WO2022032823A1 - 图像分割方法、装置、设备及存储介质 - Google Patents

图像分割方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2022032823A1
WO2022032823A1 PCT/CN2020/117826 CN2020117826W WO2022032823A1 WO 2022032823 A1 WO2022032823 A1 WO 2022032823A1 CN 2020117826 W CN2020117826 W CN 2020117826W WO 2022032823 A1 WO2022032823 A1 WO 2022032823A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
image segmentation
current high
result
image
Prior art date
Application number
PCT/CN2020/117826
Other languages
English (en)
French (fr)
Inventor
吴剑煌
倪佳佳
Original Assignee
中国科学院深圳先进技术研究院
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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2022032823A1 publication Critical patent/WO2022032823A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30041—Eye; Retina; Ophthalmic
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30101—Blood vessel; Artery; Vein; Vascular

Definitions

  • the embodiments of the present application relate to the field of medical image processing, for example, to an image segmentation method, apparatus, device, and storage medium.
  • Image segmentation technology for segmenting the lesion area from medical images emerges as the times require.
  • Image segmentation methods in related technologies are mainly divided into machine learning methods and deep learning methods, and deep learning methods in related technologies usually have problems of low universality.
  • the embodiments of the present application provide an image segmentation method, apparatus, device, and storage medium, which solve the problem of low universality of the deep learning method in the related art.
  • an image segmentation method which includes:
  • Image segmentation is performed on the image to be segmented by the trained image segmentation model to obtain the target segmentation area, wherein the image segmentation model is configured to calculate the self-attention result corresponding to the current high-level feature and the low-level corresponding to the current high-level feature.
  • the product of features to obtain an initial channel attention result and combining the initial channel attention result with the current high-level feature to obtain a channel attention result, and updating the current high-level feature according to the channel attention result
  • the size of the updated current high-level feature is larger than the size of the current high-level feature before the update.
  • an embodiment of the present application further provides an image segmentation device, where the image segmentation device includes:
  • an acquisition module configured to acquire the image to be segmented
  • the output module is configured to perform image segmentation on the to-be-segmented image through the trained image segmentation model to obtain the target segmentation area, wherein the decoding unit of the image segmentation model is used to calculate the self-attention result corresponding to the current high-level feature and The product of the low-level features corresponding to the current high-level feature to obtain an initial channel attention result, and the initial channel attention result and the current high-level feature are combined to obtain a channel attention result, and according to the channel attention As a result, the current high-level feature is updated, and the size of the updated current high-level feature is larger than the size of the current high-level feature before the update.
  • an embodiment of the present application further provides an image segmentation device, the device comprising:
  • a storage device for storing programs
  • the processor When the program is executed by the processor, the processor implements the image segmentation method described in any of the embodiments.
  • an embodiment of the present application further provides a storage medium, including computer-executable instructions, where the computer-executable instructions are used to execute the image segmentation method described in any embodiment when executed by a computer processor.
  • Embodiment 1 is a flowchart of an image segmentation method provided in Embodiment 1 of the present application;
  • FIG. 2 is a schematic diagram of an image to be segmented (fundus image) provided in Embodiment 1 of the present application;
  • FIG. 3 is a schematic diagram of an image segmentation model provided in Embodiment 1 of the present application.
  • Embodiment 4 is a schematic diagram of the combination of attention and channel attention provided by Embodiment 1 of the present application;
  • FIG. 5A is a schematic diagram of fundus blood vessel image segmentation according to Embodiment 1 of the present application.
  • FIG. 5B is a schematic diagram of segmentation of an intracranial artery blood vessel image provided in Embodiment 1 of the present application;
  • 5C is a schematic diagram of segmentation of a femoral artery image provided in Embodiment 1 of the present application;
  • FIG. 6 is a schematic diagram of a pooling unit provided in Embodiment 1 of the present application.
  • FIG. 7 is a schematic diagram of an SE block provided in Embodiment 1 of the present application.
  • FIG. 8 is a structural block diagram of an image segmentation apparatus provided in Embodiment 2 of the present application.
  • FIG. 9 is a structural block diagram of an image segmentation device provided in Embodiment 3 of the present application.
  • FIG. 1 is a flowchart of an image segmentation method provided in Embodiment 1 of the present application.
  • the technical solution of this embodiment is applicable to the case where image segmentation is automatically completed by a trained image segmentation model, wherein the image segmentation model is constructed based on a self-attention mechanism and a channel attention mechanism.
  • the method may be performed by the image segmentation apparatus provided in the embodiment of the present application, and the apparatus may be implemented in software and/or hardware, and configured to be applied in a processor.
  • the method may include the following steps:
  • the images to be segmented may be CT (Computed Tomography, CT for short) images, Magnetic Resonance Imaging (MR, magnetic resonance imaging) images, PET (Positron Emission Tomography, PET for short, positron emission computer for short) images. tomography) images and other clinical medical images, which include target segmentation areas and non-target segmentation areas.
  • the target segmentation region may be a region of interest to a doctor such as a blood vessel, a bleeding region, or the like. Referring to the fundus image shown in FIG. 2 , the target segmentation area of the fundus image is the fundus blood vessel area.
  • S102 Perform image segmentation on the image to be segmented by the trained image segmentation model to obtain the target segmentation area, wherein the decoding unit of the image segmentation model is configured to calculate the self-attention result corresponding to the current high-level feature and the current high-level feature.
  • the product of the low-level features to get the initial channel attention result, and the initial channel attention result and the current high-level feature are combined to get the channel attention result, and the current high-level feature is updated according to the channel attention result, and the updated current high-level feature
  • the size of is larger than the size of the current high-level feature before the update.
  • Image segmentation may be to determine the category information of each pixel of the image to be segmented, that is, whether each pixel belongs to the target segmentation area or the non-target segmentation area. Therefore, when performing image segmentation, it is necessary to determine the location information and category information of each pixel.
  • the image segmentation model includes a feature extraction unit, a feature fusion unit and a decoding unit, as shown in FIG. 3 .
  • the feature extraction unit adopts the method of multi-layer convolution and batch normalization to complete the feature extraction of the image to be segmented.
  • the weight of the pre-training model may or may not be added, which can be determined according to the actual use.
  • the feature extraction unit completes the feature extraction of the to-be-segmented image by means of four-layer convolution and batch normalization, so as to obtain a feature map with a size of 25 ⁇ 25 or a feature map with a size of 32 ⁇ 32.
  • the decoding unit is configured to restore the feature map reduced by the feature extraction unit step by step to the size of the image to be segmented through the decoding operation, thereby ensuring that the entire image segmentation process is an end-to-end manner.
  • the decoding unit of this embodiment completes the decoding operation by combining the self-attention mechanism and the channel attention mechanism.
  • the steps performed by the decoding unit of the image segmentation model may include: calculating the product of the self-attention result corresponding to the current high-level feature and the low-level feature corresponding to the current high-level feature to obtain the initial channel attention result, and the initial channel attention result. Combine with the current high-level feature to obtain the channel attention result, and update the current high-level feature according to the channel attention result, and the size of the updated current high-level feature is larger than the size of the current high-level feature before the update.
