WO2025199682A1 - 三维图像拼接方法、装置、计算机设备及存储介质 - Google Patents

三维图像拼接方法、装置、计算机设备及存储介质

Info

Publication number
WO2025199682A1
WO2025199682A1 PCT/CN2024/083577 CN2024083577W WO2025199682A1 WO 2025199682 A1 WO2025199682 A1 WO 2025199682A1 CN 2024083577 W CN2024083577 W CN 2024083577W WO 2025199682 A1 WO2025199682 A1 WO 2025199682A1
Authority
WO
WIPO (PCT)
Prior art keywords
deep learning
learning model
dimensional
target
dimensional 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.)
Pending
Application number
PCT/CN2024/083577
Other languages
English (en)
French (fr)
Inventor
范晓苇
徐放
肖彦洋
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
Priority to PCT/CN2024/083577 priority Critical patent/WO2025199682A1/zh
Publication of WO2025199682A1 publication Critical patent/WO2025199682A1/zh
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images

Definitions

  • the present invention relates to the field of image stitching technology, and in particular to a three-dimensional image stitching method, device, computer equipment and storage medium.
  • High-resolution imaging of large biological specimens is a critical requirement for neuroscience and biological research.
  • imaging larger specimens such as mouse and macaque brains
  • a common slicing strategy is to scan and image a layer first, then slice it off.
  • overlapping imaging can be achieved by slicing the sample slightly thinner than the imaging thickness. The overlapping portion provides important reference information for subsequent registration and stitching.
  • VISoR Synchronized 3D Microscopy
  • a three-dimensional image stitching method comprising:
  • the training data comprising three-dimensional images of each layer of a biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images;
  • a three-dimensional image stitching device comprising:
  • an acquisition module configured to acquire training data, wherein the training data includes three-dimensional images of each layer of a biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images;
  • a prediction module configured to predict a deformation field based on each of the three-dimensional images and an initial deep learning model to obtain a deformation field, wherein the deep learning model is a convolutional neural network;
  • a stitching module configured to stitch the three-dimensional images using the deformation field and the spatial transformation network to obtain a first target three-dimensional image
  • a training module is used to adjust the parameters and weights in the initial deep learning model based on the first target three-dimensional image, the three-dimensional images of each layer, the matching points and the target loss function.
  • a computer device comprising 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 above-mentioned three-dimensional image stitching method when executing the computer program.
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned three-dimensional image stitching method are implemented.
  • FIG1 is a diagram illustrating an application environment of a three-dimensional image stitching method according to an embodiment
  • FIG2 is a flow chart of a three-dimensional image stitching method according to one embodiment
  • FIG3 is a structural block diagram of a three-dimensional image stitching device according to an embodiment
  • FIG4 is a block diagram of a computer device according to an embodiment
  • FIG5 is a structural block diagram of a computer device in another embodiment.
  • references herein to "embodiments” mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application.
  • the appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.
  • the three-dimensional image stitching method provided by the embodiment of the present invention can be applied in an application environment as shown in FIG. 1 , wherein a client 110 communicates with a server 120 via a network.
  • the server 120 can receive training data through the client 110, and the training data includes three-dimensional images of each layer of the biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images, and then the server 120 predicts the deformation field based on each of the three-dimensional images and the initial deep learning model to obtain the deformation field, wherein the deep learning model is a convolutional neural network, so that the server 120 splices the three-dimensional images through the deformation field and the spatial transformation network to obtain a first target three-dimensional image.
  • the server 120 adjusts the parameters and weights in the initial deep learning model based on the first target three-dimensional image, the three-dimensional images of each layer, the matching points and the target loss function, and can train the initial deep learning model through the matching points of the three-dimensional images of each layer of the biological sample, the spatial transformation network, and the first target three-dimensional image to obtain a trained deep learning model, and then use the deep learning model to predict the target deformation field of the three-dimensional image to be spliced, and use the target deformation field to achieve efficient and accurate splicing of the three-dimensional images.
  • the client 110 may include, but is not limited to, various personal computers, laptops, smartphones, tablet computers, and portable wearable devices.
  • the server 120 may be implemented as an independent server or a server cluster consisting of multiple servers. The present invention will be described in detail below through specific embodiments.
  • FIG. 2 is a flow chart of a three-dimensional image stitching method according to an embodiment of the present invention, including the following steps:
  • Step S101 Acquire training data, wherein the training data includes three-dimensional images of each layer of a biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images;
  • the biological sample may be a mouse brain, a macaque brain, etc.
  • the biological sample is cut into several slices and imaged layer by layer to obtain a three-dimensional image of each layer of the biological sample.
  • image registration refers to aligning the pixels of two (or more images) to obtain a spatially consistent image.
  • it is divided into two-dimensional image and three-dimensional image registration according to the different image dimensions.
  • existing methods usually aim to make the pixel distribution of the two images as consistent as possible in two-dimensional space.
  • adjacent two-dimensional surfaces with no overlapping gaps.
  • their optimal registration target should not be completely consistent in two-dimensional distribution, but should be relatively similar in two dimensions and have a certain continuity in three dimensions.
  • the reference image and the image to be registered are two different images, and they need to be aligned for subsequent image processing and analysis.
  • the reference image is usually regarded as a standard or original image, and the image to be registered is the image that needs to be aligned. It is related to The reference image may be deformed, rotated, translated, scaled, and so on. Therefore, when performing image registration, the image to be registered needs to be transformed into the coordinate system of the reference image so that the two images can match.
