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

Abstract

The application discloses a kind of synergistically optimized multi-task non-enhanced CT blood vessel segmentation devices, including image acquisition module, image pre-processing module, network construction module, multi-task multi-scale joint loss function construction module, branch loss construction module, joint training module and branch training module.The application can greatly reduce the parameter quantity while promoting the synergistic optimization of double tasks by sharing the encoder, and by task branch parameter independent fine-tuning, the task branch parameters are optimized, and then the blood vessel morphology mask and blood vessel structure mask can be effectively segmented from the non-enhanced CT image.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing and artificial intelligence, and particularly relates to a multi-task non-enhanced CT blood vessel segmentation device optimized in cooperation. BACKGROUND

[0002] Due to the low distinction of non-enhanced CT (NCCT) for blood vessel tissue, there are significant challenges in capturing the subtle structural changes of blood vessels. Therefore, it is particularly important to develop an efficient segmentation model to enhance the distinction of blood vessel tissue in NCCT. In addition, in clinical practice, it is often necessary to observe the morphology and anatomical structure of the same blood vessel site to determine the cause and develop a treatment plan. However, most of the 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.] as representative medical image segmentation models focus on single tasks, which makes it necessary to develop multiple independent models in the multi-task blood vessel segmentation scenario. This approach not only leads to a multiplicative increase in model parameters with the number of tasks, but also ignores the potential synergistic relationship between different tasks, thereby limiting the further improvement of the performance of each segmentation model. SUMMARY

[0003] The purpose of the present application is to provide a multi-task non-enhanced CT blood vessel segmentation device optimized in cooperation to overcome the above-mentioned problems existing in the prior art. The synergistic optimization effect between multi-tasks is fully utilized, and the inherent limitations of NCCT are overcome. The present application takes Mamba-CNN hybrid network as the core and constructs a multi-task segmentation network, wherein the first task branch focuses on blood vessel morphology segmentation, and the second task branch is responsible for blood vessel structure segmentation. Through the design of a shared encoder, this method significantly reduces the parameter quantity of double tasks, and at the same time, with the help of the synergistic optimization mechanism between multi-tasks, the overall performance is greatly improved. In addition, the present application also introduces an independent parameter fine-tuning optimization strategy to further adjust the parameters of each task branch on the basis of synergistic optimization, ensuring that each task branch can achieve optimal performance. In summary, the present application aims to provide a more accurate, fast and safe multi-task blood vessel segmentation tool.

[0004] The above-mentioned purposes of the present application are achieved by the following technical means:

[0005] A multi-task non-enhanced CT blood vessel segmentation device optimized in cooperation, comprising the following modules:

[0006] an image acquisition module: acquiring a non-enhanced CT image and a corresponding enhanced CT image;

[0007] an image preprocessing module: preprocessing the non-enhanced CT image and the corresponding enhanced CT image, obtaining a registered non-enhanced CT image block as a training data set for generating a sample, and obtaining a blood vessel morphology mask block and a blood vessel structure mask block as corresponding labels of the sample;

[0008] a network construction module: constructing a multi-task segmentation network, the multi-task segmentation network comprising an encoder, a bottleneck layer, a skip module, a first task branch, and a second task branch, the first task branch comprising a first decoder and a first multi-level target output module, and the second task branch comprising a second decoder and a second multi-level target output module;

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

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

[0011] a joint training module: using a training data set to jointly train the multi-task segmentation network based on minimizing the multi-task multi-scale joint loss function;

[0012] a branch training module: fixing network parameters other than the first task branch for the multi-task segmentation network after joint training, using a training data set to train network parameters of the first task branch based on minimizing the first task branch loss, and fixing network parameters other than the second task branch for the multi-task segmentation network after training the network parameters of the first task branch, using a training data set to train network parameters of the second task branch based on minimizing the second task branch loss.

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

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

[0015] the resampled enhanced CT image is registered to the space of the non-enhanced CT image to obtain a registered enhanced CT image and a registered non-enhanced CT image;

[0016] the blood vessel morphology and the blood vessel structure are respectively masked based on the registered enhanced CT image to obtain a blood vessel morphology mask and a blood vessel structure mask;

[0017] the registered non-enhanced CT image is normalized;

[0018] The registered non-enhanced CT image and the corresponding registered enhanced CT image and registered non-enhanced CT image are cropped to a target size to obtain a registered non-enhanced CT image block and a corresponding actual target segmentation mask block, and the registered non-enhanced CT image block is used as a sample to generate a training data set, and the corresponding actual target segmentation mask block is used as a corresponding label.

[0019] The cropping of the registered non-enhanced CT image and the corresponding registered enhanced CT image and registered non-enhanced CT image as described above comprises the following steps:

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

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

[0022] The encoder comprises L encoding layers, each encoding layer comprising a down-sampling convolution layer, a first convolution block, a Mamba block and a second convolution block, and the input feature of the encoding layer is processed by the down-sampling convolution layer, the first convolution block, the Mamba block and the second convolution block in sequence to obtain an output feature;

[0023] The bottleneck layer comprises r residual-like modules connected in sequence, and the output feature of the last encoding layer of the encoder is used as the input feature of the bottleneck layer, and the input feature of the bottleneck layer is processed by each residual-like module in sequence to obtain an output feature of the bottleneck layer;

[0024] The first decoder and the second decoder each comprise L decoding layers: each decoding layer comprises an up-sampling 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 (L-l)th up-sampling convolution layer, the input feature of the first decoding layer is the output feature of the bottleneck layer, the output feature of the decoding layer is input to the 3D transposed convolution output layer of the corresponding layer, and each layer 3D transposed convolution output layer outputs a prediction mask of a corresponding hierarchical scale, the 3D transposed convolution output layer of each decoding layer of the first decoder constitutes a first multi-hierarchical target output module, and the 3D transposed convolution output layer of each decoding layer of the second decoder constitutes a second multi-hierarchical target output module.

