WO2024216672A1 - 基于扩散模型及3d transformer的三维分割方法、系统、设备及介质 - Google Patents

基于扩散模型及3d transformer的三维分割方法、系统、设备及介质 Download PDF

Info

Publication number
WO2024216672A1
WO2024216672A1 PCT/CN2023/091750 CN2023091750W WO2024216672A1 WO 2024216672 A1 WO2024216672 A1 WO 2024216672A1 CN 2023091750 W CN2023091750 W CN 2023091750W WO 2024216672 A1 WO2024216672 A1 WO 2024216672A1
Authority
WO
WIPO (PCT)
Prior art keywords
transformer
diffusion model
dimensional
dimensional segmentation
image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/091750
Other languages
English (en)
French (fr)
Inventor
胡颖
谭志强
王琼
赵宇
李世博
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Institute of Advanced Technology of CAS
Original Assignee
Shenzhen Institute of Advanced Technology of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen Institute of Advanced Technology of CAS filed Critical Shenzhen Institute of Advanced Technology of CAS
Publication of WO2024216672A1 publication Critical patent/WO2024216672A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T19/00—Manipulating three-dimensional [3D] models or images for computer graphics
    • G06T19/20—Editing of three-dimensional [3D] images, e.g. changing shapes or colours, aligning objects or positioning parts
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Definitions

