Multi-stage segmentation method, device and storage medium for narrow cavity of human body

Abstract

Embodiments of the present application provide a multi-stage segmentation method, device and storage medium for a narrow cavity of a human body, the method comprising: acquiring preprocessed computed tomography (CT) data; inputting the preprocessed CT data into a multi-stage segmentation model to obtain a segmentation result output by the multi-stage segmentation model; and the multi-stage segmentation model comprising a first stage, a second stage and a third stage. The multi-stage segmentation method, device and storage medium for the narrow cavity of the human body provided by the embodiments of the present application can improve the cavity accuracy, sensitivity and branch detection rate by segmenting the cavity of the human body through the multi-stage segmentation model, and each stage has a respective training task and focus according to a corresponding loss function, thereby obtaining the segmentation result of the narrow cavity of the human body.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a multi-stage segmentation method, apparatus and storage medium for narrow cavities in the human body. Background Technology

[0002] Ideal cavity segmentation needs to satisfy robustness, accuracy, and integrity, while minimizing computational and storage requirements; however, existing methods have shortcomings in various aspects.

[0003] Traditional methods use manually designed features and rules for image segmentation, which makes them computationally simple and efficient, while requiring relatively little data. However, due to the lack of semantic features, they are prone to marking a large number of leaky areas when there are problems such as noise interference and blurred boundaries in computed tomography (CT), resulting in low accuracy and completeness.

[0004] Existing deep learning methods can reduce leakage and increase the length of detected cavities, but they are more prone to fragmentation than traditional methods. Traditional methods typically rely on strength constraints, thus ensuring connectivity; however, in deep learning methods, taking tracheal segmentation as an example, inter-class imbalance (trachea vs. background) and intra-class imbalance (large trachea vs. small trachea) in lung CT images can cause gradient submersion in some small tracheal regions during backpropagation, resulting in fragmentation in the prediction results. Furthermore, due to the complex structure of some cavities in the human body, maintaining topological integrity while ensuring segmentation accuracy is a challenging task. Summary of the Invention

[0005] To address the aforementioned technical problems, embodiments of this application provide a multi-stage segmentation method, apparatus, and storage medium for narrow cavities in the human body.

[0006] In a first aspect, embodiments of this application provide a multi-stage segmentation method for narrow cavities in the human body, including:

[0007] Acquire preprocessed computed tomography (CT) data;

[0008] The preprocessed CT data is input into a multi-stage segmentation model to obtain the segmentation results output by the multi-stage segmentation model.

[0009] The multi-stage segmentation model includes a first stage, a second stage, and a third stage;

[0010] The first stage is used to obtain preliminary results based on a first loss function, which is used to perform semantic segmentation on the CT data;

[0011] The second stage is used to obtain improved results based on the first loss function, the second loss function, and the preliminary results. The second loss function is used to solve the intra-class imbalance problem in the preliminary results.

[0012] The third stage is used to obtain the segmentation result based on the first loss function, the second loss function, the third loss function and the improved result, and the third loss function is used to repair the cavity fracture in the improved result.

[0013] In some embodiments, the second loss function is determined based on a first static weight and a dynamic weight, wherein the first static weight is calculated based on local imbalance of the cavity before the second stage, and the dynamic weight is calculated based on the predicted performance of voxels in the second stage.

[0014] In some embodiments, the third loss function is determined based on a second static weight and a dynamic weight, wherein the second static weight is calculated based on a first static weight, a weight of undetected cavity voxels in the second stage based on the distance from the central skeleton line, and a weight of cavity voxels at the fracture point in the second stage based on the cavity size.

[0015] In some embodiments, the multi-stage segmentation model is determined based on a multi-scale residual feature extraction module, which is used to coordinate image features learned from cavity regions of different scales.

[0016] In some embodiments, acquiring the preprocessed computed tomography (CT) data includes:

[0017] Acquire CT data of the area where the cavity is located;

[0018] The CT values ​​are uniformly truncated into a single channel, and the preprocessed CT data is obtained based on normalization.

[0019] In some embodiments, obtaining the segmentation result output by the multi-stage segmentation model includes:

[0020] The segmentation result is determined from the feature map predicted by the network based on a dual-threshold iterative algorithm.

