Mri three-dimensional surface joint segmentation method, system, electronic device and readable storage medium

Abstract

The present application relates to a kind of MRI three-dimensional surface joint segmentation method, system, electronic equipment and readable storage medium, the method is first with the subject identification, examination identification and sequence name as the grouping key construction orthogonal three-dimensional surface image group, different dimensional surface image identification is independent and not as cross-dimensional surface pairing condition;Follow the subject level data isolation principle and divide training set and test set, prevent data leakage.Parameter sharing segmentation network is handled three-dimensional surface image, through the mask guided foreground weighting pooling extraction each dimensional surface foreground semantic prototype, without cross-dimensional surface pixel level registration condition, the consistency loss between prototype is calculated and network training is constrained.Inference stage output region of interest mask and map back to original DICOM space to calculate morphological parameter and imageomic feature.The present application reduces the fluctuation of three-dimensional surface segmentation performance, retains more stable imageomic feature in subject level sensitivity analysis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an MRI three-dimensional joint segmentation method, system, electronic device, and readable storage medium. Background Technology

[0002] Magnetic resonance imaging (MRI) offers advantages such as a large field of view and excellent soft tissue contrast. For the same subject and in the same examination, it can acquire images in different planes (i.e., imaging planes), including sagittal, coronal, and transverse / axial views, as well as different imaging sequences such as HASTE, TRUFI, and T2-weighted images. Clinical studies have shown that the morphology, thickness, volume, signal characteristics, and texture features of the region of interest (ROI) can be used to evaluate the physiological and pathological state and health status of related organs. Radiomics methods can extract high-dimensional quantitative features such as morphology, first-order grayscale, and texture from the ROI, which can assist in clinical assessment and diagnosis.

[0003] Currently, the annotation masks for regions of interest mainly rely on manual layer-by-layer or image-by-image drawing. This method is not only time-consuming and labor-intensive, but also easily affected by the operator's experience. Differences in the drawing boundaries can further affect morphological parameters such as area and thickness, as well as the extraction results of downstream radiomics features. While using deep learning for automatic segmentation can improve processing efficiency, existing methods still have several problems:

[0004] (1) The basis for cross-plane pairing is unreasonable: Image numbers, slice numbers or derived numbers of different planes are usually only valid within their respective sequences. There is no strict one-to-one correspondence between the images of the three orthogonal planes of sagittal, coronal and transverse. If the same image ID is used as the condition for cross-plane pairing, valid data will be incorrectly excluded, and images with the same number by chance may be misjudged as the corresponding anatomical images.

[0005] (2) There is a risk of data leakage at the level of the tested object: If the training set and the validation set are randomly divided by a single image, a single sequence or a combination of three planes, different plane images or different sequence images of the same tested object may appear in the training set and the validation set at the same time, so that the validation results are affected by individual-specific anatomical morphology, scanning parameters and noise characteristics, resulting in the model's generalization ability being overestimated.

[0006] (3) The single-plane training paradigm cannot explicitly utilize the three-plane association: Existing methods mostly treat each two-dimensional image as an independent training sample, or directly sum and optimize the segmentation loss of the three planes. The latter is essentially still a joint optimization of three independent segmentation tasks, which cannot guarantee the consistency of the high-level semantic representation of the three planes of the same object.

[0007] (4) Simple pixel-level consistency constraints cannot be applied to different planes: Sagittal, sagittal and transverse planes come from different viewing directions, and the pixel coordinates, target shape and area of ​​the image are all different. Directly subtracting the prediction mask or feature map pixel by pixel requires strict registration and may incorrectly suppress reasonable plane-specific features.

[0008] (5) Lack of evaluation system: Existing automatic segmentation results are often evaluated only by geometric indicators such as single image Dice, intersection-over-union ratio (IoU) or mean average precision (mAP), lacking the evaluation dimension of the three-dimensional segmentation performance balance of the same subject, and also lacking the verification of whether the automatic ROI (i.e. the region of interest mask obtained by the automatic segmentation algorithm) can retain downstream radiomics information.

[0009] (6) Disconnect between geometric accuracy evaluation and downstream quantitative stability: Existing methods usually select models based solely on geometric indices such as Dice and IoU. However, segmentation models with better geometric indices do not necessarily retain more stable radiomic features, and selecting models based solely on geometric indices may compromise the reliability of subsequent quantitative analysis. Summary of the Invention

[0010] The purpose of this invention is to provide a method, system, electronic device, and readable storage medium for joint segmentation of three MRI planes, aiming to solve problems such as inconsistent image numbering across different planes, data leakage at the subject level, failure of single-plane models to utilize intra-subject correlations, inability to directly register pixels across planes, and insufficient stability of downstream radiomics features in automatic ROI.

[0011] To achieve the above objectives, the present invention provides a method for joint three-dimensional segmentation of MRI, comprising:

[0012] MRI images and their metadata for the same subject, the same examination, and the same sequence targeting the same region of interest are acquired. Orthogonal three-dimensional image groups are constructed using the subject identifier, examination identifier, and sequence name as grouping keys. Each orthogonal three-dimensional image group includes one sagittal image, one coronal image, and one transverse image for the same subject, the same examination, and the same sequence targeting the same region of interest. The image identifiers of different planes are independent of each other and are not used as cross-plane pairing conditions.

[0013] Several orthogonal three-dimensional image groups were divided into training set, validation set and test set according to the identification of the inspected object;

[0014] The segmentation network that shares the input parameters of each facet image in the orthogonal three-facet image group is used to obtain the segmentation prediction and feature map of each facet image.

[0015] The segmentation loss of each facet image is calculated based on the segmentation prediction results and the corresponding labeled mask;

[0016] Foreground semantic prototypes of each facet image are extracted from the feature map by mask-guided foreground weighted pooling, and the consistency loss between the foreground semantic prototypes of each facet image is calculated.

[0017] The segmentation loss of each facet image is weighted and summed with the consistency loss to obtain the total loss, and the segmentation network is iteratively optimized and trained based on the total loss;

[0018] The trained segmentation network is used to automatically segment the MRI image under test to obtain a region of interest mask, and the region of interest mask is mapped back to the original DICOM space to output morphological parameters and radiomics features.

[0019] Optionally, when constructing the orthogonal three-dimensional image group, for multiple candidate images of the same surface in the same sequence of the same subject, the image with the largest effective region of interest area is selected as the representative slice of that surface.