  • Traditional registration methods include feature point-based methods and global optimization-based methods.
  • Step S102 Predicting a deformation field based on each of the three-dimensional images and an initial deep learning model to obtain a deformation field, wherein the deep learning model is a convolutional neural network;
  • the 3D image is preprocessed and data augmented; the processed 3D image is input into the initial deep learning model for deformation field prediction, thereby obtaining a deformation field.
  • automatic contrast adjustment and grayscale equalization are performed on the 3D image to enhance image detail and achieve a more uniform brightness distribution.
  • the original 16-bit depth 3D image is reduced to 8-bit depth to reduce data volume and computational burden while preserving key image information.
  • advanced image processing tools such as ants or elastix are used for preliminary affine pre-registration, converting the image to be registered to the spatial coordinate system of the standard image. During this process, the coordinates of the annotation points in the image to be registered are also transformed to ensure that the annotation information remains accurate after the transformation.
  • the 3D image is projected in the Z direction, compressing the complex 3D information into a 2D image.
  • this process results in a certain degree of 3D information loss, this can be partially compensated by adding projections of the image in two other directions (e.g., X and Y).
  • the choice of projection method will be based on the specific brightness characteristics of the image data. For example, maximum projection is used to highlight bright nerve fiber signals, minimum projection is used to extract darker blood vessel signals, and standard deviation projection can effectively emphasize these two different structural features.
  • a series of data augmentation techniques can be used.
  • the training data set is enriched by random displacement and random deformation processing.
  • Random displacement can effectively simulate the possible spatial offset between different layers of images, while random deformation is used to imitate the morphological changes that may occur in the sample during the actual imaging process. These treatments not only improve the model's adaptability to various possible deformations, but also enhance its accuracy and reliability in practical applications. While performing these transformations, special attention is paid to the annotation points of the three-dimensional image. In order to maintain data consistency and annotation accuracy, any form of displacement and deformation processing performed on the three-dimensional image will be applied accordingly to its annotation points. This step ensures that the labeling The annotation information remains unchanged throughout the data augmentation process, thus providing accurate supervision signals for subsequent deep learning training.
  • the preprocessed and data augmented 3D images can be divided into training and test sets, which can then be further divided into training and validation sets.
  • the training set is used to train the model
  • the validation set is used to debug hyperparameters during training
  • the test set is used to test the final model registration results.
  • Step S103 splicing the three-dimensional images using the deformation field and the spatial transformation network to obtain a first target three-dimensional image
  • Each pixel value in the deformation field indicates the distance the corresponding pixel has moved along the x and y directions.
  • the structure of the initial deep learning model can be adjusted appropriately based on the preprocessed image morphology. For example, a U-net network is a good choice for the initial deep learning model, but modifications may be required for the following different situations:
  • the initial deep learning model input is a 3D image
  • the initial deep learning model should be able to generate a 2D deformation field.
  • the initial deep learning model input is a 2D projection image
  • the initial deep learning model should also generate a 2D deformation field.
  • the spatial transformation network can apply the predicted deformation field to the three-dimensional image to generate a registered image as the first target three-dimensional image.
  • Step S104 Based on the first target three-dimensional image, the three-dimensional images of each layer, the matching points, and the target loss function, the parameters and weights in the initial deep learning model are adjusted.
  • the parameters and weights in the initial deep learning model are adjusted based on the first target three-dimensional image, the three-dimensional images of each layer, and the target loss function, wherein the target loss function includes matching loss, regularization loss, and semi-supervised loss function; it is determined whether the initial deep learning model has reached the training end condition, and if so, the initial deep learning model that has reached the training end condition is used as the trained deep learning model.
  • the matching loss is for the matching degree of the supervised information
  • the regularization loss is used to control the smoothness of the deformation field.
  • the unsupervised loss function is combined to take into account the two-dimensional similarity and three-dimensional continuity of the image.
  • a target loss function can be calculated through the first target 3D image, the 3D images of each layer, and the matching points, so as to adjust the parameters and weights of the initial deep learning model using the target loss function.
  • the initial deep learning model adopts a U-net network.
  • the following steps are included:
  • Step S201 obtaining the three-dimensional images to be stitched of each layer
  • Step S202 Based on the three-dimensional images to be stitched and the trained deep learning model, a deformation field prediction is performed on each layer of the three-dimensional images to be stitched to obtain a target deformation field, and the target deformation field is interpolated into the three-dimensional space;
  • Step S203 based on the interpolated target deformation field, stitching the to-be-stitched three-dimensional images of each layer to obtain a second target three-dimensional image.
  • a trained deep learning model is used to predict the deformation field of each layer of the three-dimensional image to be stitched.
  • This prediction process involves the model analyzing each layer of the three-dimensional image to be stitched to calculate the two-dimensional deformation field that needs to be applied.
  • This deformation field represents the specific movement that the image pixels need to make in the two-dimensional space.
  • the predicted two-dimensional deformation field needs to be interpolated into the three-dimensional space. This step is critical because it extends the two-dimensional deformation field to the entire three-dimensional image volume, thereby providing the necessary spatial transformation information for the subsequent stitching process.
  • the interpolated deformation field is applied to the three-dimensional image to be stitched to obtain a complete and high-precision three-dimensional image as the second target three-dimensional image.