[0021] Secondly, embodiments of this application also provide a multi-stage segmentation device for narrow cavities in the human body, comprising:

[0022] The first acquisition module is used to acquire preprocessed computed tomography (CT) data.

[0023] The second acquisition module is used to input the preprocessed CT data into the multi-stage segmentation model and acquire the segmentation result output by the multi-stage segmentation model.

[0024] The multi-stage segmentation model includes a first stage, a second stage, and a third stage;

[0025] The first stage is used to obtain preliminary results based on a first loss function, which is used to perform semantic segmentation on the CT data;

[0026] The second stage is used to obtain improved results based on the first loss function, the second loss function, and the preliminary results. The second loss function is used to solve the intra-class imbalance problem in the preliminary results.

[0027] The third stage is used to obtain the segmentation result based on the first loss function, the second loss function, the third loss function and the improved result, and the third loss function is used to repair the cavity fracture in the improved result.

[0028] In some embodiments, the second loss function is determined based on a first static weight and a dynamic weight, wherein the first static weight is calculated based on local imbalance of the cavity before the second stage, and the dynamic weight is calculated based on the predicted performance of voxels in the second stage.

[0029] In some embodiments, the third loss function is determined based on a second static weight and a dynamic weight, wherein the second static weight is calculated based on a first static weight, a weight of undetected cavity voxels in the second stage based on the distance from the central skeleton line, and a weight of cavity voxels at the fracture point in the second stage based on the cavity size.

[0030] In some embodiments, the multi-stage segmentation model is determined based on a multi-scale residual feature extraction module, which is used to coordinate image features learned from cavity regions of different scales.

[0031] In some embodiments, the first acquisition module includes:

[0032] The first acquisition submodule is used to acquire CT data of the area where the cavity is located;

[0033] The second acquisition submodule is used to uniformly truncate CT values ​​into a unified channel and acquire the preprocessed CT data based on normalization.

[0034] In some embodiments, the second acquisition module includes:

[0035] The first determining submodule is used to determine the segmentation result from the feature map predicted by the network based on a dual-threshold iterative algorithm.

[0036] Thirdly, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-stage segmentation method for narrow human cavities as described above.

[0037] Fourthly, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-stage segmentation method for human stenotic cavities as described above.

[0038] Fifthly, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the multi-stage segmentation method for narrow cavities of the human body as described above.

[0039] The multi-stage segmentation method, apparatus, and storage medium for narrow human cavities provided in this application segment human cavities through a multi-stage segmentation model. Each stage has its own training tasks and priorities based on the corresponding loss function, thereby obtaining the segmentation results of narrow human cavities and improving the accuracy, sensitivity, and branch detection rate of cavities. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating the multi-stage segmentation method for narrow cavities in the human body provided in an embodiment of this application;

[0042] Figure 2 This is a schematic diagram of the structure of the multi-stage segmentation device for narrow human cavities provided in the embodiments of this application;

[0043] Figure 3 This is a schematic diagram of the physical structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0044] For CT-based segmentation of narrow cavities in the human body, the commonly used methods are mainly divided into traditional methods and deep learning-based methods.

[0045] Traditional CT image segmentation methods primarily segment images based on physical features such as texture, shape, and structure. Based on their segmentation principles, they can be broadly categorized into region-growing, morphological, and tubular structure detection methods: 1) Region-growing is a classic image processing technique that leverages local features and pixel similarities in CT images. It selects a seed point within a tubular region as the starting point for growth and uses growth criteria to progressively add adjacent pixels to the tubular region until a stopping condition is met. 2) Morphological methods adjust and grow shapes based on the shape features of cavities and the attributes of pixel neighborhoods. In two-dimensional CT image slices, darker cavities are surrounded by brighter cavity walls, forming local extrema. Morphological methods can be used to detect these local extrema, thereby extracting the contours of small cavities. 3) Tubular structure detection refers to using tubular structure filters to detect objects resembling tubular structures in an image. Because the narrow cavities of the human body (such as the trachea and blood vessels) have natural tubular anatomical structures, the tubular structures in the image can be detected first based on the local or overall tubular structural features of the cavity, and then the detection results can be finely segmented to achieve the segmentation of the human body cavity.

