A pelvis bone and fracture block CT image segmentation method and system based on double decoding

Abstract

The application discloses a pelvis bone and fracture block CT image segmentation method and system based on double decoding, which adopts an encoder-double decoder architecture network model, the encoder comprises a plurality of encoding submodules connected in sequence, the double decoder comprises an anatomical structure decoder and a fracture segmentation decoder, the fracture segmentation decoder integrates a dynamic convolution long short-term memory module and a dynamic spatial feature enhancement module; the encoder extracts multi-scale features of an original CT image through the plurality of encoding submodules, and sends the features extracted by each encoding submodule to the double decoder through a skip connection; the anatomical structure decoder generates an anatomical structure segmentation result according to the features extracted by each encoding submodule; and the fracture segmentation decoder generates a fracture segmentation result according to the features extracted by each encoding submodule. Through the shared encoder and the two task-specific decoders respectively designed, the application realizes the collaborative modeling and efficient segmentation of the anatomical structure and the fracture area.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, specifically to a method and system for segmenting CT images of pelvic bones and fracture fragments based on dual decoding. Background Technology

[0002] Pelvic fracture segmentation is a crucial step in the diagnosis and treatment planning of pelvic trauma. Accurate segmentation of the pelvic fracture site in medical images, typically CT images, provides clinicians with detailed cross-sectional information, helping them accurately assess the fracture type, degree of displacement, and location of fracture fragments, thereby improving the accuracy of preoperative planning and postoperative evaluation. However, the complex anatomy of the pelvis, the diversity of fracture types, and significant individual differences among patients present numerous challenges to automated segmentation. Furthermore, pelvic fractures are often accompanied by other soft tissue injuries or bleeding, and CT images may contain tomographic artifacts, noise, and interference from metal implants; these factors all place higher demands on the robustness of the segmentation algorithm.

[0003] In medical image segmentation research, current work mainly focuses on the automatic segmentation of overall anatomical structures such as the pelvis, ribs, spine, and skull. Regarding pelvic imaging, Liu et al. proposed a cascaded 3D UNet framework for automatically segmenting key skeletal regions such as the hip, sacrum, and lumbar vertebrae from CT scans. However, in pelvic fracture scenarios, accurately extracting fracture fragments rather than complete anatomical regions is more clinically valuable. Current methods for fracture fissure segmentation can be broadly categorized into three types: intensity-based methods, geometry-based methods, and deep learning-based methods.

[0004] Intensity-based methods, such as fixed or adaptive thresholding, watershed algorithms, and region growing, typically rely on the similarity of gray-level intensities and the continuity of edge gradients for region segmentation. However, due to the significant intensity differences between cortical and cancellous bone, and the overlap in gray-level intensity ranges between cancellous bone and soft tissue, these methods often face segmentation difficulties. This challenge frequently leads to gaps in the segmentation results, thus affecting the accurate representation of skeletal structures in terms of morphology and topology.

[0005] Geometric methods, such as graph cut and template matching, rely on the geometric connectivity or morphological similarity between fracture fragments. Yuan et al. proposed a semi-automatic graph cut segmentation method based on continuous maximum flow, which achieves fine segmentation of pelvic fractures by predefining the number of fragments and selecting seed points on each fragment. Similarly, Wang et al. proposed an automatic maximum flow segmentation method combined with boundary enhancement filters. Although graph cut methods perform well when processing non-contact fracture fragments, their effectiveness significantly decreases when fragments come into contact or are compressed.

[0006] Deep learning-based fracture segmentation remains a relatively underdeveloped field, but some studies have demonstrated its significant potential. Yang et al. applied a two-stage Mask R-CNN model to locate and segment intertrochanteric fractures in 2D images. However, this method ignores the 3D spatial background and overall features, which may lead to misinterpretations of the fracture.

[0007] However, existing technologies have the following drawbacks and shortcomings:

[0008] (1) Currently, a common strategy in pelvic fracture image segmentation is to use a two-stage segmentation method. This method typically uses a model to coarsely segment the complete pelvic anatomical structure first, and then performs fine segmentation of the fracture fragments in the second stage. Although there is a natural semantic connection between pelvic structure segmentation and fracture fragment segmentation, in existing two-stage methods, the second stage does not utilize the structural segmentation results of the first stage as input or prior knowledge, resulting in the two tasks exhibiting approximately independent learning processes during training and inference. This lack of coupling makes it difficult to leverage the guiding role of pelvic anatomy in fracture area localization and discrimination, limiting the improvement of the overall segmentation performance of the model. In addition, the two-stage structure requires training two models or stages separately, resulting in high training costs, fragmented optimization processes, and an inability to achieve end-to-end joint learning, which is detrimental to model convergence and obtaining the global optimal result.

