Spine myeloma MRI three-dimensional image segmentation model
By performing three-dimensional spatial discretization and feature stitching on MRI images of spinal myeloma, combined with standardization and loss function training, the accuracy and efficiency issues of three-dimensional image segmentation of spinal MRI images of multiple myeloma were solved, achieving high-precision segmentation and automatic generation of structured reports.
Patent Information
- Application Number
- CN202511752042.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-03
AI Technical Summary
Existing technologies for segmenting three-dimensional images of the spine in multiple myeloma MRI suffer from problems such as high segmentation difficulty, low accuracy, high computational resource consumption, and inability to automatically generate structured reports.
The system employs a data processing unit for 3D spatial discretization and image cropping, combined with progressive downsampling and upsampling operations of the spinal segmentation unit. It utilizes feature stitching from the encoder and decoder, and trains the system using Z-score normalization and skip connections of the tumor segmentation unit, along with Diceloss and cross-entropy loss, to generate a high-precision probability distribution of myeloma tissue.
It improves the accuracy and efficiency of 3D image segmentation for spinal myeloma MRI, reduces background noise interference, can automatically generate structured reports, and reduces computational resource consumption.
Smart Images
Figure CN121600262A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image segmentation technology, and more specifically, to a three-dimensional image segmentation model for spinal myeloma MRI. Background Technology
[0002] While some progress has been made in the field of tumor image segmentation technology, segmentation of multiple myeloma lesions still faces many challenges. On the one hand, the characteristics of multiple myeloma lesions increase the difficulty of segmentation. On the other hand, traditional image processing methods rely on manually designed features and rules, which require a lot of professional knowledge and experience, and are sensitive to image noise and artifacts, resulting in limited segmentation accuracy. Currently, the interpretation of medical images is mainly done by radiologists, but the increasing demand for image reading has brought a heavy workload to doctors and affected the timeliness of patients' medical treatment. In recent years, with the rapid development of artificial intelligence, multimodal research based on images and text has attracted attention. AIGC systems have gradually acquired the ability to process medical image data (such as CT, MRI, etc.) and generate analysis reports. Existing techniques for segmenting three-dimensional MRI images of the spine in multiple myeloma have significant shortcomings. For example, inconsistent MRI image scanning specifications and the presence of interfering information in the images affect the accuracy and reliability of image segmentation. Traditional segmentation models with convolutional neural networks or Transformer structures are difficult to effectively extract key features of spinal myeloma and have a high risk of missed detection and misclassification in the segmentation of small lesions and lesion boundaries. At the same time, existing technologies cannot automatically generate structured reports and consume too much computational resources. Therefore, we propose a three-dimensional MRI image segmentation model for spinal myeloma. Summary of the Invention
[0003] The purpose of this invention is to provide a three-dimensional MRI image segmentation model for spinal myeloma to solve the problems mentioned in the background art.
[0004] To achieve the above objectives, the present invention provides a three-dimensional MRI image segmentation model for spinal myeloma, comprising a data processing unit, a spinal segmentation unit, and a tumor segmentation unit; The data processing unit acquires MRI image data and performs three-dimensional spatial discretization processing to obtain signal data for each voxel. The signal data for each voxel is stored to form a three-dimensional volume dataset for transposition, resampling, and sagittal slicing. The sagittal slice data is then processed for image cropping and separation of bone and soft tissue to obtain processed MRI image data. The spinal segmentation unit is used to receive MRI image data processed in the data processing unit, input the processed MRI image data into the encoder for progressive downsampling and downsampling operations, thereby obtaining encoded features and encoder feature maps of the corresponding level, and performing splicing and three-dimensional convolution operations to obtain fused features. By performing preliminary segmentation of the entire spine using fused and processed MRI image data, individual vertebral instances are obtained to calculate their centroids and determine the spatial locations of key anatomical landmarks. The centroids of other vertebrae are numbered according to the anatomical order of the spine, resulting in MRI image data of the numbered vertebrae. The tumor segmentation unit is used to receive MRI image data after vertebral numbering in the spinal segmentation unit. The Z-score method is used to standardize the MRI image data after vertebral numbering and the preliminary segmentation results, and then the training patch is extracted to introduce skip connections and obtain the true label. The true labels are input into the 3D image segmentation model, and the probability distribution of the predicted results and the true distribution are output. The 3D image segmentation model is then trained to obtain the probability of the trained model outputting myeloma tissue.