[0020] Optionally, when dividing several orthogonal 3D image groups into training set, validation set and test set according to the object being inspected, all orthogonal 3D images of the same object being inspected are placed in the same subset of the training set, validation set and test set.

[0021] Optionally, a feature enhancement module is provided before the backbone network or segmentation head of the parameter-sharing segmentation network.

[0022] Optionally, when extracting the foreground semantic prototype of each facet image from the feature map through mask-guided foreground weighted pooling, the labeled mask or high-confidence prediction mask is scaled to be aligned with the size of the feature map, and the features of the foreground region are weighted by the aligned mask values ​​to obtain the foreground semantic prototype of each facet image.

[0023] Optionally, the calculation of the consistency loss between the foreground semantic prototypes of each polygon image is selected from any one of the following or a combination thereof:

[0024] Calculate the cosine distances between each pair of foreground semantic prototypes in the sagittal, coronal, and transverse images of the orthogonal three-dimensional image group and take the average.

[0025] First, calculate the mean of the foreground semantic prototypes of the sagittal, coronal, and transverse images in the orthogonal three-plane image group as a shared prototype, and then constrain the foreground semantic prototypes of each plane image to be close to the shared prototype.

[0026] After mapping the foreground semantic prototypes of each plane image through a plane-specific projection layer, the smooth L1 distance or mean square error between the mapped prototypes is calculated.

[0027] A comparative learning approach that uses the foreground semantic prototypes of three images from the same sequence of the same subject as positive samples and the foreground semantic prototypes of different subjects as negative samples.

[0028] Optionally, during the inference phase, the MRI image to be tested includes a single-plane image or a group of orthogonal three-plane images of the same subject.

[0029] When the input is a single-plane image, the output is the region of interest mask and segmentation confidence of that plane image;

[0030] When the input is an orthogonal three-dimensional image group of the same object under inspection, the region of interest mask of each image is output, and at least one consistency index between the three-dimensional images is output, selected from the Dice difference within the three-dimensional image group, the correlation coefficient within the group, the symmetry relative error and its combination.

[0031] Based on the same inventive concept, the present invention also provides an MRI three-dimensional joint segmentation system, comprising:

[0032] The data acquisition module is configured to acquire MRI images and their metadata for the same subject, the same examination, and the same sequence targeting the same region of interest.

[0033] The grouping key construction module is configured to construct an orthogonal three-dimensional image group using the subject identifier, the examination identifier, and the sequence name as grouping keys. The orthogonal three-dimensional image group includes one sagittal image, one coronal image, and one transverse image of the same subject, the same examination, and the same sequence for the same region of interest. The image identifiers of different planes are independent of each other and are not used as cross-plane pairing conditions.

[0034] The data partitioning module is configured to divide several orthogonal three-dimensional image groups into training set, validation set and test set according to the identification of the inspected object;

[0035] The joint training module is configured to share the input parameters of each facet image in the orthogonal three-facet image group with the segmentation network to obtain the segmentation prediction and feature map of each facet image; calculate the segmentation loss of each facet image based on the segmentation prediction result and the corresponding labeled mask; extract the foreground semantic prototype of each facet image from the feature map through mask-guided foreground weighted pooling, and calculate the consistency loss between the foreground semantic prototypes of each facet image; and sum the segmentation loss of each facet image with the consistency loss to obtain the total loss, and iteratively optimize and train the segmentation network based on the total loss.

[0036] The inference output module is configured to automatically segment the MRI image under test using the trained segmentation network, obtain a region of interest mask, map the region of interest mask back to the original DICOM space, and output morphological parameters and radiomics features.

[0037] Based on the same inventive concept, the present invention also provides an electronic device, comprising:

[0038] Memory, which stores computer programs;

[0039] The processor, which is communicatively connected to the memory, executes the MRI three-dimensional joint segmentation method as described above when calling the computer program;

[0040] The display, communicatively connected to the processor and the memory, is used to display a GUI interface related to the MRI three-dimensional joint segmentation method.

[0041] Based on the same inventive concept, the present invention also provides a readable storage medium having a computer program stored thereon, which, when executed, can implement the MRI three-dimensional joint segmentation method as described above.

[0042] The MRI three-dimensional joint segmentation method, system, electronic device, and readable storage medium provided by this invention have at least one of the following beneficial effects:

[0043] (1) By constructing an orthogonal three-dimensional image group using the subject identifier, examination identifier and sequence name as grouping keys, and clarifying that the image identifiers of different planes are independent of each other and not used as cross-plane pairing conditions, the image numbers of different imaging planes are avoided as having an anatomical correspondence, which is in line with the imaging logic of MRI multi-plane scanning and improves the medical rationality of training sample construction.

[0044] (2) By dividing several orthogonal three-dimensional image groups according to the identification of the tested object, it is ensured that all data of the same tested object exists only in the same subset of the training set, validation set and test set. This tested object-level data isolation strategy effectively reduces the impact of tested object information leakage on generalization evaluation and helps to more realistically reflect the generalization ability of the model.

[0045] (3) By using a parameter-sharing segmentation network to process the three-dimensional images of the orthogonal three-dimensional image group, and introducing the consistency loss calculation of the foreground semantic prototype, the network can utilize the association information of different viewing directions of the same subject during training. This constraint at the feature level achieves cross-plane consistency constraint of the subject through the foreground semantic prototype without the need for cross-plane registration and pixel-level correspondence.

[0046] (4) During the inference stage, this invention not only supports segmentation of single-plane images, but also outputs consistency indicators such as intra-plane Dice difference, intra-plane correlation coefficient, and symmetry relative error when the input is an orthogonal three-plane image group of the same subject. These indicators provide clinical users with a quantitative reference for the stability of segmentation results, which helps to identify potential low-confidence segmentation situations and improves the credibility of automatic segmentation results in clinical applications. Attached Figure Description

[0047] Those skilled in the art will understand that the accompanying drawings are provided to better understand the invention and do not constitute any limitation on the scope of the invention. Wherein:

[0048] Figure 1 This is a flowchart of an MRI three-dimensional joint segmentation method provided in an embodiment of the present invention;

[0049] Figure 2 This is a representative comparison image of automatic segmentation results provided by an embodiment of the present invention, manually labeled results, and nnU-Net results.