  • the decoding unit completes the feature extraction of the current high-level features through two attention mechanisms to improve the feature extraction capability of the entire image segmentation model.
  • the calculation method of the self-attention result is shown in Figure 4.
  • the first feature map F(x), the second feature map G(X) and the third feature map H are extracted from the current high-level features through different preset convolution operations.
  • the multiplication operation of the feature map makes any pixel in the self-attention result a weighted sum of the values of all positions of the current high-level feature, which realizes the enhancement of the position feature and the acquisition of long-dependent feature information, that is, the
  • the self-attention result carries the contextual feature information of the image to be segmented.
  • different preset convolution operations are the corresponding feature extraction operations performed for the current high-level features using convolution kernels of different sizes.
  • the size of each convolution kernel can be set according to the usage scenario in actual use.
  • the image segmentation method further includes: performing global average pooling on the self-attention result to obtain the global average pooling result, that is, the updated self-attention result,
  • the global average pooling result is K 1 ⁇ 1 feature maps, where K is the number of channels, and the product of the global average pooling result and the low-level feature corresponding to the current high-level feature is calculated to obtain the initial channel attention result.
  • the skip connection between the self-attention result and the corresponding low-level feature is realized, and the contextual feature information carried by the self-attention result is used to guide the low-level feature to obtain the location information and category information of the pixel.
  • the global average pooling method can be selected as L2 regularization. Understandably, each pixel in the initial channel attention result is the weighted sum of each pixel of the low-level feature and all channels of the self-attention result.
  • the feature map with a smaller scale has a higher level
  • a feature map with a larger scale has a smaller level
  • the low-level feature adjacent to the current high-level feature is used as the low-level feature corresponding to the current high-level feature.
  • the 32 ⁇ 32 feature map has a higher level than the 64 ⁇ 64 feature map. If the current high-level feature is a 32 ⁇ 32 feature map, then its corresponding low-level feature is a 64 ⁇ 64 feature map. If the current high-level feature is a 64 ⁇ 64 feature map, then its corresponding low-level feature is a 128 ⁇ 128 map.
  • the decoding unit After obtaining the initial channel attention result, the decoding unit combines the initial channel attention result with the current high-level feature, such as adding, to obtain the channel attention result. Then, the current high-level feature is updated according to the channel attention result, and thus the decoding operation of the current high-level feature is completed.
  • the decoding unit repeatedly performs the above-mentioned decoding operations four times, that is, the target feature image is obtained after four upsampling mechanisms.
  • the target feature image has the same size as the image to be segmented. It is understood that the number of decoding operations is the same as the number of feature extractions performed by the feature extraction unit.
  • the decoding unit After the decoding unit obtains the target feature image, it uses softmax to perform a classification operation on the pixels in the target feature image to obtain the target segmented region, see FIGS. 5A , 5B and 5C .
  • the feature fusion unit includes multiple parallel channels, and optionally, different channels have different expansion rates.
  • Each channel includes an SE block (Squeeze-and-Excitation, SE for short, compression and decompression network), and the SE block is configured to perform compression and decompression operations on the feature extraction result output by the feature extraction unit.
  • SE block Sequeeze-and-Excitation, SE for short, compression and decompression network
  • the feature fusion unit includes four channels with different dilation rates.
  • the first channel performs feature extraction on the feature extraction result output by the feature extraction unit through a 1 ⁇ 1 convolution kernel (equivalent to an expansion rate of 1) to obtain the first extracted feature, and then passes through the SE block (compression and decompression network) Perform compression and decompression operations on the first extracted feature to obtain corresponding compression and decompression results; the second channel, the third channel, and the fourth channel respectively perform expansion operations on the feature extraction results.
  • the expansion rates are respectively 6, 12 and 12, then each channel uses the SE block to compress and decompress the expansion operation results to obtain the corresponding compression and decompression results; after the compression and decompression results are obtained for the four channels, all use 1 ⁇ 1
  • the convolution kernel performs feature extraction on the compression and decompression results to obtain corresponding feature maps; and after obtaining the feature maps of each channel, feature fusion is performed on the feature maps of each channel to obtain a feature fusion result.
  • FIG. 7 is a schematic diagram of compression and decompression of SE blocks.
  • F tr represents a transformation operation, such as a standard convolution operation.
  • a bypass branch is branched out after the standard convolution operation, in which the squeezing (Squeeze) operation is first performed, that is, the F sq ( ⁇ ) operation in the figure, which is used to adopt the global average pooling operation for each
  • the feature maps are compressed so that the C feature maps finally become a 1 ⁇ 1 ⁇ C real number sequence.
  • the global average pooling operation enables U (multiple feature maps) to have a global receptive field, so that the lower layers of the network can also utilize global information.
  • the decompression (Excitation) operation is performed in this bypass branch, that is, the F ex ( ) operation in the graph to generate weights for each feature channel through the parameter w, so as to fully capture the dependencies (or channels) between channels importance to each other).
  • F scale is used to multiply each value in U by a preset scalar. After the weight of each channel and the product of each value in U and a preset scalar are obtained, channel information corresponding to each value in U can be obtained. Since the compression and decompression methods can make full use of the global information of the feature map and the dependencies between each channel, the robustness of the image segmentation model can be enhanced.
  • the optimization function in the image segmentation model can optionally use the Adam algorithm.
  • the loss function includes a main function and an auxiliary function.
  • the loss function can be expressed as:
  • ⁇ can be optionally 0.5
  • L dice is the main function, and its form can be:
  • N is the total number of pixels of the image to be segmented
  • q(k,i) ⁇ [0,1] respectively represent the probability and gold standard obtained by classification.
  • L r is an auxiliary function, which is a weighted cross-entropy function, which can be:
  • y represents the actual value
  • TP represents the true value, that is, the prediction is positive, and the actual value is also positive
  • TN represents the true negative, that is, the prediction is negative, and the actual value is also negative
  • NP represents the target segmentation area
  • N n represents the non-segmentation area.
  • image segmentation model can be trained by using the model training method in the related art to obtain the trained image segmentation model.
  • the algorithm of the image segmentation method described in the embodiments of the present application is implemented by using the python language based on the disclosed Keras platform and the back end of tensorflow.
  • the configuration information of the computer running the algorithm includes: the operating system is Ubuntu 16.04, the GPU is an NVIDIA Titan XP graphics card, and the memory is 12GB.
  • the image to be segmented is segmented by the trained image segmentation module to obtain the target segmentation area
  • the decoding unit of the image segmentation model is configured to calculate The product of the self-attention result corresponding to the current high-level feature and the low-level feature corresponding to the current high-level feature to obtain the initial channel attention result, and the combination of the initial channel attention result and the current high-level feature to obtain the channel attention
  • the current high-level feature is updated according to the channel attention result, and the size of the updated current high-level feature is larger than the size of the current high-level feature before the update.
  • the initial channel attention result can use the contextual feature information to guide shallow layer features to obtain the position information and category information of pixels, and the channel attention results not only include the initial channel attention results, but also the current high-level features, which make the channel attention results more accurate and applicable to various scenarios, thus making the Image segmentation models have high generality, not just for specific medical image data.