  • accurate alignment and seamless stitching of each layer of images can be achieved, thereby ultimately constructing a complete and high-precision three-dimensional image.
  • the three-dimensional image stitching method proposed in this embodiment obtains training data, wherein the training data includes three-dimensional images of each layer of a biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images, and then deformation field prediction is performed based on each of the three-dimensional images and an initial deep learning model to obtain a deformation field, wherein the deep learning model is a convolutional neural network, thereby
  • the three-dimensional images are spliced through the deformation field and the spatial transformation network to obtain a first target three-dimensional image.
  • the three-dimensional images of each layer, the matching points and the target loss function, the parameters and weights in the initial deep learning model are adjusted.
  • the initial deep learning model can be trained through the matching points of the three-dimensional images of each layer of the biological sample, the spatial transformation network, and the first target three-dimensional image to obtain a trained deep learning model.
  • the deep learning model can then be used to predict the target deformation field of the three-dimensional image to be spliced, and the target deformation field can be used to achieve efficient and accurate splicing of the three-dimensional images to be spliced.
  • a three-dimensional image stitching device comprising: an acquisition module 10 for acquiring training data, the training data comprising three-dimensional images of various layers of a biological sample, wherein matching points are annotated between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images;
  • a prediction module 20 is configured to predict a deformation field based on each of the three-dimensional images and an initial deep learning model to obtain a deformation field, wherein the deep learning model is a convolutional neural network;
  • a stitching module 30 is configured to stitch the three-dimensional images using the deformation field and the spatial transformation network to obtain a first target three-dimensional image
  • the training module 40 is used to adjust the parameters and weights in the initial deep learning model based on the first target three-dimensional image, the three-dimensional images of each layer, the matching points and the target loss function.
  • the training module 40 is configured to adjust parameters and weights in the initial deep learning model based on the first target three-dimensional image, the three-dimensional images of each layer, and a target loss function, wherein the target loss function includes a matching loss, a regularization loss, and a semi-supervised loss function;
  • the prediction module 20 is used to perform preprocessing and data enhancement processing on the three-dimensional image; the processed three-dimensional image is input into the initial deep learning model to perform deformation field prediction to obtain the deformation field.
  • the initial deep learning model adopts a U-net network.
  • the stitching module 30 is used to: obtain the three-dimensional images to be stitched at each layer; perform deformation field prediction on the three-dimensional images to be stitched at each layer based on the three-dimensional images to be stitched and a trained deep learning model to obtain a target deformation field, and interpolate the target deformation field into the three-dimensional space; and stitch the three-dimensional images to be stitched at each layer based on the interpolated target deformation field to obtain a second target three-dimensional image.
  • a computer device which may be a server, and its internal structure diagram may be as shown in FIG4 .
  • the computer device includes a processor, a memory, a network interface, and a database connected via a system bus.
  • the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile and/or volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, a computer program, and a database.
  • the internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium.
  • the network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a three-dimensional image stitching method server side.
  • a computer device which may be a client, and its internal structure diagram may be as shown in FIG5 .
  • the computer device includes a processor, a memory, a network interface, a display screen, and an input device connected via a system bus.
  • the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system and a computer program.
  • the internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium.
  • the network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps on the client side of a three-dimensional image stitching method.
  • a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computing When programming the machine, perform the following steps:
  • the training data comprising three-dimensional images of each layer of a biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images;
  • the three-dimensional images of each layer, the matching points and the target loss function, the parameters and weights in the initial deep learning model are adjusted.
  • the three-dimensional image stitching method proposed in this embodiment obtains training data, wherein the training data includes three-dimensional images of each layer of a biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images, and then a deformation field is predicted based on each of the three-dimensional images and an initial deep learning model to obtain a deformation field, wherein the deep learning model is a convolutional neural network, and the three-dimensional images are stitched together using the deformation field and a spatial transformation network to obtain a first target three-dimensional image. Finally, based on the first target three-dimensional image, the three-dimensional images of each layer, the matching points, and a target loss function, the parameters and weights in the initial deep learning model are adjusted.
  • the three-dimensional image stitching method proposed in this embodiment obtains training data, wherein the training data includes three-dimensional images of each layer of a biological sample, wherein matching points are marked between adjacent two-dimensional surfaces of two adjacent layers of the three-dimensional images, and then a deformation field is predicted based on each of the three-dimensional images and an initial deep learning model to obtain a deformation field, wherein the deep learning model is a convolutional neural network, and the three-dimensional images are stitched together using the deformation field and a spatial transformation network to obtain a first target three-dimensional image. Finally, based on the first target three-dimensional image, the three-dimensional images of each layer, the matching points, and a target loss function, the parameters and weights in the initial deep learning model are adjusted.
  • Non-volatile Volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM random access memory
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDR SDRAM double data rate SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM synchronous link DRAM
  • RDRAM RAMbus direct RAM
  • DRAM direct RAMbus dynamic RAM
  • RDRAM RAMbus dynamic RAM

