Artificial intelligence-based intelligent determination system for preoperative magnetic resonance imaging of unicompartmental knee arthroplasty

Abstract

The application relates to the technical field of medical image analysis and artificial intelligence, and discloses an AI-based intelligent judgment system for preoperative magnetic resonance imaging of unicompartmental knee arthroplasty, which comprises data preprocessing, anatomical structure segmentation, feature extraction, multi-modal feature fusion and surgery indication evaluation modules. The system first standardizes the image data; the anatomical structure segmentation module generates a mask using deep learning, and constructs a cylindrical virtual region of interest based on the anatomical anchor points of the femur and tibia when detecting the absence of the anterior cruciate ligament; the feature extraction module extracts imaging and clinical features; the multi-modal feature fusion module uses an attention network to weight and splice heterogeneous features; and the surgery indication evaluation module generates a judgment result through a classification neural network. The application solves the problem of feature extraction interruption caused by the absence of anatomical structures in pathological states, and improves the accuracy of preoperative evaluation.

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis and artificial intelligence technology, specifically to an AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty. Background Technology

[0002] Unicompartmental knee arthroplasty is a minimally invasive procedure for treating unicompartmental osteoarthritis of the knee. Its clinical efficacy highly depends on the accurate preoperative assessment of surgical indications, particularly the extent of cartilage damage, meniscus condition, and anterior cruciate ligament functional integrity. Currently, clinical assessment mainly relies on the physician's subjective interpretation of MRI images and comprehensive analysis of clinical data. This manual assessment model is limited by the physician's experience and interpretation skills, making it difficult to guarantee the objectivity and consistency of multi-dimensional indicator determination. Furthermore, manually analyzing image sequences layer by layer and integrating clinical data is time-consuming, reducing diagnostic and treatment efficiency.

[0003] While deep learning technology has been increasingly applied to image analysis of the musculoskeletal system, existing techniques still have limitations in practical applications. First, MRI lacks absolute quantitative standards like CT values. Images generated by different manufacturers (such as Siemens and United Imaging) or using different scanning protocols exhibit significant differences in grayscale distribution and spatial resolution (domain shift). Directly inputting multi-center data into the model leads to decreased robustness of feature extraction. Second, conventional medical image segmentation algorithms typically assume the spatial existence of anatomical structures. However, in late-stage osteoarthritis cases, the anterior cruciate ligament is often missing due to rupture or absorption. In such cases, existing algorithms cannot generate effective segmentation masks, leading to interruptions in subsequent radiomics feature extraction based on the masked region. The system cannot acquire texture information of effusion or scar tissue within this critical area, resulting in lost feature input. Furthermore, existing multimodal diagnostic models often employ simple feature stitching methods when fusing image and clinical data, lacking adaptive weight adjustment mechanisms for different modal feature channels. This makes it difficult to fully exploit the differential contributions of heterogeneous data to the judgment results, limiting the accuracy of surgical indication assessment.

[0004] Therefore, this invention proposes an AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based intelligent assessment system for preoperative MRI images in unicompartmental arthroplasty, enabling automated and objective evaluation of surgical indications, thereby solving the technical problems of low efficiency, strong subjectivity, and poor consistency in traditional manual assessment methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty, comprising: The data preprocessing module is used to receive raw MRI image data and clinical text data, and process the raw MRI image data into standardized three-dimensional image data. The anatomical structure segmentation module is used to receive the three-dimensional image data and generate a segmentation mask for the anatomical structure using a deep learning semantic segmentation network. The anatomical structure segmentation module is configured to construct a cylindrical virtual region of interest based on the anatomical anchor points of the femur and tibia when the volume of the anterior cruciate ligament segmentation mask is less than a preset threshold, and generate Boolean markers for structural defects. The feature extraction module is used to extract radiomics features within the range defined by the segmentation mask or the cylindrical virtual region of interest, and to convert the clinical text data into clinical feature vectors. The multimodal feature fusion module is used to concatenate the radiomics features with the clinical feature vector, calculate the weight values ​​of the feature channels using an attention network, and multiply the weight values ​​element-wise with the concatenated feature vector to generate a fused feature vector. The surgical indication assessment module is used to input the fused feature vector into a classification neural network to generate a judgment result on whether the indications for unicompartmental arthroplasty are met.

[0007] Preferably, the anatomical structure segmentation module constructs the cylindrical virtual region of interest in the following manner: the anatomical structure segmentation module locates the femoral anchor point on the segmentation mask of the femur and the tibial anchor point on the segmentation mask of the tibia; the anatomical structure segmentation module uses the line segment connecting the femoral anchor point and the tibial anchor point as the central axis, filters out the set of voxels in space whose shortest Euclidean distance to the central axis is less than a preset radius, and defines the set of voxels as the cylindrical virtual region of interest.

[0008] Preferably, the specific method by which the anatomical structure segmentation module locates the femoral anchor point and the tibial anchor point is as follows: obtaining the minimum bounding box of the segmentation mask of the femur, truncating the central region of the minimum bounding box in the coronal plane, identifying the surface voxel of the medial wall of the lateral femoral condyle within the central region, and calculating the geometric centroid of the surface voxel as the femoral anchor point; obtaining the proximal planar surface contour of the segmentation mask of the tibia, identifying the protruding structure corresponding to the intercondylar spine of the tibia located in the central region of the tibial plateau, and extracting the coordinates of the vertex of the protruding structure as the tibial anchor point.

[0009] Preferably, the specific method by which the feature extraction module extracts the radiomics features includes: discretizing the gray values ​​in the segmentation mask or the cylindrical virtual region of interest to construct a gray-level co-occurrence matrix; calculating energy feature values ​​and contrast feature values ​​based on the normalized gray-level co-occurrence matrix; converting the structural missing Boolean marker into a numerical discrete feature, which together with the energy feature value and the contrast feature value forms a radiomics feature vector.

[0010] Preferably, the specific method by which the feature extraction module converts the clinical text data into the clinical feature vector includes: for the set of continuous data such as age, body mass index, duration of illness, and visual analog pain score, the minimum-maximum normalization method is used to map the continuous data to a numerical range of zero to one; for the set of categorical variables such as gender and side, one-hot encoding is used to convert the categorical variables into binary vectors.