[0025] The skip connection is established between the Mamba block of the first encoding layer and the corresponding L-lth upsampling convolutional layer, including:

[0026] After the input feature is input to the decoding layer, the input feature is processed by the upsampling convolutional layer to obtain an upsampling feature. The output feature of the Mamba block of the corresponding encoding layer of the layer is output after passing through the skip module. The skip feature is added to the output feature of the upsampling convolutional layer of the decoding layer of the corresponding layer, and then input to the decoding convolutional layer of the decoding layer of the corresponding layer. The output feature of the decoding convolutional layer is the output feature of the decoding layer. The output feature of the decoding layer is input to the 3D transpose convolution output layer of the corresponding layer and the decoding layer of the next layer.

[0027] The input feature of the upsampling convolutional layer of the last decoding layer includes the output feature of the decoding layer of the previous stage and the corresponding skip feature of the input feature of the encoder. The input feature of the encoder is output as the corresponding skip feature after passing through the skip module. The upsampling convolutional layer of the last decoding layer is a Unetr upsampling module. The output feature of the Unetr upsampling module is input to the decoding convolutional layer of the corresponding layer. The output feature of the decoding convolutional layer is input to the 3D transpose convolution output layer of the corresponding layer.

[0028] 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] Wherein: L seg is a 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 supervised loss of the lth decoding layer of the first decoder at the corresponding scale,

[0037] Loss1 DSCl is the segmentation loss at the scale corresponding to the lth decoding layer of the first decoder, CEl is the cross-entropy loss at the scale corresponding to the lth decoding layer of the first decoder,

[0038] X1 l is the predicted mask at the scale corresponding to the lth decoding layer of the first decoder, Y1 l is the vessel morphology mask patch at the scale corresponding to the lth 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 lth decoding layer of the first decoder, Y1 li is the i-th voxel point of the vessel morphology mask patch at the scale corresponding to the lth decoding layer of the first decoder, N is the total number of voxel points, and || and are the number of voxel points,

[0040] L2 Sbl is the supervision loss at the scale corresponding to the lth decoding layer of the second decoder,

[0041] 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,

[0042] X2 l is the predicted mask at the scale corresponding to the lth decoding layer of the second decoder, Y2 l is the vessel structure mask patch at the scale corresponding to the lth 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 lth decoding layer of the second decoder, Y2 li is the i-th voxel point of the vessel structure mask patch at the scale corresponding to the lth decoding layer of the second decoder.

[0044] The first task branch loss is as follows:

[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:

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

[0054] L1 Sazf is the task branch supervision loss at the scale corresponding to the fth decoding layer of the first decoder, Loss1 CEf is the cross-entropy loss at the scale corresponding to the fth decoding layer of the first decoder, Loss1 DSCf is the segmentation loss at the scale corresponding to the fth decoding layer of the first decoder, Loss1 IoUf is the target detection loss IOU at the scale corresponding to the fth 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,

[0055] X1 f is the predicted mask at the scale corresponding to the fth decoding layer of the first decoder, Y1 f is the blood vessel morphology mask patch at the scale corresponding to the fth decoding layer of the first decoder; X1 fi is the i-th voxel point of the predicted mask at the scale corresponding to the fth decoding layer of the first decoder, Y1 fi is the i-th voxel point of the blood vessel morphology mask patch at the scale corresponding to the fth 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 Lth decoding layer of the second decoder, L2 Sbz(L-1)a task branch supervision loss under a scale corresponding to a decoding layer of an L-1 layer of the second decoder,

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

[0058] X2 f a predicted mask under a scale corresponding to a decoding layer of an f decoding layer of the second decoder, Y2 f a blood vessel structure mask patch under a scale corresponding to a decoding layer of an f decoding layer of the second decoder; X2 fi an i-th voxel point of a predicted mask under a scale corresponding to a decoding layer of an f decoding layer of the second decoder, Y2 fi an i-th voxel point of a blood vessel structure mask patch under a scale corresponding to a decoding layer of an f decoding layer of the second decoder.

[0059] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor realizes the functions of each module of the above segmentation device when executing the computer program.

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

[0061] Compared with the prior art, the present application has the following beneficial effects:

[0062] 1. The present application designs a multi-task non-enhanced CT blood vessel segmentation device with a Mamba-CNN hybrid structure, which fully utilizes the long-distance modeling capability of Mamba and the ability of convolution to effectively capture local features, and realizes more accurate segmentation of blood vessel target tissue.

[0063] 2. The multi-task segmentation network is constructed, the task correlation of blood vessel shape segmentation and blood vessel structure segmentation is utilized, and shared encoders and branch task optimization learning are used to realize compression of model parameters while ensuring the effectiveness of multi-task segmentation.

[0064] 3. A multi-task multi-scale joint loss function is constructed, multi-task branches are cooperatively learned, and the shared encoders can learn the common knowledge of different tasks, thereby realizing performance improvement.

[0065] 4、In the embodiment, according to the characteristics of the shape and structure of the blood vessels, a series of preprocessing and model training steps are designed to ensure the efficiency and accuracy of the entire segmentation process. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 The structure schematic diagram of the multi-task segmentation network of the application.