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

本申请涉及图像拼接技术领域，揭示了一种三维图像拼接方法、装置、计算机设备及存储介质，方法包括：获取生物样本的各层三维图像，接着基于各个三维图像以及初始深度学习模型中进行形变场预测，得到形变场；通过形变场、空间变换网络，对三维图像进行拼接，得到第一目标三维图像；基于第一目标三维图像、各层三维图像、匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整，能够通过生物样本的三维图像的匹配点、空间变换网络、第一目标三维图像，对初始深度学习模型进行模型训练，得到训练好的深度学习模型，而后利用深度学习模型可以预测得到待拼接三维图像的目标形变场，利用该目标形变场可以实现对待拼接三维图像高效准确的拼接。

Description

三维图像拼接方法、装置、计算机设备及存储介质 技术领域
本发明涉及图像拼接技术领域，尤其涉及一种三维图像拼接方法、装置、计算机设备及存储介质。
背景技术
对大型生物样本的高分辨率成像是神经科学和生物学研究的重要需求。然而，成像较大体积的样本(如小鼠脑、猕猴脑等等)时，由于当前成像技术和设备的限制，往往需要将样本切分成若干片，逐层成像。常见的一种切片策略是：先扫描、成像一层，再切掉一层。在此种策略下，只要切片时注意切掉部分比成像厚度稍微薄一些，则可实现有重叠的成像。其中的重叠部分就为后续配准和拼接提供了重要参考信息。
而为了成倍提高成像速度，降低较长的成像时间对研究进展的桎梏，有的成像技术(如VISoR(同步飞扫三维显微成像))采取了另一项策略：预先将样品切分成等厚的若干片，同步扫描。此时，成像出来的不同片样品之间必然是无重叠的，这就为后续的三维重构提出了一个新的挑战：无重叠信息的三维图像的高精度拼接，即如何将成像出来的无重叠信息的各层三维图像拼接成一个完整的三维图像。
发明内容
基于此，有必要针对现有技术的生物样本的各层三维图像拼接效果较差的技术问题，提出了一种三维图像拼接方法、装置、计算机设备及存储介质。
第一方面，提供了一种三维图像拼接方法，所述方法包括：
获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
第二方面，提供了一种三维图像拼接装置，所述装置包括：
获取模块，用于获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
预测模块，用于基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
拼接模块，用于通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
训练模块，用于基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
第三方面，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上述三维图像拼接方法的步骤。
第四方面，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上述三维图像拼接方法的步骤。
本发明提出的三维图像拼接方法，通过获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间 进行匹配点标注，接着基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络，从而通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像，最后基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整，能够通过生物样本的各层三维图像的匹配点、空间变换网络、第一目标三维图像，对初始深度学习模型进行模型训练，得到训练好的深度学习模型，而后利用深度学习模型可以预测得到待拼接三维图像的目标形变场，利用该目标形变场可以实现对待拼接三维图像高效准确的拼接。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
其中：
图1为一个实施例中三维图像拼接方法的应用环境图；
图2为一个实施例中三维图像拼接方法的流程图；
图3为一个实施例中三维图像拼接装置的结构框图；
图4为一个实施例中计算机设备的结构框图；
图5为另一个实施例中计算机设备的结构框图。
具体实施方式
除非另有定义，本文所使用的所有的技术和科学术语与属于本申请的技术领域的技术人员通常理解的含义相同；本文中在申请的说明书中所使用的术语只是 为了描述具体的实施例的目的，不是旨在于限制本申请；本申请的说明书和权利要求书及上述附图说明中的术语“包括”和“具有”以及它们的任何变形，意图在于覆盖不排他的包含。本申请的说明书和权利要求书或上述附图中的术语“第一”、“第二”等是用于区别不同对象，而不是用于描述特定顺序。
在本文中提及“实施例”意味着，结合实施例描述的特定特征、结构或特性可以包含在本申请的至少一个实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例，也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是，本文所描述的实施例可以与其它实施例相结合。
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，所描述的实施例是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
本发明实施例提供的三维图像拼接方法，可应用在如图1的应用环境中，其中，客户端110通过网络与服务端120进行通信。服务端120可以通过客户端110接收训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注，接着服务端120基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络，从而服务端120通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像，最后服务端120基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整，能够通过生物样本的各层三维图像的匹配点、空间变换网络、第一目标三维图像，对初始深度学习模型进行模型训练，得到训练好的深度学习模型，而后利用深度学习模型可以预测得到待拼接三维图像的目标形变场，利用该目标形变场可以实现对待拼接三维图像高效准确 的拼接。其中，客户端110可以但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备。服务端120可以用独立的服务器或者是多个服务器组成的服务器集群来实现。下面通过具体的实施例对本发明进行详细的描述。