[0011] Preferably, the multimodal feature fusion module performs the splicing in the following manner: using two independent fully connected layers to perform feature mapping on the radiomics features and the clinical feature vectors respectively, mapping the heterogeneous features to a unified dimensional space; concatenating the two mapped vectors along the channel dimension to construct an initial joint feature vector.

[0012] Preferably, the multimodal feature fusion module calculates the weight values ​​using the attention network in the following manner: the initial joint feature vector is compressed into the hidden layer space using the first layer weight matrix, and the hidden layer feature vector is obtained by processing it through the hyperbolic tangent activation function; the hidden layer feature vector is restored to its original dimension using the second layer weight matrix, and the weight value corresponding to each feature dimension is output by processing it through the normalization function.

[0013] Preferably, the classification neural network in the surgical indication assessment module includes an input layer, at least two hidden layers, and an output layer; the hidden layer includes a linear transformation layer for feature weighted combination, a layer normalization layer for standardizing feature distribution, and an activation function layer that introduces nonlinear factors; the output layer uses a fully connected layer to map features to scalar values, and uses a sigmoid activation function to convert the scalar values ​​into probability prediction values; the surgical indication assessment module compares the probability prediction values ​​with a preset decision threshold to generate a binarized judgment result.

[0014] Preferably, the specific method by which the data preprocessing module processes the raw MRI image data includes: mapping the grayscale distribution of the image to a unified reference distribution using a histogram matching method or a Z-score normalization method; performing linear interpolation resampling on the image using a trilinear interpolation algorithm to generate isotropic resolution three-dimensional voxel data; and performing image registration based on rigid transformation using a T1-weighted imaging sequence as a fixed reference image and a T2-weighted / PD-weighted imaging sequence as a floating image.

[0015] Preferably, the system further includes a result output module, which is configured to: when the structural missing Boolean marker represents the presence of a ligament, fill the area corresponding to the segmentation mask with a preset color using a semi-transparent solid and overlay it on the original image; when the structural missing Boolean marker represents the absence of a ligament, extract the edge contour of the cylindrical virtual region of interest and outline the edge contour using a dashed line.

[0016] This invention provides an AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty. It has the following beneficial effects: 1. This invention constructs a cylindrical virtual region of interest based on anatomical anchor points of the femur and tibia when the detected anterior cruciate ligament segmentation mask volume is too small. This technique overcomes the problem of conventional segmentation algorithms failing to generate masks when anatomical structures are broken, absorbed, or blurred, thus interrupting feature extraction. It ensures that even in the pathological state of ligament loss, the system can still perform texture analysis on residual fluid or scar tissue within the anatomical region, maintaining the integrity of feature input.

[0017] 2. This invention maps radiomics features and clinical text features to a unified dimension through a multimodal feature fusion module and uses an attention network to calculate the weight values ​​of feature channels. This mechanism enables the system to automatically adjust the weights based on the contribution of features to the classification results, effectively combining the supplementary role of clinical indicators such as patient age and BMI on image features, avoiding the limitations of single-modal information, and improving the reliability of surgical indication assessment.

[0018] 3. This invention implements histogram matching or Z-score standardization and trilinear interpolation resampling through a data preprocessing module. This standardized processing eliminates inconsistencies in grayscale distribution and spatial resolution caused by differences in imaging parameters of different MRI scanners, reduces the interference of data heterogeneity on feature extraction of deep learning models, and makes the system more robust in multi-center clinical application scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram illustrating the segmentation effect of the U-Net model of the present invention on key anatomical structures of the knee joint; Figure 4 This is a schematic diagram of the attention mechanism structure for multimodal feature fusion in this invention; Figure 5 This is a schematic diagram of the ROC curve for evaluating the performance of the system of the present invention; Figure 6 This is a schematic diagram of the anterior cruciate ligament adaptive completion and feature weighting processing flow of the present invention.

[0020] The module includes: 100, data preprocessing module; 200, anatomical structure segmentation module; 300, feature extraction module; 400, multimodal feature fusion module; 500, surgical indication assessment module; and 600, result output module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 , Figure 1 This is a schematic diagram of an AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to an embodiment of the present invention. The present invention provides an AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty, the system comprising: The data preprocessing module 100 is used to acquire and process raw MRI image data and clinical text data. Due to the differences in imaging parameters of acquisition equipment in different medical institutions, the data preprocessing module 100 first converts the acquired DICOM format images into a unified NIfTI format and performs Gaussian filtering noise reduction processing.

[0023] The data preprocessing module 100 also performs signal domain normalization operations, mapping the grayscale distribution of images from different devices to a unified reference distribution through histogram matching or Z-score normalization methods. In addition, the data preprocessing module 100 performs linear interpolation resampling on the images to generate isotropic resolution three-dimensional voxel data, and performs rigid registration of multiple sequence images based on bony landmarks to output standardized three-dimensional image data.

[0024] The anatomical structure segmentation module 200 receives standardized three-dimensional image data. The anatomical structure segmentation module 200 uses a deep learning semantic segmentation network to perform voxel-level classification of the medial femoral condyle, medial tibial plateau, meniscus, and anterior cruciate ligament in the image, and generates segmentation masks for each anatomical structure.

[0025] The anatomical structure segmentation module 200 has an adaptive completion function. When the volume of the segmented anterior cruciate ligament is less than a preset threshold, the anatomical structure segmentation module 200 constructs a cylindrical virtual region of interest in a predetermined anatomical region based on the anatomical anchor points of the femur and tibia, and marks this virtual region of interest as a structural missing region, which is then output along with other normal segmentation masks.

[0026] For image data, the feature extraction module 300 calculates first-order statistical features, morphological features, and texture features within the range defined by the segmentation mask or virtual region of interest. For clinical text data, the feature extraction module 300 converts non-image information such as age, body mass index, disease duration, and pain score into numerical vectors and performs normalized encoding.

[0027] The multimodal feature fusion module 400 is used to receive radiomics feature vectors and clinical feature vectors. The multimodal feature fusion module 400 first maps feature vectors of different dimensions to the same dimension and then concatenates them.

[0028] The multimodal feature fusion module 400 calculates the weight values ​​of each feature dimension in the concatenated feature vector using an attention network. These weight values ​​reflect the degree of contribution of different features to the determination of surgical indications. The multimodal feature fusion module 400 then multiplies the calculated weight values ​​element-wise with the original feature vector to generate a weighted fused feature vector.