[0009] (2) Accurate segmentation of pelvic fracture regions in computed tomography (CT) images is a challenging task, primarily due to the complexity of pelvic anatomy, the irregularity of fracture fragment morphology, and the often blurred or discontinuous fracture boundaries. Commonly used loss functions in medical image segmentation, such as Cross-Entropy Loss and Dice loss, mainly focus on the overlap between predicted and ground truth labels. However, these region-level losses lack modeling of boundary accuracy, making it difficult to effectively constrain the localization accuracy of fracture lines, thus limiting their application potential in downstream tasks such as preoperative planning or intraoperative navigation. Furthermore, fracture fragments typically exhibit significant variability in size, shape, and spatial distribution, especially smaller fragments, which are easily overlooked during training due to class imbalance, leading to lower model sensitivity to small target detection. Simultaneously, structural errors such as false positives or broken connections in the prediction results may still yield low loss values ​​under Dice metrics, thus masking actual segmentation defects. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention provides a method and system for CT image segmentation of pelvic bones and fracture fragments based on dual decoding. By sharing an encoder and designing two task-specific decoders, the method handles the segmentation tasks of pelvic anatomical structures and fracture fragments respectively, thereby achieving collaborative segmentation of structurally clear and fragmented complex regions.

[0011] This invention provides a method for CT image segmentation of pelvic bones and fracture fragments based on dual decoding. The method employs an encoder-dual decoder architecture network model. The encoder includes multiple sequentially connected encoding sub-modules, each including a convolutional layer and a downsampling layer. The dual decoders include an anatomical structure decoder and a fracture segmentation decoder. The fracture segmentation decoder integrates a dynamic convolutional long short-term memory module and a dynamic spatial feature enhancement module. The method includes:

[0012] The encoder extracts multi-scale features of the original CT image through multiple coding sub-modules, and sends the features extracted by each coding sub-module to the dual decoder through skip connections.

[0013] The anatomical structure decoder generates anatomical structure segmentation results based on the features extracted from each coding submodule;

[0014] The fracture segmentation decoder generates fracture segmentation results based on the features extracted from each encoding submodule.

[0015] Preferably, the encoder includes four encoding sub-modules, the dynamic spatial feature enhancement module includes three enhancement sub-modules, and the fracture segmentation decoder further includes three fracture segmentation decoding sub-modules;

[0016] The first encoding submodule receives the original CT image at its input and its output is connected to the input of the second encoding submodule and the input of the first enhancement submodule; the output of the second encoding submodule is connected to the input of the third encoding submodule and the input of the second enhancement submodule; the output of the third encoding submodule is connected to the input of the fourth encoding submodule and the input of the third enhancement submodule; the output of the fourth encoding submodule is connected to the input of the dynamic convolutional long short-term memory module.

[0017] The output of the dynamic convolutional long short-term memory module is concatenated with the output of the third enhancement submodule and used as the input of the third fracture segmentation decoding submodule. The output of the third fracture segmentation decoding submodule is concatenated with the output of the second enhancement submodule and used as the input of the second fracture segmentation decoding submodule. The output of the second fracture segmentation decoding submodule is concatenated with the output of the first enhancement submodule and used as the input of the first fracture segmentation decoding submodule. The output of the first fracture segmentation decoding submodule outputs the fracture segmentation result.

[0018] Preferably, the anatomical structure decoder includes three anatomical structure decoding sub-modules;

[0019] The input and output of the fourth encoding submodule are concatenated and used as the input of the third anatomical structure decoding submodule. The output of the third anatomical structure decoding submodule is concatenated with the output of the second encoding submodule and used as the input of the second anatomical structure decoding submodule. The output of the second anatomical structure decoding submodule is concatenated with the output of the first encoding submodule and used as the input of the first anatomical structure decoding submodule. The output of the first anatomical structure decoding submodule outputs the anatomical structure segmentation result.

[0020] Preferably, each fracture segmentation decoding submodule includes a convolutional layer and an upsampling layer.

[0021] Preferably, each anatomical structure decoding submodule includes a convolutional layer and an upsampling layer.