  • the present invention belongs to the technical field of image processing, and relates to a three-dimensional segmentation method, system, equipment and medium based on a diffusion model and a 3D Transformer.
  • Medical imaging data is usually three-dimensional data, such as CT, MRI, etc.
  • scoliosis itself, which we are concerned about, is a three-dimensional deformity.
  • Another important problem is that the contextual relationship between CT images is lost through training with a single slice data. Therefore, the current research dynamics have begun to focus on using three-dimensional convolution to directly perform three-dimensional segmentation.
  • Medical imaging has always been an important branch of the medical field. It provides doctors with information about the tissues and organs of patients and analyzes and diagnoses some pathological conditions. In fact, medical image segmentation has always been a bottleneck problem in the medical field. From a technical point of view, the three-dimensional reconstruction and quantitative analysis of normal and pathological tissues are all based on image segmentation. Medical image segmentation segments the target area of interest in the image to lay the foundation for a series of subsequent analytical operations. Therefore, the quality of segmentation directly affects a series of subsequent operations. From another perspective, the accuracy of medical image segmentation is of vital importance for doctors to judge the specific situation of the disease and make a diagnosis, as well as some subsequent image operations. Medical images have the characteristics of blur, unevenness, individual differences, complexity and diversity.
  • convolutional neural networks have made great achievements in medical image processing.
  • deep convolutional neural networks based on U-shaped structure contraction paths and expansion paths and jump connections have been widely used in various medical image processing tasks.
  • convolutional neural networks have achieved good performance, due to the locality of convolution operations, the model cannot learn global information and contextual semantic interaction information well.
  • some scholars have begun to combine the Transformer model (spatial transformation model) to perform medical image segmentation tasks (Chen J, Lu Y, Yu Q, et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation[J]. 2021.; Cao H, Wang Y, Chen J, et al.
  • Swin-Unet Unet-like Pure Transformer for Medical Image Segmentation[J]. 2021.).
  • the Transformer model has unrestricted position association operations, strong modeling capabilities, strong versatility, strong scalability, and better parallel operations.
  • the number of operations required for the Transformer model to calculate the association between two positions does not increase with distance.
  • the self-attention mechanism can produce a more interpretable model, in which each attention head of the attention mechanism can learn to perform different tasks and has stronger modeling capabilities.
  • Transformer can establish the relationship between graph nodes and has good versatility: no matter how heterogeneous the nodes are, the relationship between them can be established by projecting them into a comparable space and calculating the similarity.
  • the current Transformer model is mainly based on two-dimensional images.
  • the purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a three-dimensional segmentation method, system, device and medium based on a diffusion model and a 3D Transformer, which can perform three-dimensional segmentation of medical images.
  • the present invention adopts the following technical scheme:
  • the present invention provides a three-dimensional segmentation method based on a diffusion model and a 3D Transformer, comprising:
  • the image to be segmented is input into a three-dimensional segmentation network model based on a diffusion model and a 3D Transformer to obtain a three-dimensional segmentation result of the image to be segmented.
  • the three-dimensional segmentation method based on the diffusion model and 3D Transformer of the present invention is further improved in that:
  • the three-dimensional segmentation network model based on the diffusion model and 3D Transformer is composed of a CNN-Transformer hybrid architecture, a decoder and a diffusion model.
  • the CNN-Transformer hybrid architecture is composed of a stack of 3D NCB modules and 3D NTB modules, wherein the 3D NCB module is used to learn local representations in the image to be segmented; and the 3D NTB module is used to capture multi-frequency signals in a lightweight mechanism.
  • the 3D NCB module is constructed by a 3D MHCA module and a 3D MLP module.
  • the decoder is constructed based on the three-dimensional deconvolution form of the Inception structure.
  • the Inception structure consists of a three-dimensional point-wise convolution layer with a kernel size of 1 ⁇ 1 ⁇ 1 and two three-dimensional deconvolution layers with kernel sizes of 3 ⁇ 3 ⁇ 3 and 5 ⁇ 5 ⁇ 5.
  • the diffusion model adds random noise to the semantic label, learns the random noise component on the semantic label, and then denoises the semantic label to obtain relevant semantic information.
  • the present invention provides a three-dimensional segmentation system based on a diffusion model and a 3D Transformer, comprising:
  • An acquisition module used for acquiring an image to be segmented
  • the segmentation module is used to input the image to be segmented into a three-dimensional segmentation network model based on a diffusion model and a 3D Transformer to obtain a three-dimensional segmentation result of the image to be segmented.
  • the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the three-dimensional segmentation method based on the diffusion model and the 3D Transformer when executing the computer program.
  • the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the three-dimensional segmentation method based on the diffusion model and the 3D Transformer are implemented.
  • the three-dimensional segmentation method, system, device and medium based on the diffusion model and 3D Transformer described in the present invention adopts the three-dimensional segmentation network model constructed by the diffusion model and 3D Transformer to perform three-dimensional segmentation of the image to be segmented.
  • the three-dimensional Transformer module based on the present invention is used for subsequent three-dimensional segmentation tasks of medical images to realize three-dimensional segmentation of medical images.
  • the diffusion model can well capture pixel-level semantic information, so that the three-dimensional segmentation network model can better segment the lesion area.
  • FIG1 is a framework diagram of a three-dimensional segmentation network model based on a diffusion model and a 3D Transformer in the present invention.
  • FIG. 2 is a schematic diagram of the method framework of the present invention.
  • FIG. 3 is a schematic diagram of the diffusion process in the present invention.
  • Figure 4a is a real image in the simulation experiment.
  • FIG. 4 b is a diagram showing the segmentation effect of the present invention.
  • Figure 4c is a diagram of the segmentation effect of the Transformer-based segmentation model.
  • Figure 4d is a diagram showing the segmentation effect of the segmentation model based on the convolutional network.
  • FIG5 is a system structure diagram of the present invention.
  • the three-dimensional segmentation method based on the diffusion model and the 3D Transformer of the present invention comprises the following steps:
  • step 1) The specific process of step 1) is:
  • the present invention adds the Transformer structure and the diffusion model to the existing 3D segmentation network model to form a 3D segmentation network model based on the diffusion model and the 3D Transformer, specifically:
  • the Tansformer structure is an improvement over the Next-ViT structure, following a hierarchical pyramid network structure.
  • a patch embedding layer and a series of convolutional or Transformer blocks are configured at each stage, gradually reducing its three-dimensional spatial resolution to 1/32 of the original, and then expanding the channel dimension at different stages.
  • a novel CNN-Transformer hybrid architecture is constructed by stacking 3D NCB modules and 3D NTB modules. It should be noted that the present invention reorganizes the NCB module and the NTB module, and changes the NCB module and the NTB module into three-dimensional modules, namely, 3D NCB module and 3D NTB module, so as to simulate the short-term and long-term dependencies in the spinal CT sequence image data. Among them, the 3D NTB module can also fuse local and global information to further improve the modeling capability.