[0005] As a further improvement to this technical solution, the data processing unit includes the following method steps: MRI image data is acquired, magnetic resonance signal data is collected from the MRI image data using a three-dimensional imaging sequence, and then the imaging region in the magnetic resonance signal data is discretized in three-dimensional space to obtain the signal data of each voxel. The signal data of each voxel is stored to form a three-dimensional volume dataset. Acquire coronal plane data, perform data extraction and image reconstruction along the coronal plane to obtain coronal images, and convert the coronal images into sagittal images; By transposing and resampling the three-dimensional volume dataset, the axes of the data are rearranged to generate sagittal slices.
[0006] As a further improvement to this technical solution, the transpose operation, resampling, and sagittal slicing operations specifically include the following key steps: Transpose operation: By rearranging the axes of the three-dimensional volume dataset, transposed three-dimensional MRI data is obtained; Resampling: The transposed 3D MRI data were resampled using linear interpolation to obtain resampled 3D MRI data; Sagittal slices: Sagittal slices are extracted from the resampled 3D MRI data. The extraction is performed at fixed intervals to obtain sagittal slice data.
[0007] As a further improvement to this technical solution, the data processing unit further includes the following method steps: By cropping the sagittal slice data, the cropped spinal MRI data is obtained, and non-spinal regions are excluded. The Otsu thresholding method is used to calculate the optimal threshold based on the cropped spinal MRI data. The cropped spinal MRI data is then processed to separate bones and soft tissues based on the optimal threshold, thereby automatically separating bones and soft tissues and obtaining processed MRI image data.
[0008] As a further improvement to this technical solution, the spinal segmentation unit includes the following method steps: The processed MRI image data is input into the encoder for progressive downsampling; Instance normalization is used to normalize each sample in the processed MRI image data to obtain normalized MRI image data; The downsampling operation uses a 2×2×2 three-dimensional max pooling layer to downsample the normalized MRI image data, thereby obtaining the encoded features and the corresponding level of encoder feature map.
[0009] As a further improvement to this technical solution, the fusion feature includes the following method steps: The coded features and the corresponding encoder feature map are upsampled by a 2×2×2 deconvolution layer. The deconvolution layer can map the low-resolution feature map to a high-resolution space to obtain the decoder feature map. The feature map of the decoder is concatenated with the feature map of the corresponding level in the encoder, and then two 3×3×3 three-dimensional convolution operations are performed to obtain the decoded features. The feature concatenation operation is used to fuse the features of the corresponding layer of the encoder with the features of the decoder to obtain the fused features.
[0010] As a further improvement to this technical solution, the spinal segmentation unit also includes the following method steps: TotalSpineSeg was used to perform preliminary segmentation of the entire spine on the fused and processed MRI image data, identify the overall structure of the spine, remove surrounding irrelevant tissues, and obtain preliminary spine segmentation results. Based on the initial segmentation results, the model identifies the overall region of the vertebrae, uses connected component analysis to segment the region into individual vertebrae instances, and marks interconnected pixel regions as independent objects to obtain individual vertebrae instances. The centroid of a single vertebra is calculated, and the spatial location of the key anatomical markers is determined by the three-dimensional image segmentation model using the key anatomical markers identified in the preliminary segmentation results. Based on the spatial location of key anatomical landmarks, the centroids of other vertebrae are numbered according to the anatomical order of the spine, resulting in MRI image data of the numbered vertebrae.
[0011] As a further improvement to this technical solution, the tumor segmentation unit includes the following method steps: The Z-score method is usually used to standardize the MRI image data and preliminary segmentation results after the vertebrae are numbered, converting the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, thus obtaining the preprocessed image data. The training patch is obtained by using a sliding window strategy to extract training patches from preprocessed image data. The training patch supports multi-scale input and the patch size is automatically set by the framework. By introducing skip connections, the encoded features and the corresponding level of encoder feature maps are obtained again. The decoder upsamples through training patches to gradually restore the spatial resolution, and uses skip connections to fuse low-level features to obtain decoded features and ground truth labels again.