[0022] Preferably, the network model includes two branches: anatomical structure segmentation and fracture fragment segmentation. During model training, the two branches are respectively introduced with corresponding depth supervision. The output of the fracture segmentation decoder also outputs fracture segmentation prediction results and fracture segmentation mask. The output of the anatomical structure decoder also outputs anatomical structure segmentation prediction results and anatomical structure segmentation mask.

[0023] The training optimization objective of the network model is defined as the sum of the losses of the two branches:

[0024]

[0025] Among them, L anat L represents the loss of anatomical segmentation and branching. frac This indicates the loss of branches due to the fracture fragments.

[0026] Preferably, the anatomical structure segmentation branch loss is defined as the sum of the Dice loss and the cross-entropy loss:

[0027]

[0028] Among them, L Dice L represents Dice loss. CE This represents the cross-entropy loss.

[0029] Preferably, the loss of segmentation branches of the fracture fragments is defined as:

[0030]

[0031] Among them, L Dice L represents Dice loss. CE L represents the cross-entropy loss.BWS This represents the boundary-weighted loss calculated using a distance-weighted mechanism, where λ is the balancing weight parameter.

[0032] Preferably, the boundary-weighted loss is defined as:

[0033]

[0034] Among them, P i D represents the probability value of the i-th pixel predicted by the model. i This represents the distance map value from the i-th pixel to the target boundary, and N represents the total number of pixels involved in the calculation. i and N are positive integers.

[0035] Based on the same inventive concept, the present invention also provides a CT image segmentation system for pelvic bones and fracture fragments based on dual decoding. The system adopts an encoder-dual decoder architecture network model. The encoder includes multiple encoding sub-modules connected in sequence. Each encoding sub-module includes a convolutional layer and a downsampling layer. The dual decoder includes an anatomical structure decoder and a fracture segmentation decoder. The fracture segmentation decoder integrates a dynamic convolutional long short-term memory module and a dynamic spatial feature enhancement module.

[0036] The encoder is used to extract multi-scale features of the original CT image through multiple coding sub-modules, and to send the features extracted by each coding sub-module to the dual decoder through skip connections;

[0037] The anatomical structure decoder is used to generate anatomical structure segmentation results based on the features extracted from each coding submodule;

[0038] The fracture segmentation decoder is used to generate fracture segmentation results based on the features extracted from each coding submodule.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This invention employs a dual-decoder architecture, with the overall process comprising a shared encoder and two task-specific decoding branches. The encoder extracts multi-scale feature representations while preserving rich spatial and semantic information; the anatomical structure branch and the fracture branch perform fine segmentation of the target region, respectively. The anatomical structure decoder focuses on restoring key structural regions such as the left and right iliac bones and sacrum, ensuring anatomical consistency; while the fracture segmentation decoder combines a dynamic convolutional LSTM with a dynamic spatial feature enhancement module to strengthen the modeling ability for local fracture details and long-range dependencies. Through shared feature representations and task-customized decoding strategies, this invention achieves collaborative modeling and optimization of anatomical structures and fracture regions, fully demonstrating its powerful capability for multi-target joint segmentation tasks in 3D pelvic CT. Attached Figure Description

[0041] Figure 1 A flowchart illustrating a method for segmenting CT images of pelvic bones and fracture fragments based on dual decoding, provided by the present invention.

[0042] Figure 2 This is a schematic diagram of the encoder-dual decoder architecture network model provided by the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0044] The present invention will now be described in further detail with reference to the accompanying drawings:

[0045] like Figure 1 As shown, this embodiment of the invention provides a method for CT image segmentation of pelvic bones and fracture fragments based on dual decoding. This method employs an encoder-dual decoder architecture network model. The encoder includes multiple sequentially connected encoding sub-modules, each including a convolutional layer and a downsampling layer. The dual decoders include an anatomical structure decoder and a fracture segmentation decoder. The fracture segmentation decoder integrates a dynamic convolutional long short-term memory module and a dynamic spatial feature enhancement module. The method includes:

[0046] S1. The encoder extracts multi-scale features of the original CT image through multiple coding sub-modules and sends the features extracted by each coding sub-module to the dual decoder through skip connections.

[0047] S2. The anatomical structure decoder generates anatomical structure segmentation results based on the features extracted from each coding submodule.

[0048] S3, the fracture segmentation decoder generates fracture segmentation results based on the features extracted from each encoding submodule.