请参阅图2所示，图2为本发明一实施例提供的三维图像拼接方法的一个流程示意图，包括如下步骤：
步骤S101：获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
其中，生物样本可以是小鼠脑、猕猴脑等，将生物样本切分成若干片，逐层成像，可以得到生物样本的各层三维图像。
在本实施例中，需要对生物样本相邻两层切片(三维图像)的相邻二维表面之间的匹配点进行标注，这些匹配点代表了三维结构(如血管、神经纤维等)在不同层面上的对应关系，匹配点将作为监督学习过程中的关键信息。这些匹配点的获取可以通过精细的人工标注实现，或者采用先进的算法自动提取。目的是为初始深度学习模型提供准确的地标数据，以指导其学习过程，确保配准和拼接的精度。
需要说明解释的是，图像配准是指将两幅(或多幅图像)的像素点进行对齐以获得空间一致的图像。通常根据图像维度的不同，分为二维图像和三维图像配准。以二维图像配准为例，现有方法通常意在使两幅图像的像素分布在二维空间内尽可能一致。但是对于无重叠信息的三维图像的高精度拼接的任务而言，需要配准的是存在无重叠缝隙的的相邻二维表面。这两个表面之间存在第三个未知的维度，换言之，它们最优配准目标，也不应是二维分布完全一致，而应该是二维上较为相似，同时在三维上具有一定连续性。在图像配准中，参考图像和待配准图像是两个不同的图像，需要将它们进行对齐，以便后续的图像处理和分析。参考图像通常被视为标准或原始图像，待配准图像则是需要进行对齐的图像，它与 参考图像可能存在形变、旋转、平移、缩放等不同的变换。因此，在进行图像配准时，需要将待配准图像变换到参考图像的坐标系中，以便两个图像能够匹配。传统的的配准方法包括基于特征点的方法和基于全局优化的方法。
步骤S102：基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
一实施例中，对所述三维图像进行预处理以及数据增强处理；将处理后的三维图像输入至所述初始深度学习模型中进行进行形变场预测，得到形变场。具体地，对三维图像实施自动对比度调整和灰度均衡化处理，旨在增强图像细节并实现更均匀的亮度分布。其次，将原始的16位深三维图像降低至8位深，以减少数据量和计算负担，同时保留关键图像信息。进一步，采用先进的图像处理工具如ants或elastix进行初步的仿射预配准，即将待配准图像转换至标准图像的空间坐标系。在此过程中，同样对待配准图像的标注点进行坐标变换，以确保标注信息在变换后仍保持其准确性。此外，为了降低训练资源的消耗并加速训练过程，对三维图像进行Z方向的投影处理，从而将复杂的三维信息压缩为二维图像。这一处理虽然会导致一定程度的三维信息丢失，但可以通过增加图像在其他两个方向(如X方向和Y方向)的投影来部分补偿。投影方法的选择将基于图像数据的特定亮度特征，例如，最大值投影用于突出高亮的神经纤维信号，最小值投影用于提取较暗的血管信号，而标准差投影则可以有效地强调这两种不同的结构特征。为了增强初始深度网络的泛化能力和鲁棒性，可以采用了一系列数据增强技术，特别地，通过随机位移和随机形变处理来丰富训练数据集。随机位移能有效模拟不同层图像之间可能存在的空间偏移，而随机形变则用于模仿样本在实际成像过程中可能出现的形态变化。这些处理不仅提升了模型对各种可能形变的适应能力，还增强了其在实际应用中的准确性和可靠性。在进行这些变换的同时，还特别注意到三维图像的标注点。为保持数据一致性和标注准确性，对三维图像进行的任何形式的位移和形变处理都会相应地应用于其标注点。这一步骤确保了标 注信息在整个数据增强过程中保持不变，从而为后续深度学习训练提供了准确的监督信号。
作为一示例，还可以将经过预处理以及数据增强处理后的三维图像进行划分，得到训练集和测试集，训练集中再进一步划分出训练集和验证集。将训练集用于训练模型，验证集用于调试训练过程中的超参，测试集用于最终模型配准效果的测试。
步骤S103：通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
其中，形变场的每个像素值指示了相应像素沿x和y方向的移动距离。针对预处理后的图像形态，初始深度学习模型的结构可进行适当调整。例如，初始深度学习模型采用U-net网络是一个很好的选择，但可能需要针对以下不同情况进行修改：当初始深度学习模型的输入为三维图像时，初始深度学习模型应能生成二维形变场。当初始深度学习模型的输入为二维投影图时，初始深度学习模型同样应生成二维形变场。
在本实施例中，将空间变换网络可以将预测的形变场应用于三维图像，以生成配准后的图像，作为第一目标三维图像。
步骤S104：基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
一实施例中，基于所述第一目标三维图像、各层所述三维图像以及目标损失函数，对所述初始深度学习模型中的参数和权重进行调整，其中，所述目标损失函数包括匹配损失、正则损失以及半监督损失函数；判断所述初始深度学习模型是否达到训练结束条件，若是，则将达到训练结束条件的所述初始深度学习模型作为训练好的深度学习模型。其中，匹配损失针对监督信息的匹配度，正则损失用于控制形变场的平滑性，在监督损失函数的基础上，结合无监督损失函数来兼顾图像的二维相似性和三维连续性。常见的无监督损失函数包括均方误差、结构 相似性指数(SSIM)等。在本实施例中，通过第一目标三维图像、各层所述三维图像、所述匹配点，可以计算出目标损失函数，从而采用目标损失函数对初始深度学习模型进行参数和权重调整。
一实施例中，所述初始深度学习模型采用U-net网络。
一实施例中，所述基于所述第一目标三维图像、各层所述三维图像以及目标损失函数，对初始深度学习模型中的参数和权重进行调整的步骤之后，包括：
步骤S201：获取各层的待拼接三维图像；
步骤S202：基于所述待拼接三维图像以及训练好的深度学习模型，对每一层的待拼接三维图像进行形变场预测，得到目标形变场，所述目标形变场被插值到三维空间中；
步骤S203：基于插值后的所述目标形变场，对各层的待拼接三维图像进行拼接，得到第二目标三维图像。
具体地，首先，使用训练好的深度学习模型对每一层的待拼接三维图像进行形变场的预测。这个预测过程涉及模型对每层待拼接三维图像的分析，以计算出需要应用的二维形变场，这个形变场表示了图像像素在二维空间内需要进行的具体移动。随后，预测得到的二维形变场需要被插值到三维空间中。这一步是关键的，因为它将二维形变场扩展到整个三维图像体积，从而为接下来的拼接过程提供了必要的空间转换信息。最后，将插值后的形变场应用于待拼接三维图像，从而得到完整且高精度的三维图像，作为第二目标三维图像。通过这种方式，可以实现各层图像的准确对齐和无缝拼接，从而最终构建出一个完整且高精度的三维图像。