  • FIG. 8 is a structural block diagram of an image segmentation apparatus provided in Embodiment 2 of the present application.
  • the apparatus is configured to execute the image segmentation method provided in any of the foregoing embodiments, and the apparatus can be optionally implemented in software or hardware.
  • the device includes:
  • an acquisition module 11 configured to acquire an image to be segmented
  • the output module 12 is configured to perform image segmentation on the to-be-segmented image through the trained image segmentation model to obtain the target segmentation area, wherein the decoding unit of the image segmentation model is configured to calculate the self-attention corresponding to the current high-level feature
  • the product of the result and the low-level feature corresponding to the current high-level feature to obtain the initial channel attention result, and combining the initial channel attention result with the current high-level feature to obtain the channel attention result, and according to the channel attention
  • the force result updates the current high-level feature, and the size of the updated current high-level feature is larger than the size of the current high-level feature before the update.
  • the decoding unit is configured to extract the first feature map, the second feature map and the third feature map from the current high-level features through different convolution operations; calculate the product of the first feature map and the second feature map, and Perform a classification operation on the product result to obtain the classification information of each pixel; calculate the product of the classification information and the third feature map, and the product of the product result and the current high-level feature to obtain the self-attention result.
  • the decoding unit is further configured to perform global average pooling on the self-attention result to update the self-attention result, and the updated self-attention result is K 1 ⁇ 1 feature maps, where K is the number of channels .
  • the image segmentation model further includes a feature extraction unit and a feature fusion unit; the feature extraction unit is configured to perform feature extraction on the image to be segmented to obtain a feature extraction result; The feature map of the corresponding scale is extracted from the result, and the compression and decompression operations are performed on the feature map of the corresponding scale, and the feature fusion is performed on the output compression and decompression results of all channels to obtain the determined feature fusion result.
  • the loss function of the image segmentation model includes a main function and an auxiliary function; wherein, the auxiliary function is a weighted cross-entropy function.
  • the image to be segmented is segmented by the trained image segmentation model to obtain the target segmentation area
  • the decoding unit of the image segmentation model is configured to calculate the self-attention corresponding to the current high-level feature.
  • the product of the force result and the low-level feature corresponding to the current high-level feature to obtain the initial channel attention result, and the initial channel attention result and the current high-level feature are combined to obtain the channel attention result, and according to the channel
  • the attention result updates the current high-level feature, and the size of the updated current high-level feature is larger than the size of the current high-level feature before the update.
  • the self-attention result Since the self-attention result carries the contextual feature information of the image to be segmented, it is based on the self-attention result and the size of the current high-level feature. After the low-level feature corresponding to the current high-level feature determines the initial channel attention result, the initial channel attention result can use the contextual feature information to guide the shallow feature to obtain pixel position information and category information, and the channel attention result not only includes the initial channel.
  • the attention results also contain the current high-level features, which make the channel attention results more accurate and applicable to various scenarios, so that the image segmentation model has high generality, not only for specific medical image data.
  • the image segmentation method device provided in the embodiment of the present application can execute the image segmentation method provided by any embodiment of the present application, and has corresponding functional modules and beneficial effects of the execution method.
  • FIG. 9 is a schematic structural diagram of a medical image segmentation device provided in Embodiment 3 of the present application.
  • the device includes a processor 201, a memory 202, an input device 203, and an output device 204; the number of processors 201 in the device can be is one or more, and a processor 201 is taken as an example in FIG. 9; the processor 201, memory 202, input device 203 and output device 204 in the device can be connected by a bus or in other ways. In FIG. 9, the connection by bus is example.
  • the memory 202 can be used to store software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to the image segmentation method in the embodiments of the present application (for example, the acquisition module 11 and the output module 12). ).
  • the processor 201 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 202 , that is, to implement the above-mentioned image segmentation method.
  • the memory 202 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created according to the use of the terminal, and the like. Additionally, memory 202 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device. In some examples, memory 202 may include memory located remotely from processor 201, which may be connected to the device through a network. Examples of such networks may include the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
  • the input device 203 may be used to receive input numerical or character information, and to generate key signal input related to user settings and function control of the device.
  • the output device 204 may include a display device such as a display screen, for example, a display screen of a user terminal.
  • Embodiment 4 of the present application further provides a storage medium, including computer-executable instructions, where the computer-executable instructions are used to execute an image segmentation method when executed by a computer processor, and the method includes:
  • Image segmentation is performed on the image to be segmented by the trained image segmentation model to obtain the target segmentation area, wherein the decoding unit of the image segmentation model is configured to calculate the self-attention result corresponding to the current high-level feature corresponding to the current high-level feature
  • a storage medium containing computer-executable instructions provided by the embodiments of the present application can perform the above-mentioned method operations, and can also perform the image segmentation methods provided in any embodiment of the present application. related operations.
  • the present application can be implemented by means of software and necessary general-purpose hardware, and certainly can also be implemented by hardware.
  • the technical solutions of the present application can be embodied in the form of software products in essence or the parts that make contributions to related technologies, and the computer software products can be stored in a computer-readable storage medium, such as a computer floppy disk, Read-Only Memory (ROM for short), Random Access Memory (RAM for short), Flash Memory (FLASH), hard disk or CD, etc., including several instructions to make a computer device (which can be a personal computer, server, or network device, etc.) to execute the image segmentation method described in each embodiment of the present application.
  • a computer-readable storage medium such as a computer floppy disk, Read-Only Memory (ROM for short), Random Access Memory (RAM for short), Flash Memory (FLASH), hard disk or CD, etc.
  • the units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized;
  • the names of the functional units are also only used to distinguish them from each other.
  • the image to be segmented is segmented by the trained image segmentation module to obtain the target segmentation area
  • the decoding unit of the image segmentation model is configured to calculate The product of the self-attention result corresponding to the current high-level feature and the low-level feature corresponding to the current high-level feature to obtain the initial channel attention result, and the combination of the initial channel attention result and the current high-level feature to obtain the channel attention
  • the self-attention result carries the contextual feature information of the image to be segmented, Therefore, after the initial channel attention result is determined based on the low-level feature corresponding to the self-attention result and the current high-level feature, the initial channel attention result can use the contextual feature information to guide the low-level feature to obtain pixel position information