[0012] As a further improvement to this technical solution, the tumor segmentation unit also includes the following method steps: The decoded features and ground truth labels are input into the 3D image segmentation model. The 3D image segmentation model uses Diceloss to measure the degree of overlap between the decoded features and ground truth labels, and outputs the probability distribution of the predicted results and the ground truth distribution. Cross-entropy loss is used to measure the difference between the probability distribution of the predicted result and the true distribution. Combining the two can better balance the weights of different categories. The weights are used to train the 3D image segmentation model to obtain a trained model. The trained model is used to perform block prediction on the preprocessed image data using a sliding window strategy, removes false positive regions, and outputs a segmentation mask or probability of myeloma tissue with the same size as the input image.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: In this 3D MRI image segmentation model for spinal myeloma, the spinal segmentation unit performs preliminary segmentation of the entire spine using fused features and processed MRI image data. It identifies the overall spinal structure, removes surrounding irrelevant tissue, and obtains preliminary spinal segmentation results. Based on these preliminary segmentation results, the model identifies the overall region of the vertebrae. Connectivity component analysis is used to segment the region into individual vertebral instances, and interconnected pixel regions are marked as independent objects, resulting in individual vertebral instances. The centroid of each individual vertebral instance is calculated. Using key anatomical markers identified in the preliminary segmentation results, the 3D image segmentation model determines the spatial location of these key anatomical markers. Based on the spatial location of these key anatomical markers, the centroids of other vertebrae are numbered according to the anatomical order of the spine. Preliminary segmentation removes surrounding irrelevant tissue, focusing the processing on the spinal region, reducing background noise interference, and enabling the model to more accurately identify the overall spinal structure. Subsequent segmentation of the entire vertebral region into individual vertebral instances, and further refining the segmentation results based on key anatomical markers and numbering the centroids of the vertebrae, allows the model to more accurately determine the boundaries and positions of each vertebra, improving overall segmentation accuracy. Attached Figure Description
[0014] Figure 1 This is a block diagram of the overall system structure of the present invention.
[0015] The meanings of the labels in the diagram are as follows: 1. Data processing unit; 2. Spinal segmentation unit; 3. Tumor segmentation unit. Detailed Implementation
[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1
[0017] This invention provides a three-dimensional image segmentation model for spinal myeloma MRI. Please refer to [link / reference]. Figure 1 It includes a data processing unit 1, a spine segmentation unit 2, and a tumor segmentation unit 3; Data Processing Unit 1: In the field of medical imaging, MRI (Magnetic Resonance Imaging) is a powerful diagnostic tool, playing a crucial role, especially in the diagnosis of spinal diseases. In MRI, magnetic resonance signal data is acquired from MRI image data using three-dimensional imaging sequences. The imaging region within the magnetic resonance signal data is then discretized in three-dimensional space, divided into small voxels. The signal data for each voxel is obtained and stored, forming a three-dimensional volumetric dataset. Each voxel has a defined location in three-dimensional space. The three-dimensional volumetric dataset typically has isotropic resolution, meaning it has isotropic resolution in all three dimensions. The voxel sizes are the same, for example, the voxel size is ; However, different imaging needs may require images to be displayed in different directions. Coronal data is obtained, and data extraction and image reconstruction are performed along the coronal plane (i.e., the plane that divides the human body into front and back parts in the left and right direction) to obtain coronal images. The coronal images are then converted into sagittal images. This conversion process actually involves reselecting the slice orientation in the three-dimensional volume dataset. By transposing and resampling the three-dimensional volume dataset, the axes of the data can be rearranged to generate sagittal slices. The specific operation includes the following key steps: Transpose operation: When performing sagittal transformation, the axes of the three-dimensional volume dataset must first be rearranged to obtain the transposed three-dimensional MRI data; for example, the original... The axis order becomes This operation is not just a simple axis swap; it also requires flipping the three-dimensional volume dataset to match the correct anatomical orientation, because the human body's anatomical structure has a specific orientation, ensuring that the three-dimensional volume dataset is consistent with the actual anatomical orientation. Resampling: Ensure that the spatial dimensions of all voxels are consistent, such as achieving... The standard uses linear interpolation to resample the transposed 3D MRI data to obtain the resampled 3D MRI data. Linear interpolation is a simple and fast interpolation method that estimates the value of unknown points based on the linear relationship between adjacent data points. Sagittal slices: After rearrangement and resampling, sagittal slices can be extracted from the resampled 3D MRI data. To ensure image integrity, fixed intervals are set, and extraction is performed at