[0050] Figure 3 The representative consistency analysis results provided in an embodiment of the present invention are obtained after extracting radiomics features from the same original DICOM image using a manually annotated region of interest mask and the automatic region of interest mask of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0052] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0053] In the description of this invention, it should be understood that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0054] Furthermore, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0055] Please refer to Figure 1 This invention provides a three-dimensional joint segmentation method for MRI, comprising the following steps:

[0056] S1. Obtain MRI images and their metadata for the same subject, the same examination, and the same sequence targeting the same region of interest. Construct an orthogonal three-dimensional image group using the subject identifier, examination identifier, and sequence name as grouping keys. The orthogonal three-dimensional image group includes one sagittal image, one coronal image, and one transverse image for the same subject, the same examination, and the same sequence targeting the same region of interest. The image identifiers of different planes are independent of each other and are not used as cross-plane pairing conditions.

[0057] S2. Divide several orthogonal three-dimensional image groups into training set, validation set and test set according to the identification of the inspected object;

[0058] S3. Input each facet image in the orthogonal three-facet image group into a segmentation network with shared parameters to obtain segmentation predictions and feature maps for each facet image;

[0059] S4. Calculate the segmentation loss of each facet image based on the segmentation prediction results and the corresponding labeled mask;

[0060] S5. Extract the foreground semantic prototypes of each facet image from the feature map through mask-guided foreground weighted pooling, and calculate the consistency loss between the foreground semantic prototypes of each facet image.

[0061] S6. The segmentation loss and consistency loss of each face image are weighted and summed to obtain the total loss, and the segmentation network is iteratively optimized and trained based on the total loss;

[0062] S7. The trained segmentation network is used to automatically segment the MRI image under test to obtain the region of interest mask, and the region of interest mask is mapped back to the original DICOM space to output morphological parameters and radiomics features.

[0063] This invention provides a three-dimensional joint segmentation method for MRI. Its core idea is not based on a one-to-one correspondence between pixels or image numbers on the three planes, but rather on establishing three-dimensional relationships between sagittal, coronal, and transverse images of the same region of interest (ROI) under the same subject, examination, and sequence. Foreground semantic prototypes of each plane image are obtained through mask-guided foreground feature pooling, and consistency constraints are applied in a shared semantic space. This method is particularly suitable for high-precision joint segmentation and quantitative parameter extraction of ROIs with complex anatomical structures (such as the placenta, tumors, and uterus).

[0064] First, execute S1 to acquire MRI images and their metadata for the same subject, the same examination, and the same sequence targeting the same region of interest. Then, construct an orthogonal three-dimensional image group using the subject identifier, examination identifier, and sequence name as grouping keys. The orthogonal three-dimensional image group includes one sagittal image, one coronal image, and one transverse image for the same subject, the same examination, and the same sequence targeting the same region of interest. The image identifiers of different planes are independent of each other and are not used as cross-plane pairing conditions.

[0065] In this embodiment, the image can be DICOM, NIfTI, JPG, PNG, or a two-dimensional image exported from DICOM; manual ROI annotations (i.e., annotation masks) can also be obtained during the training phase. Metadata mainly refers to DICOM header file information, which includes key parameters such as image spatial resolution (Pixel Spacing), layer thickness, and imaging orientation. To construct training samples, this invention introduces the concept of "orthogonal three-dimensional image group". Unlike simply stacking three images, this invention emphasizes using the subject identifier (e.g., medical record number), examination identifier (e.g., examination number), and sequence name as grouping keys. This means that the system will accurately select one sagittal, coronal, and transverse image each belonging to the same subject, the same examination, and the same sequence, and combine them into a training sample. It is worth noting that this invention emphasizes that the image identifiers (e.g., SOP InstanceUID) of different planes are independent of each other and are not used as cross-plane pairing conditions. This prevents the algorithm from erroneously attempting to align images of different planes at the pixel level, but rather guides the algorithm to fuse at the feature semantic level. For example, an abdominal MRI examination may contain hundreds of transverse images, but this invention selects only the most representative transverse image and combines it with a corresponding sagittal and coronal image to form a structured input unit.

[0066] In constructing an orthogonal three-dimensional image group, when multiple candidate images exist for the same object and the same sequence within the same region (e.g., a transverse section), this invention provides a preferred selection strategy: selecting the image with the largest effective region of interest (ROI) area as the representative slice for that region. In specific implementation, the system can traverse all transverse section images within the same sequence, using simple thresholding or edge detection algorithms to initially estimate the area of ​​the ROI in each image, and then select the image with the largest area. This is because the slice with the largest area typically contains the most complete anatomical information of the ROI, providing the richest information for representing the semantic prototype of that region.

[0067] Alternatively, the image with the largest coarsely segmented region of interest can be selected. This involves first obtaining a coarse segmentation result quickly using a pre-trained model, and then selecting the slice with the largest coarse segmentation area. Alternatively, the image with the highest target integrity score can be selected, which comprehensively considers factors such as edge sharpness and signal uniformity. Or, the image with the highest image quality score can be selected, based on metrics such as signal-to-noise ratio and contrast. Those skilled in the art can also construct a weighted combined scoring function incorporating the above metrics to select the optimal representative slice, based on actual needs. This selection mechanism ensures that each image input into the network is the most informative "essence" within that dimension, thereby improving the accuracy and robustness of subsequent segmentation.

[0068] For example, let the object under inspection be identified as P, the inspection identifier as E, the sequence name as S, and the geoplanar image p∈{sag, cor,tra}. A group of three orthogonal geoplanar images is represented as:

[0069]

[0070] in These are representative images selected from the same subject, the same examination, and the same sequence, targeting the same region of interest. The image IDs of the three representative images can be different.

[0071] The slice selection function can be represented as:

[0072]

[0073] Where C p Let be the set of candidate images for the plane image p, and Q(I) be the representative score. Q(I) can be the artificial ROI area, the coarse segmented ROI area, the target integrity score, the image quality score, or a weighted combination of the above indicators.

[0074] In this embodiment, only the group that simultaneously contains sagittal, coronal, and transverse images and whose image annotations for each plane are valid is used as the orthogonal three-plane image group for subsequent training; incomplete groups can be used only for unit plane segmentation supervision and do not participate in cross-plane consistency loss.

