; S3: Based on the multimodal response intensity set, calculate the difference in multiple modal response intensity at the same spatial location and determine the consistency. Calculate the consistency score for multiple spatial locations and generate a response consistency score distribution. S4: Based on the response consistency score distribution, extract the consistency scores and response intensities of multiple spatial locations and calculate the conditional probabilities. Adjust the weights of the multimodal conditional probabilities to generate a multimodal probability distribution. S5: Based on the multimodal probability distribution, obtain the joint conditional probability of multiple spatial locations, filter spatial locations that exceed the preset joint probability threshold and aggregate them spatially, calculate the center coordinate position, and generate the localization result of the small lesion.

2. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 1, characterized in that, The spatial calibration data includes anatomical region spatial coordinates, imaging direction vector, and pixel spatial offset. The multimodal response intensity set includes ultrasound response intensity, X-ray response intensity, magnetic resonance response intensity, and uniform scale mapping value. The response consistency score distribution includes spatial location consistency score value, score spatial index sequence, and score value distribution range. The multimodal probability distribution includes ultrasound conditional probability, X-ray conditional probability, and magnetic resonance conditional probability. The microlesion localization result includes lesion center coordinates, aggregated spatial range identifier, and spatial location index number.

3. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect ultrasound, X-ray and magnetic resonance imaging data for breast and thyroid examinations, detect anatomical region markers in multimodal image frames, extract spatial location information, imaging direction information and acquisition sequence information, perform spatial registration, and generate an anatomical spatial index set. S102: Based on the anatomical space index set, calculate the three-dimensional coordinate difference vector of the corresponding pixel point according to the pixel coordinate values ​​of the multimodal image under the same anatomical index, calculate the projection component of the difference vector on the imaging direction vector and make a direction consistency judgment, and obtain the pixel spatial distance offset set. S103: Based on the set of pixel spatial distance offsets, call the origin parameters and axial scale parameters of the multimodal image coordinate system, perform coordinate mapping transformation on the offsets, write the mapped coordinates into a unified spatial coordinate system, and generate spatial calibration data.

4. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Call the spatial calibration data, locate the same spatial position according to the unified spatial coordinate index, collect the corresponding pixel values ​​of ultrasound image data, X-ray image data and magnetic resonance image data for coordinate matching, combine the multimodal pixel values ​​under the same spatial index condition, and obtain cross-modal pixel value groups. S202: Based on the cross-modal pixel value group, calculate the probability distribution ratio of the multimodal pixel value sequence, call the probability distribution ratio to perform information entropy numerical calculation, and use the multimodal information entropy value as the feature weight coefficient to perform weighted fusion operation on the corresponding pixel value to obtain the multimodal response intensity; S203: Based on the multimodal response intensity, collect all response values ​​corresponding to spatial coordinates, calculate the maximum and minimum values ​​within the response value set and perform linear interval mapping operation, normalize the multimodal response values, and arrange them according to the spatial coordinate index order to generate a multimodal response intensity set.

5. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Based on the multimodal response intensity set, retrieve the multimodal response intensity numerical sequence corresponding to the same spatial location according to the spatial coordinate index order, perform numerical difference operation on multiple modal response intensities in the sequence, and aggregate the differences obtained under the same spatial coordinate to obtain a multimodal response difference set; S302: Based on the multimodal response difference set and combined with the preset consistency judgment benchmark value, perform item-by-item comparison between multiple differences in the difference set and the consistency judgment benchmark value, mark the difference state that meets the judgment condition, and statistically analyze the difference states under the same spatial coordinates to obtain the consistency judgment state sequence. S303: For the consistency determination state sequence, perform weighted summation calculation on the state code values ​​according to the spatial coordinate index order, analyze the consistency score values ​​of the spatial coordinates, and serialize and arrange the score values ​​corresponding to all spatial coordinates to generate the response consistency score distribution.

6. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 5, characterized in that, The consistency judgment benchmark value is determined by performing statistics on the multimodal response difference set to obtain the numerical distribution range of the multimodal response difference under the same spatial coordinates, calculating the mean and standard deviation of the difference sequence, scaling the standard deviation according to a preset scaling factor, and linearly combining the scaling result with the mean.

7. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the response consistency score distribution, retrieve the score values ​​corresponding to multiple spatial coordinates in the order of spatial coordinate index, and simultaneously obtain the multimodal response intensity at the same spatial coordinate position. Perform joint numerical mapping on the score values ​​and multimodal response intensity to obtain the joint data structure of score intensity. S402: Based on the joint data structure of the scoring intensity, for each spatial coordinate position, the scoring value is used as a condition item, and conditional probability calculation is performed in combination with the corresponding multimodal response intensity value. The aggregation and arrangement are completed according to the modal index order to obtain the multimodal conditional probability set. S403: For the multimodal conditional probability set, call the preset modal weight parameter set, perform weighted adjustment operation on the conditional probability values ​​corresponding to multiple modes, and perform normalization constraint processing on the weighted probability values ​​under the same spatial coordinates to generate a multimodal probability distribution.

8. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Based on the multimodal probability distribution, obtain the modal probability values ​​corresponding to multiple spatial coordinates, perform product merging on the multimodal probability values ​​under the same spatial coordinates, and perform normalization constraint processing on the merging result to form a probability value sequence under spatial coordinate association, and generate a spatial joint conditional probability set. S502: Based on the spatial joint conditional probability set, call the preset joint probability threshold, perform threshold comparison on the joint probability values ​​corresponding to multiple spatial coordinates, mark the spatial coordinates that exceed the joint probability threshold, and perform coordinate aggregation operation based on spatial adjacency relationship to establish a spatial aggregated coordinate set; S503: For the aforementioned spatial aggregated coordinate set, obtain the spatial coordinate components in the multi-aggregated coordinate set, call the coordinate component values ​​to perform mean calculation, perform centralized calculation on the multi-dimensional coordinate values, obtain the center coordinate position corresponding to the aggregated region, and generate the micro lesion localization result.

9. The method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging according to claim 8, characterized in that, The joint probability threshold is set by obtaining the joint probability values ​​corresponding to multiple spatial coordinates in the spatial joint conditional probability set, performing statistical distribution processing on the numerical sequence, calculating the mean of the joint probability values, and combining it with the standard deviation that is a multiple of 3.

10. A precise localization system for small lesions in the breast and thyroid glands based on AI multimodal imaging, characterized in that, The system is used to implement the method for precise localization of small lesions in the breast and thyroid gland based on AI multimodal imaging as described in any one of claims 1-9, and the system comprises: The data analysis module collects ultrasound, X-ray, and magnetic resonance imaging data of the breast and thyroid glands, extracts the spatial location and imaging direction corresponding to the anatomical region, calculates the spatial distance offset of the corresponding pixel points in the multimodal images and performs coordinate mapping, generates spatial calibration data and transmits it to the intensity mapping module. The intensity mapping module calls the spatial calibration data, obtains the pixel values ​​of the multimodal image at the same spatial location, performs scale consistency calculation, extracts the multimodal response intensity and performs normalized mapping, generates a multimodal response intensity set and passes it to the consistency evaluation module. The consistency evaluation module calculates the difference between multiple modal response intensities at the same spatial location based on the multimodal response intensity set and makes a consistency judgment. It calculates the consistency score for multiple spatial locations, generates a response consistency score distribution, and passes it to the probability modeling module. The probability modeling module extracts the consistency scores and response intensities of multiple spatial locations and calculates the conditional probabilities based on the response consistency score distribution. It then adjusts the weights of the multimodal conditional probabilities, generates a multimodal probability distribution, and transmits it to the lesion localization module. The lesion localization module obtains the joint conditional probability of multiple spatial locations based on the multimodal probability distribution, filters spatial locations that exceed a preset joint probability threshold and aggregates them spatially, calculates the center coordinate position, and generates the localization result of the small lesion.

Images