3D ultrasonic uterus malformation intelligent auxiliary diagnosis method and device based on multi-source feature fusion
By employing a 3D ultrasound diagnostic method for uterine malformations that integrates multi-source feature fusion, and combining deep learning and radiomics features, this method overcomes the shortcomings of existing technologies that rely on physician experience and single AI modalities, achieving highly accurate and efficient diagnosis of uterine malformations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE FIRST PEOPLES HOSPITAL OF FOSHAN
- Filing Date
- 2026-04-16
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies in 3D ultrasound diagnosis rely heavily on doctors' experience, making it easy to miss or misdiagnose cases. Furthermore, single-modality radiomics and deep learning methods suffer from overfitting and loss of subtle texture information, and lack highly sensitive automated screening and diagnostic tools.
A multi-source feature fusion method is adopted, which combines deep learning and image omics features. Through feature concatenation and machine learning classifiers, a high-dimensional joint feature vector is obtained and then dimensionality is reduced to output prediction results and early warning reminders.
It achieves highly accurate diagnosis of uterine malformations, reduces computational resource consumption, is suitable for deployment in primary hospitals, solves the overfitting problem under small sample data, and improves robustness and processing efficiency.
Smart Images
Figure CN122392885A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to a 3D ultrasound intelligent auxiliary diagnosis method and device for uterine malformations based on multi-source feature fusion. Background Technology
[0002] Currently, congenital uterine malformations (CUA) are a significant cause of female infertility and recurrent miscarriage. Clinical diagnosis primarily relies on three-dimensional ultrasound (3D-US) coronal imaging. However, current technology has the following significant limitations: Reliance on physician experience (highly subjective): The reconstruction and interpretation of the coronal plane of 3D ultrasound is highly dependent on the experience of the ultrasound physician. In primary hospitals or when dealing with complex malformations (such as the differentiation between incomplete septum and arcuate uterus), it is easy to miss or misdiagnose.
[0003] Limitations of Single AI Modalities: The shortcomings of pure radiomics: Traditional radiomics features are extremely sensitive to image noise and prone to overfitting, meaning they perform well on the training set but their accuracy drops significantly on new data. The shortcomings of pure deep learning: While using only convolutional neural networks (such as ResNet) offers good robustness, it is prone to losing subtle texture and edge information that is crucial for identifying minute deformities.
[0004] Clinical pain point: For high-risk groups such as infertility, there is currently a lack of highly sensitive automated screening and diagnostic tools that can simultaneously combine overall morphology and local texture features. Summary of the Invention
[0005] The purpose of this invention is to at least address one of the shortcomings of the prior art and provide a 3D ultrasound intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: Specifically, a 3D ultrasound intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion is proposed, including the following: A set of 3D ultrasound images of the patient's uterus is obtained, and the 3D ultrasound images are fused to obtain a 3D ultrasound image to be detected. The 3D ultrasound image to be detected is converted into a preset format and the ROI region is delineated to obtain a processed 3D ultrasound image. Depth feature vector and radiomics feature vector were obtained by performing depth feature extraction and radiomics feature extraction on the processed 3D ultrasound images, respectively. The deep feature vector and the radiomics feature vector are fused to obtain a high-dimensional joint feature vector, and the high-dimensional joint feature vector is then reduced in dimensionality to obtain a subset of key features. The subset of key features is input into a pre-trained machine learning classifier, which then makes a decision to obtain a prediction result. The prediction results will be visualized and output, and early warnings will be issued.
[0007] Furthermore, specifically, a set of 3D ultrasound images of the patient's uterus is acquired, and the 3D ultrasound images are fused to obtain the 3D ultrasound image to be detected, including... N 3D ultrasound images of the patient's uterus were acquired under the same conditions to form a 3D ultrasound image set. Then, all 3D ultrasound images in the 3D ultrasound image set are scanned synchronously under the same conditions. It is assumed that the pixel values of the N images corresponding to the same pixel point at each scanning step size movement constitute a set {pix_1, pix_2, ..., pix_N}, and Max{pix_1, pix_2, ..., pix_N} is used as the pixel value of the pixel point at that position. The image formed by combining the updated pixel values of all locations after the scan is completed is the fused 3D ultrasound image to be detected.