[0075] Next, step S2 is executed, dividing several orthogonal 3D image groups into training, validation, and test sets according to the object identifier. To ensure the realism and generalization ability of the model evaluation, this invention strictly adheres to the "individual-level isolation" principle when dividing the training, validation, and test sets. Specifically, the system divides several orthogonal 3D image groups according to the object identifier. For example, 70% of all data for the object (including its corresponding orthogonal 3D image group) is assigned to the training set, 15% to the validation set, and the remaining 15% to the test set. Crucially, all orthogonal 3D images of the same object must be placed in the same subset. This partitioning method effectively prevents data leakage. If different 3D images of the same object are randomly assigned to different subsets, the model may acquire some information from the test set during training due to the high correlation between these images, leading to inflated evaluation results. Through individual-level isolation, this invention can more realistically simulate the model's performance on unseen objects, ensuring the objectivity and scientific rigor of the evaluation metrics.

[0076] Next, step S3 is executed, inputting each plane image from the orthogonal three-plane image group into a parameter-sharing segmentation network to obtain segmentation predictions and feature maps for each plane image. One of the core architectures of this invention is the use of a parameter-sharing segmentation network to process the orthogonal three-plane image group. In specific implementations, the sagittal, coronal, and transverse images are each input into the same segmentation network. This network is preferably based on a YOLO11-seg architecture, but can also be any deep learning model suitable for medical image segmentation, such as U-Net or Mask R-CNN. Parameter sharing means that the weights of the network's backbone and segmentation head are shared during the image processing of the three planes. This design forces the network to learn a universal region of interest feature representation that is independent of the imaging orientation, reducing the number of model parameters, lowering the risk of overfitting, and improving training efficiency.

[0077] To further enhance the feature extraction capability of the network, this invention preferably places a feature enhancement module before the backbone network or the segmentation head. For example, the DeBiFormerPlus module can be used. This module, based on dynamic routing or sparse attention, establishes adaptive information interaction between local details and global context. This module focuses on enhancing the expression of region of interest boundaries, internal textures, and surrounding organizational relationships, and can be placed after the backbone network, in the feature pyramid, or before the segmentation head.

[0078] Next, S4 is executed to calculate the segmentation loss of each facet image based on the segmentation prediction results and the corresponding labeled masks. Then, S5 is executed to extract the foreground semantic prototypes of each facet image from the feature map through mask-guided foreground weighted pooling, and calculate the consistency loss between the foreground semantic prototypes of each facet image.

[0079] In this embodiment, after obtaining the feature map output by the segmentation network, the present invention extracts the foreground semantic prototype of each facet image through mask-guided foreground weighted pooling, which is one of the key innovations of the present invention. The specific process is as follows: First, obtain the labeled mask (in the early stage of training) or the high-confidence prediction mask (in the later stage of training or inference) corresponding to the current image. Here, the "high-confidence prediction mask" refers to the binary mask formed by the regions with probability values ​​greater than a preset threshold (e.g., 0.85) in the probability map predicted by the network. Then, scale the mask to align with the size of the feature map. Next, use the aligned mask values ​​as weights to perform weighted global average pooling on the feature map. Its mathematical expression can be expressed as:

[0080]

[0081] Among them, F p M represents the feature map of the plane image p. pLet Z represent the labeled mask or high-confidence prediction mask corresponding to the plane image p, i represent the spatial location of the feature map, and ε is a constant to prevent the denominator from being zero. To reduce the influence of differences in feature amplitudes across different planes, Z can be further optimized. p Perform L2 normalization.

[0082] The advantage of this pooling method lies in the fact that it doesn't simply average the entire feature map, but selectively weights the features of the foreground region with the mask value as the weight. This means that regions with higher mask values ​​(i.e., regions that the network considers more likely to be regions of interest) contribute more to the final feature prototype. Compared to traditional global average pooling, the method of this invention can more accurately focus on the target region, effectively filter out background noise interference, and extract a compact feature vector rich in semantic information, which can more accurately represent the essential features of the region of interest in that plane.

[0083] In this embodiment, the consistency loss between the foreground semantic prototypes of each polygon image is calculated, and is selected from any one or a combination thereof:

[0084] (1) Calculate the cosine distances between each pair of foreground semantic prototypes in the sagittal, coronal, and transverse images of the orthogonal three-dimensional image group and take the average. That is, calculate the cosine distances between the three foreground semantic prototype pairs of sagittal-coronal, sagittal-transverse, and coronal-transverse, and then take the average. The cosine distance measures the difference in direction between two feature vectors, and this loss function can constrain the directions of the three prototypes in the feature space to tend to be consistent.

[0085] (2) First, calculate the mean of the foreground semantic prototypes of the sagittal, coronal, and transverse images in the orthogonal three-plane image group as a shared prototype (i.e., a virtual center point representing the average features of the region of interest of the examined object). Then, constrain the foreground semantic prototypes of each plane image to be close to the shared prototype. This centralization constraint can reduce the dispersion between different plane prototypes within the same examined object and improve the consistency of features.

[0086] (3) After mapping the foreground semantic prototypes of each plane image through a plane-specific projection layer, calculate the smoothed L1 distance or mean square error between the mapped prototypes. Since images of different planes have inherent differences in imaging mechanisms and appearances, directly comparing their features may not be fair. Therefore, the foreground semantic prototypes of each plane image can be mapped to a unified feature space through a plane-specific projection layer (e.g., a small fully connected network with independent network parameters for each plane), and then the smoothed L1 distance or mean square error between the mapped prototypes can be calculated. This method can eliminate systematic biases between planes, making consistency comparisons more reasonable.

[0087] (4) A contrastive learning approach that uses the foreground semantic prototypes of three images from the same sequence of the same subject as positive samples and the foreground semantic prototypes of different subjects as negative samples. For example, the InfoNCE loss function is used to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs. This self-supervised learning approach can further enhance the discriminative power of features, enabling the model to better distinguish the region of interest features of different subjects.

[0088] These consistency losses, along with segmentation losses (such as Dice loss and cross-entropy loss), are weighted and summed to guide the network's optimized training, ensuring consistency of cross-plane features while pursuing segmentation accuracy per unit plane.

[0089] Then, S6 is executed to sum the segmentation loss and consistency loss of each facet image with weights to obtain the total loss, and the segmentation network is iteratively optimized and trained based on the total loss.

[0090] A preferred pairwise cosine consistency loss is:

[0091]

[0092]

[0093] in, , , These represent the feature vectors of the sagittal, coronal, and transverse images, respectively. This represents the cosine similarity between vectors a and b.