[0067] Figure 2 The prediction mask result schematic diagram of each stage of the application. After the joint training of the multi-task segmentation network, relatively complete blood vessel shape masks and blood vessel structure masks are obtained, but the aneurysm part of the ascending aorta is not complete enough. After further branch training, the segmentation accuracy of the blood vessel shape mask and the blood vessel structure mask of the aneurysm part of the ascending aorta is further improved, and the result mask is highly consistent with the gold standard mask. DETAILED DESCRIPTION

[0068] In order to facilitate those skilled in the art to understand and implement the application, the application will be further described in detail below in conjunction with the embodiments. The embodiments described herein are only used to illustrate and explain the application, and are not a limitation on the application.

[0069] Embodiment 1:

[0070] As shown in Figure 1 A synergistically optimized multi-task non-enhanced CT blood vessel segmentation method is used to segment blood vessel shape masks and blood vessel structure masks from non-enhanced CT images. Taking the aorta of a non-enhanced CT image as an example, the method includes the following steps:

[0071] Step 1, obtaining non-enhanced CT images and corresponding enhanced CT images: preferably, data inclusion and exclusion rules are formulated before image acquisition, and the target tissue non-enhanced CT images and corresponding enhanced CT images are collected according to the formulated rules. The inclusion and exclusion rules can be one or more of the following standards:

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

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

[0074] (3) The non-enhanced CT images and the enhanced CT images correspond to each other without significant misplacement;

[0075] (4) The quality of the non-enhanced CT images and the enhanced CT images meets the conventional diagnostic requirements, and there is no significant artifact, distortion or deformation;

[0076] Step 2, preprocessing the non-enhanced CT images and the corresponding enhanced CT images:

[0077] Step 2.1, image resample all non-enhanced CT images and corresponding enhanced CT images to (1.25, 0.67, 0.67) resolution, that is, the preset axial resampling resolution is 1.25 mm, and the preset transverse resampling resolution is 0.67 mm;

[0078] Step 2.2, register 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, so as to ensure that the mask drawn from the enhanced CT image corresponds to the blood vessels of the non-enhanced CT. In this embodiment, the registration is performed by using the Python library PyElastix of the Elastix registration software, and the registration data without serious deformation is saved;

[0079] Step 2.3, based on the registered enhanced CT image, the blood vessel shape and the blood vessel structure are respectively masked to obtain an actual target segmentation mask, and the actual target segmentation mask includes a blood vessel shape mask and a blood vessel structure mask. The mask drawing can use the growth threshold method function module or the clipping function module or other mask making function modules of 3D slicer to draw the target on the registered enhanced CT image, and respectively obtain the actual target segmentation mask (blood vessel shape mask and blood vessel structure mask). The blood vessel shape mask is an aneurysm shape mask, or a blood vessel dissection mask, or a stenosis shape blood vessel mask, or a normal shape 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, the registered non-enhanced CT image is normalized for use in training the multi-task segmentation network. Here, only the registered non-enhanced CT image needs to be normalized, and the registered enhanced CT image and the corresponding drawn actual target segmentation mask do not need to be normalized.

[0081] Step 2.5, according to the specified rule, the registered non-enhanced CT image and the corresponding drawn actual target segmentation mask are clipped to a target size.

[0082] According to the set proportion of target segmentation region voxels-non-target segmentation region voxels, the registered non-enhanced CT image and the corresponding actual target segmentation mask (including the blood vessel shape mask and the blood vessel structure mask) are clipped in the same clipping mode, and the clipping mode is randomly selected. The registered non-enhanced CT image block and the corresponding actual target segmentation mask block are obtained. The registered non-enhanced CT image block and the corresponding actual target segmentation mask block are expanded by random rotation, flipping, jittering and scaling. The registered non-enhanced CT image block is used as a sample to generate a training data set, and the corresponding actual target segmentation mask block (including the blood vessel shape mask block and the blood vessel structure mask block) is used as a label corresponding to the registered non-enhanced CT image block.

[0083] In this embodiment, the registered non-contrast CT images and the corresponding actual target segmentation masks are cropped in the same cropping manner according to a ratio of 1:1 of the number of target segmentation region voxels to the number of non-target segmentation region voxels, and the cropping manner is randomly selected, to obtain non-contrast CT image blocks and actual target segmentation mask blocks with an image depth x image width x image height of 128x192x192, and the number of target segmentation region voxels to the number of non-target segmentation region voxels in the non-contrast CT image blocks and the corresponding target tissue masks is 1:1.

[0084] For each cropping manner, the block size and the block overlap degree are the same, and the registered non-contrast CT images and the corresponding actual target segmentation masks are cropped in the same cropping manner according to the specified block size and the specified block overlap degree, to obtain registered non-contrast CT image blocks and corresponding actual target segmentation mask blocks, respectively, the registered non-contrast CT image blocks are used as sample training data sets, and the corresponding actual target segmentation mask blocks are used as labels corresponding to the registered non-contrast CT image blocks.

[0085] In this embodiment, the block overlap degree is 0.25, and registered non-contrast CT image blocks and actual target segmentation mask blocks with an image depth x image width x image height of 128x192x192 are obtained.