[0029] The surgical indication assessment module 500 inputs the weighted fused feature vector into a classification neural network. This classification neural network maps high-dimensional features into scalar outputs through fully connected layers and converts them into probability values ​​via an activation function. The surgical indication assessment module 500 compares this probability value with a preset decision threshold to generate a binary classification result regarding whether the patient meets the indications for unicompartmental arthroplasty.

[0030] The results output module 600 summarizes the judgment results, segmentation mask image, and feature weight information. The results output module 600 overlays the segmented region or virtual completion region on the original image, generates a feature importance chart, and finally outputs a comprehensive evaluation report containing anatomical evidence and judgment conclusions.

[0031] See attached document Figure 2 , Figure 2This is a schematic flowchart of an AI-based intelligent assessment method for preoperative MRI images in unicompartmental arthroplasty according to an embodiment of the present invention. The present invention provides an AI-based intelligent assessment method for preoperative MRI images in unicompartmental arthroplasty, comprising the following steps: S100 acquires and processes raw MRI image data and clinical text data, converts the images into a unified format and performs noise reduction, normalizes the image grayscale distribution through histogram matching or Z-score normalization, and resamples the images into isotropic resolution three-dimensional voxel data. S200 uses a deep learning semantic segmentation network to perform voxel-level classification of the medial femoral condyle, medial tibial plateau, meniscus, and anterior cruciate ligament in three-dimensional voxel data to generate a segmentation mask. When the volume of the anterior cruciate ligament is detected to be less than a preset threshold, a cylindrical virtual region of interest is constructed based on the anatomical anchor points of the femur and tibia and marked as a structurally missing region. S300 calculates first-order statistical features, morphological features, and texture features within the defined range of a segmentation mask or virtual region of interest, and converts clinical text data into numerical vectors for normalization encoding to generate radiomics feature vectors and clinical feature vectors. S400 maps radiomics feature vectors and clinical feature vectors to the same dimension and concatenates them. It calculates the weight values ​​of each feature dimension in the concatenated vector through an attention network, and multiplies the weight values ​​element-wise with the original feature vector to generate a weighted fused feature vector. S500 inputs the fused feature vector into the classification neural network, maps the features into scalar outputs through a fully connected layer and converts them into probability values, compares the probability values ​​with a preset decision threshold to generate a binary classification result on whether it meets the indications for unicompartmental arthroplasty; S600 summarizes the judgment results, segmentation mask and feature weight information, overlays the segmented region or virtual completion region on the original image and generates a feature importance chart, and outputs a comprehensive evaluation report containing anatomical evidence and judgment conclusions.

[0032] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0033] In one embodiment of the present invention, the data preprocessing module 100 is used to clean, standardize, and register multi-source heterogeneous MRI image data acquired from different medical institutions. Because MRI signals are significantly affected by magnetic field homogeneity, coil sensitivity, and manufacturer-specific reconstruction algorithms, the raw data exhibits significant differences in spatial resolution and signal intensity. The data preprocessing module 100 eliminates data heterogeneity introduced by differences in acquisition equipment and imaging protocols through a series of standardization steps, constructing a data benchmark with a unified spatial scale and signal distribution.

[0034] During the data cleaning and format conversion stage, the data preprocessing module 100 receives the raw MRI image sequence in DICOM format. Since the DICOM format is primarily used for storing and transmitting two-dimensional slices and lacks a three-dimensional overall topological description, the data preprocessing module 100 reads the direction cosine matrix and pixel spacing information from the DICOM header file, reconstructing the discrete two-dimensional slices into three-dimensional volumetric data in NIfTI format. The NIfTI format can preserve the absolute position of the image in an anatomical coordinate system (such as RAS or LPS coordinate systems), ensuring spatial orientation accuracy in subsequent processing. In this process, to suppress electronic thermal noise generated during imaging and preserve the edge information of bones and soft tissues, the data preprocessing module 100 uses a three-dimensional Gaussian filter to smooth the image. This filter uses a Gaussian kernel function to perform convolution operations with the three-dimensional image data. In this embodiment, the standard deviation parameter σ of the Gaussian kernel is set in the range of 0.5 to 1.5, with a preferred value of 1.0; the convolution kernel size is set to 3×3×3 voxels to avoid blurring of small anatomical structures (such as the meniscus edge) while removing high-frequency noise.

[0035] In the signal domain normalization stage, taking into account the lack of absolute physical dimensions in MRI image grayscale values ​​similar to CT values, the data preprocessing module 100 uses the Z-score normalization algorithm to normalize the voxel intensity of the entire image. Different models of MRI equipment may exhibit order-of-magnitude differences in signal intensity ranges within the same anatomical region; this domain shift can interfere with the numerical sensitivity of deep learning models to features. The data preprocessing module 100 maps the image grayscale distribution to a standard normal distribution by calculating global statistics for a single patient's image sequence.

[0036] Specifically, for the input 3D image data Normalized image data The calculation formula is as follows: ; In the formula, Represents the original image in spatial coordinates The voxel gray value at that location; This represents the arithmetic mean of the gray values ​​of all voxels in the currently processed image sequence. This represents the standard deviation of all voxel gray values ​​in the currently processed image sequence. After this step, the mean of the background signal from images from different sources is corrected to 0, and the variance is corrected to 1, achieving relative uniformity in signal intensity.

[0037] During the spatial resampling and registration stage, the data preprocessing module 100 performs isotropic resampling. Because the original clinical scans often employ anisotropic scanning modes with thicker slices (e.g., 3mm to 5mm) and higher in-plane resolution (e.g., 0.5mm), the information density of voxels in the Z-axis direction is lower than that in the X and Y axes, affecting the accuracy of 3D texture feature calculation. The data preprocessing module 100 uses a trilinear interpolation algorithm to calculate the grayscale value at the new grid point based on the preset target spatial resolution of 1mm × 1mm × 1mm. Trilinear interpolation involves sequentially performing linear interpolation along three orthogonal coordinate axes, using the weighted sum of the grayscale values ​​of the eight adjacent voxels surrounding the target point to obtain the resampled value.