[0008] Furthermore, specifically, depth feature extraction is performed on the processed 3D ultrasound image to obtain a depth feature vector, including: The processed 3D ultrasound image is input into a pre-trained convolutional neural network, and the deep feature vector before the fully connected layer is extracted to obtain the deep feature vector, which represents the global geometric structure and high-level semantic information of the uterine cavity.
[0009] Furthermore, specifically, radiomics feature vectors are obtained by extracting radiomics features from the processed 3D ultrasound images, including: High-throughput computation was performed on the processed 3D ultrasound images to extract radiomics feature vectors, including first-order statistics, texture features, shape features, transformation features, Hessian matrix features, topological features, and fractal features. These feature vectors characterize the microscopic texture heterogeneity and local connectivity of uterine tissue.
[0010] Furthermore, specifically, the methods for fusing the deep feature vector and the radiomics feature vector to obtain a high-dimensional joint feature vector include serial splicing, feature interaction in the middle layer of the network, or multi-channel superposition at the input end.
[0011] Furthermore, specifically, the high-dimensional joint feature vector is reduced in dimensionality using LASSO regression to obtain a subset of key features.
[0012] Furthermore, specifically, pre-trained machine learning classifiers include LightGBM, SVM, XGBoost, or Random Forest.
[0013] Furthermore, specifically, the prediction results include the probability values of a normal uterus and uterine malformations, and an early warning is issued when the probability value of uterine malformations exceeds a preset probability threshold.
[0014] This invention also proposes a 3D ultrasound intelligent auxiliary diagnostic device for uterine malformations based on multi-source feature fusion, which applies the steps of the above method. The device includes the following: The data acquisition module is used to acquire a set of 3D ultrasound images of the patient's uterus, fuse the 3D ultrasound images to obtain a 3D ultrasound image to be detected, convert the 3D ultrasound image to be detected into a preset format and delineate the ROI region to obtain a processed 3D ultrasound image. The feature extraction module is used to extract depth features and radiomics features from the processed 3D ultrasound images to obtain depth feature vectors and radiomics feature vectors, respectively. The feature fusion module is used to fuse the deep feature vector and the radiomics feature vector to obtain a high-dimensional joint feature vector, and to reduce the dimensionality of the high-dimensional joint feature vector to obtain a key feature subset. The result prediction module is used to input the subset of key features into a pre-trained machine learning classifier, and the machine learning classifier makes a decision to obtain the prediction result. The result output module is used to visualize the prediction results and provide early warnings.
[0015] The beneficial effects of this invention are as follows: This invention proposes a 3D ultrasound intelligent auxiliary diagnostic method and device for uterine malformations based on multi-source feature fusion. On the one hand, considering that deep learning features are good at capturing overall shape and radiomics is good at capturing subtle textures, the method achieves complementary macroscopic and microscopic information through feature stitching, resulting in more accurate prediction results. Furthermore, the overall computational efficiency is high, and clinical implementation is easy. Compared to complex dual-stream neural network training, this invention adopts an architecture of "feature extraction + stitching + LightGBM classification," which consumes less computational resources, runs faster, and is more suitable for deployment on ordinary computers in primary hospitals, solving the overfitting problem under small sample conditions. Addressing the characteristic of limited data (small sample size) for infertility uterine malformations, the method directly stitches features and uses a machine learning classifier, which is more stable and robust than training a large end-to-end network (test set AUC remains at 0.923). On the other hand, the 3D ultrasound images input to the model are preprocessed before feature extraction. By acquiring a set of multiple 3D ultrasound images under the same conditions and fusing them, a 3D ultrasound image with globally optimal pixel values is obtained, making the subsequent feature extraction results more accurate. The entire process only requires one scan of all images, resulting in high processing efficiency. Attached Figure Description
[0016] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 The diagram shown illustrates the structural principle of the feature processing network used in the 3D ultrasound intelligent auxiliary diagnosis method for uterine malformations based on multi-source feature fusion, as described in this invention. Figure 2 The figure shows the ROC curves of the proposed method model, the simple deep learning model (DL), and the simple omics model (Rad) on the training set. Figure 3 The figure shows the ROC curves of the proposed method model, the simple deep learning model (DL), and the simple omics model (Rad) on the validation set. Detailed Implementation
[0017] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.