[0086] The sample batch number of this example is batch = 1, the channel number is 1, the input is a registered non-contrast CT image block with a size of 1x1x128x192x192 (sample batch number x channel number x image depth x image width x image height), and the multi-level target output is one-hot encoding with sizes of 1x1x8x12x12, 1x1x16x24x24, 1x1x32x48x48, 1x1x64x96x96, and 1x1x128x192x192. During the training of the multi-task segmentation network, the actual target segmentation mask blocks (blood vessel morphology mask and blood vessel structure mask) of different levels are generated by the data online processing module of nnU-Net for multi-level size adjustment, and the sizes of the actual target segmentation mask blocks (blood vessel morphology mask and blood vessel structure mask) of each level are kept consistent to match the corresponding multi-level output mask. Specifically, the actual target segmentation mask blocks (blood vessel morphology mask and blood vessel structure mask) of level 4 are equivalent to 1 / 2 of the actual target segmentation mask blocks (blood vessel morphology mask and blood vessel structure mask) of level 3, 1 / 4 of the actual target segmentation mask blocks of level 2, and 1 / 8 of the actual target segmentation mask blocks of level 1. Among them, the actual target segmentation mask blocks of level 1 are actual target segmentation mask blocks without size adjustment.

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

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

[0089] Wherein, the encoder contains L (L>1) layers of encoding layers, each layer of encoding layer includes down-sampling convolution layer, first convolution block, Mamba block and second convolution block, the input feature of the encoding layer is processed by the down-sampling convolution layer, the first convolution block, the Mamba block and the second convolution block in turn to obtain the output feature, and the first convolution block, the Mamba block and the second convolution block of each layer of the encoding layer cooperatively complete the preliminary encoding and down-sampling of the input feature.

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

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

[0092] The first decoder and the second decoder have the same structure, both including L layers of decoding layers, each decoding layer of the first decoder corresponds to each encoding layer of the encoder, each decoding layer of the second decoder corresponds to each encoding layer of the encoder, and all are L (L>1) layers, each decoding layer includes sequentially connected up-sampling convolution layers and decoding convolution layers, which realize step-by-step decoding and up-sampling restoration of image features; in addition, a skip connection is established between the Mamba block of the lth (l<L) encoding layer and the corresponding L-lth up-sampling convolution layer, that is, for each decoding layer: after the input feature is input into the decoding layer, the up-sampling convolution layer is processed to obtain the up-sampling feature, the output feature of the Mamba block of the corresponding encoding layer is output after passing through the skip module, and then the skip feature is added to the output feature of the up-sampling convolution layer of the corresponding decoding layer and input into the decoding convolution layer of the corresponding decoding layer, the output feature of the decoding convolution layer is taken as the output feature of the decoding layer, and the output feature of the decoding layer is input into the 3D transpose convolution output layer of the corresponding layer and the decoding layer of the next layer.

[0093] The input feature of the first decoding layer is the output feature of the bottleneck layer.

[0094] The up-sampling convolution layer of the last decoding layer is an Unetr up-sampling module, the input feature of the up-sampling convolution layer of the last decoding layer includes the output feature of the previous decoding layer and the corresponding skip feature of the input feature of the encoder, the input feature of the encoder is output after passing through the skip module, the output feature of the Unetr up-sampling module is input into the decoding convolution layer of the corresponding layer, and the output feature of the decoding convolution layer is input into the 3D transpose convolution output layer of the corresponding layer.

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

[0096] The first decoder and the second decoder correspond to the first multi-level target output module and the second multi-level target output module respectively, and the first multi-level target output module and the second multi-level target output module each include an l-layer 3D transpose convolution output layer, and each layer of the 3D transpose convolution output layer outputs a prediction mask of a corresponding level scale to provide all-around supervision and optimization for the segmentation result.

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

[0098] The encoder:

[0099] In the embodiment, the encoder includes four encoding layers, and the input feature of the first encoding layer is a single-channel registered non-enhanced CT image block. Specifically, the channel change of the entire encoder is 1->48->96->192->384, and the input feature of each encoding layer and the output feature of the last encoding layer have sizes of 1x1x128x192x192, 1x48x64x96x96, 1x96x32x48x48, 1x192x16x24x24, and 1x384x8x12x12, respectively. For each encoding layer, the down-sampling convolution layer includes a 3D convolution layer with a convolution kernel of 1 and a convolution step of 2. The first convolution block includes two 3D convolution blocks with a convolution kernel of 3 and a step of 1, and after the 3D convolution blocks, 3D InstanceNorm normalization and ReLu nonlinear conversion are performed. The state space expansion factor of the Mamba block is 16, the block expansion factor is 2, the local convolution width is 4, and the normalization parameter is 3D InstanceNorm. The second convolution block includes two 3D convolution blocks with a convolution kernel of 1 and a convolution step of 1.

[0100] The registered non-enhanced CT image block is input to the encoder, and after the first convolution down-sampling, the channel expansion is performed to increase the feature expression capability. Then, after the convolution feature extraction, the Mamba sequence modeling, and the convolution feature induction, the features of different levels and different scales are extracted. The output feature of the last encoding layer is input to the bottleneck layer. Then, the output feature of the bottleneck layer is gradually restored to the pixel space through the up-sampling convolution layer (transpose convolution) and the decoding convolution layer of the decoder.

[0101] The bottleneck layer:

[0102] The bottleneck layer includes a plurality of residual-like modules (MedNeXt convolution blocks), and the number of residual-like modules (MedNeXt convolution blocks) used in the bottleneck layer of the present example is 8. The channel expansion rate of the bottleneck layer is 8, the inter-layer normalization strategy is a standardization group, the convolution kernel is 3, the convolution step 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 to the up-sampling convolution layer of the first decoding layer of the first decoder and the second decoder.