[0049] like Figure 2As shown in the figure, the overall architecture of the Dual-Decoder Pelvis & Fracture Network (DDPF-Net) proposed in this embodiment of the invention consists of a shared encoder and two task-specific decoders, used for segmenting pelvic anatomical structures and fracture fragments, respectively. The encoder is responsible for extracting multi-scale feature representations and passing low-level semantic information to the two decoders through skip connections. The anatomical structure decoder focuses on segmenting major pelvic regions such as the left and right iliac bones and sacrum to ensure the integrity and consistency of the overall structure; while the fracture segmentation decoder integrates a Dynamic Convolutional Long Short-Term Memory (DyCoLSTM) module and a Dynamic Spatial Feature Enhancement (DSFE) module, used to model long and short-term dependencies and improve spatial awareness, respectively, further enhancing the ability to identify the boundaries of complex fracture fragments. The two decoders are collaboratively optimized, and through a multi-task joint training strategy, efficient joint segmentation of anatomical structures and fracture fragments is achieved.

[0050] from Figure 2 As can be seen, the encoder includes four encoding sub-modules, the dynamic spatial feature enhancement module includes three enhancement sub-modules, the fracture segmentation decoder includes three fracture segmentation decoding sub-modules, and the anatomical structure decoder includes three anatomical structure decoding sub-modules.

[0051] The first encoding submodule receives the raw CT image at its input and its output is connected to the input of the second encoding submodule and the input of the first enhancement submodule. The output of the second encoding submodule is connected to the input of the third encoding submodule and the input of the second enhancement submodule. The output of the third encoding submodule is connected to the input of the fourth encoding submodule and the input of the third enhancement submodule. The output of the fourth encoding submodule is connected to the input of the dynamic convolutional long short-term memory module.

[0052] The output of the dynamic convolutional long short-term memory module is concatenated with the output of the third enhancement submodule and used as the input of the third fracture segmentation decoding submodule. The output of the third fracture segmentation decoding submodule is concatenated with the output of the second enhancement submodule and used as the input of the second fracture segmentation decoding submodule. The output of the second fracture segmentation decoding submodule is concatenated with the output of the first enhancement submodule and used as the input of the first fracture segmentation decoding submodule. The fracture segmentation result is output from the output of the first fracture segmentation decoding submodule.

[0053] The input and output of the fourth encoding submodule are concatenated and used as the input of the third anatomical structure decoding submodule. The output of the third anatomical structure decoding submodule is concatenated with the output of the second encoding submodule and used as the input of the second anatomical structure decoding submodule. The output of the second anatomical structure decoding submodule is concatenated with the output of the first encoding submodule and used as the input of the first anatomical structure decoding submodule. The output of the first anatomical structure decoding submodule outputs the anatomical structure segmentation result.

[0054] like Figure 2 As shown, each encoding submodule includes a convolutional layer and a downsampling layer, each fracture segmentation decoding submodule includes a convolutional layer and an upsampling layer, and each anatomical structure decoding submodule includes a convolutional layer and an upsampling layer.

[0055] like Figure 2 As shown, the network model includes two branches: anatomical structure segmentation and fracture fragment segmentation. During model training, the two branches are respectively introduced with corresponding depth supervision. The output of the fracture segmentation decoder also outputs fracture segmentation prediction and fracture segmentation mask, and the output of the anatomical structure decoder also outputs anatomical structure segmentation prediction and anatomical structure segmentation mask.

[0056] The training optimization objective of the network model is defined as the sum of the losses of the two branches:

[0057]

[0058] Among them, L anat L represents the loss of anatomical segmentation and branching. frac This indicates the loss of branches due to the fracture fragments.

[0059] In this embodiment of the invention, the anatomical structure segmentation branch loss is defined as the sum of the Dice loss and the cross-entropy loss:

[0060]

[0061] Among them, L Dice L represents Dice loss. CE This represents the cross-entropy loss.

[0062] In this embodiment of the invention, the loss of fracture fragment segmentation branches is defined as:

[0063]

[0064] Among them, L Dice L represents Dice loss. CE L represents the cross-entropy loss. BWS This represents the boundary-weighted loss calculated using a distance-weighted mechanism, where λ is the balancing weight parameter.

[0065] In this embodiment of the invention, the boundary-weighted loss is defined as:

[0066]

[0067] Among them, P i D represents the probability value of the i-th pixel predicted by the model. i This represents the distance map value from the i-th pixel to the target boundary, and N represents the total number of pixels involved in the calculation. i and N are positive integers.