  • the 3D NCB module is constructed by the 3D MHCA module and the 3D MLP module.
  • the 3D NCB module can obtain outstanding performance as a Transformer block while ensuring the deployment advantages of the BottleNeck block.
  • the 3D NCB module follows the general architecture of the Transformer.
  • the present invention adopts multi-head convolutional attention (MHCA) as an efficient token mixer with deployment-friendly convolution operations.
  • MHCA multi-head convolutional attention
  • 3D MHCA module To solve the problem of high latency of existing attention-based token mixers, the present invention adopts a new attention mechanism with efficient convolution operation, namely 3D convolutional attention (3D CA), to achieve fast inference speed.
  • 3D convolutional attention 3D CA
  • the present invention adopts a multi-head paradigm to construct a convolutional attention mechanism, which jointly pays attention to information from different representation subspaces at different positions to achieve effective local representation learning.
  • the representation of the multi-head convolutional attention (MHCA) is:
  • MHCA captures information from n parallel representation subspaces.
  • 3D CA as a single-head convolutional attention mechanism can be expressed as:
  • 3D CA can learn the correlation between different Tokens in the local receptive field by iteratively optimizing the trainable parameter w .
  • 3D NTB module The 3D NCB module can effectively learn local representations in images, but the capture of global information of the image also needs to be solved.
  • the Transformer module has a strong ability to capture low-frequency signals, thereby providing global information, such as feature shapes and structures in images.
  • the Transformer module may deteriorate high-frequency information to a certain extent, such as local texture information. Signals in different frequency bands are indispensable in the visual system and will be fused in a specific way to extract more basic and obvious features. Therefore, the present invention changes the Next Transformer Block (NTB) module to a 3D NTB module to capture multi-frequency signals in a lightweight mechanism.
  • the 3D NTB module as a multi-frequency signal mixer, can further enhance the overall modeling capability.
  • E-MHSA multi-head self-attention mechanism
  • x [ x 1 ,x 2 ,...,x n ] means that the input feature value x is divided into multiple forms in the channel dimension.
  • SA is a spatial reduction self-attention operator expressed as:
  • New Inception module adds a three-dimensional deconvolution form (New Inception) of the Inception structure to the three-dimensional segmentation network model as a decoder of the three-dimensional segmentation network model, and the Inception structure is composed of a three-dimensional point-wise convolution with a kernel size of 1 ⁇ 1 ⁇ 1 and two three-dimensional deconvolutions with kernel sizes of 3 ⁇ 3 ⁇ 3 and 5 ⁇ 5 ⁇ 5.
  • the three-dimensional point-wise convolution layer of the deconvolution module can superimpose more deconvolutions on the same size of the generated field, so as to obtain more details in the image to be segmented, so as to perform more refined segmentation.
  • the three-dimensional point-wise convolution layer also plays a vital role in dimensionality reduction.
  • the New Inception module consists of two branch networks, each of which uses filters of different sizes for deconvolution, wherein the branch network can generate information of different scales.
  • the principle used by the New Inception module is to decompose the sparse matrix into a dense matrix for calculation, and decompose the feature dimension into multiple densely distributed sub-feature sets, in which highly correlated features are clustered together and irrelevant features are weakened. Finally, the sub-features are spliced on the feature dimension and kept consistent with the input dimension to reduce the computational cost and ensure that the final training results are not affected.
  • the present invention adds a diffusion model to the three-dimensional segmentation network model for semantic segmentation, wherein the diffusion model, when working, adds random noise to the semantic label, learns the random noise component on the semantic label, and then denoises the semantic label to obtain relevant semantic information, so that the three-dimensional segmentation network model can better segment the lesion area.
  • the inspiration of the diffusion model is mainly derived from non-equilibrium thermodynamics.
  • a Markov chain of diffusion steps is set, and random noise is gradually added to the real data (forward process), and then the reverse diffusion process (reverse diffusion process) is learned to obtain the required data sample results from the noise.
  • the diffusion model is divided into a forward process and a reverse process.
  • the forward process does not contain learnable parameters, whose values increase with t, and the final value distribution is a Gaussian distribution that is independent in all directions.
  • the distribution of real data is In the forward process, a small Gaussian noise is gradually added for a total of T steps, thus generating a series of noisy samples , the mean and variance of the added noise are given by Determine, its value is in the range (0,1), and given the condition , it can be seen that the amount of noise added to the diffusion process is getting larger and larger, so the forward process is defined as:
  • the reverse process of the diffusion process is to recover the original data from the random Gaussian noise. Since the noise value added each time in the forward process is very small, it can be As a Gaussian distribution, its process can be fitted using a deep neural network. Therefore, the inverse process can also be a Markov chain process, so the inverse process is expressed as:
  • the present invention can not only be used for three-dimensional segmentation of the spine, but as a general algorithm, the present invention can also be used for other medical image segmentation tasks, such as breast cancer segmentation, liver segmentation, etc.
  • the three-dimensional segmentation system based on the diffusion model and the 3D Transformer of the present invention includes:
  • An acquisition module 1 is used to acquire an image to be segmented
  • the segmentation module 2 is used to input the image to be segmented into a three-dimensional segmentation network model based on a diffusion model and a 3D Transformer to obtain a three-dimensional segmentation result of the image to be segmented.
  • a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor.
  • the memory may include a memory, such as a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk memory, etc.
  • the processor, the network interface, and the memory are interconnected through an internal bus.
  • the internal bus may be an industrial standard architecture bus, a peripheral component interconnection standard bus, an extended industrial standard architecture bus, etc.
  • the bus may be divided into an address bus, a data bus, a control bus, etc.
  • the memory is used to store programs.
  • the program may include a program code, and the program code includes computer operation instructions.
  • the memory may include a memory and a non-volatile memory, and provide instructions and data to the processor.
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the three-dimensional segmentation method based on the diffusion model and 3D Transformer are implemented.
  • the computer-readable storage medium includes, but is not limited to, for example, a volatile memory and/or a non-volatile memory.
  • the volatile memory may include a random access memory (RAM) and/or a cache memory, etc.
  • the non-volatile memory may include a read-only memory (ROM), a hard disk, a flash memory, an optical disk, a magnetic disk, etc.
  • the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
  • a computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.
  • These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Public Health (AREA)
  • Primary Health Care (AREA)
  • Medical Informatics (AREA)
  • Architecture (AREA)
  • Computer Graphics (AREA)
  • Epidemiology (AREA)
  • Computer Hardware Design (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Multimedia (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了一种基于扩散模型及3D Transformer的三维分割方法、系统、设备及介质，包括：获取待分割图像；将待分割图像输入到基于扩散模型及3D Transformer的三维分割网络模型中，得到待分割图像的三维分割结果，方法、系统、设备及介质能够对医学影像进行三维分割。