[0038] Furthermore, since T1-weighted imaging sequences (emphasizing bone structures) and T2-weighted / PD-weighted imaging sequences (emphasizing synovial fluid and ligaments) are typically performed in separate scans, patients may experience minor positional shifts between scans. The data preprocessing module 100 performs image registration based on rigid transformation. In this embodiment, the data preprocessing module 100 sets the T1-weighted imaging sequence with well-defined bony structures as a fixed reference image and the T2-weighted / PD-weighted imaging sequence as a floating image. The registration algorithm uses normalized mutual information (NMI) as a similarity measure function, iteratively optimizing a rigid transformation matrix containing three translation parameters and three rotation parameters until the mutual information value of the two images reaches its maximum. This process ensures that signals from the same anatomical point in different modalities can accurately correspond to the same voxel coordinates, eliminating spatial misalignment caused by positional changes.

[0039] In one embodiment of the present invention, the anatomical structure segmentation module 200 uses deep learning semantic segmentation technology to process the standardized three-dimensional image data, realize the automated localization and extraction of key anatomical structures, and execute an adaptive completion strategy for the absence of anterior cruciate ligament in pathological conditions.

[0040] The anatomical structure segmentation module 200 first inputs the preprocessed 3D voxel data into an improved 3D U-Net (3D U-Net) deep neural network. This network architecture includes an encoder path and a decoder path: the encoder path consists of multiple consecutive 3×3×3 convolutional layers, batch normalization layers, and max-pooling layers, used to progressively extract deep semantic features from the image and compress spatial dimensions; the decoder path gradually restores the spatial resolution of the feature map through upsampling operations and stitches high-resolution texture features from the corresponding encoder layers into the decoder feature map through skip connections to compensate for the spatial information loss caused by downsampling.

[0041] like Figure 3As shown, this U-Net network structure, through multi-level feature extraction and skip connections, can accurately reconstruct the three-dimensional morphology of the femur, tibia, and anterior cruciate ligament from multimodal images, outputting a high-precision segmentation mask. Considering that the anterior cruciate ligament occupies a very small proportion of the overall knee joint image, to address the model convergence difficulties caused by extreme class imbalance between foreground and background, this embodiment uses the Dice coefficient loss function as the optimization objective for network training. The Dice loss function mainly focuses on the degree of overlap between the predicted region and the ground truth region, and its calculation formula is defined as follows: ; In the formula, This represents the total number of voxels in the input 3D image block; Indicates the network prediction of the first The probability value of an individual element belonging to the target anatomical structure (i.e., femur, tibia, meniscus or anterior cruciate ligament), with a value range of [0,1]. Indicates the first The gold standard true label value corresponding to the individual element, which takes the value 0 or 1; For the smoothing term, the value in this embodiment is 1×10. -5 Its function is to prevent the denominator from being zero and to improve the numerical stability of gradient calculation. The trained model outputs probability maps of each anatomical structure and generates a binarized segmentation mask by setting a threshold (e.g., 0.5).

[0042] See attached document Figure 6 The anatomical structure segmentation module 200 then performs an integrity check and adaptive completion process for the anterior cruciate ligament (ACL). Because severe osteoarthritis may be accompanied by ACL rupture, absorption, or liquefaction, the ligament mask output by the segmentation network may be extremely small or even empty. Directly extracting features based on this would cause subsequent texture analysis to fail. In this embodiment, the system first calculates the physical volume of the ACL segmentation mask. : ; In the formula, The physical volume of a single voxel (e.g., 1 mm) 3 ); This is a binary mask. The system compares this volume with a preset minimum effective volume threshold. Compare. The threshold. It was determined through statistical analysis of a pre-constructed dataset of knee MRI data from healthy individuals, for example, by taking the 5th percentile or mean value of the anterior cruciate ligament volume distribution in the dataset minus twice the standard deviation (μ). 2σ) to ensure that the vast majority of normal variation ranges are covered.

[0043] when At that time, the system determined that the ligament structure was intact and preserved the original segmentation mask. When the system determines that a ligament structure is missing, it triggers an adaptive virtual region of interest (ROI) construction procedure. This procedure, based on the rigid topological relationships of knee joint anatomy, uses a stabilized segmented femoral and tibial mask to infer the spatial location where the anterior cruciate ligament should exist.

[0044] Specifically, in order to accurately locate the anatomical anchor point, the anatomical structure segmentation module 200 executes the following geometric search algorithm: First, locate the femoral anchor point. By obtaining the minimum bounding box of the femoral segmentation mask, the central 1 / 3 region of this bounding box in the coronal plane (corresponding to the intercondylar fossa region) is extracted. Within this sub-region, surface voxels of the medial wall of the lateral femoral condyle are identified, and the geometric centroid of this set of surface voxels is calculated as... Secondly, locate the tibial anchor point. By acquiring the surface contour of the proximal plane of the tibial segmentation mask, a protruding structure located in the central region of the tibial plateau (corresponding to the intercondylar spine of the tibia) is identified, and the coordinates of the vertex of this protruding structure are extracted as... .

[0045] Subsequently, the system uses line segments Construct a cylindrical virtual region of interest with the central axis as the axis. This region is represented by the following set of voxels: ; In the formula, Represents spatial voxel coordinates; Voxel representation The shortest Euclidean distance to the line segment connecting the two anchor points. This distance constraint ensures that the virtual region is limited to the two bones and does not extend indefinitely. The preset radius is set to 5mm in this embodiment.

[0046] Furthermore, to transform the qualitative fact of "ligament loss" into a quantitative feature that can be processed by the subsequent neural network, the anatomical structure segmentation module 200 generates a Boolean-type label variable, Missing_Flag. When virtual region construction is triggered, Missing_Flag is assigned a value of 1; otherwise, it is assigned a value of 0. This label variable is then transmitted to the feature extraction module 300 as a key index to distinguish between normal anatomical structures and virtual completed structures.

[0047] In one embodiment of the present invention, the feature extraction module 300 is connected to the anatomical structure segmentation module 200, and is used to convert high-dimensional three-dimensional image data and unstructured clinical text data into computer-processable quantitative feature vectors. This feature extraction module 300 uses radiomics analysis technology to map the biological heterogeneity of anatomical tissues into a multidimensional feature set in numerical space.