[0068] The joint loss function proposed in this invention aims to balance the segmentation needs of two different features: pelvic anatomical structures and fracture fragments. The anatomical structure segmentation branch primarily optimizes Dice loss and cross-entropy loss to ensure the consistency and accuracy of overall structural region segmentation. For the fracture fragment segmentation branch, in addition to Dice loss and cross-entropy loss, a boundary-weighted loss (BWSLoss) is also introduced. BWSLoss, through the fusion of distance weighting mechanisms, effectively enhances the model's sensitivity to fracture fragment boundaries, significantly improving the boundary refinement effect of fracture segmentation under complex morphologies.

[0069] BWSLoss uses the target's boundary distance map as weights and optimizes the element-wise weighted product of the predicted probability map and the boundary distance map, thereby significantly improving the ability to perceive the boundary region of fracture fragments.

[0070] This invention proposes a dual-decoding-based CT image segmentation method for pelvic bones and fracture fragments. Addressing the heterogeneity of pelvic anatomical structures and fracture fragments, a task-specific decoding branch is designed, enabling collaborative modeling and efficient segmentation of anatomical structures and fracture regions. This method not only effectively solves the problems of target confusion and performance imbalance that arise when a single decoder is used for multi-task segmentation, but also significantly improves the modeling capability for structures at different scales. By extracting unified feature representations through a shared encoder and combining them with a task-customized decoding strategy, this approach maintains the overall structural consistency of the anatomical region while improving the accuracy of fracture fragment boundary recognition. This provides solid technical support for auxiliary fracture diagnosis and preoperative planning in clinical practice, demonstrating good application potential and promotional value.

[0071] This invention also provides a CT image segmentation system for pelvic bones and fracture fragments based on dual decoding. The system adopts an encoder-dual decoder architecture network model. The encoder includes multiple encoding sub-modules connected in sequence. Each encoding sub-module includes a convolutional layer and a downsampling layer. The dual decoder includes an anatomical structure decoder and a fracture segmentation decoder. The fracture segmentation decoder integrates a dynamic convolutional long short-term memory module and a dynamic spatial feature enhancement module.

[0072] The encoder is used to extract multi-scale features of the original CT image through multiple coding sub-modules, and sends the features extracted by each coding sub-module to the dual decoder through skip connections;

[0073] The anatomical structure decoder is used to generate anatomical structure segmentation results based on the features extracted from each coding submodule;

[0074] The fracture segmentation decoder is used to generate fracture segmentation results based on the features extracted from each coding submodule.

[0075] Compared with existing technologies, the advantages of this invention are:

[0076] 1. This invention proposes a joint segmentation framework based on a dual-decoder architecture. It constructs pelvic anatomy branches and fracture branches on top of a shared encoder to handle pelvic anatomy segmentation and fracture fragment segmentation tasks respectively. This structure maintains task decoupling while achieving semantic interaction through shared deep features, providing an efficient pelvic fracture segmentation framework.

[0077] 2. This invention designs a joint loss function that combines region overlap optimization with boundary sensitivity modeling to further improve the model's performance in pelvic anatomy and fracture segmentation tasks, especially in boundary localization accuracy and small target recognition capabilities. Specifically, the loss function consists of three parts: the pelvic anatomy branch mainly uses Dice loss and cross-entropy loss to enhance the model's predictive consistency for the overall anatomical region; the fracture branch, in addition to this, introduces a boundary-aware surface loss term, which assigns a higher penalty to misclassified regions near the boundary through a distance map based on the true labels, thereby significantly improving the model's segmentation accuracy for fracture edges. This joint loss provides differentiated supervision mechanisms for the pelvic anatomy and fracture decoder branches during training, enabling the model to simultaneously possess the ability to understand the global anatomical structure and express local fine-grained targets under a unified feature representation, thus achieving a synergistic improvement in segmentation accuracy and boundary robustness.

[0078] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dual-decoding-based CT image segmentation method for pelvic bones and fracture fragments, characterized in that, The method adopts an encoder-decoder architecture network model, the encoder comprises a plurality of encoding sub-modules connected in sequence, each encoding sub-module comprises a convolution layer and a down-sampling layer, the decoder comprises an anatomical structure decoder and a fracture segmentation decoder, the fracture segmentation decoder integrates a dynamic convolution long short-term memory module and a dynamic spatial feature enhancement module; The method comprises: The encoder extracts multi-scale features of the original CT image through a plurality of encoding sub-modules, and sends the features extracted by each encoding sub-module to the decoder through a skip connection; The anatomical structure decoder generates an anatomical structure segmentation result according to the features extracted by each encoding sub-module; The fracture segmentation decoder generates a fracture segmentation result according to the features extracted by each encoding sub-module.