[0046] Deep learning methods can acquire more semantic features of images. Recently, more patents have applied neural networks to cavity segmentation, assisting in the removal of leaks or directly obtaining the final segmentation result end-to-end: 1) Leak detection is modeled as a classification problem, using Convolutional Neural Networks (CNNs) trained in a supervised learning manner to perform this classification task. 2) Using CT images as input, the CNN is trained to learn feature representations in the image. The last layer of the CNN uses a pixel-level classifier to classify each pixel, obtaining the probability that the pixel belongs to a human cavity, thus outputting the cavity segmentation label. 3) Segmenting human cavities in CT images based on 3D U-Net, mainly using a fixed-step sliding window method to sample cavity images into the network, predicting cavity locations through an encoder-decoder symmetric structure, and finally combining all extracted candidate regions into an integrated cavity structure. 4) Combining other algorithms and modules on the basis of 3D U-Net segmentation to improve the accuracy or completeness of cavity segmentation.

[0047] To address the aforementioned technical issues, the multi-stage segmentation method for narrow cavities in the human body provided in this application inputs preprocessed preoperative CT data into a depth network to complete multi-stage voxel-level prediction of cavities. Post-processing generates connected 0 / 1 binary cavity labels, thereby providing a visualized cavity model for clinical surgery.

[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0049] Figure 1 This is a flowchart illustrating the multi-stage segmentation method for narrow cavities in the human body provided in an embodiment of this application, as shown below. Figure 1 As shown, this application provides a multi-stage segmentation method for narrow cavities in the human body, including:

[0050] Step 101: Obtain the preprocessed computed tomography (CT) data.

[0051] Step 102: Input the preprocessed CT data into the multi-stage segmentation model to obtain the segmentation result output by the multi-stage segmentation model.

[0052] The multi-stage segmentation model includes a first stage, a second stage, and a third stage.

[0053] The first stage is used to obtain preliminary results based on a first loss function, which is used to perform semantic segmentation on the CT data.

[0054] The second stage is used to obtain improved results based on the first loss function, the second loss function, and the preliminary results. The second loss function is used to solve the intra-class imbalance problem in the preliminary results.

[0055] The third stage is used to obtain the segmentation result based on the first loss function, the second loss function, the third loss function and the improved result, and the third loss function is used to repair the cavity fracture in the improved result.

[0056] Specifically, the region of interest in the cavity is first extracted using traditional image processing methods to obtain CT data of the region of interest. Then, the CT data is preprocessed to obtain preprocessed CT data.

[0057] The preprocessed CT data is input into a multi-stage segmentation model. The multi-stage model consists of three phases: Phase 1: The CT data is processed using only the Random Crop strategy and a first loss function, such as Dice Loss, for initial network training. Phase 2: Hard mining Crop and Small airway Crop strategies are added, focusing on training the segmentation of cavities not extracted in Phase 1. A second loss function, such as General Union Loss, is added to address intra-class imbalance. Phase 3: Breakage Crop strategy is added, focusing on repairing the location of cavity fractures from Phase 2. A third loss function, such as a weighted break-aware loss function, is added, combined with Dice Loss and General Union Loss to train the deep network, emphasizing improving cavity length and branch detection rate.

[0058] During each training stage, corresponding weights were designed and added to the loss function to handle intra-class imbalance and cavity disconnection issues: the General Union Loss loss function combines static prior weights based on local cavity imbalance with dynamic weights based on voxel-level classification difficulty and importance, focusing on training the small cavity structures not extracted in the first stage; the Break-aware Loss loss function uses weights based on centerline distance and cavity size, respectively applied to repair the cavity locations and fracture locations not extracted in the second stage.

[0059] It is worth mentioning that the multi-stage segmentation model is a network model DMR-UNet formed by fusing the WingNet network model with a deformed multi-scale residual feature extraction module (MultiRes Block) structure. The several convolutional blocks of the encoder and decoder in the original network are replaced by a Residual Concatenation Module (RCM) structure, and the input data is used to complete the residual calculation in the form of Short-Middle-Long Skip Connection.