[0094] Accordingly, the total loss function for:

[0095]

[0096] in, , , These represent the segmentation losses for sagittal, coronal, and transverse images, respectively, including masking loss and the classification, bounding box, or distribution focus loss required by the network. denoted as the plane weights of the sagittal, coronal, and transverse images, respectively, and λ is the consistency loss weight.

[0097] To avoid early training errors where incorrect masks interfere with feature consistency, the prototype can be extracted using only labeled masks in the first few rounds, or a warm-up strategy can be adopted to gradually increase the consistency loss weights. Alternatively, λ can be dynamically adjusted based on prediction confidence, region of interest area ratio, or segmentation quality estimation. For example, when the network's prediction confidence for a certain plane is low, the weight of the segmentation loss for that plane can be reduced, or the weight of the consistency loss can be increased to utilize reliable information from other planes to assist in the segmentation of that plane. Based on this total loss, the system iteratively optimizes and trains the parameter-sharing segmentation network using the backpropagation algorithm, updating the network weights until the model converges.

[0098] Finally, S7 is executed, which uses the trained segmentation network to automatically segment the MRI image under test, obtains the region of interest mask, maps the region of interest mask back to the original DICOM space, and outputs morphological parameters and radiomics features.

[0099] During the inference phase, this invention demonstrates good flexibility. The MRI image to be tested includes a single-plane image or a group of orthogonal three-plane images of the same subject; when the input is a single-plane image, the region of interest mask and segmentation confidence of that plane image are output; when the input is a group of orthogonal three-plane images of the same subject, the region of interest masks of each plane image are output respectively.

[0100] More importantly, this invention also outputs inter-plane consistency indices to assess the reliability of the segmentation results. These indices are selected from at least one of the following: intra-plane Dice difference, intra-plane correlation coefficient (ICC), symmetry relative error, and combinations thereof. For example, the intra-plane Dice difference calculates the Dice coefficients between each pair of the three planes and takes their standard deviation or range; the smaller the value, the better the consistency of the three-plane segmentation results. The intra-plane correlation coefficient (ICC) measures the repeatability of the three-plane measurement results. The symmetry relative error is used to assess the relative deviation between the three-plane measurements. These indices provide valuable quality control information for clinicians. If the consistency index exceeds a preset threshold, the system can issue an early warning, prompting the doctor to note that the segmentation result may have uncertainties. Finally, the system maps the region of interest mask back to the original DICOM pixel matrix through nearest neighbor interpolation or spatial transformation based on DICOM geometric information, and calculates physical scale parameters based on the pixel spacing in the DICOM header information, outputting morphological parameters including the region of interest area, perimeter, major and minor axes, roundness, thickness, etc., as well as first-order grayscale features and texture features extracted based on the original image or Wavelet transformed image. These quantitative parameters provide solid data support for subsequent radiomics analysis and clinical diagnosis.

[0101] The technical concept of the present invention will be further illustrated below through some specific examples.

[0102] First, the MRI dataset of the subjects was exported from the original PACS system. When constructing the dataset, the images were organized and archived strictly according to three dimensions: subject identification, examination or sequence information, and plane (i.e., sagittal, coronal, and transverse).

[0103] Since the thickness and coverage of scan slices may vary across different planes, it is not mandatory for image IDs to be identical across different planes. To avoid data redundancy and potential leakage risks, this implementation adopts the "maximum ROI principle" for candidate slice selection: for each candidate image of the same object and the same sequence, the area of ​​its effective region of interest (ROI) is calculated, and the image with the largest area is selected as the representative slice of that plane.

[0104] After screening, only orthogonal three-dimensional image sets containing complete sagittal, coronal, and transverse images were retained. To ensure the rigor of model evaluation, all images of the same subject were only allowed to enter the same data subset (training set, validation set, or test set) throughout the entire experimental process, thereby eliminating the risk of subject-level data leakage.

[0105] This implementation is based on the Python 3.10 environment, using the PyTorch deep learning framework and the Ultralytics YOLO11-seg architecture. The specific configuration for network training is as follows: the number of training epochs is set to 150, the input image size is uniformly adjusted to 640×640 pixels, and the hardware platform selected is an NVIDIA GeForce RTX 4090 D GPU.

[0106] During training, for an orthogonal 3D image group containing three planes, each 3D image is input into the network in the same iteration. To fully utilize the correlation of anatomical structures within the examined object, this implementation employs a parameter-sharing segmentation network. Specifically, the three 2D images—sagittal, coronal, and transverse—are forward-propagated using the same set of network weights to obtain their respective segmentation predictions and intermediate layer feature maps. Foreground semantic prototypes of each plane image are extracted from the feature maps using mask-guided foreground weighted pooling. The consistency loss of the 3D foreground semantic prototypes is calculated and jointly optimized with the segmentation loss of each plane.

[0107] Meanwhile, the YOLO11 baseline, the DeBiFormerPlus baseline without consistency loss, and the nnU-Net 2D strong baseline were trained with the same data partitioning and subject-level isolation method for comparison and validation.

[0108] On 42 segmentation validation images, the DeBiFormerPlus model of this invention (150 epochs) achieved a macroscopic average Dice of 0.9153 and a macroscopic average IoU of 0.8505. Among them, 39 images were successfully matched with the original DICOM and used for geometric indices and radiomics analysis. These 39 images were from 7 subjects.

[0109] Table 1 below compares the geometric segmentation metrics of the proposed method (DeBiFormerPlus-Consistency) with those of the YOLO11 baseline, DeBiFormerPlus baseline, and nnU-Net 2D strong baseline on the aforementioned 39 DICOM matching images (HD50 and HD95 are in mm, calculated based on DICOM in-plane spacing).

[0110] Table 1

[0111]

[0112] To evaluate the performance uniformity of different planes of the same tested object, the |Dice| within each complete orthogonal three-plane image group is calculated. sag -Dice cor |、|Dice sag -Dice tra |and|Dice cor -Dice tra The average was then taken. The results show that, using the three-dimensional consistency loss model, the average absolute difference across all dimensions decreased from 0.0887 in the YOLO11 baseline to 0.0699, a reduction of approximately 21.2%; compared to 0.0787 in the DeBiFormerPlus baseline, it decreased by approximately 11.2%. Simultaneously, the fluctuation range of the number of CI85 stable feature candidates for each of the three dimensions decreased from 35 in the YOLO11 baseline to 18 in the method of this invention, and the range of the median ICC of the dimensions decreased from 0.142 to 0.039.