[0048] In the conventional radiomics extraction steps, the feature extraction module 300 calculates first-order statistical features and three-dimensional morphological features for the three-dimensional region of interest (ROI) defined by the segmentation mask generated by the anatomical structure segmentation module 200. The first-order statistical features are statistical descriptions based on voxel gray-level histograms, without considering the spatial relationships between voxels. In this embodiment, the feature extraction module 300 calculates the mean, variance, skewness, and kurtosis of all voxel gray-level values ​​within the ROI to quantify the signal intensity level and the dispersion of its distribution within the anatomical tissue. For the three-dimensional morphological features, the feature extraction module 300 calculates the physical volume, maximum three-dimensional diameter, and surface area-to-volume ratio of each anatomical structure based on the physical spacing between voxels. Particularly for the meniscus structure, to numerically characterize whether there are edge irregularities due to wear or tear, this embodiment calculates "morphological compactness." The calculation formula is as follows: ; In the formula, Represents the physical volume of the region of interest in the meniscus; This represents the surface area of ​​the region of interest. This feature is dimensionless; when the meniscus edge becomes rough or fragmented due to degeneration, its surface area relative to volume increases significantly, leading to… The value decreased.

[0049] In the texture feature extraction step, the feature extraction module 300 performs Gray-Level Co-occurrence Matrix (GLCM) analysis. GLCM is a second-order statistical texture analysis method used to describe the spatial correlation characteristics of brightness values ​​in an image. This step is crucial for the evaluation of the anterior cruciate ligament (ROI), especially when dealing with "virtual regions of interest" (ROIs). Since ROI dysfunction in unicompartmental arthroplasty contraindications is often accompanied by joint effusion, this region appears as a high-signal, homogeneous fluid-like feature on T2-weighted images, while normal ligaments appear as low-signal features with fibrous, cord-like textures. To eliminate noise interference and reduce computational load, before constructing the matrix, the feature extraction module 300 first discretizes the gray values ​​within the ROI, using an equal-width binning method to quantize the original continuous gray values. There are discrete gray levels. In this embodiment, The setting is 32 levels. Subsequently, the feature extraction module 300 constructs a gray-level co-occurrence matrix. To ensure the rotation invariance of the features, the calculation parameters are set as follows: spatial distance. Voxels; directions encompass 13 principal directions in three-dimensional space (i.e., combinations of 0°, 45°, 90°, and 135° on different cut surfaces), and the final texture feature value is the arithmetic mean of the calculation results of these 13 directions.

[0050] Based on the normalized gray-level co-occurrence matrix, the feature extraction module 300 calculates two core texture parameters: "energy" and "contrast." The energy feature (also known as the second moment of the angle) is used to measure the uniformity of the texture, and its calculation formula is as follows: ; In the formula, Indicates the number of gray levels after quantization (32 levels); Represents grayscale level and The joint probability value in the normalized gray-level co-occurrence matrix satisfies When processing virtual regions of interest, if the region is uniformly filled with synovial fluid, the grayscale distribution is highly concentrated, and the diagonal elements of the matrix have large values, the calculated energy value will be significantly higher than that of normal ligament tissue. Contrast features are used to measure the local changes in matrix values ​​and their distribution in the image; their calculation formula is as follows: ; In the formula, This represents the squared weight of the gray-level difference between two pixels. When there are complex fibrous tears or uneven scar tissue in the region of interest, the gray-level difference between adjacent pixels is large, and the calculated contrast value will be significantly increased. Furthermore, the feature extraction module 300 directly receives structural missing Boolean tags from the anatomical structure segmentation module 200. These tags are converted into numerical discrete features (0 or 1) and used as prior information to be concatenated into the radiomics feature vector, prompting the model to pay attention to the current anatomical integrity state in subsequent fusion stages.

[0051] In the clinical data quantification step, the feature extraction module 300 encodes the patient's age, body mass index (BMI), visual analog scale (VAS) pain score, and disease duration data. Since these data have different dimensions, directly inputting them into the neural network would lead to unstable gradient updates. The feature extraction module 300 uses a min-max normalization method to map continuous clinical variables to the [0,1] interval. The normalized clinical variables... The calculation formula is as follows: ; In the formula, These are the original clinical data values; and These represent the minimum and maximum bounding values ​​of this type of clinical indicator in a pre-defined reference dataset. The reference dataset is derived from a historical case database used during model training. For example, in this embodiment, for the target population for unicompartmental arthroplasty, the age indicator... Set to age 40 Set to 90 years old; BMI indicator Set to 15kg / m 2 , Set to 40kg / m 2 For categorical variables such as gender and laterality (left knee / right knee), the feature extraction module 300 uses one-hot encoding to convert them into binary vectors. Finally, the feature extraction module 300 outputs a multidimensional feature vector set containing both radiomics and clinical features.

[0052] In one embodiment of the present invention, the multimodal feature fusion module 400 is used to receive radiomics feature vectors and clinical feature vectors from the feature extraction module 300, and to achieve effective fusion of heterogeneous data through an adaptive weighting mechanism based on channel attention.

[0053] To address the differences in feature dimensions and numerical distribution among different modalities, the multimodal feature fusion module 400 first performs feature mapping and concatenation operations. Since radiomics features and clinical features are extracted from different data sources, direct concatenation would lead to discontinuous feature space distribution, affecting the convergence performance of subsequent classifiers. In this embodiment, the multimodal feature fusion module 400 utilizes two independent deep fully connected layers to process the input image feature vectors. and clinical feature vector Feature mapping is performed. Both fully connected layers are configured with the ReLU activation function to map heterogeneous features to a unified dimensional space. In this embodiment, The dimensions are set to 128 or 256. Subsequently, the system concatenates the two mapped vectors along the channel dimension to construct an initial joint feature vector. .at this time, The total dimension is 2D.