[0060] Furthermore, the backbone network of the multi-stage segmentation model can all be DMR-Unet, and the backbone network used in each stage can also be different. Specific networks can be designed according to the task of each training stage, such as nnUNet, VNet and Attention UNet.

[0061] The multi-stage segmentation method for narrow human cavities provided in this application segmentes human cavities through a multi-stage segmentation model. Each stage has its own training tasks and priorities based on the corresponding loss function, thereby obtaining the segmentation results of narrow human cavities and improving the accuracy, sensitivity and branch detection rate of cavities.

[0062] In some embodiments, acquiring the preprocessed computed tomography (CT) data includes:

[0063] Acquire CT data of the area where the cavity is located;

[0064] The CT values ​​are uniformly truncated into a single channel, and the preprocessed CT data is obtained based on normalization.

[0065] Specifically, traditional image processing methods are first used to extract the regions of interest (ROIs) to improve the efficiency of subsequent network training. Since the contrast between the ROI region and the background region in CT scans is relatively significant, resulting in a bimodal histogram, histogram-based thresholding of the CT scan is first employed. Then, morphological knowledge is used to optimize the binary region extraction to obtain the ROI mask, which is then used to crop the CT scan and its corresponding ground truth label.

[0066] Secondly, due to the individual differences in the distribution of CT scan data, CT values ​​are uniformly truncated to a single channel and normalized to the range [0,1]. To reduce the GPU space usage and computational complexity during network training, CT scans are cropped into small 2D / 3D cubes and input into the network.

[0067] In addition, to improve the robustness of the segmentation model, we use random rotation and flipping for data augmentation.

[0068] It is worth mentioning that ROI extraction in data preprocessing can be performed using deep learning methods, such as network structures like UNet and FCN.

[0069] The multi-stage segmentation method for narrow cavities in the human body provided in this application can reduce the GPU space occupation and computational complexity during network training and improve the robustness of the segmentation model by preprocessing CT data.

[0070] In some embodiments, the multi-stage segmentation model is determined based on a multi-scale residual feature extraction module, which is used to coordinate image features learned from cavity regions of different scales.

[0071] Specifically, the multi-stage segmentation model in this embodiment is composed of a WingNet fused and deformed MultiResBlock. Several convolutional blocks of the encoder and decoder are replaced by an RCM. The input data is used to complete the residual calculation in the form of Short-Middle-LongSkip Connection. The specific structure is as follows.

[0072] WingNet comprises a total of 18 convolutional blocks (ConvBlocks), with 12 serving as encoders and 6 as decoders. Each ConvBlock undergoes convolution, normalization, and activation functions to compute the data passed from the previous ConvBlock and pass the output to the next ConvBlock. It is also upsampled to its original resolution and added to the feature pyramid for group supervision.

[0073] To address the issue of scale variations in certain cavities, this application incorporates a deformed MultiRes Block in the WingsNet encoder to coordinate image features learned from cavity regions of different scales, thus mitigating the gradient vanishing problem in small structures within the cavities. The MultiRes Block is an extension of the Residual Block, inspired by the inception block. In the MultiRes Block, the results of three 3x3 convolutions are concatenated to form a combined feature map, which is then added to the result of a 1x1 convolution of the input feature map.

[0074] This application proposes RCM, which completes multiple convolutional block calculations of its input data and concatenation of the convolution results. The original CT data blocks are convolved at different resolutions using 1x1x1 and then added to the corresponding RCM result feature map. That is, the original image information participates in the supplementation and understanding of the encoder information flow through Small-Middle-Long skip connection. The decoder only uses RCM to stitch and fuse multi-scale information.

[0075] It is worth mentioning that 3D CT data can be cropped into cube inputs of different sizes during different training stages. Skip connections between the encoder and decoder can enhance channel attention and spatial attention mechanisms.

[0076] The multi-stage segmentation method for narrow cavities in the human body provided in this application can introduce multi-scale information by fusing deformed MultiRes Block structures during the feature extraction stage of WingNet, thereby improving the network's ability to perceive features at different scales and helping to alleviate the gradient vanishing problem in small cavities.