2. The method of claim 1, wherein, The encoder comprises four encoding sub-modules, the dynamic spatial feature enhancement module comprises three enhancement sub-modules, and the fracture segmentation decoder further comprises three fracture segmentation decoding sub-modules; The input end of the first encoding sub-module receives the original CT image, and the output end is connected to the input end of the second encoding sub-module and the input end of the first enhancement sub-module; the output end of the second encoding sub-module is connected to the input end of the third encoding sub-module and the input end of the second enhancement sub-module; the output end of the third encoding sub-module is connected to the input end of the fourth encoding sub-module and the input end of the third enhancement sub-module; and the output end of the fourth encoding sub-module is connected to the input end of the dynamic convolution long short-term memory module; The output of the dynamic convolution long short-term memory module and the output of the third enhancement sub-module are spliced as the input of the third fracture segmentation decoding sub-module, the output of the third fracture segmentation decoding sub-module and the output of the second enhancement sub-module are spliced as the input of the second fracture segmentation decoding sub-module, the output of the second fracture segmentation decoding sub-module and the output of the first enhancement sub-module are spliced as the input of the first fracture segmentation decoding sub-module, and the output end of the first fracture segmentation decoding sub-module outputs the fracture segmentation result.

3. The method of claim 2, wherein, The anatomical structure decoder comprises three anatomical structure decoding sub-modules; The input and output of the fourth encoding sub-module are spliced as the input of the third anatomical structure decoding sub-module, the output of the third anatomical structure decoding sub-module and the output of the second encoding sub-module are spliced as the input of the second anatomical structure decoding sub-module, the output of the second anatomical structure decoding sub-module and the output of the first encoding sub-module are spliced as the input of the first anatomical structure decoding sub-module, and the output end of the first anatomical structure decoding sub-module outputs the anatomical structure segmentation result.

4. The method of claim 2, wherein, Each fracture segmentation decoding sub-module comprises a convolution layer and an up-sampling layer.

5. The method of claim 3, wherein, Each anatomical structure decoding sub-module comprises a convolution layer and an up-sampling layer.

6. The method of claim 1, wherein, The network model comprises two branches of anatomical structure segmentation and fracture fragment segmentation, in the model training process, the two branches are respectively introduced into corresponding deep supervision, the output end of the fracture segmentation decoder also outputs fracture segmentation prediction results and fracture segmentation mask, the output end of the anatomical structure decoder also outputs anatomical structure segmentation prediction results and anatomical structure segmentation mask; The training optimization target of the network model is defined as the sum of two branch losses: where L anat represents the anatomical structure segmentation branch loss, L frac represents the fracture fragment segmentation branch loss.

7. The method of claim 6, wherein, The anatomical structure segmentation branch loss is defined as the sum of Dice loss and cross-entropy loss: where L Dice represents the Dice loss, L CE represents the cross-entropy loss.

8. The method of claim 6, wherein, The fracture fragment segmentation branch loss is defined as: where L Dice represents the Dice loss, L CE represents the cross-entropy loss, L BWS represents the boundary weighted loss calculated by the distance weighting mechanism, and λ is a balance weight parameter.

9. The method of claim 8, wherein, The boundary weighted loss is defined as: where P i represents the probability value of the i-th pixel predicted by the model, D i represents the distance map value of the i-th pixel to the target boundary, and N represents the number of all pixels participating in the calculation, i and N being positive integers.

10. A dual-decoding based pelvis bone and fracture fragment CT image segmentation system for implementing the method of any one of claims 1-9, characterized in that, The system adopts an encoder-double-decoder architecture network model, the encoder comprises a plurality of encoding sub-modules connected in sequence, each encoding sub-module comprises a convolution layer and a down-sampling layer, the double decoder comprises an anatomical structure decoder and a fracture segmentation decoder, the fracture segmentation decoder integrates a dynamic convolution long short-term memory module and a dynamic spatial feature enhancement module; The encoder is used to extract multi-scale features of the original CT image through a plurality of encoding sub-modules, and send the features extracted by each encoding sub-module to the double decoder through a jump connection; The anatomical structure decoder is used to generate anatomical structure segmentation results according to the features extracted by each encoding sub-module; The fracture segmentation decoder is used to generate fracture segmentation results according to the features extracted by each encoding sub-module.

Images