[0103] The skip connection module includes a convolutional layer and a Unetr basic block, wherein the convolutional kernel of the convolutional layer is 1, the convolutional step is 1, the input channel is 1, and the output channel is 48. The convolutional kernel of the Unetr basic block is 3, the convolutional step is 1, and the input and output channels are 48. In addition to the skip connection between the first (l (l < L)) down-sampling convolutional layer and the corresponding L-l up-sampling convolutional layer, the input feature of the encoder is also connected to the last up-sampling convolutional layer.

[0104] The decoder (first decoder and second decoder):

[0105] The decoder includes 4 decoding layers, each decoding layer includes a decoding convolutional layer and a corresponding up-sampling convolutional layer, and there is no channel change between the decoding convolutional layer and the up-sampling convolutional layer in 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 a MedNeXt convolutional block, and the number of MedNeXt convolutional blocks of the decoding convolutional layer is 8 in the first decoding layer, 8 in the second decoding layer, 4 in the third decoding layer, and 3 in the fourth decoding layer. The channel expansion rate of the decoding convolutional layer is also set to 8, 8, 4 and 3, and other parameter settings are consistent with the bottleneck layer. The last up-sampling convolutional layer of the decoder uses a Unetr up-sampling module, and other up-sampling convolutional layers have the same parameter settings as the down-sampling convolutional layers of the encoder, except that the down-sampling convolutional layers in the encoder are replaced by up-sampling convolutional layers of transposed convolution.

[0106] The decoder includes 4 decoding layers, each decoding layer includes a decoding convolutional layer and a corresponding up-sampling convolutional layer, and there is no channel change between the decoding convolutional layer and the up-sampling convolutional layer in 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 a MedNeXt convolutional block, and the number of MedNeXt convolutional blocks of the decoding convolutional layer is 8 in the first decoding layer, 8 in the second decoding layer, 4 in the third decoding layer, and 3 in the fourth decoding layer. The channel expansion rate of the decoding convolutional layer is also set to 8, 8, 4 and 3, and other parameter settings are consistent with the bottleneck layer. The last up-sampling convolutional layer of the decoder uses a Unetr up-sampling module, and other up-sampling convolutional layers have the same parameter settings as the down-sampling convolutional layers of the encoder, except that the down-sampling convolutional layers in the encoder are replaced by up-sampling convolutional layers of transposed convolution.

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

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

[0109] The first multi-level target output module and the second multi-level target output module in this embodiment each include L 3D transpose convolution output layers (4 in this embodiment). The convolution kernel size and the step length of each 3D transpose convolution output layer are both 1, and no activation layer is included. The final number of channels of the output features of each of the 3D transpose convolution output layers (in this embodiment, the 3D transpose convolution output layers Sa1 to Sa4 of the first level and the 3D transpose convolution output layers Sb1 to Sb4 of the first level) is 1. The output features of the decoding convolution layers of each level of the decoder generate prediction masks of different scales through the corresponding 3D transpose convolution output layers. Specifically, for the first decoder and the second decoder: the side length (width and height) of the output features of the decoding convolution layer of the fourth decoding layer is twice the side length of the output features of the decoding layer of the third level; the side length of the output features of the decoding layer of the third level is twice the side length of the output features of the decoding layer of the second level; the side length of the output features of the decoding layer of the second level is twice the side length 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 from multiple scales of the registered non-enhanced CT image blocks, thereby achieving efficient target segmentation.

[0110] Step 4, constructing a multi-task multi-scale joint loss function for collaborative training of the multi-task segmentation network:

[0111] The training loss function used in the collaborative training of 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] wherein, L seg is a multi-task multi-scale joint loss function, L seg1 and L seg2 are a first multi-scale loss and a second multi-scale loss respectively, and both loss calculation methods are exactly the same.

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

[0115]

[0116] wherein, L1 Sal is a supervised loss at a scale corresponding to the lth 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 transpose convolution output layers of the multi-level target output module.

[0117] L1 Sal = Loss1 DSCl + Loss1 CEl

[0118]

[0119] wherein, L1 Sal is a supervised loss at a scale corresponding to the lth decoding layer of the first decoder, X1 l is a predicted mask at a scale corresponding to the lth decoding layer of the first decoder, Y1 l is a blood vessel morphology mask patch at a scale corresponding to the lth decoding layer of the first decoder, Y1 l and X1 l are the same size at the same scale, X1 li is the i th voxel point of the predicted mask at a scale corresponding to the lth decoding layer of the first decoder, Y1 li is the i th voxel point of the blood vessel morphology mask patch at a scale corresponding to the lth decoding layer of the first decoder, N is the total number of voxel points, || and are 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 intersection voxel points of X1 l and Y1 l , Loss1 DSCl is a segmentation loss at a scale corresponding to the lth decoding layer of the first decoder, Loss1CEl the cross-entropy loss at the scale corresponding to the lth decoding layer of the first decoder;

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

[0121]

[0122] wherein L2 Sbl is the supervised loss at the scale corresponding to the lth decoding layer of the second decoder, L is the total number of decoding layers of the second decoder.