[0077] In some embodiments, the second loss function is determined based on a first static weight and a dynamic weight. The first static weight is calculated based on the local imbalance of the cavity before the second stage, and the dynamic weight is calculated based on the prediction performance of the voxels in the second stage.

[0078] Specifically, the segmentation in the first stage of the multi-stage segmentation model is trained using the first loss function Dice Loss (DL):

[0079]

[0080] where N is the total number of voxels, p i is the predicted value of a certain voxel, and g i is the ground truth of this voxel. Since there are size differences at different positions of some cavities, if the input data patch contains both large and small cavities, using DL training will exacerbate the gradient erosion of the small cavities. In the embodiments of this application, the training in the second stage adopts a mixed form of DL and the second loss function GeneralUnion Loss (GUL):

[0081]

[0082] where x and y are hyperparameters that control the degree of alleviating the intra-class imbalance. GUL is specifically:

[0083]

[0084] where w i is the weight assigned to each voxel. 0 < r < 1 achieves voxel-level focusing of the gradient ratio, and α + β = 1.

[0085] In this solution, w i is defined as two parts. One part is the static weight w s of each voxel calculated based on the local imbalance of the cavity before training, that is, the first static weight w s , and the other part is the dynamic weight w d calculated according to the prediction performance of the voxels during training.

[0086] w i = w s + w d

[0087] Among them, the first static weight w s :

[0088] w s = (1 - μ)(-log 10 FR p ) θ + μ

[0089]

[0090] Where 0 < μ < 1, 2 < θ < 3 are hyperparameters, B is a predefined neighborhood of N voxels centered at p, and FR p It is the local foreground ratio of voxel p, which can reflect the size of the branch in which the voxel is located. s This results in cavities with smaller branch sizes having a greater weight.

[0091] Dynamic weight w d :

[0092] w d =-n t (1-p t ) v log(p t )

[0093]

[0094]

[0095] Where 2<ν<3 is the modulating factor for focal loss, and n is set as the ratio of the number of non-cavity voxels to the number of cavity voxels. d Using Focal Loss results in high weights for voxels in cavities that are difficult to classify.

[0096] The multi-stage segmentation method for narrow cavities in the human body provided in this application adopts a multi-stage deep network training process and designs appropriate weighted loss functions at each stage, so that each stage can complete its own training focus in a targeted manner. At the same time, the training results of each stage are connected to each other, and the parameters of the multi-stage segmentation model can be continuously fine-tuned, which can improve the accuracy, sensitivity and branch detection rate of cavities.

[0097] In some embodiments, the third loss function is determined based on a second static weight and a dynamic weight, wherein the second static weight is calculated based on a first static weight, a weight of undetected cavity voxels in the second stage based on the distance from the central skeleton line, and a weight of cavity voxels at the fracture point in the second stage based on the cavity size.

[0098] Specifically, small cavities are difficult to detect and often have disconnections, affecting the overall connectivity of the cavities. In this application, the third stage is trained using a hybrid approach of DL, GUL, and the third loss function Break-aware Loss (BAL).

[0099]

[0100]

[0101] Where C is the total number of voxels on the cavity skeleton line. It is the weight assigned to each voxel.

[0102] In the third phase, It consists of a second static weight and a dynamic weight, wherein the second static weight It will add new components:

[0103]

[0104]

[0105] Where 0 < ρ < 1 controls the degree of attention paid to fracture repair, w hm The undetected cavity voxels in the second stage are weighted based on their distance from the central skeletal line, w br The cavity voxels at the fracture site in the second stage are weighted based on cavity size.

[0106] in, The expression is:

[0107]

[0108] Where d i dmax is the shortest distance from the cavity voxel to the central skeleton line, and dmax is the maximum cavity radius not detected in the second stage. hm This results in voxels closer to the center line having a greater weight in hard-to-segment samples.

[0109] w br The expression is:

[0110]

[0111] in It is the shortest distance from the cavity voxel to the cavity wall, and K is a constant. br This results in a greater weight for voxels in fractured samples that are located in thicker branches and closer to the center line.

[0112] The method for detecting fractures in hard-to-segment samples is to perform convolution operations on the unextracted skeleton points. Points with consistent convolution results are considered fractured skeleton lines.