[0018] Example 1, referring to Figure 1 This invention proposes a 3D ultrasound intelligent auxiliary diagnosis method for uterine malformations based on multi-source feature fusion, including the following: A set of 3D ultrasound images of the patient's uterus is obtained, and the 3D ultrasound images are fused to obtain a 3D ultrasound image to be detected. The 3D ultrasound image to be detected is converted into a preset format and the ROI region is delineated to obtain a processed 3D ultrasound image. Depth feature vector and radiomics feature vector were obtained by performing depth feature extraction and radiomics feature extraction on the processed 3D ultrasound images, respectively. The deep feature vector and the radiomics feature vector are fused to obtain a high-dimensional joint feature vector, and the high-dimensional joint feature vector is then reduced in dimensionality to obtain a subset of key features. The subset of key features is input into a pre-trained machine learning classifier, which then makes a decision to obtain a prediction result. The prediction results will be visualized and output, and early warnings will be issued.
[0019] In this embodiment 1, on the one hand, considering that deep learning features are good at capturing overall shape and radiomics is good at capturing subtle textures, the macroscopic and microscopic information is complemented by feature stitching, resulting in more accurate prediction results. Moreover, the overall computational efficiency is high and it is easy to implement in clinical practice. Compared with complex dual-stream neural network training, this invention adopts an architecture of "feature extraction + stitching + LightGBM classification", which consumes less computational resources, runs faster, and is more suitable for deployment on ordinary computers in primary hospitals, solving the overfitting problem under small sample conditions. For the characteristic of small data volume (small sample) of uterine malformation in infertility, features are directly stitched together and a machine learning classifier is used, which is more stable and robust than training a large end-to-end network (the AUC on the test set is maintained at 0.923). On the other hand, the 3D ultrasound images of the input model are preprocessed before feature extraction. By collecting a set of 3D ultrasound images under the same conditions, the images are fused to obtain the 3D ultrasound images with the global optimal pixel values, so that the subsequent feature extraction results are more accurate. Moreover, the whole process only requires scanning the entire image once, which is highly efficient.
[0020] In a preferred application, the specific process is as follows: Step 1: Data Acquisition and ROI Determination. Acquire 3D ultrasound images of the patient's uterus, convert them into nii.gz format, and have the doctor delineate the region of interest (ROI) within the uterine cavity.
[0021] Step 2: Independent Feature Extraction. This system performs feature extraction on the same ROI image using two different dimensions: Deep feature extraction: The ROI is input into a pre-trained convolutional neural network (ResNet50) to extract deep feature vectors before the fully connected layers. These feature vectors represent the global geometric structure and high-level semantic information of the uterine cavity.
[0022] Radiomics feature extraction: High-throughput computation is performed on the ROI to extract features including first-order statistics, texture features (GLCM, etc.), shape features, transform features (wavelet transform), Hessian matrix features (for edge enhancement), topological features, and fractal features. This feature vector characterizes the microscopic texture heterogeneity and local connectivity of uterine tissue.
[0023] Step 3: Feature Concatenation and Dimensionality Reduction. The "deep feature vector" and "radiomics feature vector" obtained in Step 2 are concatenated sequentially to construct a high-dimensional joint feature vector containing multimodal information. Due to the excessively high dimensionality after concatenation, LASSO regression is used to reduce the dimensionality of the joint feature vector, eliminating redundant features and retaining the most discriminative key feature subset.
[0024] Step 4: Classification and Diagnosis. The filtered joint feature vector is input into the LightGBM machine learning classifier. The classifier makes a decision based on the fused features, outputting the probability value that the case belongs to "normal uterus" or "uterine malformation".
[0025] Step 5: Output Results. Output the diagnostic results and confidence levels, and provide early warnings for high-risk cases (e.g., probability > 0.35).
[0026] See also Figure 2 as well as Figure 3 The method proposed in this invention was compared and verified with conventional methods, and the results showed that it is significantly superior to conventional methods.