Description

基于扩散模型及3D Transformer的三维分割方法、系统、设备及介质 技术领域
本发明属于图像处理技术领域，涉及一种基于扩散模型及3D Transformer的三维分割方法、系统、设备及介质。
 背景技术
医学影像数据通常是三维数据，如CT、MRI等。具体来说，对我们所关注的脊柱侧弯本身，其就是三维畸形。若从二维的脊柱侧弯图像分割模型去提取出脊柱区域也有两种方式，第一种则是从X光数据进行分割，但分割得到的数据时二维层面的分割图，第二种则是从CT的每一张切片数据进行分割，但是这种方法需要将脊柱侧弯的CT影像数据的所有切片提取出来，然后单独一张张地输入到模型进行训练，在这种情况下会存在一个效率问题。另一个重要的问题则是通过单张切片数据进行训练丢失了CT影像之间的上下文关联关系。因此，当前的研究动态开始着眼于利用三维卷积来直接进行三维分割。
医学影像一直都是医学领域重要的分支，其提供给医生病人的组织器官信息，对一些病变情况进行分析诊断。实际上，医学影像分割，一直都是医学领域一个瓶颈问题。从技术角度来说，正常以及病变组织的三维重建、定量分析等操作都是建立在影像分割的基础上，医学影像分割通过将图像中的感兴趣目标区域分割出来，以便为后续的一系列分析性操作打下基础。因此，分割的好坏程度直接影响了后续一系列的操作。从另外一个角度来看，医学影像分割的准确性对医生判断疾病的具体情况并做出诊断以及后续一些影像上的操作具有至关重要的意义。医学图像具有模糊、不均匀、个体差异、复杂多样等特点，也正是由于这个特点使得医学影像的分割显得尤为重要。深度学习技术发展至今，许多领域都依托于深度学习技术得到了长足的发展，医学影像分割也不例外，Ronneberger等人在ISBI cell tracking challenge 2015 使用U-net网络（Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation[J]. Springer, Cham, 2015.）以较大的优势赢得了比赛，并成为医学影像领域广为人知的模型。
Lee等人（Lee H, Kim J, Lee S, Kim H, Ro Y. Structure Boundary Preserving Segmentation for Medical Image With Ambiguous Boundary[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, 4816-4825.）提出了一种新颖的结构边界保留分割框架。此外，为了将专家知识嵌入全自动分割中，该论文提出了一种新颖的形状边界感知评估器（SBE），其具有专家指示的地面真实结构信息。所提出的SBE可以基于结构边界关键点向分割网络提供反馈。另一方面，医学影像通常难以获取，当前许多学者也致力于研究小样本甚至零样本在医学影像中的应用。Khoi等人（Nguyen K, Todorovic S. FAPIS: A Few-shot Anchor-free Part-based Instance Segmenter[J]. 2021.）提出了一个小样本无锚实例分割框架，通过对训练对象类别之间的共享隐藏对象部分进行明确建模，从而促进测试中新样本的小样本学习，该方法在基准COCO20i数据集上得到很好的评估结果。针对脊柱分割任务，AL Arif等人利用U-net与FCN设计出一个全自动分割颈柱的框架，将颈柱分割任务分成了三个子任务，通过分别制作相应的数据集完成了任务并且达到了很好的效果。并且其在DICE测度下，精度能达到95.7%（Al, Arif, S, et al. Fully automatic cervical vertebrae segmentation framework for X-ray images[J]. Computer Methods and Programs in Biomedicine: An International Journal Devoted to the Development, Implementation and Exchange of Computing Methodology and Software Systems in Biomedical Research and Medical Practice, 2018, 157:95-111.）。Yu等人通过整合不确定性估计来进行半监督学习，通过结合蒙特卡洛法来进行不确定性图构建并指导模型进行逐步学习，可以得到不错的效果（Yu L, Wang S, Li X, et al. Uncertainty-aware Self-ensembling Model for Semi-supervised 3D Left Atrium Segmentation[J]. 2019.）。Sun等人（Lsa B, Cl A, Xd A , et al. Few-shot medical image segmentation using a global correlation network with discriminative embedding[J]. Computers in Biology and Medicine, 2021, 140.）利用片段式(episodic)训练的深度卷积网络构造了一个小样本图像分割器。基于医学图像的空间一致性和规律性，其开发了一个高效的全局相关模块来捕获支持和查询图像之间的相关性，并将其纳入到全局相关网络中，其实验结果在小样本医学分割中得到了不错的结果验证。
在过去的几年时间中，卷积神经网络在医学影像处理中取得了非常大的成就。其中，尤为突出的是基于U型结构的收缩路径和扩展路径以及跳跃连接的深度卷积神经网络在各类医学图像处理任务中得到了广泛的应用。然而，尽管卷积神经网络已经取得了很好的性能，但由于卷积运算的局部性，使得该模型不能很好地学习全局信息以及上下文语义交互信息。目前已有一些学者开始结合Transformer模型（空间变换模型）来进行医学影像分割任务（Chen J, Lu Y, Yu Q, et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation[J]. 2021.；Cao H, Wang Y, Chen J, et al. Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation[J]. 2021.）。相比卷积神经网络，Transformer模型位置关联操作不受限，建模能力强，通用性强，可扩展性强，能更好的进行并行运算。另一方面，Transformer模型计算两个位置之间的关联所需的操作次数不随距离增长。而且其中的自注意力机制可以产生更具可解释性的模型，其中注意力机制的各个注意头(attention head)可以学会执行不同的任务，有更强的建模能力。另外，Transformer 可以建立图节点之间的关系，具有较好的通用性：无论节点多么异构，它们之间的关系都可以通过投影到一个可以比较的空间里计算相似度来建立。