[0113] The multi-stage segmentation method for narrow cavities in the human body provided in this application adopts a multi-stage deep network training process and designs appropriate weighted loss functions at each stage, so that each stage can complete its own training focus in a targeted manner. At the same time, the training results of each stage are connected to each other, and the parameters of the multi-stage segmentation model can be continuously fine-tuned, which can improve the accuracy, sensitivity and branch detection rate of cavities.

[0114] In some embodiments, obtaining the segmentation result output by the multi-stage segmentation model includes:

[0115] The segmentation result is determined from the feature map predicted by the network based on a dual-threshold iterative algorithm.

[0116] Specifically, the segmentation result is usually determined by the predicted probability map with a threshold of 0.5. Considering that there are some weak cavity voxels whose scores are slightly less than the threshold on the probability map, in order to enhance the connectivity between voxels, this scheme adopts the dual threshold iterative algorithm (DTI) to effectively segment the predicted probability map.

[0117] The multi-stage segmentation method for narrow human cavities provided in this application segmentes human cavities through a multi-stage segmentation model. Each stage has its own training tasks and priorities based on the corresponding loss function, thereby obtaining the segmentation results of narrow human cavities and improving the accuracy, sensitivity and branch detection rate of cavities.

[0118] Figure 2 This is a schematic diagram of the structure of the multi-stage segmentation device for narrow human cavities provided in the embodiments of this application, as shown below. Figure 2 As shown, the multi-stage segmentation device for narrow cavities in the human body provided in this application embodiment includes a first acquisition module 201 and a second acquisition module 202, wherein:

[0119] The first acquisition module 201 is used to acquire preprocessed computed tomography (CT) data.

[0120] The second acquisition module 202 is used to input the preprocessed CT data into the multi-stage segmentation model and acquire the segmentation result output by the multi-stage segmentation model.

[0121] The multi-stage segmentation model includes a first stage, a second stage, and a third stage;

[0122] The first stage is used to obtain preliminary results based on a first loss function, which is used to perform semantic segmentation on the CT data;

[0123] The second stage is used to obtain improved results based on the first loss function, the second loss function, and the preliminary results. The second loss function is used to solve the intra-class imbalance problem in the preliminary results.

[0124] The third stage is used to obtain the segmentation result based on the first loss function, the second loss function, the third loss function and the improved result, and the third loss function is used to repair the cavity fracture in the improved result.

[0125] In some embodiments, the second loss function is determined based on a first static weight and a dynamic weight, wherein the first static weight is calculated based on local imbalance of the cavity before the second stage, and the dynamic weight is calculated based on the predicted performance of voxels in the second stage.

[0126] In some embodiments, the third loss function is determined based on a second static weight and a dynamic weight, wherein the second static weight is calculated based on a first static weight, a weight of undetected cavity voxels in the second stage based on the distance from the central skeleton line, and a weight of cavity voxels at the fracture point in the second stage based on the cavity size.

[0127] In some embodiments, the multi-stage segmentation model is determined based on a multi-scale residual feature extraction module, which is used to coordinate image features learned from cavity regions of different scales.

[0128] In some embodiments, the first acquisition module includes:

[0129] The first acquisition submodule is used to acquire CT data of the area where the cavity is located;

[0130] The second acquisition submodule is used to uniformly truncate CT values ​​into a unified channel and acquire the preprocessed CT data based on normalization.

[0131] In some embodiments, the second acquisition module includes:

[0132] The first determining submodule is used to determine the segmentation result from the feature map predicted by the network based on a dual-threshold iterative algorithm.

[0133] Specifically, the multi-stage segmentation device for human stenotic cavities provided in this application embodiment can realize all the method steps implemented in the multi-stage segmentation method embodiment for human stenotic cavities and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0134] Figure 3 This is a schematic diagram of the physical structure of the electronic device provided in the embodiments of this application, such as... Figure 3 As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a multi-stage segmentation method for narrow cavities in the human body, the method including:

[0135] Acquire preprocessed computed tomography (CT) data;

[0136] The preprocessed CT data is input into a multi-stage segmentation model to obtain the segmentation results output by the multi-stage segmentation model.

