Collaborative optimization multi-task non-enhanced CT blood vessel segmentation device

Abstract

The invention discloses a collaborative optimization multi-task non-enhanced CT blood vessel segmentation device. The collaborative optimization multi-task non-enhanced CT blood vessel segmentation device comprises an image acquisition module, an image preprocessing module, a network construction module, a multi-task multi-scale joint loss function construction module, a branch loss construction module, a joint training module and a branch training module. According to the method, collaborative optimization of double tasks can be promoted while the parameter quantity is greatly reduced in an encoder sharing mode, branch parameters of the tasks are independently and finely adjusted, so that the branch parameters of the tasks are optimal, and the blood vessel form mask and the blood vessel structure mask can be effectively segmented from the non-enhanced CT image.

Description

Technical Field

[0001] The present invention belongs to the technical fields of medical image processing and artificial intelligence, and particularly relates to a multi-task non-enhanced CT vascular segmentation device with collaborative optimization. Background Art

[0002] Since non-enhanced CT (NCCT) has low discrimination for vascular tissues, it faces significant challenges in capturing subtle changes in vascular structures. Therefore, it is particularly important to develop an efficient segmentation model to enhance the discrimination ability of vascular tissues in NCCT. In addition, in clinical practice, it is usually necessary to observe the morphology and anatomical structure of the same vascular site to clarify the cause and formulate a treatment plan. However, most of the current existing segmentation methods, such as nnU-Net [Isensee F, et al. Nature Methods. 2021; 18(2):203-211] and MedNeXt [Roy S, et al. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2023:405-415.], which are representative medical image segmentation models, focus on single tasks. This makes it necessary to develop multiple independent models in multi-task vascular segmentation scenarios. This approach not only leads to a multiple increase in the number of model parameters with the number of tasks but also ignores the potential collaborative relationships between different tasks, thus limiting the further improvement of the performance of each segmentation model. Summary of the Invention

[0003] The purpose of the present invention is to provide a multi-task non-enhanced CT vascular segmentation device with collaborative optimization for the above problems existing in the prior art. Give full play to the collaborative optimization effect between multi-tasks and overcome the inherent limitations of NCCT. The present invention constructs a multi-task segmentation network with the Mamba-CNN hybrid network as the core. Among them, the first task branch focuses on vascular morphology segmentation, and the second task branch is responsible for vascular structure segmentation. Through the design of a shared encoder, this method significantly reduces the number of parameters for dual tasks. At the same time, with the help of the collaborative optimization mechanism between multi-tasks, the overall performance is greatly improved. In addition, the present invention also introduces an independent parameter fine-tuning optimization strategy to further adjust the parameters of each task branch on the basis of collaborative optimization to ensure that each task branch can achieve optimal performance. To sum up, the present invention aims to provide a more accurate, fast and safe vascular multi-task segmentation tool.

[0004] The above object of the present invention is achieved by the following technical means:

[0005] A multi-task non-enhanced CT vascular segmentation device with collaborative optimization, comprising the following modules:

[0006] Image acquisition module: acquire non-enhanced CT images and corresponding enhanced CT images;

[0007] Image preprocessing module: preprocess the non-enhanced CT images and corresponding enhanced CT images, obtain registered non-enhanced CT image patches as samples to generate a training dataset, and obtain vascular morphology mask patches and vascular structure mask patches as labels corresponding to the samples;

[0008] Network construction module: construct a multi-task segmentation network, which includes an encoder, a bottleneck layer, a skip module, a first task branch, and a second task branch. The first task branch includes a first decoder and a first multi-level target output module, and the second task branch includes a second decoder and a second multi-level target output module;

[0009] Multi-task multi-scale joint loss function construction module: construct a multi-task multi-scale joint loss function;

[0010] Branch loss construction module: construct a first task branch loss and a second task branch loss;

[0011] Joint training module: use the training dataset to jointly train the multi-task segmentation network based on minimizing the multi-task multi-scale joint loss function;

[0012] Branch training module: for the multi-task segmentation network after joint training, fix the network parameters except those of the first task branch, use the training dataset, and train the network parameters of the first task branch based on minimizing the first task branch loss; for the multi-task segmentation network after training the network parameters of the first task branch, fix the network parameters except those of the second task branch, use the training dataset, and train the network parameters of the second task branch based on minimizing the second task branch loss.

[0013] The preprocessing of the non-enhanced CT images and corresponding enhanced CT images as described above includes the following steps:

[0014] Perform image resampling on all non-enhanced CT images and corresponding enhanced CT images;

[0015] Register the resampled enhanced CT images to the space of the non-enhanced CT images to obtain registered enhanced CT images and registered non-enhanced CT images;

[0016] Based on the registered enhanced CT images, outline the masks for vascular morphology and vascular structure respectively to obtain a vascular morphology mask and a vascular structure mask;

[0017] Perform normalization processing on the registered non-enhanced CT images;

[0018] The registered non-enhanced CT image, the corresponding registered enhanced CT image, and the registered non-enhanced CT image are cropped to the target size to obtain the registered non-enhanced CT image patches and the corresponding actual target segmentation mask patches. The registered non-enhanced CT image patches are used as samples to generate a training dataset, and the corresponding actual target segmentation mask patches are used as the corresponding labels.

[0019] As described above, the cropping of the registered non-enhanced CT image, the corresponding registered enhanced CT image, and the registered non-enhanced CT image includes the following steps:

[0020] The registered non-enhanced CT image, the corresponding vessel morphology mask, and the vessel structure mask are cropped in the same cropping manner according to the set ratio of the target segmentation region voxels to the non-target segmentation region voxels. The cropping manner is randomly selected to obtain the registered non-enhanced CT image patches, the corresponding vessel morphology mask patches, and the vessel structure mask patches respectively. The registered non-enhanced CT image patches, the corresponding vessel morphology mask patches, and the vessel structure mask patches are augmented by data perturbation through random rotation, flipping, jittering, and scaling.

[0021] In each cropping manner, the registered non-enhanced CT image, the corresponding vessel morphology mask, and the vessel structure mask are cropped according to the specified patch size and the specified patch overlap degree.

[0022] As described above, the encoder includes L encoding layers. Each encoding layer includes a downsampling convolutional layer, a first convolutional block, a Mamba block, and a second convolutional block. The input features of the encoding layer are sequentially processed by the downsampling convolutional layer, the first convolutional block, the Mamba block, and the second convolutional block to obtain the output features.

[0023] The bottleneck layer includes r sequentially connected class residual modules. The output features of the last encoding layer of the encoder are used as the input features of the bottleneck layer. The input features of the bottleneck layer are sequentially processed by each class residual module to obtain the output features of the bottleneck layer.

[0024] Both the first decoder and the second decoder include L decoding layers: Each decoding layer includes an upsampling convolutional layer and a decoding convolutional layer connected in sequence. A skip connection is established between the Mamba block of the l-th encoding layer and the corresponding (L - l)-th upsampling convolutional layer. The input features of the first decoding layer are the output features of the bottleneck layer. The output features of the decoding layer are input to the 3D transposed convolutional output layer of the corresponding layer. The 3D transposed convolutional output layers of each layer output the predicted masks at the corresponding hierarchical scales. The 3D transposed convolutional output layers corresponding to the respective decoding layers of the first decoder constitute the first multi-level target output module, and the 3D transposed convolutional output layers corresponding to the respective decoding layers of the second decoder constitute the second multi-level target output module.