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

一种图像分割方法、装置、设备及存储介质，该方法包括：获取待分割图像（S101）；通过已训练的图像分割模型对所述待分割图像进行图像分割，以得到目标分割区域，其中，图像分割模型的解码单元被配置为通过计算得到初始通道注意力结果，对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸（S102）。

Description

图像分割方法、装置、设备及存储介质
本公开要求在2020年08月10日提交中国专利局、申请号为202010795220.3的中国专利申请的优先权，以上申请的全部内容通过引用结合在本公开中。
技术领域
本申请实施例涉及医学图像处理领域，例如涉及一种图像分割方法、装置、设备及存储介质。
背景技术
随着医疗科学技术的发展，很多医院都配置有各种各样的医学影像设备，这些医学影像设备每天都会产生大量的医学图像数据。这些医学图像数据对于病人的病情诊断具有重要的作用。然而，限于医生的时间、精力以及临床经验，仅仅依靠医生的视觉很难准确、有效地通过影像数据完成疾病的诊断。
为了提高医生的诊断效率，用于将病变区域从医学图像中分割出来的图像分割技术应运而生。相关技术的图像分割方法主要分为机器学习方法和深度学习方法，而相关技术中的深度学习方法通常具有普适性较低的问题。
发明内容
本申请实施例提供了一种图像分割方法、装置、设备及存储介质，解决了相关技术中的深度学习方法的普适性较低的问题。
第一方面，本申请实施例提供了一种图像分割方法，该方法包括：
获取待分割图像；
通过已训练的图像分割模型对所述待分割图像进行图像分割，以得到目标分割区域，其中，图像分割模型被配置为计算当前高层特征对应的自注意力结果与所述当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
第二方面，本申请实施例还提供了图像分割装置，所述图像分割装置包括：
获取模块，被配置为获取待分割图像；
输出模块，被配置为通过已训练的图像分割模型对所述待分割图像进行图 像分割，以得到目标分割区域，其中，图像分割模型的解码单元用于计算当前高层特征对应的自注意力结果与所述当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
第三方面，本申请实施例还提供了一种图像分割设备，该设备包括：
处理器；
存储装置，用于存储程序；
当所述程序被所述处理器执行，使得所述处理器实现如任意实施例所述的图像分割方法。
第四方面，本申请实施例还提供了一种存储介质，包含计算机可执行指令，所述计算机可执行指令在由计算机处理器执行时用于执行任意实施例所述的图像分割方法。
附图说明
下面将对实施例描述中所需要使用的附图做一简单地介绍，显而易见地，下面描述中的附图是本申请的一些实施例，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。
图1是本申请实施例一提供的图像分割方法的流程图；
图2是本申请实施例一提供的待分割图像(眼底图像)示意图；
图3是本申请实施例一提供的图像分割模型的示意图；
图4是本申请实施例一提供的注意力与通道注意力结合示意图；
图5A是本申请实施例一提供的眼底血管图像分割示意图；
图5B是本申请实施例一提供的颅内动脉血管图像分割示意图；
图5C是本申请实施例一提供的腿骨动脉图像分割示意图；
图6是本申请实施例一提供的池化单元的示意图；
图7是本申请实施例一提供的SE块的示意图；
图8是本申请实施例二提供的图像分割装置的结构框图；
图9是本申请实施例三提供的图像分割设备的结构框图。
具体实施方式
以下将参照本申请实施例中的附图，通过实施方式清楚、完整地描述本申请的技术方案，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，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，都属于本申请保护的范围。
实施例一
图1是本申请实施例一提供的图像分割方法的流程图。本实施例的技术方案适用于通过已训练的图像分割模型自动完成图像分割的情况，其中，图像分割模型基于自注意力机制与通道注意力机制构建。该方法可以由本申请实施例提供的图像分割装置来执行，该装置可以采用软件和/或硬件的方式实现，并配置在处理器中应用。该方法可以包括如下步骤：
S101、获取待分割图像。
其中，待分割图像可以是CT(Computed Tomography，简称CT，电子计算机断层成像)图像、MR(Magnetic Resonance Imaging，简称MR，磁共振成像)图像、PET(Positron Emission Tomography，简称PET，正电子发射计算机断层显像)图像等临床医学图像，其包括目标分割区域和非目标分割区域。其中，目标分割区域可以是血管、出血区域等医生感兴趣区域。参见图2所示的眼底图像，该眼底图像的目标分割区域为眼底血管区域。
S102、通过已训练的图像分割模型对待分割图像进行图像分割，以得到目标分割区域，其中，图像分割模型的解码单元被配置为计算当前高层特征对应的自注意力结果与该当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对初始通道注意力结果与当前高层特征进行组合以得到通道注意力结果，以及根据通道注意力结果更新当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
图像分割可以是要确定待分割图像的每一个像素的类别信息，即每一个像素是属于目标分割区域还是属于非目标分割区域。因此在进行图像分割时，需要确定每个像素的位置信息和类别信息。
其中，图像分割模型包括特征提取单元、特征融合单元和解码单元，参见图3所示。特征提取单元采用多层卷积和批标准化的方法完成待分割图像的特征提取。而且该特征提取单元在进行特征提取时，可以添加预训练模型权重，也可以不添加预训练模型权重，实际使用时可以根据情况确定。
示例性的，参见表1，特征提取单元通过四层卷积和批标准化的方法完成待分割图像的特征提取，以得到25×25大小的特征图或32×32大小的特征图。
表1特征提取过程信息表
Figure PCTCN2020117826-appb-000001
解码单元被配置为通过解码操作，将特征提取单元缩小的特征图一步一步的恢复到待分割图像的大小，从而保证整个图像分割过程是一种端到端的方式。本实施例的解码单元通过自注意力机制和通道注意力机制相结合的方式完成解码操作。所述图像分割模型的解码单元执行的步骤可以包括：计算当前高层特征对应的自注意力结果与该当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对初始通道注意力结果与当前高层特征进行组合以得到通道注意力结果，以及根据通道注意力结果更新当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