[0137] The multi-stage segmentation model includes a first stage, a second stage, and a third stage;

[0138] The first stage is used to obtain preliminary results based on a first loss function, which is used to perform semantic segmentation on the CT data;

[0139] The second stage is used to obtain improved results based on the first loss function, the second loss function, and the preliminary results. The second loss function is used to solve the intra-class imbalance problem in the preliminary results.

[0140] The third stage is used to obtain the segmentation result based on the first loss function, the second loss function, the third loss function and the improved result, and the third loss function is used to repair the cavity fracture in the improved result.

[0141] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0142] In some embodiments, the second loss function is determined based on a first static weight and a dynamic weight, wherein the first static weight is calculated based on local imbalance of the cavity before the second stage, and the dynamic weight is calculated based on the predicted performance of voxels in the second stage.

[0143] In some embodiments, the third loss function is determined based on a second static weight and a dynamic weight, wherein the second static weight is calculated based on a first static weight, a weight of undetected cavity voxels in the second stage based on the distance from the central skeleton line, and a weight of cavity voxels at the fracture point in the second stage based on the cavity size.

[0144] In some embodiments, the multi-stage segmentation model is determined based on a multi-scale residual feature extraction module, which is used to coordinate image features learned from cavity regions of different scales.

[0145] In some embodiments, acquiring the preprocessed computed tomography (CT) data includes:

[0146] Acquire CT data of the area where the cavity is located;

[0147] The CT values ​​are uniformly truncated into a single channel, and the preprocessed CT data is obtained based on normalization.

[0148] In some embodiments, obtaining the segmentation result output by the multi-stage segmentation model includes:

[0149] The segmentation result is determined from the feature map predicted by the network based on a dual-threshold iterative algorithm.

[0150] Specifically, the electronic device provided in this application embodiment can implement all the method steps implemented by the method embodiment with the execution subject being an electronic device, and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0151] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the multi-stage segmentation method for human stenotic cavities provided by the above methods, the method comprising:

[0152] Acquire preprocessed computed tomography (CT) data;

[0153] The preprocessed CT data is input into a multi-stage segmentation model to obtain the segmentation results output by the multi-stage segmentation model.

[0154] The multi-stage segmentation model includes a first stage, a second stage, and a third stage;

[0155] The first stage is used to obtain preliminary results based on a first loss function, which is used to perform semantic segmentation on the CT data;

[0156] The second stage is used to obtain improved results based on the first loss function, the second loss function, and the preliminary results. The second loss function is used to solve the intra-class imbalance problem in the preliminary results.

[0157] The third stage is used to obtain the segmentation result based on the first loss function, the second loss function, the third loss function and the improved result, and the third loss function is used to repair the cavity fracture in the improved result.

[0158] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a multi-stage segmentation method for human stenotic cavities provided by the methods described above, the method comprising:

[0159] Acquire preprocessed computed tomography (CT) data;

[0160] The preprocessed CT data is input into a multi-stage segmentation model to obtain the segmentation results output by the multi-stage segmentation model.

[0161] The multi-stage segmentation model includes a first stage, a second stage, and a third stage;

[0162] The first stage is used to obtain preliminary results based on a first loss function, which is used to perform semantic segmentation on the CT data;

[0163] The second stage is used to obtain improved results based on the first loss function, the second loss function, and the preliminary results. The second loss function is used to solve the intra-class imbalance problem in the preliminary results.

[0164] The third stage is used to obtain the segmentation result based on the first loss function, the second loss function, the third loss function and the improved result, and the third loss function is used to repair the cavity fracture in the improved result.

[0165] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0166] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0167] It should also be noted that the terms "first," "second," etc., used in the embodiments of this application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, and the number of objects is not limited. For example, the first object can be one or more.

[0168] In the embodiments of this application, the term "and / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0169] In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar.