[0054] like Figure 4 As shown in the figure, this visually illustrates the process of stitching together radiomics features and clinical features after mapping to a unified dimension to construct fused features. Based on this fusion structure, after the initial feature stitching is completed, the multimodal feature fusion module 400 introduces a channel attention network to establish the interdependencies between feature channels and dynamically adjust the response weights of each feature dimension. This network adopts a "compression-excitation" bottleneck structure design to capture the nonlinear correlations between channels while reducing parameter computation. Specifically, the attention weight vector... The calculation process includes two steps: dimensionality reduction encoding and dimensionality increase decoding. The calculation formula is as follows: ; ; In the formula, This represents the initial joint feature vector with dimension 2D; This is the first layer weight matrix, with dimension 1. This is used to compress the feature space into the hidden layer space; In this embodiment, the reduction ratio is set to 8 or 16 to control the degree of compression of the bottleneck layer. For dimension The bias vector; These are the feature vectors of the hidden layer; It is the hyperbolic tangent activation function. This is the second-layer weight matrix, with dimensions of... This is used to restore features from the hidden layer space to the original dimension; A bias vector of dimension 2D; The normalization function ensures the output weight vector The sum of all elements in the distribution is 1, thus forming a probability distribution that represents the relative importance of each feature dimension to the final decision.

[0055] After generating the attention weights, the multimodal feature fusion module 400 performs a feature recalibration operation. The system utilizes the calculated attention weight vector... For the initial joint eigenvector Element-wise weighting is performed to generate the final fused feature vector Z. The calculation formula for this process is as follows: ; In the formula, This represents the Hadamard product, which is the element-wise multiplication. Through this operation, the model can automatically enhance the numerical response of highly relevant features based on the specificity of the input data. For example, when an anterior cruciate ligament (ACL) defect is detected and marked as 1, the weight of relevant texture features is automatically increased, while redundant noise features are suppressed. This fused feature vector Z is then transmitted to the surgical indication assessment module 500 as the input benchmark for the final binary classification decision.

[0056] In one embodiment of the present invention, the surgical indication assessment module 500 is connected to the multimodal feature fusion module 400, which receives the weighted and calibrated fusion feature vector Z and converts it into a quantitative probability score reflecting whether the patient is suitable for unicompartmental arthroplasty through nonlinear mapping, and finally outputs a binarized classification suggestion.

[0057] The surgical indication assessment module 500 constructs a classification network based on a multilayer perceptron. To handle the high-dimensional nonlinear relationship of fused feature vectors, this embodiment adopts a pyramid-shaped decreasing network architecture, aiming to abstract high-level semantic features and compress information dimensions step by step. Specifically, the classification network includes an input layer, at least two hidden layers, and an output layer. The number of nodes in the first hidden layer is set to half the input dimension, and the number of nodes in the second hidden layer is further halved, thus forming a gradual contraction of feature channels. Each fully connected layer block sequentially includes: a linear transformation layer, used to weight and combine the input features; a layer normalization layer, used to standardize the feature distribution of a single sample across all channels, making it meet the distribution characteristics of mean 0 and variance 1, thereby preventing gradient vanishing or exploding in deeper layers of the network; an activation function layer (ReLU or GeLU is selected in this embodiment), used to introduce nonlinear factors; and a random deactivation layer. During the model training phase, the random deactivation layer randomly sets the output of some neurons to zero with a preset probability p (set to 0.5 in this embodiment) to block the cooperative adaptation between neurons and enhance the generalization ability of the model; during the inference phase, this layer is automatically turned off, and all neurons participate in the calculation.

[0058] After deep feature extraction by the classification network, the last fully connected layer maps the high-dimensional features to a one-dimensional scalar value, namely the log odds. Since the range of this log odds is (… The scalar value (∞, +∞) cannot be directly interpreted as a probability. Therefore, the surgical indication assessment module 500 uses the Sigmoid activation function to map this scalar to the (0, 1) interval. Probability prediction value The calculation formula is as follows: ; In the formula, is the base of the natural logarithm; This is the feature vector output by the last hidden layer in the classification network; The weight vector of the output layer has the same dimension as... The dimensions match; For the scalar bias term of the output layer; This represents the confidence level probability that the model predicts the current patient meets the criteria for unicompartmental arthroplasty.

[0059] After obtaining the probability prediction value, the surgical indication assessment module 500 executes binary classification logic. The system presets a decision threshold. The threshold It is determined based on the Receiver Operating Characteristic (ROC) curve on the validation set during the model training phase. This embodiment calculates the Youden index on the ROC curve and selects the probability value corresponding to the point of maximum Youden index as the optimal cutoff threshold to balance the model's sensitivity and specificity. The decision logic is as follows: If... If the surgical indication assessment module 500 outputs a first judgment signal (e.g., a value "1"), it determines that the patient's knee joint condition meets the surgical indications for unicompartmental knee arthroplasty; if If the condition is not met, a second judgment signal (e.g., a value of "0") will be output, indicating that the patient has anterior cruciate ligament loss, lateral compartment cartilage wear, or other contraindications and does not meet the surgical indications for unicompartmental arthroplasty.

[0060] In one embodiment of the present invention, the result output module 600 is connected to the surgical indication assessment module 500 and the multimodal feature fusion module 400, and is used to convert the numerical algorithm calculation results into visualized auxiliary diagnostic information.

[0061] S601, in the anatomical structure visualization step, the result output module 600 performs a rendering operation based on multi-layer image fusion. The system receives the original MRI grayscale image and a binary segmentation mask generated by the anatomical structure segmentation module 200. To clearly display the spatial location of the region of interest while preserving the original anatomical details, this embodiment uses alpha blending technology to overlay the segmentation mask onto the original image. The overlaid display pixel values... The calculation formula is as follows: ; In the formula, The vector is the gray value of the original MRI image at a specific coordinate point mapped to the RGB color space, meaning that the values ​​of the R, G, and B channels are all equal to the original gray value of that point. The preset mask pseudo-color RGB values ​​are used, for example, in this embodiment, the femur is set to green encoding value [0,255,0] and the tibia is set to blue encoding value [0,0,255]; This is the transparency factor, used to control the see-through effect of the mask. In this embodiment, Set it to 0.3 to 0.5 to ensure that the underlying trabecular bone texture features are still visible while the mask coverage area is displayed.

[0062] For the anterior cruciate ligament (ACL) region, the result output module 600 executes differentiated rendering logic based on the "structural missing Boolean flag". If the flag is "0", it indicates that the ligament is present, and the system uses a preset color to fill the segmented region with a semi-transparent solid. If the flag is "1", it indicates that the ligament is missing, and the system calls the coordinate data of the virtual region of interest, uses a morphological gradient algorithm to extract the edge contour of the virtual region, and uses a red dashed line to outline the contour, without filling the interior. This differentiated display strategy is used to distinguish between the solid tissue and the virtual completed region.

[0063] S602, in the attribution analysis chart generation step, the result output module 600 is based on the attention weight vector output by the multimodal feature fusion module 400. The system generates a visualization chart of feature importance. Since radiomics features and clinical features are concatenated in a fixed order during feature extraction and fusion, there is a definite mapping relationship between the index positions in the feature vector and their specific physical meanings. This embodiment pre-defines an index mapping table, dividing the index range of the high-dimensional feature vector into different semantic categories. For example, indices 0-9 correspond to "age and BMI," indices 10-41 correspond to "ACL texture features," and indices 42-73 correspond to "meningeal morphological features," etc. The system calculates the aggregate importance score for each semantic category. The calculation formula is as follows: ; In the formula, This represents a set of feature dimension indices belonging to a specific semantic category; Indicates the first The normalized weights for each feature dimension are calculated in the attention network. After calculation, the system sorts the importance scores of each category in descending order and selects the top-ranked categories. A histogram of attribution factors was plotted for key factors attribution. In this embodiment, Set to 5. This chart is used to quantitatively display the main sources of features that influence the model's judgment results.

[0064] S603, in the comprehensive report generation step, the result output module 600 calls a preset report template and structures the generated MRI key-level images with segmentation rendering, the attribution analysis histogram, and the first or second judgment signal and probability confidence level output by the surgical indication assessment module 500. Finally, the system generates a comprehensive diagnostic report containing basic patient information, imaging analysis atlases, and judgment results, and stores or transmits it in PDF or DICOM structured report (SR) format.

[0065] To further verify the technical effectiveness and clinical application value of the AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty provided by this invention, this embodiment constructs a complete verification environment and provides a detailed description of the system's workflow and performance in conjunction with specific clinical cases and statistical data.

[0066] In this embodiment, the dataset for system training and validation was derived from real retrospective clinical cases collected between January 2018 and December 2023. This dataset included 1200 patients scheduled for unicompartmental knee arthroplasty (UKA), totaling 1560 knee joints. All enrolled cases had complete knee MRI images and corresponding clinical data; the MRI images included sagittal T1WI, sagittal PDWI / T2WI, coronal PDWI / T2WI, and axial PDWI / T2WI sequences, while the clinical data included age, sex, BMI, symptom duration, preoperative VAS pain score, and preoperative KSS functional score.

[0067] To establish the "gold standard" for model training, this embodiment employs a rigorous "double-blind judgment + arbitration" data annotation mechanism. Specifically, two senior joint surgeons with over 10 years of experience independently determine whether the MRI images and clinical data of each case meet the indications for UKA surgery, referring to the "Chinese Clinical Guidelines for Unicompartmental Joint Replacement (2021 Edition)". For cases with inconsistent judgments, a chief physician with over 20 years of experience provides final arbitration. This annotation process ensures the clinical authority and consistency of the training labels, providing a reliable benchmark for subsequent supervised learning.

[0068] Based on the dataset constructed above, this embodiment performs targeted training on the core learning module of the system.

[0069] For the anatomical structure segmentation module 200, this embodiment selected 840 cases (i.e., 1092 knees) from the dataset as the training set and 360 cases (i.e., 468 knees) as the test set. During training, an improved U-Net network architecture was adopted, using expert-drawn anatomical contours as the gold standard, and the Dice coefficient loss function was used for iterative optimization. Training results show that the mean Dice similarity coefficient (DSC) of each anatomical structure is greater than 0.88. Specifically, for the relatively fixed morphology of the medial femoral condyle, its DSC reaches 0.94±0.03; for the anterior cruciate ligament, which has large morphological variations and blurred boundaries under pathological conditions, its DSC reaches 0.88±0.06. These results indicate that the segmentation module can effectively overcome the effects of MRI slice resolution differences and part of the volume effect, achieving high-precision voxel-level classification.

[0070] For the surgical indication assessment module 500, this embodiment uses the same training set to test set ratio as the segmentation module, i.e., 7:3. The classification neural network in this module receives the fused feature vector output by the multimodal feature fusion module 400 as input. During training, this embodiment uses the Adam optimizer for parameter updates, with an initial learning rate set to 0.001. To avoid the model getting trapped in local minima and improve generalization ability, this embodiment introduces a cosine annealing strategy to dynamically adjust the learning rate. The training epochs are set to 100, and an early stopping strategy is configured: if the loss function value on the validation set does not decrease within 10 consecutive epochs, training is automatically terminated and the current optimal weights are saved. This training strategy effectively prevents overfitting and ensures that the model maintains stable predictive performance when processing unseen clinical data.

[0071] The following example, using a 58-year-old female patient (Case-0823), illustrates the automated decision-making logic of the system in this embodiment: Step 1: Data Access The system retrieved the patient's MRI images and clinical data. Clinical data showed the patient's BMI was 26.5 kg / m². 2 The VAS score is 7.

[0072] Step Two: Structural Analysis The anatomical segmentation module 200 identifies the segmented volume of the anterior cruciate ligament region. 120mm 3 The system compares this value with the preset minimum effective volume threshold. A comparison was made; in this embodiment, Set to 350mm 3 The judgment result is... The system then triggers a structural missing marker and automatically constructs a cylindrical virtual region of interest.

[0073] Step 3: Feature Fusion After extracting the texture features of the virtual region, the feature data shows that the region has high energy values ​​and low contrast values. The multimodal feature fusion module 400 assigns high weight coefficients to the texture features and the "structural missing Boolean marker" through an attention mechanism. In this embodiment, the key feature is... .

[0074] Step 4: Decision Output The surgical indication assessment module 500 calculates a probability score indicating that the patient meets the surgical indications. The value is 0.12. This is because the value is below the preset decision threshold. In this embodiment, The value was set to 0.55. The system output a "Does not meet the indications" result and generated an auxiliary diagnostic report with a red dotted line warning. This AI judgment result was consistent with the fact that the "anterior cruciate ligament was completely ruptured" was found during clinical intraoperative exploration.

[0075] Performance evaluation: The performance of the system in this embodiment was statistically evaluated in an independent test set containing 360 cases (468 knees). Experimental results showed that the model achieved an accuracy of 94.2%, a sensitivity of 92.8%, and a specificity of 95.3% in the binary classification task. Figure 5 As shown, by comparing the ROC curves of the training set and the independent test set, it can be seen that the multimodal joint model proposed in this invention outperforms the single-modal model in terms of AUC, indicating that the fused features have higher diagnostic efficacy. The area under the receiver operating characteristic curve (AUC) is 0.976, indicating that the model has extremely high discriminative ability.

[0076] Furthermore, the consistency between the model and human experts was assessed by calculating the Kappa consistency coefficient. The results showed that the Kappa coefficient between the model's judgment and that of senior physicians was 0.873 (P<0.001), with no statistically significant difference (P>0.05). Simultaneously, the model's accuracy was significantly better than the average level of junior physicians (P<0.05). These data confirm that this invention, by integrating multimodal features and an adaptive anatomical completion mechanism, can provide auxiliary diagnostic suggestions at the level of senior experts, effectively compensating for the shortcomings of primary healthcare institutions in assessing complex cases.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty, characterized in that, include: The data preprocessing module is used to receive raw MRI image data and clinical text data, and process the raw MRI image data into standardized three-dimensional image data. The anatomical structure segmentation module is used to receive the three-dimensional image data and generate a segmentation mask for the anatomical structure using a deep learning semantic segmentation network. The anatomical structure segmentation module is configured to construct a cylindrical virtual region of interest based on the anatomical anchor points of the femur and tibia when the volume of the anterior cruciate ligament segmentation mask is less than a preset threshold, and generate Boolean markers for structural defects. The feature extraction module is used to extract radiomics features within the range defined by the segmentation mask or the cylindrical virtual region of interest, and to convert the clinical text data into clinical feature vectors. The multimodal feature fusion module is used to concatenate the radiomics features with the clinical feature vector, calculate the weight values ​​of the feature channels using an attention network, and multiply the weight values ​​element-wise with the concatenated feature vector to generate a fused feature vector. The surgical indication assessment module is used to input the fused feature vector into a classification neural network to generate a judgment result on whether the indications for unicompartmental arthroplasty are met.

2. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 1, characterized in that, The specific method by which the anatomical structure segmentation module constructs the cylindrical virtual region of interest is as follows: The anatomical structure segmentation module locates the femoral anchor point on the segmentation mask of the femur and the tibial anchor point on the segmentation mask of the tibia. The anatomical structure segmentation module uses the line segment connecting the femoral anchor point and the tibia anchor point as the central axis, filters out the set of voxels in space whose shortest Euclidean distance to the central axis is less than a preset radius, and defines the set of voxels as the cylindrical virtual region of interest.

3. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 2, characterized in that, The specific method by which the anatomical structure segmentation module locates the femoral anchor point and the tibial anchor point is as follows: Obtain the minimum bounding box of the segmentation mask of the femur, extract the central region of the minimum bounding box in the coronal plane, identify the surface voxels of the medial wall of the lateral condyle of the femur within the central region, and calculate the geometric centroid of the surface voxels as the femoral anchor point. Obtain the proximal planar surface contour of the segmentation mask of the tibia, identify the protruding structure corresponding to the intercondylar spine of the tibia located in the central region of the tibial plateau, and extract the coordinates of the vertex of the protruding structure as the tibial anchor point.

4. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 1, characterized in that, The specific methods by which the feature extraction module extracts the radiomics features include: The gray values ​​of the segmentation mask or the cylindrical virtual region of interest are discretized to construct a gray-level co-occurrence matrix; Energy and contrast feature values ​​are calculated based on the normalized gray-level co-occurrence matrix; the structural missing Boolean markers are converted into numerical discrete features, which, together with the energy and contrast feature values, form an image omics feature vector.

5. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 1, characterized in that, The specific methods by which the feature extraction module transforms the clinical text data into the clinical feature vector include: For the set of continuous data including age, body mass index, duration of illness, and visual analog scale pain score, the min-max normalization method is used to map the continuous data to a numerical range of zero to one. For the categorical variables of gender and side, one-hot encoding is used to convert the categorical variables into binary vectors.

6. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 1, characterized in that, The specific method by which the multimodal feature fusion module performs the splicing is as follows: Two independent fully connected layers are used to perform feature mapping on the radiomics features and the clinical feature vectors, respectively, to map heterogeneous features to a unified dimensional space; The two mapped vectors are concatenated along the channel dimension to construct an initial joint feature vector.

7. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 6, characterized in that, The multimodal feature fusion module calculates the weight values ​​using the attention network in the following specific way: The initial joint feature vector is compressed into the hidden layer space using the first layer weight matrix, and then processed by the hyperbolic tangent activation function to obtain the hidden layer feature vector; The hidden layer feature vector is restored to its original dimension using the second layer weight matrix, and then processed by a normalization function to output the weight value corresponding to each feature dimension.

8. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 1, characterized in that, The classification neural network in the surgical indication assessment module includes an input layer, at least two hidden layers, and an output layer. The hidden layer includes a linear transformation layer for feature weighted combination, a layer normalization layer for standardizing feature distribution, and an activation function layer that introduces nonlinear factors. The output layer uses a fully connected layer to map features into scalar values, and uses a Sigmoid activation function to convert the scalar values ​​into probabilistic predicted values; the surgical indication assessment module compares the probabilistic predicted values ​​with a preset decision threshold to generate a binarized judgment result.

9. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 1, characterized in that, The specific methods by which the data preprocessing module processes the raw MRI image data include: The grayscale distribution of the image is mapped to a uniform baseline distribution using histogram matching or Z-score normalization methods. The image was resampled using a trilinear interpolation algorithm to generate isotropic resolution three-dimensional voxel data. Using T1-weighted imaging sequences as fixed reference images and T2-weighted / PD-weighted imaging sequences as floating images, image registration is performed based on rigid transformation.

10. The AI-based intelligent judgment system for preoperative MRI images in unicompartmental arthroplasty according to claim 1, characterized in that, The system also includes a result output module, which is configured to: When the missing Boolean marker of the structure represents the presence of a ligament, a preset color is used to fill the area corresponding to the segmentation mask with a semi-transparent solid and overlay it on the original image. When the Boolean flag indicating structural absence represents ligament absence, the edge contour of the cylindrical virtual region of interest is extracted, and the edge contour is outlined using a dashed line.

Images