[0027] In a preferred embodiment of the present invention, specifically, a set of 3D ultrasound images of the patient's uterus is acquired, and the 3D ultrasound image set is fused to obtain the 3D ultrasound image to be detected, including... N 3D ultrasound images of the patient's uterus were acquired under the same conditions to form a 3D ultrasound image set. Then, all 3D ultrasound images in the 3D ultrasound image set are scanned synchronously under the same conditions. It is assumed that the pixel values of the N images corresponding to the same pixel point at each scanning step size movement constitute a set {pix_1, pix_2, ..., pix_N}, and Max{pix_1, pix_2, ..., pix_N} is used as the pixel value of the pixel point at that position. The image formed by combining the updated pixel values of all locations after the scan is completed is the fused 3D ultrasound image to be detected.
[0028] As a preferred embodiment of the present invention, the image fusion is performed through the above process. Before feature extraction, the 3D ultrasound image of the input model is preprocessed. By acquiring a set of multiple 3D ultrasound images under the same conditions, the fusion is performed to obtain a 3D ultrasound image with the global optimal pixel value, so that the subsequent feature extraction results are more accurate. Moreover, the whole process only requires scanning the entire image once, which has high processing efficiency.
[0029] In a preferred embodiment of the present invention, specifically, depth feature extraction is performed on the processed 3D ultrasound image to obtain a depth feature vector, including... The processed 3D ultrasound image is input into a pre-trained convolutional neural network, and the deep feature vector before the fully connected layer is extracted to obtain the deep feature vector, which represents the global geometric structure and high-level semantic information of the uterine cavity.
[0030] Specifically, radiomics feature vectors are obtained by extracting radiomics features from the processed 3D ultrasound images, including: High-throughput computation was performed on the processed 3D ultrasound images to extract radiomics feature vectors, including first-order statistics, texture features, shape features, transformation features, Hessian matrix features, topological features, and fractal features. These feature vectors characterize the microscopic texture heterogeneity and local connectivity of uterine tissue.
[0031] In this preferred embodiment, the features extracted by deep learning are fused with radiomics features including Hessian / topological features. Applying this specific combination of "ResNet deep features + radiomics features (including Hessian matrix, topological features, fractal features, etc.)" in the diagnosis of uterine malformations yields more accurate prediction results.
[0032] In 3D ultrasound diagnosis of uterine malformations, single-type features often have information limitations. The core reason for fusing deep learning features with hand-designed radiomics features lies in the complementarity of information dimensions.
[0033] Deep features, derived from the deep abstraction of convolutional neural networks, excel at capturing the overall geometric structure and macroscopic semantic logic of the uterine cavity. However, due to the reduced spatial resolution caused by pooling operations in deep networks, these features are often not sensitive enough to subtle textural heterogeneity within tissues. Radiomics features, on the other hand, are quantitative indicators based on rigorous mathematical definitions, capable of accurately characterizing microscopic pathological changes within tissues. Through high-throughput computing, these features can reveal voxel spatial distribution patterns that are imperceptible to the human eye. Combining the two preserves deep learning's ability to model complex anatomical structures while introducing microscopic descriptors with clear physical meaning, thereby significantly improving the classifier's accuracy in identifying uterine developmental abnormalities.
[0034] The following are the extraction logic and calculation definitions of various radiomics features: First-order statistical features: These features describe the distribution of pixel intensity within the uterine cavity ROI region, without considering spatial relationships. Significance: Reflects the overall brightness level and contrast of the uterine myometrium or uterine cavity filling material. Calculation method: Based on grayscale histogram. Perform calculations. Common indicators include the mean. Variance, skewness, and entropy.
[0035] Texture features: Primarily extracted using the Gray-Level Co-occurrence Matrix (GLCM) or Gray-Level Region Size Matrix (GLSZM). Significance: Characterizes the fineness, roughness, or regularity of uterine cavity tissue. Calculation method: Taking GLCM as an example, it calculates the pixel pair under a specific displacement d and angle θ. probability of occurrence Further, indicators such as energy, contrast, and correlation were derived.
[0036] Shape characteristics: Geometric measurements are performed based on the volume of the delineated 3D ROI of the uterine cavity. Significance: This is crucial for diagnosing uterine malformations, directly reflecting the depth of the uterine contour's indentation and the overall symmetry of its shape. Calculation method: Calculation of volume V, surface area A, and sphericity. .