[0123] L2 Sbl = Loss2 DSCl + Loss2 CEl

[0124]

[0125] wherein L2 Sbl is the supervised loss at the scale corresponding to the lth decoding layer of the second decoder, X2 l is the predicted mask at the scale corresponding to the lth decoding layer of the second decoder, Y2 l is the vessel structure mask patch at the scale corresponding to the lth 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 predicted mask at the scale corresponding to the lth decoding layer of the second decoder, Y2 li is the i-th voxel point of the vessel structure mask patch at the scale corresponding to the lth decoding layer of the second decoder, N is the total number of voxel points, || and || are the number of voxel points, |X2 l | and |Y2 l | are the number of voxel points of X2 l and Y2 l , respectively, |X2 l ∩ Y2 l | is the number of intersection voxel points of X2 l and Y2 l , 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;

[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 loss L segz1 of the first task branch 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 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 (L-1)th decoding layer of the first decoder,

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

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

[0135] || and are 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] wherein, L segz2 is the second task branch loss, 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 (L-1)th decoding layer of the second decoder,

[0141] X2 f is the predicted mask at the scale corresponding to the fth decoding layer of the second decoder, Y2 f is the blood vessel structure mask patch at the scale corresponding to the fth decoding layer of the second decoder; X2 fi is the ith voxel point of the predicted mask at the scale corresponding to the fth decoding layer of the second decoder, Y2 fi is the ith voxel point of the blood vessel structure mask patch at the scale corresponding to the fth decoding layer of the second decoder,

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

[0143] || and are the number of voxel points.

[0144] Step 6, based on the minimization of the multi-task multi-scale joint loss function, the multi-task segmentation network is collaboratively trained using the training data set.

[0145] The predicted blood vessel shape mask patches are taken as the target output of the first decoder, and the predicted blood vessel structure mask patches are taken as the target output of the second decoder. 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) is collaboratively trained by using a training data set, according to the minimization of the multi-task multi-scale joint loss function in step 4, and by using an AdamW optimizer with an initial learning rate of 0.0001 and a training iteration number of 40,000.

[0146] Step 7, for the multi-task segmentation network collaboratively trained in step 6, the network parameters except the first task branch are fixed, the training data set is used, and the network parameters of the first task branch are trained based on the minimization of the first task branch loss; for the multi-task segmentation network after the network parameters of the first task branch are trained, the network parameters except the second task branch are fixed, the training data set is used, and the network parameters of the second task branch are trained based on the minimization of the second task branch loss.

[0147] After the collaborative training in step 6 is completed, each branch is de-localized optimized by using the respective loss. The branch optimization of the multi-task segmentation network is divided into two steps.

[0148] Step 7.1, the network parameters of the multi-task segmentation network obtained by training in step 6 are loaded, the network parameters of the multi-task segmentation network except the first decoder and the first multi-level target output module (the first task branch) are fixed, and the multi-task segmentation network is trained based on the minimization of the first task branch loss in step 5. The network parameters of the first decoder and the first multi-level target output module are optimized. The initial learning rate of the branch optimization here is 0.0001, and the training iteration number is 1,000.

[0149] Step 7.2, the network parameters of the multi-task segmentation network obtained by training in step 7.1 are loaded, the network parameters of the multi-task segmentation network except the second decoder and the second multi-level target output module (the second task branch) are fixed, and the multi-task segmentation network is trained based on the minimization of the second task branch loss in step 5. The network parameters of the second decoder and the second multi-level target output module are optimized. The final multi-task segmentation network is obtained.

[0150] Step 8, model inference

[0151] In the inference stage, the non-enhanced CT image to be predicted is preprocessed and cropped to obtain a to-be-segmented non-enhanced CT image block, which is input into the trained multi-task segmentation network to obtain a predicted blood vessel morphology mask block and a predicted blood vessel structure mask block through the first decoder and the second decoder, respectively. The predicted blood vessel morphology mask block and the predicted blood vessel structure mask block are subjected to reverse preprocessing and reconstruction to obtain a predicted blood vessel morphology mask block and a blood vessel structure mask block of the same size as the original non-enhanced CT image. In order to solve the poor connection of the edge part of the reconstructed mask block, the original non-enhanced CT image is cropped according to an overlap rate of 0.25 when preprocessing and cropping. Correspondingly, the mask values of the overlapping voxel points of the predicted blood vessel morphology mask block and the predicted blood vessel structure mask block are averaged, and the mask value greater than 0.5 is the corresponding effective mask value of the overlapping voxel points.

[0152] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments.

[0153] Embodiment 2:

[0154] In this embodiment, a synergistically optimized multi-task non-enhanced CT blood vessel segmentation device is also provided, which comprises:

[0155] An image acquisition module is configured to implement step 1 of embodiment 1, and acquire a non-enhanced CT image and a corresponding enhanced CT image.

[0156] An image preprocessing module is configured to implement step 2 of embodiment 1, and preprocess the non-enhanced CT image and the corresponding enhanced CT image to obtain a registered non-enhanced CT image block as a sample to generate a training data set, and obtain a blood vessel morphology mask block and a blood vessel structure mask block as a corresponding label of the sample.

[0157] A network construction module is configured to implement step 3 of embodiment 1, and construct a multi-task segmentation network. The multi-task segmentation network comprises an encoder, a bottleneck layer, a skip module, a first task branch, and a second task branch. The first task branch comprises a first decoder and a first multi-level target output module, and the second task branch comprises a second decoder and a second multi-level target output module.

[0158] A multi-task multi-scale joint loss function construction module is configured to implement step 4 of embodiment 1, and construct a multi-task multi-scale joint loss function.

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

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

[0161] Branch training module: used to implement step 7 of the above embodiment 1, to train the network parameters of the first task branch based on minimizing the first task branch loss by using the training data set, while fixing the network parameters of the multi-task segmentation network except the first task branch; to train the network parameters of the second task branch based on minimizing the second task branch loss by using the training data set, while fixing the network parameters of the multi-task segmentation network except the second task branch.