其中，解码单元通过两次注意力机制完成当前高层特征的特征提取，以提高整个图像分割模型的特征提取能力。自注意力结果的计算方法参见图4所示，通过不同的预设卷积操作从当前高层特征中提取第一特征图F(x)、第二特征图G(X)和第三特征图H(X)，然后直接计算F(x)与G(X)的乘积以得到乘积结果，并对F(x)与G(X)的乘积结果执行分类操作以得到包含每个像素分类信息的分类结果；计算该分类结果与H(X)的乘积，然后计算该分类结果与H(X)的乘积结果与该当前高层特征的乘积，以得到包含当前高层特征的每个像素的分类信息和位置信息的自注意力结果。可以理解的是，特征图的乘法操作使得自注意力结果中的任一像素都是当前高层特征的所有位置的值的加权和，实现了位置特征的加强和长依赖特征信息的获取，即使得自注意力结果携带有待分割图像的上下文特征信息。其中，不同的预设卷积操作，为使用不同大小的卷积核对当前高 层特征所进行的相应的特征提取操作。各个卷积核的大小可在实际使用时根据使用场景进行设定。
参见图4所示，在计算初始通道注意力结果之前，所述图像分割方法还包括：对自注意力结果进行全局平均池化以得到全局平均池化结果，即更新后的自注意力结果，该全局平均池化结果为K个1×1的特征图，其中，K为通道数，计算该全局平均池化结果和该当前高层特征对应的低层特征的乘积，以得到初始通道注意力结果。实现了自注意力结果与对应的低层特征的跳跃连接，以及使用自注意力结果携带的上下文特征信息来指导低层特征获取像素的位置信息和类别信息。其中，全局平均池化方式可选为L2正则化。可以理解的是，初始通道注意力结果中的每个像素是低层特征的每个像素与自注意力结果的所有通道的加权和。
可以理解的是，尺度越小的特征图的层级越高，尺度越大的特征图的层级越小。本实施例将邻近当前高层特征的低层特征作为当前高层特征对应的低层特征。如图2所示，32×32的特征图的层级高于64×64的特征图。如果当前高层特征为32×32的特征图，那么其对应的低层特征为64×64的特征图，如果当前高层特征为64×64的特征图，那么其对应的低层特征为128×128图。
解码单元在得到初始通道注意力结果之后，将该初始通道注意力结果与该当前高层特征进行组合，比如相加，以得到通道注意力结果。然后根据该通道注意力结果更新该当前高层特征，至此完成该当前高层特征的解码操作。
在一些实施例中，解码单元重复执行四次上述解码操作，即经过4次上采样机制之后得到目标特征图像。该目标特征图像与待分割图像的大小相同。可以理解的是，解码操作的次数与特征提取单元执行的特征提取的次数相同。
解码单元得到该目标特征图像之后，使用softmax对目标特征图像中的像素执行分类操作，以得到目标分割区域，参见图5A、图5B和图5C。
其中，特征融合单元包括多条并联通道，且可选地，不同通道的扩张率大小不同。每个通道均包括SE块(Squeeze-and-Excitation，简称SE，压缩与解压缩网络)，该SE块被配置为对特征提取单元输出的特征提取结果进行压缩与解压缩操作。
在一些实施例中，如图6所示，特征融合单元包括四条扩张率大小不同的通道。其中，第一通道通过1×1卷积核(相当于扩张率为1)对特征提取单元输出的特征提取结果进行特征提取以得到第一提取特征，然后通过SE块(压缩 与解压缩网络)对该第一提取特征进行压缩与解压缩操作以得到对应的压缩与解压缩结果；第二通道、第三通道和第四通道分别对特征提取结果进行扩张操作，示例性的，扩张率分别为6、12和12，然后各个通道分别使用SE块对扩张操作结果进行压缩与解压缩操作以得到相应的压缩与解压缩结果；四个通道在得到压缩与解压缩结果之后，均使用1×1卷积核对压缩与解压缩结果进行特征提取以得到相应的特征图；以及在得到各个通道特征图之后，对各个通道的特征图进行特征融合，以得到特征融合结果。
图7为SE块进行压缩与解压缩的示意图。其中，F tr表示转换操作，比如标准的卷积操作。在标准的卷积操作之后分出一个旁路分支，在该旁路分支中首先进行压缩(Squeeze)操作，即图中的F sq(·)操作，用于采用全局平均池化操作对其每个特征图进行压缩，使其C个特征图最后变成1×1×C的实数数列。全局平均池化操作使U(多个feature map)具有全局的感受野，使得网络低层也能利用全局信息。然后在该旁路分支中进行解压缩(Excitation)操作，即图中的F ex(·)操作，以通过参数w为每个特征通道生成权重，从而全面捕获通道之间的依赖性(或通道相互之间的重要性)。F scale用于将U中的每个值乘以一个预设标量。在得到每个通道的权重以及U中的每个值与预设标量的乘积之后，即可得到U中的每个值所对应的通道信息。由于压缩与解压缩方式可以充分利用特征图的全局信息，以及各个通道间的依赖性，因而能够增强图像分割模型的鲁棒性。
其中，图像分割模型中的优化函数可选使用Adam算法。损失函数包括主函数和辅助函数。损失函数可表示为：
L all＝L dice+λL r
其中，λ可选0.5，L dice为主函数，其形式可以为：
Figure PCTCN2020117826-appb-000002
其中，N为待分割图像的像素总数，p(k,i)∈[0,1]，q(k,i)∈[0,1]分别代表分类得到的概率和金标准。
其中，L r为辅助函数，其为权重交叉熵函数，其可以为：
Figure PCTCN2020117826-appb-000003
其中，y表示实际值，TP表示真正，即预测为正，实际也为正；TN表示真负，即预测为负，实际也为负，N P表示目标分割区域，N n表示非分割区域。
需要说明的是，采用相关技术中的模型训练方法对上述图像分割模型进行训练，以得到已训练的图像分割模型即可。
在一些实施例中，基于公开的Keras平台和tensorflow的后端，使用python语言实现本申请实施例所述的图像分割方法的算法。运行该算法的计算机的配置信息包括：操作系统为Ubuntu 16.04，GPU为英伟达Titan XP显卡，内存为12GB。
本申请实施例提供的图像分割方法、装置、设备及存储介质的技术方案，通过已训练的图像分割模块对待分割图像进行图像分割以得到目标分割区域，该图像分割模型的解码单元被配置为计算当前高层特征对应的自注意力结果与该当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。由于自注意力结果携带有待分割图像的上下文特征信息，因此基于自注意力结果与当前高层特征对应的低层特征确定初始通道注意力结果之后，该初始通道注意力结果可使用该上下文特征信息指导浅层特征来获取像素的位置信息和类别信息，加之通道注意力结果不仅包含初始通道注意力结果，还包含当前高层特征，这使得通道注意力结果更加准确并可适用于各种场景，从而使得该图像分割模型具有较高的普适性，而不仅仅针对特定的医学图像数据。