[0037] Transformation Features: Wavelet transform or Laplacian operator is typically used to decompose the original image at multiple scales. Significance: It can effectively remove ultrasonic noise and capture the edge information of the Region of Interest (ROI) at multiple resolutions. Calculation Method: Wavelet basis functions are used to perform high-frequency and low-frequency filtering on the image to extract features from different frequency subbands.
[0038] Hessian matrix features: Describes the local curvature of an image using a second-order partial derivative matrix. Significance: Extracts linear, tubular, or spherical structures through eigenvalue combinations, often used to enhance uterine cavity boundaries or identify subtle passages. Calculation method: Constructing the Hessian matrix. Calculate its eigenvalues , , .
[0039] Topological characteristics: Analysis based on persistent homology or spatial connectivity components. Significance: Quantifying the connectivity complexity of the uterine cavity structure, which is of unique value for identifying complex fusion malformations. Calculation method: Calculating the Betti number of the image level set at different thresholds, and statistically analyzing the evolution of holes and connected components.
[0040] Fractal features: Describe the self-similarity and space-filling ability of the ROI structure. Calculation method: The most commonly used is the fractal dimension, which is calculated by determining the variable length required to cover the ROI. Number of boxes Using the formula Solve this problem. Significance: It represents the morphological complexity of the uterine cavity; a higher fractal dimension usually indicates a more disordered tissue structure.
[0041] As a preferred embodiment of the present invention, the method of fusing the deep feature vector and the radiomics feature vector to obtain a high-dimensional joint feature vector includes serial splicing, feature interaction in the middle layer of the network, or multi-channel superposition at the input end.
[0042] In a preferred embodiment of the present invention, specifically, the high-dimensional joint feature vector is reduced in dimensionality using LASSO regression to obtain a subset of key features.
[0043] In this preferred embodiment, considering that the dimensionality is too high after splicing, LASSO regression is used to reduce the dimensionality of the joint feature vector, remove redundant features, and retain the most discriminative key feature subset.
[0044] As a preferred embodiment of the present invention, the pre-trained machine learning classifier includes LightGBM, SVM, XGBoost, or Random Forest.
[0045] In this preferred embodiment, LightGBM has been proven to be the best predictor. The classifier can be replaced with SVM, XGBoost or Random Forest, but the performance may vary slightly.
[0046] In a preferred embodiment of the present invention, the prediction result specifically includes the probability values of a normal uterus and uterine malformation, and when the probability value of uterine malformation exceeds a preset probability threshold, an early warning is issued.
[0047] Example 2: This invention also proposes a 3D ultrasound intelligent auxiliary diagnostic device for uterine malformations based on multi-source feature fusion, which applies the steps of the above method. The device includes the following: The data acquisition module is used to acquire a set of 3D ultrasound images of the patient's uterus, fuse the 3D ultrasound images to obtain a 3D ultrasound image to be detected, convert the 3D ultrasound image to be detected into a preset format and delineate the ROI region to obtain a processed 3D ultrasound image. The feature extraction module is used to extract depth features and radiomics features from the processed 3D ultrasound images to obtain depth feature vectors and radiomics feature vectors, respectively. The feature fusion module is used to fuse the deep feature vector and the radiomics feature vector to obtain a high-dimensional joint feature vector, and to reduce the dimensionality of the high-dimensional joint feature vector to obtain a key feature subset. The result prediction module is used to input the subset of key features into a pre-trained machine learning classifier, and the machine learning classifier makes a decision to obtain the prediction result. The result output module is used to visualize the prediction results and provide early warnings.
[0048] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.
[0049] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.
[0050] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.
[0051] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.