[0113] The above results demonstrate that the present invention can significantly reduce the fluctuations in segmentation and quantification performance between different MRI planes of the same subject while maintaining high segmentation accuracy.

[0114] To intuitively evaluate the consistency between the model output and manual annotation, this embodiment visually compares the segmentation results of different models within the original DICOM image space, such as... Figure 2 As shown.

[0115] Figure 2Each row corresponds to one MRI sample. The first column shows the manually annotated region of interest (ROI) mask, with blue filled areas representing manually drawn ROIs. The second column shows a comparison of the superimposed results of the model's prediction (i.e., the automatic ROI mask) and the manually annotated ROI mask. The third column shows a comparison of the superimposed results of the nnU-Net prediction and the manually annotated ROI mask. Specifically, green areas represent the overlap between the automatic segmentation and the manually annotated ROI mask, red areas represent areas where the automatic segmentation extends beyond the manually annotated ROI mask, and blue areas represent regions in the manually annotated ROI mask not covered by the model.

[0116] In the top row of samples, the manually annotated region of interest (ROI) mask has 16,699 pixels. The model of this invention achieves Dice=0.958 and IoU=0.920, with an automatically annotated ROI mask of 15,484 pixels; nnU-Net achieves Dice=0.964 and IoU=0.931, with an automatically annotated ROI mask of 15,831 pixels. Both automatic methods can cover the manually annotated ROI relatively completely, with only slight differences at local boundaries.

[0117] The bottom row of samples represents cases with higher Dice scores, and the manually annotated region of interest (ROI) mask has 25,044 pixels. The model of this invention achieves Dice = 0.978 and IoU = 0.957, with an automatically annotated ROI mask of 24,718 pixels; nnU-Net achieves Dice = 0.978 and IoU = 0.957, with an automatically annotated ROI mask of 24,486 pixels. The prediction boundary of the model of this invention highly overlaps with the manually annotated ROI mask, indicating that the proposed method can maintain high placental region segmentation accuracy under different MRI plane conditions.

[0118] The figure further illustrates that although the present invention adopts a lightweight segmentation structure oriented towards three-dimensional consistency, its segmentation results can still achieve a geometric overlap level close to that of the classic nnU-Net, thus providing a reliable basis for subsequent morphological measurement and radiomics feature extraction.

[0119] The manually annotated region of interest (ROI) mask and the automatically annotated ROI mask were mapped onto the same original DICOM image. Using the same PyRadiomics (version 3.0.1) parameters, 102 Original features were extracted, including shape2D (two-dimensional shape features), firstorder (first-order statistical features), GLCM (gray-level co-occurrence matrix features), GLRLM (gray-level run-length matrix features), GLSZM (gray-level size region matrix features), GLDM (gray-level dependence matrix features), and NGTDM (neighborhood gray-level tone difference matrix features). ICC(A,1) (two-way random effects, absolute consistency, single measurement), Spearman correlation, symmetric relative error (SRE), and Bland-Altman index were calculated.

[0120] Of the 42 validation images, 39 successfully matched the original DICOM and were included in the analysis, from 7 subjects. Under the original image type, 72 stable features were obtained by screening according to ICC ≥ 0.90, Spearman rho ≥ 0.80, FDR (False Discovery Rate) correction (p < 0.05), and median SRE ≤ 0.20. Further, after requiring a 95% confidence interval lower limit of ICC ≥ 0.85 (denoted as the CI85 standard), 65 strictly stable features were retained (4 Tier A features and 61 Tier B features). In the subject-by-subject (LOPO) sensitivity analysis, none of the 65 CI85 features showed an ICC decrease ≥ 0.10; clustering bootstrap analysis using subjects as the sampling unit still retained 60 features. In the exploratory analysis of Original + Wavelet (i.e., original image type + wavelet transform image type) (474 ​​features in total), 326 features met the CI85 standard.

[0121] This result demonstrates that automatic region of interest (ROI) masks not only possess pixel space segmentation capabilities but also stably preserve quantitative information generated by manually labeled ROI masks across most first-order grayscale and texture features. Furthermore, this conclusion is independent of individual subjects.

[0122] Please refer to Figure 3 , Figure 3The diagram illustrates the representative consistency analysis results obtained by extracting radiomics features from the same original DICOM image using a manually annotated region-of-interest (ROI) mask and the automatically annotated ROI mask of this invention. The top row shows a scatter plot of manually measured and automatically measured values, with the dashed line representing the identity line under ideal conditions where the two measurements are completely consistent. The bottom row shows the corresponding Bland–Altman analysis plot, where the blue solid line represents the mean bias, the orange dashed line represents the 95% agreement limit (LoA), and the gray horizontal line represents zero difference.

[0123] The three representative features selected are First-order Mean (first-order grayscale feature), GLCMContrast (texture feature), and Shape2D Sphericity (two-dimensional shape feature).

[0124] For the First-order Mean, the results of manual and automatic extraction show high consistency, with an ICC of approximately 0.990 and a Spearman correlation coefficient of approximately 0.985. The Bland–Altman mean bias is approximately 1.13, and the 95% consensus threshold ranges from -6.99 to 9.24. Most of the scatter points are located near the identity line, indicating that changes in the region of interest mask boundaries caused by automatic segmentation have a relatively small impact on this type of first-order grayscale statistical feature.

[0125] For GLCM Contrast, the ICC is approximately 0.986, the Spearman correlation coefficient is approximately 0.990, the mean bias is approximately -0.116, and the 95% consensus threshold is approximately -1.27 to 1.04. These results indicate that, under the automatic region-of-interest (ROI) mask of this invention, texture features reflecting local gray-level spatial relationships can still reproduce the measurement results obtained from manually annotated ROI masks quite well.

[0126] In contrast, Shape2D Sphericity exhibits relatively low consistency, with an ICC of approximately 0.639, a Spearman correlation coefficient of approximately 0.731, a mean bias of approximately -0.014, and a 95% consistency threshold ranging from -0.171 to 0.143. This phenomenon suggests that 2D shape features, which directly depend on the morphology of the region of interest mask boundary, are more sensitive to subtle changes in the segmentation boundary.