实施例二
图8是本申请实施例二提供的图像分割装置的结构框图。该装置用于执行上述任意实施例所提供的图像分割方法，该装置可选为软件或硬件实现。该装置包括：
获取模块11，被配置为获取待分割图像；
输出模块12，被配置为通过已训练的图像分割模型对所述待分割图像进行图像分割，以得到目标分割区域，其中，图像分割模型的解码单元被配置为计算当前高层特征对应的自注意力结果与该当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层 特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
可选地，解码单元被配置为通过不同的卷积操作从当前高层特征中提取第一特征图、第二特征图和第三特征图；计算第一特征图和第二特征图的乘积，并对该乘积结果执行分类操作以得到每个像素的分类信息；计算分类信息与第三特征图的乘积，以及该乘积结果与该当前高层特征的乘积，以得到自注意力结果。
可选地，解码单元还被配置为对自注意力结果进行全局平均池化以更新自注意力结果，且更新之后的自注意力结果为K个1×1的特征图，其中K为通道数。
可选地，图像分割模型还包括特征提取单元和特征融合单元；特征提取单元被配置为对待分割图像进行特征提取以得到特征提取结果；特征融合单元被配置为通过多条并联通道分别从特征提取结果中提取相应尺度的特征图并对相应尺度的特征图进行压缩与解压缩操作，以及对所有通道输出的压缩与解压缩结果进行特征融合以得到确定特征融合结果。
可选地，图像分割模型的损失函数包括主函数和辅助函数；其中，辅助函数为权重交叉熵函数。
本申请实施例提供的图像分割装置的技术方案，通过已训练的图像分割模型对待分割图像进行图像分割以得到目标分割区域，该图像分割模型的解码单元被配置为计算当前高层特征对应的自注意力结果与该当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸，由于自注意力结果携带有待分割图像的上下文特征信息，因此基于自注意力结果与当前高层特征对应的低层特征确定初始通道注意力结果之后，该初始通道注意力结果可使用该上下文特征信息指导浅层特征来获取像素的位置信息和类别信息，加之通道注意力结果不仅包含初始通道注意力结果，还包含当前高层特征，这使得通道注意力结果更加准确并可适用于各种场景，从而使得该图像分割模型具有较高的普适性，而不仅仅针对特定的医学图像数据。
本申请实施例所提供的图像分割方法装置可执行本申请任意实施例所提供的图像分割方法，具备执行方法相应的功能模块和有益效果。
实施例三
图9为本申请实施例三提供的医学图像分割设备的结构示意图，如图9所示，该设备包括处理器201、存储器202、输入装置203以及输出装置204；设备中处理器201的数量可以是一个或多个，图9中以一个处理器201为例；设备中的处理器201、存储器202、输入装置203以及输出装置204可以通过总线或其他方式连接，图9中以通过总线连接为例。
存储器202作为一种计算机可读存储介质，可用于存储软件程序、计算机可执行程序以及模块，如本申请实施例中的图像分割方法对应的程序指令/模块(例如，获取模块11和输出模块12)。处理器201通过运行存储在存储器202中的软件程序、指令以及模块，从而执行设备的各种功能应用以及数据处理，即实现上述的图像分割方法。
存储器202可主要包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需的应用程序；存储数据区可存储根据终端的使用所创建的数据等。此外，存储器202可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件、或其他非易失性固态存储器件。在一些实例中，存储器202可包括相对于处理器201远程设置的存储器，这些远程存储器可以通过网络连接至设备。上述网络的实例可以包括互联网、企业内部网、局域网、移动通信网及其组合。
输入装置203可用于接收输入的数字或字符信息，以及产生与设备的用户设置以及功能控制有关的键信号输入。
输出装置204可包括显示屏等显示设备，例如，用户终端的显示屏。
实施例四
本申请实施例四还提供了一种存储介质，包含计算机可执行指令，所述计算机可执行指令在由计算机处理器执行时用于执行一种图像分割方法，该方法包括：
获取待分割图像；
通过已训练的图像分割模型对所述待分割图像进行图像分割，以得到目标 分割区域，其中，图像分割模型的解码单元被配置为计算当前高层特征对应的自注意力结果与该当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
当然，本申请实施例所提供的一种包含计算机可执行指令的存储介质，其计算机可执行指令可以执行如上所述的方法操作，还可以执行本申请任意实施例所提供的图像分割方法中的相关操作。
通过以上关于实施方式的描述，所属领域的技术人员可以清楚地了解到，本申请可借助软件及必需的通用硬件来实现，当然也可以通过硬件实现。基于这样的理解，本申请的技术方案本质上或者说对相关技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品可以存储在计算机可读存储介质中，如计算机的软盘、只读存储器(Read-Only Memory，简称ROM)、随机存取存储器(Random Access Memory，简称RAM)、闪存(FLASH)、硬盘或光盘等，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述的图像分割方法。
值得注意的是，上述图像分割装置的实施例中，所包括的各个单元和模块只是按照功能逻辑进行划分的，但并不局限于上述的划分，只要能够实现相应的功能即可；另外，各功能单元的名称也只是为了便于相互区分。
本申请实施例提供的图像分割方法、装置、设备及存储介质的技术方案，通过已训练的图像分割模块对待分割图像进行图像分割以得到目标分割区域，该图像分割模型的解码单元被配置为计算当前高层特征对应的自注意力结果与该当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸，由于自注意力结果携带有待分割图像的上下文特征信息，因此基于自注意力结果与当前高层特征对应的低层特征确定初始通道注意力结果之后，该初始通道注意力结果可使用该上下文特征信息指导低层特征来获取像素的位置信息和类别信息，加之通道注意力结果不 仅包含初始通道注意力结果，还包含当前高层特征，这使得通道注意力结果更加准确并可适用于各种场景，从而使得该图像分割模型具有较高的普适性和鲁棒性。