本实施例提出的三维图像拼接方法，通过获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注，接着基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络，从而通 过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像，最后基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整，能够通过生物样本的各层三维图像的匹配点、空间变换网络、第一目标三维图像，对初始深度学习模型进行模型训练，得到训练好的深度学习模型，而后利用深度学习模型可以预测得到待拼接三维图像的目标形变场，利用该目标形变场可以实现对待拼接三维图像高效准确的拼接。
请参阅图3所示，在一实施例中，提供一种三维图像拼接装置，所述装置包括：获取模块10，用于获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
预测模块20，用于基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
拼接模块30，用于通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
训练模块40，用于基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
一实施例中，训练模块40，用于基于所述第一目标三维图像、各层所述三维图像以及目标损失函数，对所述初始深度学习模型中的参数和权重进行调整，其中，所述目标损失函数包括匹配损失、正则损失以及半监督损失函数；
判断所述初始深度学习模型是否达到训练结束条件，若是，则将达到训练结束条件的所述初始深度学习模型作为训练好的深度学习模型。
一实施例中，预测模块20，用于对所述三维图像进行预处理以及数据增强处理；将处理后的三维图像输入至所述初始深度学习模型中进行进行形变场预测，得到形变场。
一实施例中，所述初始深度学习模型采用U-net网络。
一实施例中，拼接模块30，用于：获取各层的待拼接三维图像；基于所述待拼接三维图像以及训练好的深度学习模型，对每一层的待拼接三维图像进行形变场预测，得到目标形变场，所述目标形变场被插值到三维空间中；基于插值后的所述目标形变场，对各层的待拼接三维图像进行拼接，得到第二目标三维图像。
在一个实施例中，提供了一种计算机设备，该计算机设备可以是服务端，其内部结构图可以如图4所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性和/或易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机程序和数据库。该内存储器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的网络接口用于与外部的客户端通过网络连接通信。该计算机程序被处理器执行时以实现一种三维图像拼接方法服务端侧的功能或步骤。
在一个实施例中，提供了一种计算机设备，该计算机设备可以是客户端，其内部结构图可以如图5所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口、显示屏和输入装置。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统和计算机程序。该内存储器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的网络接口用于与外部服务器通过网络连接通信。该计算机程序被处理器执行时以实现一种三维图像拼接方法客户端侧的功能或步骤。
在一个实施例中，提出了一种计算机设备，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 机程序时实现以下步骤：
获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
本实施例提出的三维图像拼接方法，通过获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注，接着基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络，从而通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像，最后基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整，能够通过生物样本的各层三维图像的匹配点、空间变换网络、第一目标三维图像，对初始深度学习模型进行模型训练，得到训练好的深度学习模型，而后利用深度学习模型可以预测得到待拼接三维图像的目标形变场，利用该目标形变场可以实现对待拼接三维图像高效准确的拼接。
在一个实施例中，提出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以下步骤：
获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
本实施例提出的三维图像拼接方法，通过获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注，接着基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络，从而通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像，最后基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整，能够通过生物样本的各层三维图像的匹配点、空间变换网络、第一目标三维图像，对初始深度学习模型进行模型训练，得到训练好的深度学习模型，而后利用深度学习模型可以预测得到待拼接三维图像的目标形变场，利用该目标形变场可以实现对待拼接三维图像高效准确的拼接。
需要说明的是，上述关于计算机可读存储介质或计算机设备所能实现的功能或步骤，可对应参阅前述方法实施例中，服务端侧以及客户端侧的相关描述，为避免重复，这里不再一一描述。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一非易失性计算机可读取存储介质中，该计算机程序在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失 性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。
以上所述实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的精神和范围，均应包含在本发明的保护范围之内。