[0025] Establishing a skip connection between the Mamba block of the l-th encoding layer and the corresponding (L-l)-th upsampling convolutional layer as described above includes:

[0026] After the input features are input to the decoding layer, the upsampled features are obtained through the processing of the upsampling convolutional layer. The output features of the Mamba block of the encoding layer corresponding to the layer pass through the skip module and then output the skip features. The skip features are added to the output features of the upsampling convolutional layer of the decoding layer corresponding to the layer and then input to the decoding convolutional layer of the decoding layer corresponding to the layer. The output features of the decoding convolutional layer are used as the output features of the decoding layer. The output features of the decoding layer are input to the 3D transposed convolutional output layer corresponding to the layer and the decoding layer of the next layer.

[0027] The input features of the upsampling convolutional layer of the last decoder include the output features of the previous decoding layer and the skip features corresponding to the input features of the encoder. The input features of the encoder pass through the skip module and then output the corresponding skip features. The upsampling convolutional layer of the last decoding layer is the Unetr upsampling module. The output features of the Unetr upsampling module are input to the decoding convolutional layer corresponding to the layer. The output features of the decoding convolutional layer are input to the 3D transposed convolutional output layer corresponding to the layer.

[0028] As described above, the multi-task multi-scale joint loss function is based on the following formula:

[0029] L seg＝ 0.6 * L seg1 + 0.4 * L seg2

[0030]

[0031] L1 Sal = Loss1 DSCl + Loss1 CEl

[0032]

[0033] L2 Sbl = Loss2 DSCl + Loss2 CEl

[0034]

[0035] Where: L seg is the multi-task multi-scale joint loss function, L seg1 and L seg2 are the first multi-scale loss and the second multi-scale loss respectively,

[0036] L1 Sal is the supervision loss at the scale corresponding to the l-th decoding layer of the first decoder,

[0037] Loss1 DSCl is the segmentation loss at the scale corresponding to the \(l\)-th decoding layer of the first decoder, Loss1 CEl is the cross-entropy loss at the scale corresponding to the \(l\)-th decoding layer of the first decoder

[0038] X1 l is the predicted mask at the scale corresponding to the \(l\)-th decoding layer of the first decoder, Y1 l is the vascular morphology mask patch at the scale corresponding to the \(l\)-th decoding layer of the first decoder

[0039] X1 li is the \(i\)-th voxel point of the predicted mask at the scale corresponding to the \(l\)-th decoding layer of the first decoder, Y1 li is the \(i\)-th voxel point of the vascular morphology mask patch at the scale corresponding to the \(l\)-th decoding layer of the first decoder, \(N\) is the total number of voxel points, \(\|\) and \(\sum\) are for obtaining the number of voxel points

[0040] L2 Sbl is the supervision loss at the scale corresponding to the \(l\)-th decoding layer of the second decoder

[0041] Loss2 DSCl is the segmentation loss at the scale corresponding to the \(l\)-th decoding layer of the second decoder, Loss2 CEl is the cross-entropy loss at the scale corresponding to the \(l\)-th decoding layer of the second decoder

[0042] X2 l is the predicted mask at the scale corresponding to the \(l\)-th decoding layer of the second decoder, Y2 l is the vascular structure mask patch at the scale corresponding to the \(l\)-th decoding layer of the second decoder

[0043] X2 li is the \(i\)-th voxel point of the predicted mask at the scale corresponding to the \(l\)-th decoding layer of the second decoder, Y2 li is the \(i\)-th voxel point of the vascular structure mask patch at the scale corresponding to the \(l\)-th decoding layer of the second decoder

[0044] The first task branch loss as described above is:

[0045] L segz1 = L1 SazL + 0.5 * L1 Saz(L-1)

[0046] L1 Sazf = Loss1 CEf + 0.5 * Loss1 DSCf + 0.5 * Loss1 IoUf

[0047]

[0048] The second task branch loss is as follows:

[0049] L segz2 = L2 SbzL + 0.5 * L2 Sbz(L-1)

[0050] L2 Sbzf = Loss2 CEf + 0.5 * Loss2 DSCf + 0.5 * Loss2 IoUf

[0051]

[0052]

[0053] Wherein, L segz1 is the first task branch loss, L1 SazL is the task branch supervision loss at the scale corresponding to the decoding layer of the L-th layer of the first decoder, L1 Saz(L-1) is the task branch supervision loss at the scale corresponding to the decoding layer of the (L - 1)-th layer of the first decoder,

[0054] L1 Sazf is the task branch supervision loss at the scale corresponding to the decoding layer of the f-th layer of the first decoder, Loss1 CEf is the cross-entropy loss at the scale corresponding to the f-th decoding layer of the first decoder, Loss1 DSCf is the segmentation loss at the scale corresponding to the f-th decoding layer of the first decoder, Loss1 IoUf is the target detection loss IOU at the scale corresponding to the l-th decoding layer of the first decoder, f is L or L - 1, || and is to obtain the number of voxel points, N is the total number of voxel points,

[0055] X1 f is the predicted mask at the scale corresponding to the f-th decoding layer of the first decoder, Y1 f is the vascular morphology mask tile at the scale corresponding to the f-th decoding layer of the first decoder; X1 fi is the i-th voxel point of the predicted mask at the scale corresponding to the f-th decoding layer of the first decoder, Y1 fi is the i-th voxel point of the vascular morphology mask tile at the scale corresponding to the f-th decoding layer of the first decoder,

[0056] L segz2 is the second task branch loss, L2 SbzL is the task branch supervision loss at the scale corresponding to the decoding layer of the L-th layer of the second decoder, L2 Sbz(L-1)is the task branch supervision loss at the scale corresponding to the decoding layer of the (L - 1)-th layer of the second decoder

[0057] L2 Sbzf is the task branch supervision loss at the scale corresponding to the decoding layer of the f-th layer of the second decoder, Loss2 CEf is the cross-entropy loss at the scale corresponding to the f-th decoding layer of the second decoder, Loss2 DSCf is the segmentation loss at the scale corresponding to the f-th decoding layer of the second decoder, Loss2 IoUf is the object detection loss IOU at the scale corresponding to the f-th decoding layer of the second decoder

[0058] X2 f is the predicted mask at the scale corresponding to the f-th decoding layer of the second decoder, Y2 f is the vascular structure mask tile at the scale corresponding to the f-th decoding layer of the second decoder; X2 fi is the i-th voxel point of the predicted mask at the scale corresponding to the f-th decoding layer of the second decoder, Y2 fi is the i-th voxel point of the vascular structure mask tile at the scale corresponding to the f-th decoding layer of the second decoder

[0059] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it realizes the functions of each module of the above-mentioned segmentation device

[0060] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it realizes the functions of each module of the above-mentioned segmentation device

[0061] The present invention has the following beneficial effects compared with the prior art

[0062] 1. The present invention designs a multi-task non-enhanced CT vascular segmentation device with a Mamba-CNN hybrid structure, which makes full use of the long-distance modeling ability of Mamba and the ability of convolution to effectively capture local features to achieve more accurate segmentation of vascular target tissues

[0063] 2. The constructed multi-task segmentation network utilizes the task correlation between vascular morphology segmentation and vascular structure segmentation, and through shared encoder and branch task tuning learning, while compressing model parameters, it ensures the effectiveness of multi-task segmentation

[0064] 3. A multi-task multi-scale joint loss function is constructed to enable multi-task branches to cooperate and learn, enabling the shared encoder to learn the common knowledge of different tasks, thereby achieving performance improvement

[0065] 4. In the embodiment, according to the characteristics of the blood vessel morphology and structure, a series of preprocessing and model training steps are designed to ensure the efficiency and accuracy of the entire segmentation process. Description of the Drawings

[0066] Figure 1 It is a schematic structural diagram of the multi-task segmentation network of the present invention.