Claims (9)

  1. 一种图像分割方法，包括：
    获取待分割图像；
    通过已训练的图像分割模型对所述待分割图像进行图像分割，以得到目标分割区域，其中，所述图像分割模型的解码单元被配置为计算当前高层特征对应的自注意力结果与所述当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
  2. 根据权利要求1所述的图像分割方法，其中，获取自注意力结果的方法包括：
    通过不同的卷积操作从当前高层特征中提取第一特征图、第二特征图和第三特征图；
    计算所述第一特征图和所述第二特征图的乘积，并对所述第一特征图和所述第二特征图的乘积结果执行分类操作以得到包含每个像素分类信息的分类结果；
    计算所述分类结果与所述第三特征图的乘积，以及所述分类结果与所述第三特征图的乘积结果与所述当前高层特征的乘积，以得到自注意力结果。
  3. 根据权利要求1所述的图像分割方法，其中，在计算初始通道注意力结果之前，所述图像分割方法还包括：
    对自注意力结果进行全局平均池化以更新所述自注意力结果，且更新之后的自注意力结果为K个1×1的特征图，其中K为通道数。
  4. 根据权利要求1-3任一所述的图像分割方法，其中，所述图像分割模型还包括特征提取单元和特征融合单元；
    所述特征提取单元被配置为对所述待分割图像进行特征提取以得到特征提取结果；
    所述特征融合单元被配置为通过多条并联通道分别从所述特征提取结果中提取相应尺度的特征图并对相应尺度的特征图进行压缩与解压缩操作，以及对所有通道输出的压缩与解压缩结果进行特征融合以得到特征融合结果。
  5. 根据权利要求4所述的图像分割方法，其中，所述特征融合单元的所有通道均通过SE块完成相应尺度的特征图的压缩与解压缩操作。
  6. 根据权利要求1所述的图像分割方法，其中，所述图像分割模型的损失 函数包括主函数和辅助函数；
    其中，所述辅助函数为权重交叉熵函数。
  7. 一种图像分割装置，包括：
    获取模块，被配置为获取待分割图像；
    输出模块，被配置为通过已训练的图像分割模型对所述待分割图像进行图像分割，以得到目标分割区域，其中，图像分割模型用于计算当前高层特征对应的自注意力结果与所述当前高层特征对应的低层特征的乘积以得到初始通道注意力结果，以及对所述初始通道注意力结果与所述当前高层特征进行组合以得到通道注意力结果，以及根据所述通道注意力结果更新所述当前高层特征，且更新后的当前高层特征的尺寸大于更新前的当前高层特征的尺寸。
  8. 一种图像分割设备，所述图像分割设备包括：
    处理器；
    存储装置，用于存储程序；
    当所述程序被所述处理器执行，使得所述处理器实现如权利要求1-6中任一所述的图像分割方法。
  9. 一种存储介质，包含计算机可执行指令，所述计算机可执行指令在由计算机处理器执行时用于执行如权利要求1-6中任一所述的图像分割方法。
PCT/CN2020/117826 2020-08-10 2020-09-25 图像分割方法、装置、设备及存储介质 WO2022032823A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010795220.3 2020-08-10
CN202010795220.3A CN111951280B (zh) 2020-08-10 2020-08-10 图像分割方法、装置、设备及存储介质

Publications (1)

Publication Number Publication Date
WO2022032823A1 true WO2022032823A1 (zh) 2022-02-17