Claims (8)

  1. 一种三维图像拼接方法，其特征在于，所述三维图像拼接方法包括：
    获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
    基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
    通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
    基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
  2. 根据权利要求1所述的三维图像拼接方法，其特征在于，所述基于所述第一目标三维图像、各层所述三维图像以及目标损失函数，对初始深度学习模型中的参数和权重进行调整的步骤，包括：
    基于所述第一目标三维图像、各层所述三维图像以及目标损失函数，对所述初始深度学习模型中的参数和权重进行调整，其中，所述目标损失函数包括匹配损失、正则损失以及半监督损失函数；
    判断所述初始深度学习模型是否达到训练结束条件，若是，则将达到训练结束条件的所述初始深度学习模型作为训练好的深度学习模型。
  3. 根据权利要求1所述的三维图像拼接方法，其特征在于，所述基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络的步骤，包括：
    对所述三维图像进行预处理以及数据增强处理；
    将处理后的三维图像输入至所述初始深度学习模型中进行进行形变场预测，得到形变场。
  4. 根据权利要求1所述的三维图像拼接方法，其特征在于，所述初始深度学习模型采用U-net网络。
  5. 根据权利要求1所述的三维图像拼接方法，其特征在于，所述基于所述第一目标三维图像、各层所述三维图像以及目标损失函数，对初始深度学习模型中的参数和权重进行调整的步骤之后，包括：
    获取各层的待拼接三维图像；
    基于所述待拼接三维图像以及训练好的深度学习模型，对每一层的待拼接三维图像进行形变场预测，得到目标形变场，所述目标形变场被插值到三维空间中；
    基于插值后的所述目标形变场，对各层的待拼接三维图像进行拼接，得到第二目标三维图像。
  6. 一种三维图像拼接装置，其特征在于，所述三维图像拼接装置包括：
    获取模块，用于获取训练数据，所述训练数据包括生物样本的各层三维图像，其中，相邻的两层所述三维图像的相邻的二维表面之间进行匹配点标注；
    预测模块，用于基于各个所述三维图像以及初始深度学习模型中进行形变场预测，得到形变场，其中，所述深度学习模型是一种卷积神经网络；
    拼接模块，用于通过所述形变场、空间变换网络，对所述三维图像进行拼接，得到第一目标三维图像；
    训练模块，用于基于所述第一目标三维图像、各层所述三维图像、所述匹配点以及目标损失函数，对初始深度学习模型中的参数和权重进行调整。
  7. 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至5任一项所述三维图像拼接方法的步骤。
  8. 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至5任一项所述三维图像拼接方法的步骤。
PCT/CN2024/083577 2024-03-25 2024-03-25 三维图像拼接方法、装置、计算机设备及存储介质 Pending WO2025199682A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2024/083577 WO2025199682A1 (zh) 2024-03-25 2024-03-25 三维图像拼接方法、装置、计算机设备及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2024/083577 WO2025199682A1 (zh) 2024-03-25 2024-03-25 三维图像拼接方法、装置、计算机设备及存储介质