[0127] The above results demonstrate that the technical effectiveness of this invention is not only reflected in pixel-level geometric overlap indices such as Dice and IoU, but also in its ability to largely preserve the first-order grayscale and texture information represented by the manually annotated region of interest mask. Therefore, the automated placental region of interest mask generated by this invention can be further used for radiomics feature extraction, placental morphology quantification, and subsequent computer-aided analysis. Meanwhile, different types of radiomics features exhibit varying sensitivities to segmentation errors, with boundary-related shape features requiring more careful selection.

[0128] On the same 39 DICOM-matched images (7 subjects), the ability of automated RIO to preserve radiomics stability features was compared with that of the YOLO11 baseline, DeBiFormerPlus baseline, and the method of this invention (DeBiFormerPlus-Consistency). The results are shown in Table 2 below:

[0129] Table 2

[0130]

[0131] Compared with nnU-Net 2D, which has the highest geometric accuracy: both have 72 original stable features and 65 CI85 candidates (intersection of 57, Jaccard coefficient 0.78), with a median ICC difference of approximately -0.0003 (Wilcoxon signed-rank test p=0.64, no statistical difference); however, in the subject clustering bootstrap analysis, the method of this invention retains 60 stable features, while nnU-Net 2D retains 51.

[0132] The above comparison shows that: First, the method of the present invention improves the ability of automatic RIO to retain radiomics information compared with the YOLO11 baseline (65 vs. 41 CI85 candidates, 60 vs. 33 after clustering bootstrap); Second, the segmentation model with the best geometric index is not superior in terms of stability of downstream quantitative features. The method of the present invention retains more stable features under the sensitivity analysis at the subject level, indicating that the three-dimensional consistency constraint and the subject-level sample construction have independent value in improving the reliability of downstream quantitative analysis.

[0133] Based on the same inventive concept, this invention also proposes an MRI three-dimensional joint segmentation system, comprising:

[0134] The data acquisition module is configured to acquire MRI images and their metadata for the same subject, the same examination, and the same sequence targeting the same region of interest.

[0135] The grouping key construction module is configured to construct orthogonal three-dimensional image groups using the subject identifier, examination identifier, and sequence name as grouping keys. The orthogonal three-dimensional image group includes one sagittal image, one coronal image, and one transverse image of the same subject, the same examination, and the same sequence targeting the same region of interest. The image identifiers of different planes are independent of each other and are not used as cross-plane pairing conditions.

[0136] The data partitioning module is configured to divide several orthogonal three-dimensional image groups into training set, validation set and test set according to the identification of the inspected object;

[0137] The joint training module is configured to share the input parameters of each facet image in the orthogonal three-facet image group to obtain the segmentation prediction and feature map of each facet image; and calculate the segmentation loss of each facet image based on the segmentation prediction result and the corresponding labeled mask; then extract the foreground semantic prototype of each facet image from the feature map through mask-guided foreground weighted pooling, and calculate the consistency loss between the foreground semantic prototypes of each facet image; and finally, sum the segmentation loss and consistency loss of each facet image to obtain the total loss, and iteratively optimize and train the segmentation network based on the total loss.

[0138] The inference output module is configured to automatically segment the MRI image under test using the trained segmentation network, obtain the region of interest mask, map the region of interest mask back to the original DICOM space, and output morphological parameters and radiomics features.

[0139] Since the MRI three-dimensional joint segmentation system provided by this invention belongs to the same inventive concept as the MRI three-dimensional joint segmentation method described above, the MRI three-dimensional joint segmentation system provided by this invention has all the advantages of the MRI three-dimensional joint segmentation method described above. Therefore, the beneficial effects of the MRI three-dimensional joint segmentation system provided by this invention will not be described in detail here.

[0140] Based on the same inventive concept, this invention also proposes an electronic device, comprising:

[0141] Memory, which stores computer programs;

[0142] The processor, which communicates with the memory, executes the above-mentioned MRI three-dimensional joint segmentation method when calling computer programs.

[0143] The display, connected to the processor and memory, is used to show a GUI interface related to the MRI three-dimensional joint segmentation method.

[0144] Since the electronic device provided by this invention belongs to the same inventive concept as the MRI three-dimensional joint segmentation method described above, the electronic device provided by this invention has all the advantages of the MRI three-dimensional joint segmentation method described above. Therefore, the beneficial effects of the electronic device provided by this invention will not be described in detail here.

[0145] Based on the same inventive concept, embodiments of the present invention also propose a readable storage medium storing a computer program thereon, which, when executed, can implement the above-mentioned MRI three-dimensional joint segmentation method.

[0146] A readable storage medium can be a tangible device capable of holding and storing instructions for use by an instruction execution device, such as, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer programs described herein can be downloaded from the readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. Networks can include copper transmission cables, fiber optic transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. Each computing / processing device's network adapter card or network interface receives and forwards a computer program from the network for storage on a readable storage medium within the respective computing / processing device. The computer program used to perform the operations of this invention can be execution instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer program can execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), are personalized by utilizing state information from a computer program. These electronic circuits can execute computer-readable program instructions, thereby realizing various aspects of the present invention.

[0147] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by a computer program. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. These computer programs can also be stored in a readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the readable storage medium storing the computer program comprises an article of manufacture including instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams.

[0148] A computer program may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the computer program executing on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0149] Since the readable storage medium provided by this invention belongs to the same inventive concept as the MRI three-dimensional joint segmentation method described above, the readable storage medium provided by this invention has all the advantages of the MRI three-dimensional joint segmentation method described above. Therefore, the beneficial effects of the readable storage medium provided by this invention will not be described in detail here.

[0150] In summary, this invention provides an MRI three-dimensional joint segmentation method, system, electronic device, and readable storage medium. It innovatively constructs orthogonal three-dimensional image groups using subject identifiers, examination identifiers, and sequence names as grouping keys, eliminating reliance on pixel-level correspondence and improving the medical rationality of sample construction. To address the problem of subject-level data leakage, this invention strictly adheres to the subject-level isolation principle, ensuring that all data from the same subject exists only in a single data subset, thereby objectively evaluating the model's generalization ability. In terms of the training mechanism, this invention employs a parameter-sharing network combined with mask-guided foreground weighted pooling technology. This extracts the foreground semantic prototype without cross-plane registration and applies consistency constraints to the feature space, preserving the specific morphological differences between planes while reducing fluctuations in three-dimensional segmentation performance. Crucially, this invention overcomes the limitations of existing technologies that rely solely on geometric metrics such as Dice and IoU. Through multi-dimensional metrics such as ICC, symmetry relative error, and subject clustering bootstrap, it verifies the stability of automatic ROI in preserving quantitative radiomics information. This mechanism allows for the preservation of more stable features in subject-level sensitivity analysis even with comparable geometric accuracy, providing a highly reliable quantitative analysis basis for clinical research and possessing broad applicability to other multi-planar MRI segmentation tasks.

[0151] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure are within the protection scope of the present invention. Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A three-dimensional joint segmentation method for MRI, characterized in that, include: MRI images and their metadata for the same subject, the same examination, and the same sequence targeting the same region of interest are acquired. Orthogonal three-dimensional image groups are constructed using the subject identifier, examination identifier, and sequence name as grouping keys. Each orthogonal three-dimensional image group includes one sagittal image, one coronal image, and one transverse image for the same subject, the same examination, and the same sequence targeting the same region of interest. The image identifiers of different planes are independent of each other and are not used as cross-plane pairing conditions. Several orthogonal three-dimensional image groups were divided into training set, validation set and test set according to the identification of the inspected object; The segmentation network that shares the input parameters of each facet image in the orthogonal three-facet image group is used to obtain the segmentation prediction and feature map of each facet image. The segmentation loss of each facet image is calculated based on the segmentation prediction results and the corresponding labeled mask; Foreground semantic prototypes of each facet image are extracted from the feature map by mask-guided foreground weighted pooling, and the consistency loss between the foreground semantic prototypes of each facet image is calculated. The segmentation loss of each facet image is weighted and summed with the consistency loss to obtain the total loss, and the segmentation network is iteratively optimized and trained based on the total loss; The trained segmentation network is used to automatically segment the MRI image under test to obtain a region of interest mask, and the region of interest mask is mapped back to the original DICOM space to output morphological parameters and radiomics features.

2. The MRI three-dimensional joint segmentation method according to claim 1, characterized in that, When constructing the orthogonal three-dimensional image group, for multiple candidate images of the same object and the same sequence of the same surface, the image with the largest effective region of interest area is selected as the representative slice of that surface.

3. The MRI three-dimensional joint segmentation method according to claim 1, characterized in that, When dividing a number of orthogonal 3D image groups into a training set, a validation set, and a test set according to the identification of the tested object, all orthogonal 3D images of the same tested object are placed into the same subset of the training set, the validation set, and the test set.

4. The MRI three-dimensional joint segmentation method according to claim 1, characterized in that, The parameter-sharing segmentation network has a feature enhancement module preceding the backbone network or segmentation head.

5. The MRI three-dimensional joint segmentation method according to claim 4, characterized in that, When extracting the foreground semantic prototype of each facet image from the feature map through mask-guided foreground weighted pooling, the labeled mask or high-confidence prediction mask is scaled to be aligned with the size of the feature map. The aligned mask values ​​are used as weights to perform weighted global average pooling on the features of the foreground region to obtain the foreground semantic prototype of each facet image.

6. The MRI three-dimensional joint segmentation method according to claim 1, characterized in that, The consistency loss calculated between the foreground semantic prototypes of each polygon image is selected from any one of the following or a combination thereof: Calculate the cosine distances between each pair of foreground semantic prototypes in the sagittal, coronal, and transverse images of the orthogonal three-dimensional image group and take the average. First, calculate the mean of the foreground semantic prototypes of the sagittal, coronal, and transverse images in the orthogonal three-plane image group as a shared prototype, and then constrain the foreground semantic prototypes of each plane image to be close to the shared prototype. After mapping the foreground semantic prototypes of each plane image through a plane-specific projection layer, the smooth L1 distance or mean square error between the mapped prototypes is calculated. A comparative learning approach that uses the foreground semantic prototypes of three images from the same sequence of the same subject as positive samples and the foreground semantic prototypes of different subjects as negative samples.

7. The MRI three-dimensional joint segmentation method according to claim 1, characterized in that, During the inference phase, the MRI images to be tested include a single-plane image or a group of orthogonal three-plane images of the same subject. When the input is a single-plane image, the output is the region of interest mask and segmentation confidence of that plane image; When the input is an orthogonal three-dimensional image group of the same object under inspection, the region of interest mask of each image is output, and at least one consistency index between the three-dimensional images is output, selected from the Dice difference within the three-dimensional image group, the correlation coefficient within the group, the symmetry relative error and its combination.

8. An MRI three-dimensional joint segmentation system, characterized in that, include: The data acquisition module is configured to acquire MRI images and their metadata for the same subject, the same examination, and the same sequence targeting the same region of interest. The grouping key construction module is configured to construct an orthogonal three-dimensional image group using the subject identifier, the examination identifier, and the sequence name as grouping keys. The orthogonal three-dimensional image group includes one sagittal image, one coronal image, and one transverse image of the same subject, the same examination, and the same sequence for the same region of interest. The image identifiers of different planes are independent of each other and are not used as cross-plane pairing conditions. The data partitioning module is configured to divide several orthogonal three-dimensional image groups into training set, validation set and test set according to the identification of the inspected object; The joint training module is configured to share the input parameters of each facet image in the orthogonal three-facet image group with the segmentation network to obtain the segmentation prediction and feature map of each facet image; and calculate the segmentation loss of each facet image based on the segmentation prediction result and the corresponding labeled mask. Then, foreground semantic prototypes of each facet image are extracted from the feature map through mask-guided foreground weighted pooling, and the consistency loss between the foreground semantic prototypes of each facet image is calculated. The segmentation loss of each facet image is weighted and summed with the consistency loss to obtain the total loss, and the segmentation network is iteratively optimized and trained based on the total loss; The inference output module is configured to automatically segment the MRI image under test using the trained segmentation network, obtain a region of interest mask, map the region of interest mask back to the original DICOM space, and output morphological parameters and radiomics features.

9. An electronic device, characterized in that, include: Memory, which stores computer programs; The processor, which is communicatively connected to the memory, executes the MRI three-dimensional joint segmentation method according to any one of claims 1-7 when the computer program is invoked. The display, communicatively connected to the processor and the memory, is used to display a GUI interface related to the MRI three-dimensional joint segmentation method.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it can implement the MRI three-dimensional joint segmentation method according to any one of claims 1-7.

Images