但从当前的研究现状来看，目前的Transformer模型主要基于二维影像，目前还没有针对三维影像构建的医学影像Transformer模型，不能进行医学影像的三维分割。
 技术问题
本发明的目的在于克服上述现有技术的缺点，提供了一种基于扩散模型及3D Transformer的三维分割方法、系统、设备及介质，该方法、系统、设备及介质能够对医学影像进行三维分割。
 技术解决方案
为达到上述目的，本发明采用如下技术方案：
本发明一方面，本发明提供了一种基于扩散模型及3D Transformer的三维分割方法，包括：
获取待分割图像；
将所述待分割图像输入到基于扩散模型及3D Transformer的三维分割网络模型中，得到待分割图像的三维分割结果。
本发明所述基于扩散模型及3D Transformer的三维分割方法进一步的改进在于：
所述基于扩散模型及3D Transformer的三维分割网络模型由CNN-Transformer 混合架构、解码器及扩散模型构成。
所述CNN-Transformer 混合架构由3D NCB模块及3D NTB模块堆叠而成，其中，通过所述3D NCB模块学习待分割图像中的局部表示；通过所述3D NTB模块捕获轻量级机制中的多频信号。
所述3D NCB模块通过3D MHCA模块及3D MLP 模块构建而成。
所述解码器基于Inception结构的三维反卷积形式构建而成。
所述Inception结构由一个核大小为1×1×1的三维point-wise卷积层及两个核大小为3×3×3和5×5×5的三维反卷积层组成。
所述扩散模型通过对语义标签添加随机噪声，并学习语义标签上的随机噪声分量，再对所述语义标签进行去噪，得到相关的语义信息。
本发明二方面，本发明提供了一种基于扩散模型及3D Transformer的三维分割系统，包括：
获取模块，用于获取待分割图像；
分割模块，用于将所述待分割图像输入到基于扩散模型及3D Transformer的三维分割网络模型中，得到待分割图像的三维分割结果。
本发明三方面，本发明提供了一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现所述基于扩散模型及3D Transformer的三维分割方法的步骤。
本发明四方面，本发明提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现所述基于扩散模型及3D Transformer的三维分割方法的步骤。
 有益效果
本发明具有以下有益效果：
本发明所述的基于扩散模型及3D Transformer的三维分割方法、系统、设备及介质在具体操作时，采用扩散模型及3D Transformer构建而成的三维分割网络模型进行待分割图像的三维分割，其中，本发明基于三维的Transformer模块用于后续的医学影像三维分割任务，实现医学影像的三维分割，同时通过扩散模型可以很好的捕捉到像素级的语义信息，使得三维分割网络模型能够更好的分割出病灶区域。
 附图说明
图1为本发明中基于扩散模型及3D Transformer的三维分割网络模型的框架图。
图2为本发明的方法框架示意图。
图3为本发明中扩散过程的示意图。
图4a为仿真实验中真实图像。
图4b为本发明的分割效果图。
图4c为基于Transformer的分割模型的分割效果图。
图4d为基于卷积网络的分割模型的分割效果图。
图5为本发明的系统结构图。
 本发明的实施方式
为了使本技术领域的人员更好地理解本发明方案，下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分的实施例，不是全部的实施例，而并非要限制本发明公开的范围。此外，在以下说明中，省略了对公知结构和技术的描述，以避免不必要的混淆本发明公开的概念。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都应当属于本发明保护的范围。
在附图中示出了根据本发明公开实施例的结构示意图。这些图并非是按比例绘制的，其中为了清楚表达的目的，放大了某些细节，并且可能省略了某些细节。图中所示出的各种区域、层的形状及它们之间的相对大小、位置关系仅是示例性的，实际中可能由于制造公差或技术限制而有所偏差，并且本领域技术人员根据实际所需可以另外设计具有不同形状、大小、相对位置的区域/层。
实施例一
参考图1及图2，本发明所述的基于扩散模型及3D Transformer的三维分割方法包括以下步骤：
1）构建基于扩散模型及3D Transformer的三维分割网络模型；
步骤1）的具体过程为：
本发明将Transformer结构及扩散模型加入到现有的三维分割网络模型中，以形成基于扩散模型及3D Transformer的三维分割网络模型，具体为：
所述Tansformer结构改进于Next-ViT结构，遵循分层金字塔网络结构，在每一个阶段都配置一个patch 嵌入层及一系列卷积或 Transformer块，将其三维空间分辨率逐步降低为原来的 1/32，然后将通道维度按不同阶段进行扩展。
本发明中通过堆叠3D NCB模块及3D NTB模块构建新型CNN-Transformer 混合架构，需要说明的是，本发明将NCB模块及NTB模块进行重组，并将NCB模块及NTB模块改为三维模块，即3D NCB模块及3D NTB模块，以便模拟脊柱CT序列影像数据中的短期和长期依赖关系，其中，3D NTB模块中还能够进行局部和全局信息的融合，进一步提高建模能力。
需要说明的是，所述3D NCB模块通过3D MHCA模块及3D MLP 模块构建而成，所述3D NCB模块能够在保证BottleNeck 块的部署优势的同时，获得作为 Transformer 块的突出性能，3D NCB模块遵循Transformer的一般架构。与此同时，本发明采用多头卷积注意力（MHCA）作为具有部署友好卷积操作的高效token 混合器。
进一步，3D MHCA模块：为解决现有基于注意力的token混合器具有高延迟的问题，本发明采用新的具有高效卷积操作的注意力机制，即3D卷积注意力(3D CA)，以实现快速推理速度。同时，受 MHSA 中有效多头设计的启发，本发明采用多头范式构建卷积注意力机制，该范式联合关注来自不同位置的不同表示子空间的信息，以实现有效的局部表示学习，所述多头卷积注意力(MHCA)的表示为：
（1）
参考图1，在工作时，MHCA从n个并行表示子空间中捕获信息，为促进多个头部之间的信息交互，MHCA配备有一个投影层W P， x=[ x 1,x 2,…,x n ]表示将输入特征 x在通道维度上划分为多头形式，3D CA作为一种单头卷积注意力机制可以表示为：
（2）
其中， 及 为输入特征 x中相邻的Token， 代表具有可训练参数 w与输入Token之间的内积运算，3D CA可以通过迭代优化可训练参数 w来学习局部感受野中不同Token 之间的关联性。