[0170] In this application, "determining B based on A" means that factor A must be considered when determining B. It is not limited to "B can be determined based solely on A," but should also include: "determining B based on A and C," "determining B based on A, C, and E," "determining C based on A, and further determining B based on C," etc. It can also include using A as a condition for determining B, for example, "when A satisfies the first condition, B is determined using the first method"; or "when A satisfies the second condition, B is determined," or "when A satisfies the third condition, B is determined based on the first parameter," etc. Of course, it can also be a condition where A is a factor in determining B, for example, "when A satisfies the first condition, C is determined using the first method, and B is further determined based on C," etc.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multi-stage segmentation of a human narrow lumen, the method comprising: The method comprises: obtaining preprocessed computed tomography (CT) data; inputting the preprocessed CT data into a multi-stage segmentation model to obtain a lumen segmentation result output by the multi-stage segmentation model; the multi-stage segmentation model comprises a first stage, a second stage, and a third stage; the first stage is configured to obtain a preliminary result based on a first loss function, and the first loss function is configured to perform semantic segmentation on the CT data; the second stage is configured to obtain an improved result based on the first loss function, a second loss function, and the preliminary result, and the second loss function is configured to train a part of the lumen segmentation that is not extracted by the first stage; the third stage is configured to obtain the segmentation result based on the first loss function, the second loss function, a third loss function, and the improved result, and the third loss function is configured to repair lumen fractures in the improved result.

2. The method of claim 1, wherein, The second loss function is determined based on a first static weight and a dynamic weight, the first static weight is calculated based on lumen local imbalance before the second stage, and the dynamic weight is calculated based on the predicted performance of the voxel in the second stage; the second loss function is expressed by the following formula: where N is the total number of voxels, p i is the predicted value of a certain voxel, g i is the true value of the voxel, w i is the weight assigned to each voxel, 0 < r < 1 achieves voxel-level focusing of the gradient ratio, and a + b = 1.

3. The method of claim 2, wherein, The third loss function is determined based on a second static weight and a dynamic weight, the second static weight is calculated according to the first static weight, a weight based on the center skeleton distance of the lumen voxel that is not detected in the second stage, and a weight based on the lumen size of the lumen voxel at the fracture in the second stage; the third loss function is expressed by the following formula: where C is the total number of voxels on the cavity skeleton line, is the weight assigned to each voxel.

4. The method of claim 1, wherein, The multi-stage segmentation model is determined based on a multi-scale residual feature extraction module (MultiRes Block), and the multi-scale residual feature extraction module (MultiRes Block) is configured to coordinate image features learned from lumen regions of different scales.

5. The method of claim 1, wherein, The method for obtaining the preprocessed CT data comprises: obtaining CT data of a lumen region; uniformly truncating CT values in a uniform channel, and obtaining the preprocessed CT data based on normalization.

6. The method of claim 1, wherein, The method for obtaining the segmentation result output by the multi-stage segmentation model comprises: determining the segmentation result from a feature map predicted by a network based on a double-threshold iterative algorithm.

7. A multi-stage device for segmenting a human body's narrow lumen, comprising: The method comprises: a first obtaining module configured to obtain preprocessed computed tomography (CT) data; a second obtaining module configured to input the preprocessed CT data into a multi-stage segmentation model to obtain a lumen segmentation result output by the multi-stage segmentation model; the multi-stage segmentation model comprises a first stage, a second stage, and a third stage; the first stage is configured to obtain a preliminary result based on a first loss function, and the first loss function is configured to perform semantic segmentation on the CT data; the second stage is configured to obtain an improved result based on the first loss function, a second loss function, and the preliminary result, and the second loss function is configured to train a part of the lumen segmentation that is not extracted by the first stage; The third stage is configured to obtain the segmentation result based on a first loss function, a second loss function, a third loss function, and the improved result, wherein the third loss function is configured to repair a lumen break in the improved result.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor implements the multi-stage segmentation method of a human narrow lumen according to any one of claims 1 to 6 when executing the program. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the multi-stage segmentation method of a human narrow lumen according to any one of claims 1 to 6 when executed by the processor.

10. A computer program product comprising a computer program, characterized in that, The computer program implements the multi-stage segmentation method of a human narrow lumen according to any one of claims 1 to 6 when executed by the processor. The computer program implements the multi-stage segmentation method of a human narrow lumen according to any one of claims 1 to 6 when executed by the processor.

Images