[0162] Embodiment 3:

[0163] In this embodiment, a computer device is also provided, including a memory and a processor, the memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0164] Embodiment 4:

[0165] In this embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program implements the steps in the above method embodiments when executed by a processor.

[0166] Embodiment 5:

[0167] In this embodiment, a computer program product is provided, which includes a computer program, and the computer program implements the steps in the above method embodiments when executed by a processor.

[0168] It should be emphasized that the scope of the present application is not limited to the specific embodiments described above. These examples are only a description of the preferred embodiments of the present application and are not intended to limit the core idea. In fact, the schemes in these embodiments can be further integrated, replaced, or adjusted and improved according to the technical ideas of the present application by those skilled in the art. All these reasonable changes and optimizations are within the scope of protection of the present application. Therefore, any modification and improvement based on the spirit and principles of the present application, as long as it meets the conventional technical means in the art, should be considered as part of the present application.

Claims

1. A synergistically optimized multi-task non-contrast enhanced CT vessel segmentation device, characterized in that, The method comprises the following modules: An image acquisition module: acquiring non-enhanced CT images and corresponding enhanced CT images; An image preprocessing module: preprocessing the non-enhanced CT images and the corresponding enhanced CT images, obtaining registered non-enhanced CT image blocks as training data sets, and obtaining blood vessel morphology mask blocks and blood vessel structure mask blocks as corresponding labels; A network construction module: constructing a multi-task segmentation network, the multi-task segmentation network comprising an encoder, a bottleneck layer, a skip module, a first task branch, and a second task branch, the first task branch comprising a first decoder and a first multi-level target output module, and the second task branch comprising a second decoder and a second multi-level target output module; A multi-task multi-scale joint loss function construction module: constructing a multi-task multi-scale joint loss function; A branch loss construction module: constructing a first task branch loss and a second task branch loss; A joint training module: using a training data set to jointly train the multi-task segmentation network based on minimizing the multi-task multi-scale joint loss function; A branch training module: fixing network parameters other than the first task branch of the multi-task segmentation network after joint training, using a training data set to train network parameters of the first task branch based on minimizing the first task branch loss, and fixing network parameters other than the second task branch of the multi-task segmentation network after the network parameters of the first task branch are trained, using a training data set to train network parameters of the second task branch based on minimizing the second task branch loss.

2. The synergistically optimized multi-task non-contrast CT vessel segmentation device according to claim 1, wherein, The preprocessing of the non-enhanced CT images and the corresponding enhanced CT images comprises the following steps: Resampling all non-enhanced CT images and corresponding enhanced CT images; Registering 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; Masking the blood vessel morphology and the blood vessel structure based on the registered enhanced CT images to obtain a blood vessel morphology mask and a blood vessel structure mask; Normalizing the registered non-enhanced CT images; Cutting the registered non-enhanced CT images and the corresponding registered enhanced CT images and registered non-enhanced CT images to a target size to obtain registered non-enhanced CT image blocks and corresponding actual target segmentation mask blocks, the registered non-enhanced CT image blocks being used as training data sets, and the corresponding actual target segmentation mask blocks being used as corresponding labels.

3. The synergistically optimized multi-task non-contrast CT vessel segmentation device of claim 2, wherein, The cutting of the registered non-enhanced CT images and the corresponding registered enhanced CT images and registered non-enhanced CT images comprises the following steps: According to a set ratio of target segmentation region voxels to non-target segmentation region voxels, cutting the registered non-enhanced CT images and the corresponding blood vessel morphology mask and blood vessel structure mask in the same cutting mode, the cutting mode being randomly selected, to obtain registered non-enhanced CT image blocks and corresponding blood vessel morphology mask blocks and blood vessel structure mask blocks, and performing data perturbation expansion on the registered non-enhanced CT image blocks and the corresponding blood vessel morphology mask blocks and blood vessel structure mask blocks by random rotation, flipping, jittering, and scaling.

4. The synergistically optimized multi-task non-contrast CT vessel segmentation device according to claim 3, wherein, In each cropping manner, the registered non-enhanced CT image and the corresponding vessel morphology mask and vessel structure mask are cropped according to a specified tile size and a specified tile overlap.

5. The synergistically optimized multi-task non-contrast CT vessel segmentation device according to claim 1, wherein, The encoder comprises L encoding layers, each encoding layer comprising a down-sampling convolutional layer, a first convolutional block, a Mamba block and a second convolutional block, and the input feature of the encoding layer is sequentially processed by the down-sampling convolutional layer, the first convolutional block, the Mamba block and the second convolutional block to obtain the output feature; The bottleneck layer comprises r residual-like modules connected in sequence, and the output feature of the last encoding layer of the encoder is taken as the input feature of the bottleneck layer, and the input feature of the bottleneck layer is sequentially processed by each residual-like module to obtain the output feature of the bottleneck layer; The first decoder and the second decoder each comprise L decoding layers: each decoding layer comprises an up-sampling convolutional layer and a decoding convolutional layer connected in sequence, a skip connection is established between the Mamba block of the lth encoding layer and the corresponding (L-l)th up-sampling convolutional layer, the input feature of the first decoding layer is the output feature of the bottleneck layer, the output feature of the decoding layer is input to the 3D transposed convolutional output layer of the corresponding layer, each 3D transposed convolutional output layer outputs a prediction mask at a corresponding hierarchical scale, the 3D transposed convolutional output layers of each decoding layer of the first decoder constitute a first multi-level target output module, and the 3D transposed convolutional output layers of each decoding layer of the second decoder constitute a second multi-level target output module.