进一步，3D NTB模块：所述3D NCB模块可以有效地学习影像中的局部表示，但影像的全局信息的捕获同样需要解决。Transformer 模块具有较强的捕获低频信号的能力，从而提供全局信息，例如，影像中特征形状和结构。但是Transformer 模块可能会在一定程度上恶化高频信息，例如，局部纹理信息。不同频率段的信号在视觉系统中是必不可少的，并将以特定的方式进行融合，以提取更基本和明显的特征。因此，本发明将Next Transformer Block(NTB)模块改为3D NTB模块，以捕获轻量级机制中的多频信号。此外，所述3D NTB模块作为多频信号混频器，可以进一步提升整体建模能力，在工作时，所述3D NTB模块采用高效的多头自注意力机制(E-MHSA)捕获低频信号，所述多头自注意力机制表示为：
（3）
其中， x=[ x 1,x 2,…,x n ]表示将输入的特征值 x在通道维度上划分为多头形式，SA作为一种空间缩减自注意力算子表示为：
（4）
其中， 表示为一种具有Stride的平均池化操作，主要用于在注意力操作之前对空间维度进行降采样，从而减少计算成本。
进一步，New Inception模块：本发明在三维分割网络模型中加入Inception结构的三维反卷积形式(New Inception)来作为三维分割网络模型的解码器，所述Inception结构由一个核大小为1×1×1的三维point-wise卷积及两个核大小为3×3×3和5×5×5的三维反卷积组成。所述反卷积模块的三维point-wise卷积层能够在相同大小的生成场上叠加更多的反卷积，从而获得待分割图像中更多的细节，以便进行更为细化的分割。另外，三维point-wise卷积层在降维中也起着至关重要的作用。进行三维反卷积运算将产生大量的计算量，通过加入三维point-wise卷积层可以有效减少输入特征的数量，从而提高计算效率。New Inception模块由两个分支网络组成，每个分支网络使用不同大小的滤波器进行反卷积，其中，所述分支网络可以产生不同规模的信息。
New Inception模块使用的原理是将稀疏矩阵分解成密集矩阵进行计算，将特征维数分解为多个密集分布的子特征集，其中，高度相关的特征聚集在一起，不相关的特征则被削弱，最后将各子特征在特征维度上进行拼接，并与输入维度保持一致，以降低计算成本，保证最终的训练结果不受影响。
进一步，本发明将扩散模型加入到所述三维分割网络模型中，用于进行语义分割，其中，所述扩散模型在工作时，通过对语义标签添加随机噪声，并学习语义标签上的随机噪声分量，然后对语义标签进行去噪，得到相关的语义信息，从而使得三维分割网络模型能够更好地分割出病灶区域。
需要说明的是，参考图3，所述扩散模型的灵感主要来源于非平衡热力学，设定一个扩散步骤的马尔可夫链，通过逐步向真实数据中添加随机噪声（前向过程），然后学习反向扩散过程（逆扩散过程），从噪声中得到所需的数据样本结果。
所述扩散模型分为前向过程以及逆向过程，其中，前向过程不包含可学习参数，其值随着t不断增大，最终值的分布为各向独立的高斯分布。例如，给出真实数据的分布为 ，在前向过程中通过逐步加入一个小的高斯噪声，一共加入 T步，从而产生一系列加入有噪声的样本 ，加入噪声的均值和方差由 决定，其值在(0,1)范围内，而且给定条件 ，可以看出，向扩散过程中所添加的噪声量越来越大，因此前向过程的定义为：
（5）
扩散过程中的逆向过程是从随机高斯噪声中恢复出原始数据，由于正向过程中每次所添加的噪声值很小，所以可以将 作为一个高斯分布，其过程可以使用深度神经网络进行拟合。因此，该逆向过程也可以是一个马尔科夫链过程，因此，所述逆向过程表示为：
（6）
2）获取待分割图像，将所述待分割图像输入到所述基于扩散模型及3D Transformer的三维分割网络模型，得到待分割图像的脊柱三维分割结果。
仿真实验
本实验分别构建当前主流的基于卷积神经网络的分割网络以及本发明，对脊椎图像进行分割，分割结果如图4a、图4b、图4c及图4d所示，由图4a、图4b、图4c及图4d可知，本发明构建的分割网络模型所得到的分割结果更为精确。
需要说明的是，本发明不仅可以用于脊柱三维分割，作为一种通用型算法，本发明还可以用于其他医学影像分割任务，如乳腺癌分割，肝脏分割等。
实施例二
参考图5，本发明所述的基于扩散模型及3D Transformer的三维分割系统，包括：
获取模块1，用于获取待分割图像；
分割模块2，用于将所述待分割图像输入到基于扩散模型及3D Transformer的三维分割网络模型中，得到待分割图像的三维分割结果。
实施例三
一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现所述基于扩散模型及3D Transformer的三维分割方法的步骤，其中，所述存储器可能包含内存，例如高速随机存储器，也可能还包括非易失性存储器，例如，至少一个磁盘存储器等；处理器、网络接口、存储器通过内部总线互相连接，该内部总线可以是工业标准体系结构总线、外设部件互连标准总线、扩展工业标准结构总线等，总线可以分为地址总线、数据总线、控制总线等。存储器用于存放程序，具体地，程序可以包括程序代码、所述程序代码包括计算机操作指令。存储器可以包括内存和非易失性存储器，并向处理器提供指令和数据。
实施例四
一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现所述基于扩散模型及3D Transformer的三维分割方法的步骤，具体地，所述计算机可读存储介质包括但不限于例如易失性存储器和/或非易失性存储器。所述易失性存储器可以包括随机存储存储器（RAM）和/或高速缓冲存储器（cache）等。所述非易失性存储器可以包括只读存储器（ROM）、硬盘、闪存、光盘、磁盘等。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质（包括但不限于磁盘存储器、CD-ROM、光学存储器等）上实施的计算机程序产品的形式。
本申请是参照根据本申请实施例的方法、设备（系统）、和计算机程序产品的流程图和／或方框图来描述的。应理解可由计算机程序指令实现流程图和／或方框图中的每一流程和／或方框、以及流程图和／或方框图中的流程和／或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和／或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和／或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和／或方框图一个方框或多个方框中指定的功能的步骤。
最后应当说明的是：以上实施例仅用以说明本发明的技术方案而非对其限制，尽管参照上述实施例对本发明进行了详细的说明，所属领域的普通技术人员应当理解：依然可以对本发明的具体实施方式进行修改或者等同替换，而未脱离本发明精神和范围的任何修改或者等同替换，其均应涵盖在本发明的权利要求保护范围之内。