Family

ID=73331896

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/117826 WO2022032823A1 (zh) 2020-08-10 2020-09-25 图像分割方法、装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN111951280B (zh)
WO (1) WO2022032823A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114862881A (zh) * 2022-07-11 2022-08-05 四川大学 一种基于pet-ct的跨模态注意力肿瘤分割方法、系统
CN115456064A (zh) * 2022-09-05 2022-12-09 江汉大学 基于点云的物体分类方法及相关设备
CN116311192A (zh) * 2023-05-15 2023-06-23 中国科学院长春光学精密机械与物理研究所 空间目标定位、区域超分辨重建及类型识别的系统及方法
CN117095177A (zh) * 2023-08-23 2023-11-21 脉得智能科技(无锡)有限公司 肾脏图像定位方法、装置和电子设备
WO2024011835A1 (zh) * 2022-07-15 2024-01-18 深圳先进技术研究院 一种图像处理方法、装置、设备及可读存储介质
CN117671509A (zh) * 2024-02-02 2024-03-08 武汉卓目科技有限公司 遥感目标检测方法、装置、电子设备及存储介质
CN117095177B (zh) * 2023-08-23 2024-06-04 脉得智能科技(无锡)有限公司 肾脏图像定位方法、装置和电子设备

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113159056B (zh) * 2021-05-21 2023-11-21 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
CN113326851B (zh) * 2021-05-21 2023-10-27 中国科学院深圳先进技术研究院 图像特征提取方法、装置、电子设备及存储介质
CN113837208B (zh) * 2021-10-18 2024-01-23 北京远鉴信息技术有限公司 一种异常图像的确定方法、装置、电子设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110443813A (zh) * 2019-07-29 2019-11-12 腾讯医疗健康(深圳)有限公司 血管、眼底图像的分割方法、装置、设备及可读存储介质
CN110866938A (zh) * 2019-11-21 2020-03-06 北京理工大学 一种全自动视频运动目标分割方法
US20200126209A1 (en) * 2018-10-18 2020-04-23 Nhn Corporation System and method for detecting image forgery through convolutional neural network and method for providing non-manipulation detection service using the same
CN111080655A (zh) * 2019-12-04 2020-04-28 腾讯科技（深圳）有限公司 图像分割及其模型训练方法、装置、介质及电子设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200126209A1 (en) * 2018-10-18 2020-04-23 Nhn Corporation System and method for detecting image forgery through convolutional neural network and method for providing non-manipulation detection service using the same
CN110443813A (zh) * 2019-07-29 2019-11-12 腾讯医疗健康(深圳)有限公司 血管、眼底图像的分割方法、装置、设备及可读存储介质
CN110866938A (zh) * 2019-11-21 2020-03-06 北京理工大学 一种全自动视频运动目标分割方法
CN111080655A (zh) * 2019-12-04 2020-04-28 腾讯科技（深圳）有限公司 图像分割及其模型训练方法、装置、介质及电子设备

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114862881A (zh) * 2022-07-11 2022-08-05 四川大学 一种基于pet-ct的跨模态注意力肿瘤分割方法、系统
WO2024011835A1 (zh) * 2022-07-15 2024-01-18 深圳先进技术研究院 一种图像处理方法、装置、设备及可读存储介质
CN115456064A (zh) * 2022-09-05 2022-12-09 江汉大学 基于点云的物体分类方法及相关设备
CN115456064B (zh) * 2022-09-05 2024-02-02 江汉大学 基于点云的物体分类方法及相关设备
CN116311192A (zh) * 2023-05-15 2023-06-23 中国科学院长春光学精密机械与物理研究所 空间目标定位、区域超分辨重建及类型识别的系统及方法
CN116311192B (zh) * 2023-05-15 2023-08-22 中国科学院长春光学精密机械与物理研究所 空间目标定位、区域超分辨重建及类型识别的系统及方法
CN117095177A (zh) * 2023-08-23 2023-11-21 脉得智能科技(无锡)有限公司 肾脏图像定位方法、装置和电子设备
CN117095177B (zh) * 2023-08-23 2024-06-04 脉得智能科技(无锡)有限公司 肾脏图像定位方法、装置和电子设备
CN117671509A (zh) * 2024-02-02 2024-03-08 武汉卓目科技有限公司 遥感目标检测方法、装置、电子设备及存储介质
CN117671509B (zh) * 2024-02-02 2024-05-24 武汉卓目科技有限公司 遥感目标检测方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN111951280A (zh) 2020-11-17
CN111951280B (zh) 2022-03-15

Similar Documents

Publication Publication Date Title
WO2022032823A1 (zh) 图像分割方法、装置、设备及存储介质
Li et al. ANU-Net: Attention-based nested U-Net to exploit full resolution features for medical image segmentation
US11861829B2 (en) Deep learning based medical image detection method and related device
US10706333B2 (en) Medical image analysis method, medical image analysis system and storage medium
WO2022032824A1 (zh) 图像分割方法、装置、设备及存储介质
CN111429421B (zh) 模型生成方法、医学图像分割方法、装置、设备及介质
CN110378913B (zh) 图像分割方法、装置、设备及存储介质
CN110738235B (zh) 肺结核判定方法、装置、计算机设备及存储介质
US11967181B2 (en) Method and device for retinal image recognition, electronic equipment, and storage medium
WO2023207743A1 (zh) 图像检测方法、装置、计算机设备、存储介质及程序产品
WO2021184799A1 (zh) 医学图像处理方法、装置、设备及存储介质
CN110570394A (zh) 医学图像分割方法、装置、设备及存储介质
Tan et al. Skin lesion recognition via global-local attention and dual-branch input network
CN111292299A (zh) 一种乳腺肿瘤识别方法、装置及存储介质
US20230103262A1 (en) Image processing method and device
CN115375787A (zh) 伪影校正方法、计算机设备及可读存储介质
CN116245832A (zh) 一种图像处理方法、装置、设备及存储介质
CN113610856B (zh) 训练图像分割模型和图像分割的方法和装置
CN113888566B (zh) 目标轮廓曲线确定方法、装置、电子设备以及存储介质
CN114419375A (zh) 图像分类方法、训练方法、装置、电子设备以及存储介质
CN115147360A (zh) 一种斑块分割方法、装置、电子设备及可读存储介质
CN114649092A (zh) 基于半监督学习和多尺度特征融合的辅助诊断方法及装置
CN113256622A (zh) 基于三维图像的目标检测方法、装置及电子设备
CN117726822B (zh) 基于双分支特征融合的三维医学图像分类分割系统及方法
CN117173488A (zh) 一种分类模型的确定方法、装置、电子设备及存储介质

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: 20949320

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20949320

Country of ref document: EP

Kind code of ref document: A1