Publications (1)

Publication Number Publication Date
WO2025199682A1 true WO2025199682A1 (zh) 2025-10-02

Family

ID=97219809

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/083577 Pending WO2025199682A1 (zh) 2024-03-25 2024-03-25 三维图像拼接方法、装置、计算机设备及存储介质

Country Status (1)

Country Link
WO (1) WO2025199682A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112070781A (zh) * 2020-08-13 2020-12-11 沈阳东软智能医疗科技研究院有限公司 颅脑断层扫描图像的处理方法、装置、存储介质及电子设备
WO2022199135A1 (zh) * 2021-03-26 2022-09-29 中国科学院深圳先进技术研究院 一种基于深度学习的仰卧位和俯卧位乳腺图像配准方法
CN115830016A (zh) * 2023-02-09 2023-03-21 真健康（北京）医疗科技有限公司 医学图像配准模型训练方法及设备
CN116485853A (zh) * 2023-04-14 2023-07-25 深圳技术大学 一种基于深度学习神经网络的医学图像配准方法和装置
WO2023169575A1 (zh) * 2022-03-10 2023-09-14 中国科学院深圳先进技术研究院 核磁共振图像批量配准方法及系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112070781A (zh) * 2020-08-13 2020-12-11 沈阳东软智能医疗科技研究院有限公司 颅脑断层扫描图像的处理方法、装置、存储介质及电子设备
WO2022199135A1 (zh) * 2021-03-26 2022-09-29 中国科学院深圳先进技术研究院 一种基于深度学习的仰卧位和俯卧位乳腺图像配准方法
WO2023169575A1 (zh) * 2022-03-10 2023-09-14 中国科学院深圳先进技术研究院 核磁共振图像批量配准方法及系统
CN115830016A (zh) * 2023-02-09 2023-03-21 真健康（北京）医疗科技有限公司 医学图像配准模型训练方法及设备
CN116485853A (zh) * 2023-04-14 2023-07-25 深圳技术大学 一种基于深度学习神经网络的医学图像配准方法和装置

Similar Documents

Publication Publication Date Title
CN110599526B (zh) 图像配准方法、计算机设备和存储介质
CN110211076B (zh) 图像拼接方法、图像拼接设备和可读存储介质
KR20210021039A (ko) 이미지 처리 방법, 장치, 전자 기기 및 컴퓨터 판독 가능한 저장 매체
CN112837357B (zh) 医学影像配准方法、装置、计算机设备及存储介质
CN111192252B (zh) 一种图像分割结果优化方法、装置、智能终端及存储介质
CN111161269B (zh) 图像分割方法、计算机设备和可读存储介质
CN118195891A (zh) 三维图像拼接方法、装置、计算机设备及存储介质
WO2022194152A1 (zh) 基于图像处理模型的图像处理方法、装置、电子设备、存储介质及计算机程序产品
Li et al. Retinal vessel segmentation by using AFNet
WO2022120669A1 (zh) 手势识别方法、计算机设备及存储介质
CN115083571B (zh) 病理切片处理方法、计算机设备和存储介质
Lee et al. Enhancing keypoint detection in Y-Maze behavior test automation: introducing stadium heatmap and squared adaptive wing loss
Zhang et al. Efficient skin lesion segmentation with boundary distillation
CN113673517B (zh) 基于姿态检测的数据增强方法、装置、设备及介质
WO2025199682A1 (zh) 三维图像拼接方法、装置、计算机设备及存储介质
CN118351304B (zh) 医学图像分割方法、装置、计算机设备及存储介质
Wu et al. Integrating multi-scale information and diverse prompts in large model SAM-Med2D for accurate left ventricular ejection fraction estimation
Kumari et al. Conditional stable diffusion for distortion correction and image rectification
CN110766653B (zh) 图像分割方法、装置、计算机设备和存储介质
CN118196152B (zh) 生物样本图像的标注点获取方法、装置、设备及介质
WO2025199775A1 (zh) 生物样本图像的标注点获取方法、装置、设备及介质
CN109587469B (zh) 基于人工智能识别的影像处理方法及装置
CN109509262B (zh) 基于人工智能的智能增强建模方法及装置
CN111161239B (zh) 医学图像分析方法、装置、存储介质及计算机设备
CN119578494B (zh) 基于图像-文本对的蒸馏训练方法、装置、终端及介质

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: 24933044

Country of ref document: EP

Kind code of ref document: A1