these fixed intervals. Each sagittal slice is a two-dimensional image representing a specific point in time. Slices of the 3D volume dataset at coordinates are used to obtain sagittal slice data. Since the sagittal slice data contains tissues (including the head, limbs, and internal organs) that are not related to the segmentation of spinal myeloma, these irrelevant tissues will increase the computational burden and interfere with the learning of the three-dimensional image segmentation model, thus reducing the accuracy of segmentation. Therefore, the sagittal slice data is cropped to obtain cropped spinal MRI data. The goal of image cropping is to ensure that the spinal part below the shoulders and above the hips is completely preserved, while non-spine areas are removed. The Otsu thresholding method is used to calculate the optimal threshold based on the cropped spinal MRI data. The optimal threshold is then used to separate the bone and soft tissue in the cropped spinal MRI data, thereby automatically separating the bone and soft tissue and obtaining the processed MRI image data. The Otsu thresholding method is an automatic threshold selection method based on the image grayscale histogram. It determines the optimal threshold by maximizing the inter-class variance and can effectively divide the image into foreground (bone) and background (soft tissue). The processed MRI images contain only the complete spinal region with a resolution of 256×256. The refined cropping method not only reduces computational complexity and improves the training efficiency of the 3D image segmentation model, but also ensures the accuracy and robustness of the segmentation results. The spinal segmentation unit 2 is used to receive the MRI image data processed in the data processing unit 1 and input the processed MRI image data into the encoder for progressive downsampling to extract high-level semantic information. The encoder contains two 3×3×3 three-dimensional convolutional layers, and the LeakyReLU activation function is used between the convolutional layers. The LeakyReLU activation function can effectively alleviate the "neuron death" problem that may occur in the traditional ReLU function, making the network more stable during training. Meanwhile, instance normalization is used instead of batch normalization. Instance normalization normalizes each sample in the processed MRI image data to obtain normalized MRI image data. The downsampling operation is performed on the normalized MRI image data through a 2×2×2 three-dimensional max pooling layer. The number of feature maps is doubled after each downsampling. This can gradually reduce the spatial size of the feature maps and increase the abstraction of the features, thereby obtaining the encoded features and the corresponding level of encoder feature maps. The decoder is responsible for progressively upsampling to restore the spatial resolution of the image. It uses 2×2×2 deconvolution layers to upsample the encoded features and the corresponding level of encoder feature maps with a stride of 2. The deconvolution layer can map the low-resolution feature map to the high-resolution space to obtain the decoder feature map. After each upsampling, the decoder feature map is concatenated with the corresponding level of feature map in the encoder, and then two 3×3×3 three-dimensional convolution operations are performed to obtain the decoded features. 3DU-Net introduces skip connections in the decoder, concatenating the feature maps of corresponding levels in the encoder with the feature maps in the decoder to preserve high-resolution feature information. By using feature concatenation, the features of the corresponding layer in the encoder are fused with the features in the decoder to obtain fused features, ensuring that spatial information is preserved. By connecting the feature information of different levels in the encoder and decoder, the network can simultaneously utilize high-level semantic information and low-level detail information, thereby improving the accuracy of segmentation. In the first stage of spinal segmentation, we used TotalSpineSeg to perform preliminary segmentation of the entire spine using fused features and processed MRI image data, identify the overall structure of the spine, and remove surrounding irrelevant tissues to obtain preliminary spinal segmentation results. The three-dimensional image segmentation model can identify 9 categories, which are used to identify key anatomical structures of the spine to ensure accurate localization of subsequent tumor detection. The preliminary spinal segmentation results are saved in NIfTI format. Based on the initial segmentation, TotalSpineSeg uses an iterative algorithm to segment and number each vertebra, ensuring that subsequent tumor detection can accurately locate the specific vertebra. First, based on the preliminary segmentation results, the model identifies the overall region of the vertebra. Then, connected component analysis is used to segment the region into individual vertebral instances, and interconnected pixel regions are marked as independent objects to obtain individual vertebral instances. This step ensures that each vertebra is separated individually for subsequent numbering. The centroid (center point) of a single vertebra is calculated, and its coordinates are used to determine the vertebra's position in three-dimensional space, which is crucial for accurate numbering. Using key anatomical landmarks (such as specific intervertebral discs and the C1 vertebra) identified in the preliminary segmentation results, the three-dimensional image segmentation model determines the spatial location of the key anatomical landmarks. Then, based on the spatial location of the key anatomical landmarks, the centroids of other vertebrae are numbered according to the anatomical order of the spine, resulting in MRI image data of the numbered vertebrae. This method ensures continuous numbering from the cervical vertebrae to the sacrum. Tumor segmentation unit 3 receives MRI image data and preliminary segmentation results from vertebrae numbering in spinal segmentation unit 2. Typically, the Z-score method is used to standardize the MRI image data and preliminary segmentation results after vertebrae numbering, converting the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, resulting in preprocessed image data. Since 3D medical images are large, directly training on the entire preprocessed image data would consume a significant amount of GPU memory, potentially leading to memory shortages. Therefore, nnU-Net uses a sliding window strategy to extract training patches from the preprocessed image data. These training patches support multi-scale input, and their size is automatically set by the framework (e.g., 128×128×128), ensuring sufficient contextual information while controlling GPU memory usage. The encoder consists of multiple residual convolutional blocks used to progressively extract hierarchical features. A residual convolutional block is a special type of convolutional block structure that introduces skip connections to obtain encoded features and corresponding encoder feature maps at different levels. This allows the network to learn feature information at different levels more easily, alleviating the gradient vanishing problem during deep neural network training. The decoder performs upsampling (such as deconvolution) through training patches to progressively restore spatial resolution and uses skip connections to fuse low-level features to obtain decoded features and ground truth labels again. This structural design enables the network to extract high-level semantic information while preserving the detailed information of the image. Deep supervision signals can provide additional supervision information at different levels of the network, enabling the network to pay more attention to feature information at different scales during training. The loss function is usually a combination of Diceloss and cross-entropy loss, which effectively alleviates the class imbalance problem. It is especially suitable for tumor segmentation tasks where the lesion area accounts for a small proportion. In tumor segmentation tasks, the lesion area usually only accounts for a small part of the entire image. If only a single loss function is used, it may cause the model to tend to predict the background area and ignore the lesion area. The decoded features and ground truth labels are input into the 3D image segmentation model. The 3D image segmentation model uses Diceloss to measure the degree of overlap between the decoded features and ground truth labels and outputs the probability distribution of the predicted result and the ground truth distribution. Then, cross-entropy loss is used to measure the difference between the probability distribution of the predicted result and the ground truth distribution. Combining the two can better balance the weights of different classes. The weights are used to train the 3D image segmentation model to obtain the trained model. The trained model uses a sliding window strategy to perform block prediction on the preprocessed image data, and finally achieves fusion by averaging the probabilities of overlapping regions. Block prediction reduces the amount of computation, while averaging the probabilities of overlapping regions improves the accuracy of the prediction results. The model automatically learns and executes post-processing strategies based on the training data, such as connected component analysis to remove false positive regions. Finally, it outputs a segmentation mask with the same size as the input image, representing the probability that each voxel belongs to normal tissue or myeloma tissue.
[0018] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A three-dimensional image segmentation model for spinal multiple myeloma MRI, characterized in that: It includes a data processing unit (1), a spine segmentation unit (2), and a tumor segmentation unit (3); The data processing unit (1) acquires MRI image data and performs three-dimensional spatial discretization processing to obtain signal data of each voxel. The signal data of each voxel is stored to form a three-dimensional volume dataset for transposition, resampling and sagittal slicing. The sagittal slice data is then processed for image cropping and separation of bone and soft tissue to obtain the processed MRI image data. The spinal segmentation unit (2) is used to receive the MRI image data processed in the data processing unit (1), input the processed MRI image data into the encoder for progressive downsampling and downsampling operations, thereby obtaining the encoded features and the corresponding level encoder feature map, and performing splicing and three-dimensional convolution operations to obtain the fused features; By performing preliminary segmentation of the entire spine using fused and processed MRI image data, individual vertebral instances are obtained to calculate their centroids and determine the spatial locations of key anatomical landmarks. The centroids of other vertebrae are numbered according to the anatomical order of the spine, resulting in MRI image data of the numbered vertebrae. The tumor segmentation unit (3) is used to receive MRI image data after vertebral numbering in the spinal segmentation unit (2). The Z-score method is used to standardize the MRI image data after vertebral numbering and the preliminary segmentation results, and then the training patch is extracted to introduce skip connections and obtain real labels. The true labels are input into the 3D image segmentation model, and the probability distribution of the predicted results and the true distribution are output. The 3D image segmentation model is then trained to obtain the probability of the trained model outputting myeloma tissue.
2. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The data processing unit (1) includes the following method steps: MRI image data is acquired, magnetic resonance signal data is collected from the MRI image data using a three-dimensional imaging sequence, and then the imaging region in the magnetic resonance signal data is discretized in three-dimensional space to obtain the signal data of each voxel. The signal data of each voxel is stored to form a three-dimensional volume dataset. Acquire coronal plane data, perform data extraction and image reconstruction along the coronal plane to obtain coronal images, and convert the coronal images into sagittal images; By transposing and resampling the three-dimensional volume dataset, the axes of the data are rearranged to generate sagittal slices.
3. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The transpose operation, resampling, and sagittal slicing operations include the following key steps: Transpose operation: By rearranging the axes of the three-dimensional volume dataset, transposed three-dimensional MRI data is obtained; Resampling: The transposed 3D MRI data were resampled using linear interpolation to obtain resampled 3D MRI data; Sagittal slices: Sagittal slices are extracted from the resampled 3D MRI data. The extraction is performed at fixed intervals to obtain sagittal slice data.
4. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The data processing unit (1) further includes the following method steps: By cropping the sagittal slice data, the cropped spinal MRI data is obtained, and non-spinal regions are excluded. The Otsu thresholding method is used to calculate the optimal threshold based on the cropped spinal MRI data. The cropped spinal MRI data is then processed to separate bones and soft tissues based on the optimal threshold, thereby automatically separating bones and soft tissues and obtaining processed MRI image data.
5. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The spinal segmentation unit (2) includes the following method steps: The processed MRI image data is input into the encoder for progressive downsampling; Instance normalization is used to normalize each sample in the processed MRI image data to obtain normalized MRI image data; The downsampling operation uses a 2×2×2 three-dimensional max pooling layer to downsample the normalized MRI image data, thereby obtaining the encoded features and the corresponding level of encoder feature map.
6. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The fusion features include the following method steps: The coded features and the corresponding encoder feature map are upsampled by a 2×2×2 deconvolution layer. The deconvolution layer can map the low-resolution feature map to a high-resolution space to obtain the decoder feature map. The feature map of the decoder is concatenated with the feature map of the corresponding level in the encoder, and then two 3×3×3 three-dimensional convolution operations are performed to obtain the decoded features. The feature concatenation operation is used to fuse the features of the corresponding layer of the encoder with the features of the decoder to obtain the fused features.
7. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The spinal segmentation unit (2) also includes the following method steps: TotalSpineSeg was used to perform preliminary segmentation of the entire spine on the fused and processed MRI image data, identify the overall structure of the spine, remove surrounding irrelevant tissues, and obtain preliminary spine segmentation results. Based on the initial segmentation results, the model identifies the overall region of the vertebrae, uses connected component analysis to segment the region into individual vertebrae instances, and marks interconnected pixel regions as independent objects to obtain individual vertebrae instances. The centroid of a single vertebra is calculated, and the spatial location of the key anatomical markers is determined by the three-dimensional image segmentation model using the key anatomical markers identified in the preliminary segmentation results. Based on the spatial location of key anatomical landmarks, the centroids of other vertebrae are numbered according to the anatomical order of the spine, resulting in MRI image data of the numbered vertebrae.
8. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The tumor segmentation unit (3) includes the following method steps: The Z-score method is usually used to standardize the MRI image data and preliminary segmentation results after the vertebrae are numbered, converting the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, thus obtaining the preprocessed image data. The training patch is obtained by using a sliding window strategy to extract training patches from preprocessed image data. The training patch supports multi-scale input and the patch size is automatically set by the framework. By introducing skip connections, the encoded features and the corresponding level of encoder feature maps are obtained again. The decoder upsamples through training patches to gradually restore the spatial resolution, and uses skip connections to fuse low-level features to obtain decoded features and ground truth labels again.
9. The MRI three-dimensional image segmentation model for spinal myeloma according to claim 1, characterized in that: The tumor segmentation unit (3) further includes the following method steps: The decoded features and ground truth labels are input into the 3D image segmentation model. The 3D image segmentation model uses Diceloss to measure the degree of overlap between the decoded features and ground truth labels, and outputs the probability distribution of the predicted results and the ground truth distribution. Cross-entropy loss is used to measure the difference between the probability distribution of the predicted result and the true distribution. Combining the two can better balance the weights of different categories. The weights are used to train the 3D image segmentation model to obtain a trained model. The trained model is used to perform block prediction on the preprocessed image data using a sliding window strategy, removes false positive regions, and outputs a segmentation mask or probability of myeloma tissue with the same size as the input image.