Claims (10)

  1. 一种基于扩散模型及3D Transformer的三维分割方法，其特征在于，包括：
    获取待分割图像；
    将所述待分割图像输入到基于扩散模型及3D Transformer的三维分割网络模型中，得到待分割图像的三维分割结果。
  2. 根据权利要求1所述的基于扩散模型及3D Transformer的三维分割方法，其特征在于，所述基于扩散模型及3D Transformer的三维分割网络模型由CNN-Transformer 混合架构、解码器及扩散模型构成。
  3. 根据权利要求2所述的基于扩散模型及3D Transformer的三维分割方法，其特征在于，所述CNN-Transformer 混合架构由3D NCB模块及3D NTB模块堆叠而成，其中，通过所述3D NCB模块学习待分割图像中的局部表示；通过所述3D NTB模块捕获轻量级机制中的多频信号。
  4. 根据权利要求3所述的基于扩散模型及3D Transformer的三维分割方法，其特征在于，所述3D NCB模块通过3D MHCA模块及3D MLP 模块构建而成。
  5. 根据权利要求2所述的基于扩散模型及3D Transformer的三维分割方法，其特征在于，所述解码器基于Inception结构的三维反卷积形式构建而成。
  6. 根据权利要求5所述的基于扩散模型及3D Transformer的三维分割方法，其特征在于，所述Inception结构由一个核大小为1×1×1的三维point-wise卷积层及两个核大小为3×3×3和5×5×5的三维反卷积层组成。
  7. 根据权利要求2所述的基于扩散模型及3D Transformer的三维分割方法，其特征在于，所述扩散模型通过对语义标签添加随机噪声，并学习语义标签上的随机噪声分量，再对所述语义标签进行去噪，得到相关的语义信息。
  8. 一种基于扩散模型及3D Transformer的三维分割系统，其特征在于，包括：
    获取模块（1），用于获取待分割图像；
    分割模块（2），用于将所述待分割图像输入到基于扩散模型及3D Transformer的三维分割网络模型中，得到待分割图像的三维分割结果。
  9. 一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如权利要求1-7任一项所述基于扩散模型及3D Transformer的三维分割方法的步骤。
  10. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1-7任一项所述基于扩散模型及3D Transformer的三维分割方法的步骤。
PCT/CN2023/091750 2023-04-20 2023-04-28 基于扩散模型及3d transformer的三维分割方法、系统、设备及介质 Ceased WO2024216672A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202310465972.7A CN116402987A (zh) 2023-04-20 2023-04-20 基于扩散模型及3DTransformer的三维分割方法、系统、设备及介质
CN202310465972.7 2023-04-20

Publications (1)

Publication Number Publication Date
WO2024216672A1 true WO2024216672A1 (zh) 2024-10-24

Family

ID=87014194

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/091750 Ceased WO2024216672A1 (zh) 2023-04-20 2023-04-28 基于扩散模型及3d transformer的三维分割方法、系统、设备及介质

Country Status (2)

Country Link
CN (1) CN116402987A (zh)
WO (1) WO2024216672A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119027596A (zh) * 2024-10-28 2024-11-26 苏州共营互联网科技有限公司 一种基于物联网的客户数据智能监管系统及方法
CN119273704A (zh) * 2024-12-10 2025-01-07 安徽大学 一种显著性参考和解耦的食管肿瘤ct影像扩散分割方法
CN119722576A (zh) * 2024-11-14 2025-03-28 武汉大学 一种基于三维Transformer的核磁影像阿兹海默症诊断方法和系统
CN120198324A (zh) * 2025-05-26 2025-06-24 中国科学院深圳先进技术研究院 基于条件扩散模型与电阻抗的肌肉成像方法、介质和设备

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117078539A (zh) * 2023-07-20 2023-11-17 北京交通大学 基于CNN-Transformer的局部全局交互式图像恢复方法
CN117036696B (zh) * 2023-07-21 2024-11-12 清华大学深圳国际研究生院 图像分割方法、装置、设备和存储介质
CN116758098A (zh) * 2023-08-07 2023-09-15 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) 磁共振图像的丘脑底核分割方法及模型构建方法
CN117314733A (zh) * 2023-08-14 2023-12-29 清华大学深圳国际研究生院 基于扩散模型的视频填充方法、装置、设备及存储介质
CN117689533B (zh) * 2023-12-19 2024-12-27 南京大学 基于稳定扩散模型将背部照片转换成脊柱x光影像的方法
CN118196121B (zh) * 2024-04-08 2024-09-20 兰州交通大学 一种基于去噪扩散概率模型的乳腺超声图像分割方法
CN118570467A (zh) * 2024-05-17 2024-08-30 中国科学院深圳先进技术研究院 一种辅助脊椎置钉手术的临床脊椎图像分割方法及装置
CN119180954B (zh) * 2024-09-02 2025-05-06 重庆邮电大学 一种基于扩散模型的颈椎椎骨分割及关键点检测方法