Claims
1. A 3D ultrasound intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion, characterized in that, Including the following: A set of 3D ultrasound images of the patient's uterus is obtained, and the 3D ultrasound images are fused to obtain a 3D ultrasound image to be detected. The 3D ultrasound image to be detected is converted into a preset format and the ROI region is delineated to obtain a processed 3D ultrasound image. Depth feature vector and radiomics feature vector were obtained by performing depth feature extraction and radiomics feature extraction on the processed 3D ultrasound images, respectively. The deep feature vector and the radiomics feature vector are fused to obtain a high-dimensional joint feature vector, and the high-dimensional joint feature vector is then reduced in dimensionality to obtain a subset of key features. The subset of key features is input into a pre-trained machine learning classifier, which then makes a decision to obtain a prediction result. The prediction results will be visualized and output, and early warnings will be issued.
2. The intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion using 3D ultrasound according to claim 1, characterized in that, Specifically, a set of 3D ultrasound images of the patient's uterus is acquired, and the 3D ultrasound images are fused to obtain the 3D ultrasound image to be detected, including... N 3D ultrasound images of the patient's uterus were acquired under the same conditions to form a 3D ultrasound image set. Then, all 3D ultrasound images in the 3D ultrasound image set are scanned synchronously under the same conditions. It is assumed that the pixel values of the N images corresponding to the same pixel point at each scanning step size movement constitute a set {pix_1, pix_2, ..., pix_N}, and Max{pix_1, pix_2, ..., pix_N} is used as the pixel value of the pixel point at that position. The image formed by combining the updated pixel values of all locations after the scan is completed is the fused 3D ultrasound image to be detected.
3. The intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion using 3D ultrasound according to claim 1, characterized in that, Specifically, depth feature vectors are obtained by extracting depth features from the processed 3D ultrasound images, including: The processed 3D ultrasound image is input into a pre-trained convolutional neural network, and the deep feature vector before the fully connected layer is extracted to obtain the deep feature vector, which represents the global geometric structure and high-level semantic information of the uterine cavity.
4. The intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion using 3D ultrasound according to claim 3, characterized in that, Specifically, radiomics feature vectors are obtained by extracting radiomics features from the processed 3D ultrasound images, including: High-throughput computation was performed on the processed 3D ultrasound images to extract radiomics feature vectors, including first-order statistics, texture features, shape features, transformation features, Hessian matrix features, topological features, and fractal features. These feature vectors characterize the microscopic texture heterogeneity and local connectivity of uterine tissue.
5. The intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion using 3D ultrasound according to claim 4, characterized in that, Specifically, the methods for fusing the deep feature vector and the radiomics feature vector to obtain a high-dimensional joint feature vector include serial splicing, feature interaction in the middle layer of the network, or multi-channel superposition at the input end.
6. The intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion using 3D ultrasound according to claim 1, characterized in that, Specifically, the high-dimensional joint feature vector is reduced in dimensionality using LASSO regression to obtain a subset of key features.
7. The intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion using 3D ultrasound according to claim 1, characterized in that, Specifically, pre-trained machine learning classifiers include LightGBM, SVM, XGBoost, or Random Forest.
8. The intelligent auxiliary diagnostic method for uterine malformations based on multi-source feature fusion using 3D ultrasound according to claim 1, characterized in that, Specifically, the prediction results include the probability values of a normal uterus and uterine malformations, and an early warning is issued when the probability value of uterine malformations exceeds a preset probability threshold.
9. A 3D ultrasound intelligent auxiliary diagnostic device for uterine malformations based on multi-source feature fusion, characterized in that, The apparatus comprising the steps of the method according to any one of claims 1-8, wherein the method is applied, the apparatus includes the following: The data acquisition module is used to acquire a set of 3D ultrasound images of the patient's uterus, fuse the 3D ultrasound images to obtain a 3D ultrasound image to be detected, convert the 3D ultrasound image to be detected into a preset format and delineate the ROI region to obtain a processed 3D ultrasound image. The feature extraction module is used to extract depth features and radiomics features from the processed 3D ultrasound images to obtain depth feature vectors and radiomics feature vectors, respectively. The feature fusion module is used to fuse the deep feature vector and the radiomics feature vector to obtain a high-dimensional joint feature vector, and to reduce the dimensionality of the high-dimensional joint feature vector to obtain a key feature subset. The result prediction module is used to input the subset of key features into a pre-trained machine learning classifier, and the machine learning classifier makes a decision to obtain the prediction result. The result output module is used to visualize the prediction results and provide early warnings.