[0067] Figure 2 It is a schematic diagram of the prediction mask results at each stage of the present invention. After the joint training of the multi-task segmentation network, relatively complete blood vessel morphology masks and blood vessel structure masks are obtained, but the segmentation of the aneurysm part of the ascending aorta is not complete enough. After further branch training, the segmentation accuracy of the blood vessel morphology masks and blood vessel structure masks of the aneurysm part of the ascending aorta is further improved, and the resulting masks are highly consistent with the gold standard masks. Detailed Embodiments

[0068] To facilitate the understanding and implementation of the present invention by those of ordinary skill in the art, the present invention will be further described in detail below with reference to the embodiments. The embodiments described herein are only used to illustrate and explain the present invention, and are not intended to limit the present invention.

[0069] Embodiment 1:

[0070] As Figure 1 shown, a collaborative optimization multi-task non-enhanced CT blood vessel segmentation method for segmenting blood vessel morphology masks and blood vessel structure masks from non-enhanced CT images. Taking the aorta of non-enhanced CT images as an example, it includes the following steps:

[0071] Step 1. Obtain non-enhanced CT images and corresponding enhanced CT images: Preferably, formulate data inclusion and exclusion rules before collecting images, and collect non-enhanced CT images and corresponding enhanced CT images of the target tissue according to the formulated rules. The inclusion and exclusion rules can be one or more of the following criteria:

[0072] (1) The scanned parts corresponding to the non-enhanced CT images and the enhanced CT images are the chest and abdomen;

[0073] (2) The scanning interval time corresponding to the non-enhanced CT images and the enhanced CT images does not exceed 15 days;

[0074] (3) The corresponding layers of the non-enhanced CT images and the enhanced CT images correspond one by one without significant misalignment;

[0075] (4) The qualities of the non-enhanced CT images and the enhanced CT images meet the requirements of routine diagnosis, and there are no significant artifacts, distortions or deformations;

[0076] Step 2. Preprocess the non-enhanced CT images and the corresponding enhanced CT images:

[0077] Step 2.1: Resample all non-enhanced CT images and their corresponding enhanced CT images to a resolution of (1.25, 0.67, 0.67), that is, the preset axial resampling resolution is 1.25 mm, and the preset cross-sectional resampling resolution is 0.67 mm.

[0078] Step 2.2: Register the resampled enhanced CT images to the space of the non-enhanced CT images to obtain registered enhanced CT images and registered non-enhanced CT images, so as to ensure that the masks outlined from the enhanced CT images correspond to the blood vessels in the non-enhanced CT. In this embodiment, the Elastix registration software Python library PyElastix is used for registration processing and the registered data without serious deformation is saved.

[0079] Step 2.3: Outline masks for blood vessel morphology and blood vessel structure based on the registered enhanced CT images respectively to obtain actual target segmentation masks, where the actual target segmentation masks include blood vessel morphology masks and blood vessel structure masks. Mask outlining can use the growth threshold method function module or the cropping function module of 3D slicer or other mask-making function modules to outline the targets for the registered enhanced CT images, and actual target segmentation masks (blood vessel morphology masks and blood vessel structure masks) are obtained respectively. The blood vessel morphology mask is an aneurysm morphology mask, or a dissection mask, or a stenosis morphology blood vessel mask, or a normal morphology blood vessel mask, and the blood vessel structure mask is a lumen mask, or a calcification mask, or a non-calcification mask, or a false lumen mask.

[0080] Step 2.4: Perform normalization processing on the registered non-enhanced CT images for use in multi-task segmentation network training. Here, only the registered non-enhanced CT images need to be normalized, and there is no need to normalize the registered enhanced CT images and the corresponding outlined actual target segmentation masks.

[0081] Step 2.5: Crop the registered non-enhanced CT images and the corresponding outlined actual target segmentation masks to the target size according to the specified rules.

[0082] Among them, the registered non-enhanced CT images and the corresponding actual target segmentation masks (including blood vessel morphology masks and blood vessel structure masks) are cropped in the same cropping manner according to the set ratio of the voxel of the target segmentation area - the voxel of the non-target segmentation area. The cropping manner is randomly selected, and registered non-enhanced CT image blocks and the corresponding actual target segmentation mask blocks are obtained respectively. The registered non-enhanced CT image blocks and the corresponding actual target segmentation mask blocks are perturbed and augmented through random rotation, flipping, jittering, and scaling. The registered non-enhanced CT image blocks are used as samples to generate a training dataset, and the corresponding actual target segmentation mask blocks (including blood vessel morphology mask blocks and blood vessel structure mask blocks) are used as the labels corresponding to the registered non-enhanced CT image blocks.

[0083] In this embodiment, the registered non-enhanced CT image and the corresponding actual target segmentation mask are cropped in the same cropping manner according to the ratio of the number of voxels in the target segmentation region: the number of voxels in the non-target segmentation region being 1:1. The cropping manner is randomly selected to obtain a non-enhanced CT image patch and an actual target segmentation mask patch with an image depth × image width × image height of 128×192×192. The number of voxels in the target segmentation region: the number of voxels in the non-target segmentation region in the non-enhanced CT image patch and the corresponding target tissue mask is 1:1.

[0084] For each cropping manner, the patch size and patch overlap are the same. The registered non-enhanced CT image and the corresponding actual target segmentation mask are cropped in the same cropping manner according to the specified patch size and specified patch overlap to obtain a registered non-enhanced CT image patch and the corresponding actual target segmentation mask patch respectively. The registered non-enhanced CT image patch is used as a sample to generate a training data set, and the corresponding actual target segmentation mask patch is used as the label corresponding to the registered non-enhanced CT image patch.

[0085] In this embodiment, the patch overlap is 0.25, and a registered non-enhanced CT image patch and an actual target segmentation mask patch with an image depth × image width × image height of 128×192×192 are obtained.

[0086] In this example, the number of sample batches batch = 1, the number of channels is 1, the input is a registered non-enhanced CT image patch with a size of 1×1×128×192×192 (number of sample batches × number of channels × image depth × image width × image height), and the multi-level target outputs are one-hot encodings with sizes of 1×1×8×12×12, 1×1×16×24×24, 1×1×32×48×48, 1×1×64×96×96, and 1×1×128×192×192 respectively. During the training process of the multi-task segmentation network, the actual target segmentation mask patches (vascular morphology mask and vascular structure mask) at different levels are generated by the data online processing module of nnU-Net for multi-level size adjustment. The sizes of the actual target segmentation mask patches (vascular morphology mask and vascular structure mask) at each level are kept consistent to match the corresponding multi-level output masks. Specifically, the actual target segmentation mask patches (vascular morphology mask and vascular structure mask) at level 4 are equivalent to 1 / 2 of the actual target segmentation mask patches (vascular morphology mask and vascular structure mask) at level 3, 1 / 4 of the actual target segmentation mask patches at level 2, and 1 / 8 times of the actual target segmentation mask patches at level 1. The actual target segmentation mask patch at level 1 is the actual target segmentation mask patch without size adjustment.

[0087] Step 3. Construction of the multi-task segmentation network: In this embodiment, a multi-task segmentation network architecture with a Mamba-CNN hybrid structure is constructed using the Python language based on the Segmamba encoder [Xing Z, et al. International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 578-588.] of the MONAI open-source medical image deep learning framework and the MedNeXt convolutional block [Roy S, et al. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2023: 405-415.]. Through the encoding process of the registered non-enhanced CT image patches, deep semantic feature extraction, high-resolution semantic feature reconstruction, and multi-level target mask supervised learning, the vascular morphology mask (aortic morphology mask in this embodiment) and the vascular structure mask (aortic structure mask in this embodiment) are segmented from the registered non-enhanced CT image.

[0088] The multi-task segmentation network includes a shared encoder, a shared bottleneck layer, a shared skip module, a first decoder of the first task branch and a corresponding first multi-level target output module, and a second decoder of the second task branch and a corresponding second multi-level target output module. The 3D transposed convolution output layers (Sa1 to Sa4 in this embodiment) corresponding to the decoding layers of the first decoder constitute the first multi-level target output module, and the 3D transposed convolution output layers (Sb1 to Sb4 in this embodiment) corresponding to the decoding layers of the second decoder constitute the second multi-level target output module.

[0089] Among them, the encoder includes L (L>1) encoding layers. Each encoding layer includes a downsampling convolutional layer, a first convolutional block, a Mamba block, and a second convolutional block. The input feature of the encoding layer is processed by the downsampling convolutional layer, the first convolutional block, the Mamba block, and the second convolutional block in sequence to obtain the output feature. The first convolutional block, the Mamba block, and the second convolutional block of each encoding layer cooperate to complete the preliminary encoding and downsampling tasks of the input feature.

[0090] The bottleneck layer includes r (r≥1) sequentially connected class residual modules. The output features of the last encoding layer of the encoder serve as the input features of the bottleneck layer. The input features of the bottleneck layer are sequentially processed by each class residual module to obtain the output features of the bottleneck layer, which are responsible for further mining the deep features of the image. The output features of the bottleneck layer are respectively input into the first decoder and the second decoder;

[0091] Since the network structures of the first task branch and the second task branch are exactly the same, the decoder in the first task branch serves as the first decoder, and the decoder in the second task branch serves as the second decoder.

[0092] The first decoder and the second decoder have the same structure, both including L decoding layers. Each decoding layer of the first decoder corresponds to each encoding layer of the encoder, and each decoding layer of the second decoder corresponds to each encoding layer of the encoder. They are both L (L>1) layers. Each decoding layer includes an upsampling convolutional layer and a decoding convolutional layer connected in sequence to achieve the step-by-step decoding and upsampling restoration of image features. In addition, a skip connection is established between the Mamba block of the l (l<L) -th encoding layer and the corresponding (L - l) -th upsampling convolutional layer. That is, for each decoding layer: after the input features are input into the decoding layer, the upsampling features are obtained through the processing of the upsampling convolutional layer. The output features of the Mamba block of the corresponding encoding layer pass through the skip module to output the skip features. The skip features are added to the output features of the upsampling convolutional layer of the corresponding decoding layer and then input into the decoding convolutional layer of the corresponding decoding layer. The output features of the decoding convolutional layer serve as the output features of the decoding layer. The output features of the decoding layer are input into the 3D transposed convolutional output layer of the corresponding layer and the next decoding layer.

[0093] The input features of the first decoding layer are the output features of the bottleneck layer.

[0094] The upsampling convolutional layer of the last decoding layer is the Unetr upsampling module. The input features of the upsampling convolutional layer of the last decoder include the output features of the previous decoding layer and the skip features corresponding to the input features of the encoder. The input features of the encoder pass through the skip module to output the corresponding skip features. The output features of the Unetr upsampling module are input into the decoding convolutional layer of the corresponding layer, and the output features of the decoding convolutional layer are input into the 3D transposed convolutional output layer of the corresponding layer.

[0095] The skip module (shared skip connection) includes a convolutional layer and a Unetr basic block connected in sequence.

[0096] The first decoder and the second decoder respectively correspond to the first multi-level target output module and the second multi-level target output module. Both the first multi-level target output module and the second multi-level target output module include l layers of 3D transposed convolution output layers. Each layer of 3D transposed convolution output layer outputs a predicted mask at the corresponding hierarchical scale, providing omni-directional supervision and optimization for the segmentation result.

[0097] In this embodiment, the specific structure is as follows:

[0098] Encoder:

[0099] In this embodiment, the encoder includes 4 encoding layers. The input feature of the first encoding layer is a single-channel registered non-enhanced CT image patch. Specifically, the channel change of the entire encoder is 1->48->96->192->384. The input features corresponding to each encoding layer and the output feature size of the last encoding layer are 1×1×128×192×192, 1×48×64×96×96, 1×96×32×48×48, 1×192×16×24×24, 1×384×8×12×12 respectively. Among them, for each encoding layer: the downsampling convolutional layer includes a 3D convolutional layer with a convolutional kernel of 1 and a convolutional stride of 2. The first convolutional block includes 2 layers of 3D convolutional blocks with a convolutional kernel of 3 and a stride of 1. After performing the 3D convolutional block, 3D InstanceNorm normalization and ReLu non-linear transformation are carried out. The state space expansion factor of the Mamba block is 16, the block expansion factor is 2, the local convolutional width is 4, and the normalization parameter is 3D InstanceNorm. The second convolutional block includes 2 layers of 3D convolutional blocks with a convolutional kernel of 1 and a convolutional stride of 1.

[0100] The registered non-enhanced CT image patch is input into the encoder. After the first convolutional downsampling, through channel expansion, the feature expression ability is increased; then through convolutional feature extraction, Mamba sequence modeling, and convolutional feature induction, features at different levels and different scales are extracted; the output feature of the last encoding layer is input into the bottleneck layer; then, the output feature of the bottleneck layer gradually restores the pixel space through the upsampling convolutional layer (transposed convolution) and decoding convolutional layer of the decoder.

[0101] Bottleneck layer:

[0102] The bottleneck layer includes multiple class residual modules (MedNeXt convolutional blocks). The number of class residual modules (MedNeXt convolutional blocks) used in the bottleneck layer of this instance is 8. The channel expansion rate of the bottleneck layer is 8. The inter-layer normalization strategy is the normalization group. The convolutional kernel is 3, the convolutional stride is 1, and the activation function used is GELU. After the output feature of the deep data output by the last encoding layer passes through the bottleneck layer, the output feature of the bottleneck layer is input into the upsampling convolutional layer of the first decoding layer of the first decoder and the second decoder.

[0103] Jump module:

[0104] The jump module consists of a convolutional layer and a Unetr basic block. The convolutional layer has a kernel size of 1, a stride of 1, an input channel of 1, and an output channel of 48. The Unetr basic block has a kernel size of 3, a stride of 1, and input and output channels of 48. In addition to establishing a jump connection between the l-th (l < L) downsampling convolutional layer and the (L - l)-th upsampling convolutional layer of the corresponding layer, a jump connection is also established between the input features of the encoder and the last upsampling convolutional layer.

[0105] Decoder (the first decoder and the second decoder):

[0106] The decoder consists of 4 decoding layers. Each decoding layer includes a decoding convolutional layer and a corresponding upsampling convolutional layer, and there is no channel change between the decoding convolutional layer and the upsampling convolutional layer within the same decoding layer. The number of channels between adjacent decoding layers is halved layer by layer, and the channel change order of the entire decoder is 768 -> 384 -> 192 -> 96 -> 48. Specifically, the decoding convolutional layer uses the MedNeXt convolutional block, and the numbers are as follows: the MedNeXt convolutional block of the decoding convolutional layer of the first decoding layer is 8, the MedNeXt convolutional block of the decoding convolutional layer of the second decoding layer is 8, the MedNeXt convolutional block of the decoding convolutional layer of the third decoding layer is 4, and the MedNeXt convolutional block of the decoding convolutional layer of the fourth decoding layer is 3. The channel expansion rate of the decoding convolutional layer is also set to 8, 8, 4, and 3 accordingly, and other parameter settings are the same as those of the bottleneck layer. The last upsampling convolutional layer of the decoder uses the Unetr upsampling module, and the other upsampling convolutional layers have the same parameter settings as the downsampling convolutional layer of the encoder, except that the downsampling convolutional layer in the encoder is replaced by a transposed convolutional upsampling layer.

[0107] During the decoding process, each decoding layer of the decoder gradually performs deconvolution and transposed convolution operations on the output features of the bottleneck layer to restore the pixel space information. The output features of the decoding convolutional layers of each level of the decoding layer are respectively input into the 3D transposed convolutional output layers of the corresponding layers of the corresponding multi-level target output module. The output of the last decoding convolutional layer in this embodiment is the expected mask, and the outputs of other levels are mainly used for supervised learning. Specifically, the number of channels of the decoding layer of the first level is reduced from 768 to 384, including 8 MedNeXt convolutional blocks with a channel expansion rate of 8; the number of channels of the decoding layer of the second level is reduced from 384 to 192, including 8 MedNeXt convolutional blocks with a channel expansion rate of 8; the number of channels of the decoding layer of the third level is reduced from 192 to 96, including 4 MedNeXt convolutional blocks with a channel expansion rate of 4; the number of channels of the decoding layer of the fourth level is reduced from 96 to 48, including 3 MedNeXt convolutional blocks with a channel expansion rate of 3. Through this layer-by-layer decreasing design, the decoder can effectively restore the spatial resolution of the image.

[0108] (4) Multi-level target output module

[0109] Both the first multi-level target output module and the second multi-level target output module in this embodiment include L 3D transposed convolutional output layers (4 in this embodiment). The convolutional kernel size and stride of each 3D transposed convolutional output layer are both 1, and there is no activation layer. The final number of channels of the output features of each layer of the 3D transposed convolutional output layer (in this embodiment, the 3D transposed convolutional output layer Sa1 of the first layer to the 3D transposed convolutional output layer Sa4 of the fourth layer, and the 3D transposed convolutional output layer Sb1 of the first layer to the 3D transposed convolutional output layer Sb4 of the fourth layer) is 1. The output features of the decoding convolutional layers of each level of the decoder generate prediction masks of different scales through the corresponding 3D transposed convolutional output layers. Specifically, for the first decoder and the second decoder: the side length (width and height) of the output features of the decoding convolutional layer of the fourth layer is twice that of the output features of the decoding layer of the third layer; the side length of the output features of the decoding layer of the third layer is twice that of the output features of the decoding layer of the second layer; the side length of the output features of the decoding layer of the second level is twice that of the output features of the decoding layer of the first level; this layer-by-layer increasing design enables the multi-task segmentation network to learn to distinguish targets at multiple scales from the registered non-enhanced CT image patches, thereby achieving efficient target segmentation.

[0110] Step 4: Construct a multi-task multi-scale joint loss function for co-training the multi-task segmentation network:

[0111] The training loss function used for co-training in this example is a multi-task multi-scale joint loss function, and the formula is:

[0112] L seg＝0.6*L seg1 +0.4*L seg2

[0113] Among them, L seg is the multi-task multi-scale joint loss function, L seg1 and L seg2 are the first multi-scale loss and the second multi-scale loss respectively, and the loss calculation methods of both are exactly the same.

[0114] The first multi-scale loss L seg1 is defined as:

[0115]

[0116] Among them, L1 Sal is the supervision loss at the scale corresponding to the l-th decoding layer of the first decoder. L is the total number of decoding layers of the first decoder, which is also the total number of encoding layers of the encoder and the total number of 3D transposed convolution output layers of the multi-level target output module.

[0117] L1 Sal = Loss1 DSCl + Loss1 CEl

[0118]

[0119] Among them, L1 Sal is the supervision loss at the scale corresponding to the l-th decoding layer of the first decoder, X1 l is the prediction mask at the scale corresponding to the l-th decoding layer of the first decoder, Y1 l is the vascular morphology mask patch at the scale corresponding to the l-th decoding layer of the first decoder. Y1 l and X1 l are of the same size at the same scale. X1 li is the i-th voxel point of the prediction mask at the scale corresponding to the l-th decoding layer of the first decoder, Y1 li is the i-th voxel point of the vascular morphology mask patch at the scale corresponding to the l-th decoding layer of the first decoder. N is the total number of voxel points, || and are used to calculate the number of voxel points, |X1 l | and |Y1 l | respectively represent the number of voxel points of X1 l and Y1 l |X1 l ∩Y1 l | represents the number of voxel points in the intersection of X1 l and Y1 l Loss1 DSCl is the segmentation loss at the scale corresponding to the l-th decoding layer of the first decoder, Loss1CEl is the cross-entropy loss at the scale corresponding to the l-th decoding layer of the first decoder;

[0120] The second multi-scale loss L seg2 is defined as:

[0121]

[0122] where, L2 Sbl is the supervision loss at the scale corresponding to the l-th decoding layer of the second decoder, and L is the total number of decoding layers of the second decoder.

[0123] L2 Sbl = Loss2 DSCl + Loss2 CEl

[0124]

[0125] where, L2 Sbl is the supervision loss at the scale corresponding to the l-th decoding layer of the second decoder, X2 l is the prediction mask at the scale corresponding to the l-th decoding layer of the second decoder, Y2 l is the vascular structure mask patch at the scale corresponding to the l-th decoding layer of the second decoder, Y2 l and X2 l are of the same size at the same scale, X2 li is the i-th voxel point of the prediction mask at the scale corresponding to the l-th decoding layer of the second decoder, Y2 li is the i-th voxel point of the vascular structure mask patch at the scale corresponding to the l-th decoding layer of the second decoder, N is the total number of voxel points, || and are used to calculate the number of voxel points, |X2 l | and |Y2 l | respectively represent the number of voxel points of X2 l and Y2 l |X2 l ∩Y2 l | represents the number of intersecting voxel points of X2 l and Y2 l Loss2 DSCl is the segmentation loss at the scale corresponding to the l-th decoding layer of the second decoder, Loss2 CEl is the cross-entropy loss at the scale corresponding to the l-th decoding layer of the second decoder;

[0126] Step 5, Construction of the branch optimization loss function of the multi-task segmentation network:

[0127] Taking the first task branch as an example, the first task branch loss L segz1 is:

[0128] L segz1 = L1 SazL + 0.5 * L1 Saz(L-1)

[0129] L1 Sazf = Loss1 CEf + 0.5 * Loss1 DSCf + 0.5 * Loss1 IoUf

[0130]

[0131]

[0132] Wherein, L segz1 is the loss of the first task branch, and L1 SazL is the task branch supervision loss at the scale corresponding to the decoding layer of the L-th layer of the first decoder, and L1 Saz(L-1) is the task branch supervision loss at the scale corresponding to the decoding layer of the (L - 1)-th layer of the first decoder,

[0133] X1 f is the predicted mask at the scale corresponding to the f-th decoding layer of the first decoder, and Y1 f is the vascular morphology mask patch at the scale corresponding to the f-th decoding layer of the first decoder; X1 fi is the i-th voxel point of the predicted mask at the scale corresponding to the f-th decoding layer of the first decoder, and Y1 fi is the i-th voxel point of the vascular morphology mask patch at the scale corresponding to the f-th decoding layer of the first decoder,

[0134] L1 Sazf is the task branch supervision loss at the scale corresponding to the decoding layer of the f-th layer of the first decoder, and Loss1 CEf is the cross-entropy loss at the scale corresponding to the f-th decoding layer of the first decoder, and Loss1 DSCf is the segmentation loss at the scale corresponding to the f-th decoding layer of the first decoder, and Loss1 IoUf is the object detection loss IOU (Jaccard) at the scale corresponding to the l-th decoding layer of the first decoder, and f is L or L - 1.

[0135] || And is to obtain the number of voxel points.

[0136] Taking the second task branch as an example, the second task branch loss L segz2 is:

[0137] L segz2 = L2 SbzL + 0.5 * L2 Sbz(L-1)

[0138] L2 Sbzf = Loss2 CEf + 0.5 * Loss2 DSCf + 0.5 * Loss2 IoUf

[0139]

[0140] where L segz2 is the loss of the second task branch, and L2 SbzL is the task branch supervision loss at the scale corresponding to the decoding layer of the L-th layer of the second decoder, and L2 Sbz(L-1) is the task branch supervision loss at the scale corresponding to the decoding layer of the (L - 1)-th layer of the second decoder,

[0141] X2 f is the predicted mask at the scale corresponding to the f-th decoding layer of the second decoder, and Y2 f is the vascular structure mask patch at the scale corresponding to the f-th decoding layer of the second decoder; X2 fi is the i-th voxel point of the predicted mask at the scale corresponding to the f-th decoding layer of the second decoder, and Y2 fi is the i-th voxel point of the vascular structure mask patch at the scale corresponding to the f-th decoding layer of the second decoder,

[0142] L2 Sbzf is the task branch supervision loss at the scale corresponding to the decoding layer of the f-th layer of the second decoder, and Loss2 CEf is the cross-entropy loss at the scale corresponding to the f-th decoding layer of the second decoder, and Loss2 DSCf is the segmentation loss at the scale corresponding to the f-th decoding layer of the second decoder, and Loss2 IoUf is the object detection loss IOU (Jaccard) at the scale corresponding to the f-th decoding layer of the second decoder, where f is L or L - 1.

[0143] || and is to obtain the number of voxel points.

[0144] Step 6: Use the training dataset to co-train the multi-task segmentation network based on minimizing the multi-task multi-scale joint loss function;

[0145] Predict the vascular morphology mask patch as the target output of the first decoder, and predict the vascular structure mask patch as the target output of the second decoder. Using the training dataset, according to the multi-task multi-scale joint loss function in step 4, with the AdamW optimizer, the initial learning rate is 0.0001, and the number of training iterations is 40,000 times, co-train the multi-task segmentation network (including the shared encoder, the shared bottleneck layer, the shared skip module, the first decoder of the first task branch and the corresponding first multi-level target output module, the second decoder of the second task branch and the corresponding second multi-level target output module).

[0146] Step 7: For the multi-task segmentation network completed in co-training in step 6, fix the network parameters except those of the first task branch. Using the training dataset, based on minimizing the loss of the first task branch, train the network parameters of the first task branch; for the multi-task segmentation network after the network parameters of the first task branch are trained, fix the network parameters except those of the second task branch. Using the training dataset, based on minimizing the loss of the second task branch, train the network parameters of the second task branch.

[0147] After completing the co-training in step 6, perform de-local optimization on each branch using its respective loss. The branch optimization of the multi-task segmentation network is divided into two steps.

[0148] Step 7.1: Load the network parameters of the multi-task segmentation network obtained in the training of step 6, fix the network parameters of the multi-task segmentation network except for the first decoder and the first multi-level target output module (outside the first task branch), and train the multi-task segmentation network based on minimizing the loss of the first task branch in step 5. Optimize the network parameters of the first decoder and the first multi-level target output module. Here, the initial learning rate for branch optimization is 0.0001, and the number of training iterations is 1000 times.

[0149] Step 7.2: Load the network parameters of the multi-task segmentation network obtained in the training of step 7.1, fix the network parameters of the multi-task segmentation network except for the second decoder and the second multi-level target output module (outside the second task branch), and train the multi-task segmentation network based on minimizing the loss of the second task branch in step 5. Optimize the network parameters of the second decoder and the second multi-level target output module. Obtain the final multi-task segmentation network.

[0150] Step 8: Model inference

[0151] In the inference stage, the preprocessed and cropped non-enhanced CT image patches to be segmented of the non-enhanced CT image to be predicted are input into the trained multi-task segmentation network. Through the first decoder and the second decoder, the predicted vascular morphology mask patches and the predicted vascular structure mask patches are obtained respectively. The predicted vascular morphology mask patches and the predicted vascular structure mask patches are respectively subjected to reverse preprocessing and reconstruction to obtain the predicted vascular morphology mask blocks and vascular structure mask blocks of the same size as the original non-enhanced CT image. To solve the problem of poor connection at the edge of the reconstructed mask blocks, the original non-enhanced CT image is cropped according to an overlap rate of 0.25 during preprocessing and cropping. Correspondingly, during reconstruction, the mask values of the overlapping voxel points of the predicted vascular morphology mask patches and the predicted vascular structure mask patches are averaged, and the mask value with an average mask value greater than 0.5 is the corresponding effective mask value of the overlapping voxel points.

[0152] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0153] Embodiment 2:

[0154] In this embodiment, a multi-task non-enhanced CT vascular segmentation device with collaborative optimization is further provided, including:

[0155] Image acquisition module: used to implement step 1 of the above Embodiment 1 to acquire non-enhanced CT images and corresponding enhanced CT images;

[0156] Image preprocessing module: used to implement step 2 of the above Embodiment 1 to preprocess non-enhanced CT images and corresponding enhanced CT images, obtain registered non-enhanced CT image patches as a sample to generate a training data set, and obtain the vascular morphology mask patches and vascular structure mask patches as labels corresponding to the samples;

[0157] Network construction module: used to implement step 3 of the above Embodiment 1 to construct a multi-task segmentation network. The multi-task segmentation network includes an encoder, a bottleneck layer, a jump module, a first task branch, and a second task branch. The first task branch includes a first decoder and a first multi-level target output module, and the second task branch includes a second decoder and a second multi-level target output module;

[0158] Multi-task multi-scale joint loss function construction module: used to implement step 4 of the above Embodiment 1 to construct a multi-task multi-scale joint loss function;

[0159] Branch loss construction module: used to implement step 5 of the above-mentioned embodiment 1, and construct the first task branch loss and the second task branch loss;

[0160] Joint training module: used to implement step 6 of the above-mentioned embodiment 1, and use the training data set to jointly train the multi-task segmentation network based on minimizing the multi-task multi-scale joint loss function;

[0161] Branch training module: used to implement step 7 of the above-mentioned embodiment 1. For the multi-task segmentation network after joint training, fix the network parameters except for the first task branch, and use the training data set to train the network parameters of the first task branch based on minimizing the first task branch loss; for the multi-task segmentation network after the network parameters of the first task branch are trained, fix the network parameters except for the second task branch, and use the training data set to train the network parameters of the second task branch based on minimizing the second task branch loss.

[0162] Embodiment 3:

[0163] In this embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented.

[0164] Embodiment 4:

[0165] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0166] Embodiment 5:

[0167] In this embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0168] It should be emphasized that the scope of the present invention is not limited to the specific embodiments described above. These examples are only descriptions of the preferred embodiments of the present invention and are not intended to limit its core idea. In fact, the various solutions in these embodiments can be further integrated, substituted, or adjusted and improved according to the various adjustments and improvements made by those skilled in the art based on the technical idea of the present invention. All these reasonable changes and optimizations fall within the protection scope of the present invention. Therefore, any modifications and improvements based on the spirit and principles of the present invention, as long as they conform to the conventional technical means in the art, should be regarded as part of the present invention.

Claims

1. A collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device, characterized in that: Includes the following modules: Image acquisition module: acquires non-enhanced CT images and corresponding enhanced CT images; Image preprocessing module: preprocess the non-enhanced CT image and the corresponding enhanced CT image, obtain the registered non-enhanced CT image blocks as samples to generate training data sets, and obtain the vascular morphology mask blocks and vascular structure mask blocks as labels corresponding to the samples; Network construction module: construct a multi-task segmentation network, which includes an encoder, a bottleneck layer, a jump module, a first task branch and a second task branch. The first task branch includes a first decoder and a first multi-level target output module, and the second task branch includes a second decoder and a second multi-level target output module. Multi-task multi-scale joint loss function construction module: construct a multi-task multi-scale joint loss function; Branch loss construction module: construct the first task branch loss and the second task branch loss; Joint training module: Use the training dataset to jointly train the multi-task segmentation network based on minimizing the multi-task multi-scale joint loss function; Branch training module: for the multi-task segmentation network completed by joint training, fix the network parameters except the first task branch, use the training data set, and train the network parameters of the first task branch based on minimizing the loss of the first task branch; for the multi-task segmentation network after the network parameter training of the first task branch is completed, fix the network parameters except the second task branch, and use the training data set to train the network parameters of the second task branch based on minimizing the loss of the second task branch.

2. The collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device according to claim 1, characterized in that: The preprocessing of the non-enhanced CT image and the corresponding enhanced CT image comprises the following steps: Perform image resampling on all non-enhanced CT images and corresponding enhanced CT images; Registering the resampled enhanced CT image to the space of the non-enhanced CT image to obtain a registered enhanced CT image and a registered non-enhanced CT image; Based on the registered enhanced CT images, the vascular morphology and vascular structure are masked and outlined respectively to obtain the vascular morphology mask and the vascular structure mask; Normalize the registered non-enhanced CT images; The registered non-enhanced CT image and the corresponding registered enhanced CT image and registered non-enhanced CT image are cropped to the target size to obtain the registered non-enhanced CT image blocks and the corresponding actual target segmentation mask blocks. The registered non-enhanced CT image blocks are used as samples to generate a training data set, and the corresponding actual target segmentation mask blocks are used as corresponding labels.

3. The collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device according to claim 2, characterized in that: The cropping of the registered non-enhanced CT image and the corresponding registered enhanced CT image and the registered non-enhanced CT image comprises the following steps: According to the set ratio of voxels in the target segmentation area to voxels in the non-target segmentation area, the registered non-enhanced CT image and the corresponding vascular morphology mask and vascular structure mask are cropped in the same way. The cropping method is randomly selected to obtain the registered non-enhanced CT image blocks and the corresponding vascular morphology mask blocks and vascular structure mask blocks, respectively. The registered non-enhanced CT image blocks and the corresponding vascular morphology mask blocks and vascular structure mask blocks are expanded by random rotation, flipping, jittering and scaling for data perturbation.

4. The collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device according to claim 3, characterized in that: In each cropping method, the registered non-enhanced CT image and the corresponding vascular morphology mask and vascular structure mask are cropped according to the specified tile size and the specified tile overlap.

5. The collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device according to claim 1, characterized in that: The encoder comprises L encoding layers, each encoding layer comprises a downsampling convolution layer, a first convolution block, a Mamba block and a second convolution block, and the input features of the encoding layer are processed by the downsampling convolution layer, the first convolution block, the Mamba block and the second convolution block in sequence to obtain output features; The bottleneck layer includes r sequentially connected residual-like modules. The output features of the last encoding layer of the encoder are used as the input features of the bottleneck layer. The input features of the bottleneck layer are processed by each residual-like module in sequence to obtain the output features of the bottleneck layer. The first decoder and the second decoder both include L decoding layers: each decoding layer includes an upsampling convolution layer and a decoding convolution layer connected in sequence, a jump connection is established between the Mamba block of the lth encoding layer and the corresponding Llth upsampling convolution layer, the input features of the first decoding layer are the output features of the bottleneck layer, the output features of the decoding layer are input to the 3D transposed convolution output layer of the corresponding layer, each 3D transposed convolution output layer outputs a prediction mask of the corresponding level scale, the 3D transposed convolution output layers corresponding to each decoding layer of the first decoder constitute a first multi-level target output module, and the 3D transposed convolution output layers corresponding to each decoding layer of the second decoder constitute a second multi-level target output module.

6. The collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device according to claim 5, characterized in that: A jump connection is established between the Mamba block of the lth coding layer and the corresponding Llth upsampling convolutional layer. include: After the input features are input to the decoding layer, they are processed by the upsampling convolution layer to obtain upsampling features. The output features of the Mamba block of the encoding layer of the corresponding layer are output through the jump module to output jump features. The jump features are then added to the output features of the upsampling convolution layer of the decoding layer of the corresponding layer and input to the decoding convolution layer of the decoding layer of the corresponding layer. The output features of the decoding convolution layer are used as the output features of the decoding layer. The output features of the decoding layer are input to the 3D transposed convolution output layer of the corresponding layer and the decoding layer of the next layer. The input features of the upsampling convolution layer of the last decoder layer include the output features of the previous decoding layer and the jump features corresponding to the input features of the encoder. The input features of the encoder output the corresponding jump features after passing through the jump module. The upsampling convolution layer of the last decoding layer is the Unetr upsampling module. The output features of the Unetr upsampling module are input to the decoding convolution layer of the corresponding layer, and the output features of the decoding convolution layer are input to the 3D transposed convolution output layer of the corresponding layer.

7. The collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device according to claim 6, characterized in that: The multi-task multi-scale joint loss function is based on the following formula: L seg =0.6*L seg1 +0.4*L seg2 L1 Sal =Loss1 BSCl +Loss1 CEl <h2 style=";text-align:left;direction:ltr">L2<h2 style=";text-align:left;direction:ltr"> Sbl <h2 style=";text-align:left;direction:ltr"> =Loss2<h2 style=";text-align:left;direction:ltr"> DSCl <h2 style=";text-align:left;direction:ltr"> +Loss2<h2 style=";text-align:left;direction:ltr"> CEl Where: L seg is the multi-task multi-scale joint loss function, L seg1 and L seg2 are the first multi-scale loss and the second multi-scale loss, respectively. L1 Sal is the supervision loss at the scale corresponding to the lth decoding layer of the first decoder, Loss1 DSCl is the segmentation loss at the scale corresponding to the lth decoding layer of the first decoder, Loss1 CEl is the cross entropy loss at the scale corresponding to the lth decoding layer of the first decoder, X1 l is the prediction mask at the scale corresponding to the lth decoding layer of the first decoder, Y1 l is the blood vessel morphology mask block at the scale corresponding to the l-th decoding layer of the first decoder, X1 li is the i-th voxel point of the prediction mask at the scale corresponding to the l-th decoding layer of the first decoder, Y1 li is the i-th voxel point of the vascular morphology mask block at the scale corresponding to the l-th decoding layer of the first decoder, N is the total number of voxel points, || and are the number of voxel points. L2 Sbl is the supervision loss at the scale corresponding to the lth decoding layer of the second decoder, Loss2 DSCl is the segmentation loss at the scale corresponding to the lth decoding layer of the second decoder, Loss2 CEl is the cross entropy loss at the scale corresponding to the lth decoding layer of the second decoder, X2 l is the prediction mask at the scale corresponding to the lth decoding layer of the second decoder, Y2 l is the vascular structure mask block at the scale corresponding to the l-th decoding layer of the second decoder, X2 li is the i-th voxel point of the prediction mask at the scale corresponding to the l-th decoding layer of the second decoder, Y2 li is the i-th voxel point of the vascular structure mask block at the scale corresponding to the l-th decoding layer of the second decoder.

8. The collaboratively optimized multi-task non-enhanced CT blood vessel segmentation device according to claim 6, characterized in that: The loss of the first task branch is: L segz1 =L1 SazL +0.5*L1 Saz(L-1) L1 Sazf =Loss1 CEf +0.5*Loss1 DSCf +0.5*Loss1 IoUf The second task branch loss is: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> segz2 <h2 style=";text-align:left;direction:ltr"> =L2<h2 style=";text-align:left;direction:ltr"> SbzL <h2 style=";text-align:left;direction:ltr"> +0.5*L2<h2 style=";text-align:left;direction:ltr"> Sbz(L-1) <h2 style=";text-align:left;direction:ltr">L2<h2 style=";text-align:left;direction:ltr"> Sbzf <h2 style=";text-align:left;direction:ltr"> =Loss2<h2 style=";text-align:left;direction:ltr"> CEf <h2 style=";text-align:left;direction:ltr"> +0.5*Loss2<h2 style=";text-align:left;direction:ltr"> DSCf <h2 style=";text-align:left;direction:ltr"> +0.5*Loss2<h2 style=";text-align:left;direction:ltr"> IoUf Among them, L segz1 is the first task branch loss, L1 SazL is the task branch supervision loss at the scale corresponding to the Lth decoding layer of the first decoder, L1 Saz(L-1) is the task branch supervision loss at the scale corresponding to the decoding layer of the L-1th layer of the first decoder, L1 Sazf is the task branch supervision loss at the scale corresponding to the decoding layer of the fth layer of the first decoder, Loss1 CEf is the cross entropy loss at the scale corresponding to the f-th decoding layer of the first decoder, Loss1 DSCf is the segmentation loss at the scale corresponding to the f-th decoding layer of the first decoder, Loss1 IoUf is the target detection loss IOU at the scale corresponding to the lth decoding layer of the first decoder, f is L or L-1, || and are the number of voxel points, N is the total number of voxel points, X1 f is the prediction mask at the scale corresponding to the f-th decoding layer of the first decoder, Y1 f X1 is the blood vessel morphology mask block at the scale corresponding to the f-th decoding layer of the first decoder; fi is the i-th voxel point of the prediction mask at the scale corresponding to the f decoding layer of the first decoder, Y1 fi is the i-th voxel point of the vascular morphology mask block at the scale corresponding to the f-th decoding layer of the first decoder, L segz2 is the loss of the second task branch, L2 SbzL is the task branch supervision loss at the scale corresponding to the Lth decoding layer of the second decoder, L2 Sbz(L-1) is the task branch supervision loss at the scale corresponding to the decoding layer of the L-1th layer of the second decoder, L2 Sbzf is the task branch supervision loss at the scale corresponding to the decoding layer of the fth layer of the second decoder, Loss2 CEf is the cross entropy loss at the scale corresponding to the f-th decoding layer of the second decoder, Loss2 DSCf is the segmentation loss at the scale corresponding to the f-th decoding layer of the second decoder, Loss2 IoUf is the target detection loss IOU at the scale corresponding to the f-th decoding layer of the second decoder, X2 f is the prediction mask at the scale corresponding to the f-th decoding layer of the second decoder, Y2 f X2 is the blood vessel structure mask block at the scale corresponding to the f-th decoding layer of the second decoder; fi is the i-th voxel point of the prediction mask at the scale corresponding to the f decoding layer of the second decoder, Y2 fi is the i-th voxel point of the vascular structure mask block at the scale corresponding to the f-th decoding layer of the second decoder.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the functions of the modules of the segmentation device according to any one of claims 1 to 8 are realized.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the functions of each module of the segmentation device according to any one of claims 1 to 8 are realized.

Images