6. The synergistically optimized multi-task non-enhanced CT vessel segmentation device according to claim 5, wherein, A skip connection is established between the Mamba block of the lth encoding layer and the corresponding (L-l)th up-sampling convolutional layer Comprise: After the input feature is input to the decoding layer, the up-sampling convolutional layer processes the input feature to obtain an up-sampled feature, the output feature of the Mamba block of the corresponding encoding layer is output as a skip feature after being processed by a skip module, the skip feature is added to the output feature of the up-sampling convolutional layer of the decoding layer of the corresponding layer and then input to the decoding convolutional layer of the decoding layer of the corresponding layer, the output feature of the decoding convolutional layer is taken as the output feature of the decoding layer, and the output feature of the decoding layer is input to the 3D transposed convolutional output layer of the corresponding layer and the decoding layer of the next layer, The input feature of the up-sampling convolutional layer of the last decoding layer comprises the output feature of the previous decoding layer and the skip feature corresponding to the input feature of the encoder, the input feature of the encoder is output as the corresponding skip feature after being processed by a skip module, and the up-sampling convolutional layer of the last decoding layer is an Unetr up-sampling module, the output feature of the Unetr up-sampling module is input to the decoding convolutional layer of the corresponding layer, and the output feature of the decoding convolutional layer is input to the 3D transposed convolutional output layer of the corresponding layer.

7. The synergistically optimized multi-task non-contrast CT vessel segmentation device of claim 6, wherein, 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 L2 Sbl = Loss2 DSCl + Loss2 CEl wherein: L seg is a multi-task multi-scale joint loss function, L seg1 and L seg2 are a first multi-scale loss and a second multi-scale loss, respectively, L1 Sal L1is the supervised loss at the scale corresponding to the lth decoding layer of the first decoder, Loss1 DSCl Loss1 CEl Loss1 X1 l is a prediction mask at the scale corresponding to the lth decoding layer of the first decoder, Y1 l is a vessel morphology mask tile at the scale corresponding to the lth decoding layer of the first decoder, Xi li Yi is the ith voxel point of the prediction mask at the scale corresponding to the lth decoding layer of the first decoder, li Yi is the ith voxel point of the vessel shape mask tile at the scale corresponding to the lth decoding layer of the first decoder, N is the total number of voxel points, and || and || are the number of voxel points. L2 Sbl the supervised loss at the scale corresponding to the l-th decoding layer of the second decoder, Loss2 DSCl Loss2 CEl Loss2 X2 l Y2 is a prediction mask at the scale corresponding to the lth decoding layer of the second decoder, l Y2 is a vessel structure mask tile at the scale corresponding to the lth decoding layer of the second decoder, X2 li Y2 is the i-th voxel point of the prediction mask at the scale corresponding to the l-th decoding layer of the second decoder li Y2 is the i-th voxel point of the prediction mask at the scale corresponding to the l-th decoding layer of the second decoder 8. The synergistically optimized multi-task non-contrast CT vessel segmentation device of claim 6, wherein, The first task branch loss 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: L segz2 = L2 SbzL + 0.5 * L2 Sbz(L-1) L2 Sbzf = Loss2 CEf + 0.5*Loss2 DSCf + 0.5*Loss2 IoUf wherein 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 (L-1)th decoding layer of the first decoder, L1 Sazf Loss1is the task branch supervision loss of the corresponding scale of the fth decoding layer of the first decoder, CEf Loss1is the cross-entropy loss of the corresponding scale of the fth decoding layer of the first decoder, DSCf Loss1is the segmentation loss of the corresponding scale of the fth decoding layer of the first decoder, IoUf Loss1is the target detection loss IOU of the corresponding scale of the fth decoding layer of the first decoder, f is L or L-1, || and are the number of voxel points, and N is the total number of voxel points. X1 f is a prediction mask at a scale corresponding to the fth decoding layer of the first decoder, Y1 f is a vessel morphology mask tile at a scale corresponding to the fth decoding layer of the first decoder; X1 fi is an i-th voxel of the prediction mask at a scale corresponding to the fth decoding layer of the first decoder, Y1 fi is an i-th voxel of the vessel morphology mask tile at a scale corresponding to the fth decoding layer of the first decoder, L segz2 is a second task branch loss, L2 SbzL is a task branch supervision loss at a scale corresponding to a decoding layer of an Lth layer of the second decoder, L2 Sbz(L-1) is a task branch supervision loss at a scale corresponding to a decoding layer of an L-1th layer of the second decoder, L2 Sbzf the task branch supervision loss of the corresponding scale of the f-th decoding layer of the second decoder, Loss2 CEf the cross-entropy loss of the corresponding scale of the f-th decoding layer of the second decoder, Loss2 DSCf the segmentation loss of the corresponding scale of the f-th decoding layer of the second decoder, Loss2 IoUf the target detection loss IOU of the corresponding scale of the f-th decoding layer of the second decoder, X2 f Y2is a prediction mask at a scale corresponding to the fth decoding layer of the second decoder, f X2is a vessel structure mask tile at a scale corresponding to the fth decoding layer of the second decoder; fi Y2is an i-th voxel of the prediction mask at a scale corresponding to the fth decoding layer of the second decoder, fi X2is an i-th voxel of the vessel structure mask tile at a scale corresponding to the fth decoding layer of the second decoder. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor implements the functions of each module of the segmentation device of any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the functions of each module of the segmentation device of any one of claims 1 to 8 when executed by the processor.

Images