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115018809A (zh) * 2022-06-28 2022-09-06 华中科技大学 Ct图像的靶区分割识别方法及系统
CN115239637A (zh) * 2022-06-28 2022-10-25 中国科学院深圳先进技术研究院 一种ct胰腺肿瘤自动分割方法、系统、终端以及存储介质
CN115375698A (zh) * 2022-08-22 2022-11-22 武汉理工大学重庆研究院 医学图像分割方法、装置、电子设备及存储介质
WO2022242131A1 (zh) * 2021-05-21 2022-11-24 中国科学院深圳先进技术研究院 图像分割方法、装置、设备及存储介质
CN115409733A (zh) * 2022-09-02 2022-11-29 山东财经大学 一种基于图像增强和扩散模型的低剂量ct图像降噪方法
CN115457009A (zh) * 2022-09-24 2022-12-09 福建农林大学 基于Transformer和卷积神经网络的三维医学图像分割方法
CN115619810A (zh) * 2022-12-19 2023-01-17 中国医学科学院北京协和医院 一种前列腺分区分割方法、系统及设备
CN115880312A (zh) * 2022-11-18 2023-03-31 重庆邮电大学 一种三维图像自动分割方法、系统、设备和介质
WO2023059699A1 (en) * 2021-10-05 2023-04-13 Google Llc Vector-quantized image modeling

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022242131A1 (zh) * 2021-05-21 2022-11-24 中国科学院深圳先进技术研究院 图像分割方法、装置、设备及存储介质
WO2023059699A1 (en) * 2021-10-05 2023-04-13 Google Llc Vector-quantized image modeling
CN115018809A (zh) * 2022-06-28 2022-09-06 华中科技大学 Ct图像的靶区分割识别方法及系统
CN115239637A (zh) * 2022-06-28 2022-10-25 中国科学院深圳先进技术研究院 一种ct胰腺肿瘤自动分割方法、系统、终端以及存储介质
CN115375698A (zh) * 2022-08-22 2022-11-22 武汉理工大学重庆研究院 医学图像分割方法、装置、电子设备及存储介质
CN115409733A (zh) * 2022-09-02 2022-11-29 山东财经大学 一种基于图像增强和扩散模型的低剂量ct图像降噪方法
CN115457009A (zh) * 2022-09-24 2022-12-09 福建农林大学 基于Transformer和卷积神经网络的三维医学图像分割方法
CN115880312A (zh) * 2022-11-18 2023-03-31 重庆邮电大学 一种三维图像自动分割方法、系统、设备和介质
CN115619810A (zh) * 2022-12-19 2023-01-17 中国医学科学院北京协和医院 一种前列腺分区分割方法、系统及设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
JIASHI LI; XIN XIA; WEI LI; HUIXIA LI; XING WANG; XUEFENG XIAO; RUI WANG; MIN ZHENG; XIN PAN: "Next-ViT: Next Generation Vision Transformer for Efficient Deployment in Realistic Industrial Scenarios", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 8 August 2022 (2022-08-08), 201 Olin Library Cornell University Ithaca, NY 14853, XP091289536 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119027596A (zh) * 2024-10-28 2024-11-26 苏州共营互联网科技有限公司 一种基于物联网的客户数据智能监管系统及方法
CN119722576A (zh) * 2024-11-14 2025-03-28 武汉大学 一种基于三维Transformer的核磁影像阿兹海默症诊断方法和系统
CN119273704A (zh) * 2024-12-10 2025-01-07 安徽大学 一种显著性参考和解耦的食管肿瘤ct影像扩散分割方法
CN120198324A (zh) * 2025-05-26 2025-06-24 中国科学院深圳先进技术研究院 基于条件扩散模型与电阻抗的肌肉成像方法、介质和设备

Also Published As

Publication number Publication date
CN116402987A (zh) 2023-07-07

Similar Documents

Publication Publication Date Title
WO2024216672A1 (zh) 基于扩散模型及3d transformer的三维分割方法、系统、设备及介质
Wang et al. Automatic brain tumor segmentation based on cascaded convolutional neural networks with uncertainty estimation
CN115760874B (zh) 一种基于联合空间域的多尺度U-Net医学图像分割方法
Chen et al. Generative adversarial U-Net for domain-free medical image augmentation
CN118967706B (zh) 一种基于医学高光谱图像分割网络的图像分割方法
Rastogi et al. Deep learning and big datatechnologies in medical image analysis
CN117670703A (zh) 一种基于多任务联合训练的磁共振图像恢复方法
Amin et al. Detection of brain space-occupying lesions using quantum machine learning
Zakaria et al. Deep learning in medical imaging: A review
Saxena et al. An intelligent system for segmenting an abdominal image in multi core architecture
Li et al. Plantformer: plant point cloud completion based on local–global feature aggregation and spatial context-aware transformer
Hu et al. HAB‐Net: Hierarchical asymmetric convolution and boundary enhancement network for brain tumor segmentation
Chen et al. Hybrid‐attention densely connected U‐Net with GAP for extracting livers from CT volumes
Xiao et al. A novel semi-supervised domain adaptive method for cross-modality medical image segmentation
Asma-Ull et al. Data efficient segmentation of various 3d medical images using guided generative adversarial networks
Xiang et al. Turbopixel segmentation using eigen-images
Wiehman et al. Unsupervised pre-training for fully convolutional neural networks
Yang et al. T-CADiff: Conditional guidance based diffusion model for medical image segmentation
Gammoudi et al. 3D Brain Tumor Segmentation Using Modified U-Net Architecture
Sobirov et al. Super Images-A New 2D Perspective on 3D Medical Imaging Analysis
CN110706209A (zh) 一种网格网络的大脑磁共振图像肿瘤定位方法
Hou et al. InstantTrace: fast parallel neuron tracing on GPUs
Wang et al. MGMFormer: Multi‐Scale Attentional Medical Image Segmentation Network for Semantic Feature Enhancement
Zhang et al. Multi-task learning with hierarchical guidance for locating and stratifying submucosal tumors
Ramzan et al. CLAIM: Clinically-guided LGE augmentation for realistic and diverse myocardial scar synthesis and segmentation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23933548

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE