WO2023102995A1 - 颅内出血区域检测方法、装置、计算机设备及存储介质 - Google Patents

颅内出血区域检测方法、装置、计算机设备及存储介质 Download PDF

Info

Publication number
WO2023102995A1
WO2023102995A1 PCT/CN2021/138657 CN2021138657W WO2023102995A1 WO 2023102995 A1 WO2023102995 A1 WO 2023102995A1 CN 2021138657 W CN2021138657 W CN 2021138657W WO 2023102995 A1 WO2023102995 A1 WO 2023102995A1
Authority
WO
WIPO (PCT)
Prior art keywords
image data
layer
module
hemorrhage area
intracranial
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/CN2021/138657
Other languages
English (en)
French (fr)
Inventor
吴剑煌
陈铭林
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 WO2023102995A1 publication Critical patent/WO2023102995A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/12—Edge-based segmentation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; 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
    • G06T2207/30016—Brain

Definitions

  • the present application relates to the technical field of medical detection, for example, to a detection method, device, computer equipment and storage medium of an intracranial hemorrhage area.
  • Intracranial hemorrhage is a type of intracranial hemorrhage that occurs in brain tissue or ventricles, often resulting in high morbidity and mortality.
  • Computed Tomography has become the first choice for the diagnosis of intracranial hemorrhage due to its advantages of rapidity, convenience, and sensitivity to fresh bleeding.
  • CT images Through the analysis of CT images, the bleeding site can be directly checked, and the amount of intracranial hemorrhage can be estimated, which has unique advantages in formulating operation plans or treatment plans.
  • Intracranial hemorrhage region segmentation is to judge whether each voxel belongs to the hemorrhage region.
  • Gradient-weighted Class Activation Mapping Grad-CAM
  • a more accurate ICH segmentation method is obtained by supervised training of full-volume machine networks, that is, through a U-Net-based approach to segment multiple types of ICH. Through the U-Net-based method, the intracranial hemorrhage area can be segmented separately on three planes (coronal, sagittal and horizontal planes) by combining test-time augmentation and model fusion.
  • the pathological morphology is similar to that of hemorrhage, which makes it very difficult to correctly identify the hemorrhage area.
  • the hemorrhage area is an area with a high Hounsfield unit (HU) value on CT, the transition area between the hemorrhage area and the brain tissue is often blurred, and the boundary between them is even difficult to be distinguished by experts.
  • HU Hounsfield unit
  • the part with high HU value is not only caused by hemorrhage, such as pineal gland calcification and lateral ventricle choroid plexus calcification will also show high HU value in CT images, which also makes it very difficult to segment the bleeding area , the methods used in the related art are also difficult to achieve a high accuracy rate.
  • the present application provides a method, device, computer equipment and storage medium for detecting an intracranial hemorrhage area, so as to improve the accuracy of intracranial hemorrhage detection results.
  • the application provides a method for detecting an intracranial hemorrhage area, the method comprising:
  • the intracranial three-dimensional hemorrhage area data are determined according to the prediction results of the hemorrhage area in the planar image data of each layer.
  • the present application also provides a device for detecting an intracranial hemorrhage area, the device comprising:
  • the image data acquisition module is configured to acquire intracranial computer tomography image data, and convert the computer tomography image data into three-dimensional image data;
  • the data sampling module is configured to sample the 3D image data according to preset intervals and preset resolutions to obtain multi-layer planar image data;
  • the prediction result obtaining module is configured to input each layer of planar image data into a trained deep learning network that fuses spatial features and context features, so as to obtain the prediction result of the hemorrhage area in each layer of planar image data;
  • the hemorrhage area determination module is configured to determine intracranial three-dimensional hemorrhage area data according to the prediction result of the hemorrhage area in the planar image data of each layer.
  • the present application also provides a kind of computer equipment, and this computer equipment comprises:
  • processors one or more processors
  • memory configured to store one or more programs
  • the one or more processors When the one or more programs are executed by the one or more processors, the one or more processors are made to implement the above method for detecting an intracranial hemorrhage area.
  • the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the above method for detecting an intracranial hemorrhage area is realized.
  • FIG. 1 is a flow chart of a method for detecting an intracranial hemorrhage area provided in Embodiment 1 of the present application;
  • FIG. 2 is a schematic structural diagram of an exemplary deep learning network provided in Embodiment 1 of the present application;
  • FIG. 3 is a schematic structural diagram of an intracranial hemorrhage area detection device provided in Embodiment 2 of the present application;
  • FIG. 4 is a schematic structural diagram of a computer device provided in Embodiment 3 of the present application.
  • FIG. 1 is a flow chart of a method for detecting an intracranial hemorrhage area provided in Embodiment 1 of the present application.
  • This embodiment is applicable to the detection of intracranial hemorrhage area using intracranial CT scan image data.
  • the method can be implemented by the intracranial hemorrhage area detection device provided in the embodiment of the present application.
  • the device can be implemented by means of hardware and/or software. To achieve, generally can be integrated in computer equipment. As shown in Figure 1, it includes the following steps:
  • Meta Header Data MHD
  • the electronic computer tomography image data of the brain can be obtained, and then the obtained electronic computer tomography image
  • the data is converted into 3D image data for subsequent processing.
  • the three-dimensional conversion method for data may be any method in related technologies, which is not limited in this embodiment.
  • the 3D image data can be sampled according to the preset interval and the preset resolution, so that the planar image data of the specified number of layers and the specified number of points can be obtained as required without being limited by the accuracy of the CT equipment.
  • the preset interval may be 10mm ⁇ 1mm ⁇ 1mm
  • the preset resolution may be 10 ⁇ 256 ⁇ 256.
  • the sampling process includes intra-layer sampling and inter-layer sampling, wherein intra-layer sampling may be performed using a bilinear interpolation algorithm, and inter-layer sampling may be performed using a nearest neighbor algorithm.
  • the obtained planar image data can also be truncated to a range of 0-90, and the data can be normalized to be between -1 and 1 to facilitate subsequent processing.
  • each layer of planar image data can be independently passed through a deep learning network that fuses spatial features and context features, so that more accurate hemorrhage in each layer of planar image data can be output through the deep learning network Forecast results for the area.
  • the deep learning network may be a convolutional neural network.
  • the deep learning network includes a spatial module, a context module, and an attention fusion module; the spatial module is used to extract spatial features in each layer of plane image data, the context module is used to capture context features, and the The attention fusion module is used to fuse the spatial features and the context features.
  • the spatial module includes multiple convolutional layers with a kernel size of 3 ⁇ 3.
  • the spatial module does not require a large receptive field, but needs to be more sensitive to specific spatial features. Shallow convolutional networks with fewer convolutional layers are able to restore finer spatial features well. Therefore, in this embodiment, a lightweight convolutional neural network can be selected as a spatial module to extract spatial detail features, and the extracted spatial detail features can be better used to determine the boundary between the hemorrhage area and normal tissue.
  • the lightweight convolutional neural network can be composed of multiple convolutional layers with a convolution kernel size of 3 ⁇ 3, and each convolutional layer can have a batch normalization layer and a linear rectification function (Rectified Linear Unit, ReLU) activation function layer, there can also be a maximum pooling layer between adjacent convolutional layers.
  • ReLU Rectified Linear Unit
  • the context module includes a deep residual network (Deep Residual Network, ResNet)-18 network or an EfficientNet network.
  • the context module is used to distinguish hemorrhagic areas from anatomical structures such as pineal calcifications, which have similar local representations in spatial detail.
  • the deep convolutional network has a strong ability to extract context information. Therefore, in this embodiment, a ResNet-18 network or an EfficientNet network can be used as a context module, and can be added after each convolutional layer. Batch normalization layer.
  • the attention fusion module is multi-layer, and the input of the current layer in the attention fusion module includes the output of the previous layer received by the current layer or the global mean pooling of the context module output;
  • the spatial features are fused with the context features, including:
  • S t-1 represents the output of the space module
  • C t-1 represents the output of the context module
  • a t-1 represents the output of the upper layer of the current layer in the attention fusion module
  • Concat represents Feature fusion
  • f means the first feature map
  • Conv(f) means passing the first feature map f through the convolutional layer
  • T means matrix transposition
  • reshape() means matrix reorganization
  • softmax means normalized exponential function
  • a t represents the output of the current layer of the attention fusion module
  • t is the identifier of the current layer of the attention fusion module.
  • the attention fusion module can be used to fuse them to output the corresponding prediction feature map, and the attention fusion module can use the above formula for fusion.
  • S t-1 , C t-1 and A t-1 can be concatenated in the channel dimension to form the first feature map f, and then three branches can be formed, and convolutional layers with a convolution kernel size of 3 ⁇ 3 can be used respectively
  • the ReLU activation function layer acts on the first feature map f to obtain feature maps f 1 , f 2 and f 3 .
  • the feature maps f1 and f2 can be transformed into corresponding dimensions by matrix reorganization and matrix transposition, and the second feature map Map can be obtained by matrix multiplication and softmax layer.
  • the second feature map Map can be regarded as a global attention map, which can include the correlation between any two positions in the image feature, so the global context information can be captured in the second feature map Map, so that the second feature map Map
  • the first feature map f can be guided to find spatially detailed bleeding region features.
  • the feature map f 3 is transformed into the corresponding dimension through matrix reorganization, and then combined with the second feature map Map through matrix multiplication.
  • a convolution layer with a convolution kernel size of 1 ⁇ 1, a batch normalization layer, and ReLU can be used.
  • the activation function layer acts on the combined output to obtain the output of the attention fusion module of the current layer.
  • a convolutional layer with a kernel size of 1 ⁇ 1 is used to match the number of channels to ensure that the number of channels output can match the next layer in the attention fusion module.
  • the structure of the deep learning network in this embodiment can be as shown in FIG. 2, and the spatial module includes five convolutional layers (Conv 3 ⁇ 3) with a convolution kernel size of 3 ⁇ 3.
  • the context module includes a convolution layer with a convolution kernel size of 7 ⁇ 7 (Conv 7 ⁇ 7) and four ResNet-18 networks (Res-1, Res-2, Res-3 and Res-4), and correspond to Receive the output of multiple convolutional layers of the spatial module, and at the same time, the output of each layer of the spatial module is the input of the next layer of the spatial module, and finally output the global mean pooling.
  • the number of Attentional Factorization Machines (AFM) included in the attention fusion module is five layers, and corresponds to the multi-layer output of the receiving space module and the context module.
  • AMF Attentional Factorization Machines
  • the last layer of AFM receives the global mean pooling of the output of the context module. and pass the output up layer by layer. Finally, the output of the AFM on the top layer passes through the convolutional layer with a convolution kernel size of 3 ⁇ 3 and a convolution kernel size of 1 ⁇ 1 (Conv 3 ⁇ 3 and Conv 1 ⁇ 1) to obtain the final prediction result.
  • the number of filters from top to bottom of the five convolutional layers of the spatial module can be 64, 64, 128, 256, and 512, respectively.
  • the input of each layer of planar image data into the trained deep learning network that fuses spatial features and context features to obtain the prediction result of the hemorrhage area in each layer of planar image data includes: through the deep learning The network outputs the prediction result of each point in the plane image data of each layer; compares the prediction result of each point with a preset threshold, and binarizes the plane image data of each layer according to the comparison result to obtain the Describe the prediction results of the hemorrhage area in the planar image data of each layer.
  • the input of the deep learning network can be the obtained planar image data of each layer, and the output can be the prediction result of each point in the planar image data of each layer, which can be the probability that each point is a bleeding point.
  • the prediction result corresponding to each point can be compared with the preset threshold.
  • the preset threshold can be a preset probability threshold. When the prediction result is greater than the preset threshold, the corresponding The point is determined as a bleeding point, and this point can be set as 1. When the prediction result is less than or equal to the preset threshold, the corresponding point can be determined as a normal background point, and this point can be set as 0. After the determination of all points in the plane image data of one layer is completed, the prediction result of the hemorrhage area in the plane image data of this layer can be obtained.
  • the hemorrhage area predicted by multiple layers can be stacked, and the three-dimensional hemorrhage area data can be re-formed according to the same preset interval.
  • the trained deep learning network of the fusion of spatial features and context features into the planar image data of each layer, so as to obtain the prediction of the hemorrhage area in the planar image data of each layer
  • the loss function used for training is the Dice function.
  • the Dice function can be used as the loss function, the formula is:
  • £ Dice(p,y) represents the loss function
  • p represents the output of the deep learning network
  • y represents the real bleeding label
  • 2 represents the 2-norm of the matrix
  • T represents the matrix transposition. Due to the problem of class imbalance in the intracranial hemorrhage data set, that is, the number of background pixels is much greater than the number of hemorrhage pixels, and the number of non-bleeding pixels can be more than 100 times the number of hemorrhages. If the cross-entropy function is used as the loss function, the network will fall into a local maximum An optimal solution such that the network always predicts the background without any bleeding pixels. By using the Dice function as the loss function, this class imbalance problem can be well solved.
  • the intracranial computerized tomography image data is obtained and converted into three-dimensional image data, and then the three-dimensional image data is sampled according to a preset interval and a preset resolution to obtain multiple Each layer of planar image data, and then input each layer of planar image data into the trained deep learning network that integrates spatial features and context features to obtain the prediction result of the bleeding area in each layer of planar image data, and finally according to each layer of planar image data
  • the prediction results of the hemorrhage area in the middle determine the three-dimensional intracranial hemorrhage area data.
  • Fig. 3 is a schematic structural diagram of an intracranial hemorrhage area detection device provided in Embodiment 2 of the present application.
  • the device can be realized by hardware and/or software, and generally can be integrated into a computer device, and is configured to execute any implementation of the present application.
  • the detection method of the intracranial hemorrhage area provided by the example. As shown in Figure 3, the device includes:
  • the image data acquiring module 31 is configured to acquire intracranial computerized tomography image data, and convert the computerized tomography image data into three-dimensional image data;
  • the data sampling module 32 is configured to The three-dimensional image data is sampled at a high rate to obtain multi-layer planar image data;
  • the prediction result obtaining module 33 is configured to input each layer of planar image data into a trained deep learning network that fuses spatial features and context features to obtain the obtained The prediction result of the hemorrhage area in the planar image data of each layer is described;
  • the bleeding area determination module 34 is configured to determine the three-dimensional intracranial hemorrhage area data according to the prediction result of the hemorrhage area in the planar image data of each layer.
  • the intracranial computerized tomography image data is obtained and converted into three-dimensional image data, and then the three-dimensional image data is sampled according to a preset interval and a preset resolution to obtain multiple Each layer of planar image data, and then input each layer of planar image data into the trained deep learning network that integrates spatial features and context features to obtain the prediction result of the bleeding area in each layer of planar image data, and finally according to each layer of planar image data
  • the prediction results of the hemorrhage area in the middle determine the three-dimensional intracranial hemorrhage area data.
  • the deep learning network includes a spatial module, a context module, and an attention fusion module; the spatial module is used to extract spatial features in each layer of plane image data, and the context module is used to capture context features, and the attention fusion module is used to fuse the spatial features and the context features.
  • the spatial module includes a plurality of convolution layers with a convolution kernel size of 3 ⁇ 3.
  • the context module includes a ResNet-18 network or an EfficientNet network.
  • the attention fusion module is multi-layered, and the input of the current layer in the attention fusion module includes the current layer receiving the output of the previous layer or the context module Global mean pooling of the output; the attention fusion module is used for:
  • S t-1 represents the output of the space module
  • C t-1 represents the output of the context module
  • a t-1 represents the output of the upper layer of the current layer in the attention fusion module
  • Concat represents Feature fusion
  • f represents the first feature map
  • Conv(f) represents passing the first feature map f through the convolutional layer
  • () T represents matrix transposition
  • reshape() represents matrix reorganization
  • softmax represents normalized exponential function
  • a t represents the output of the current layer of the attention fusion module
  • t is the identifier of the current layer of the attention fusion module.
  • the intracranial hemorrhage area detection device further includes:
  • the training module is configured to, before inputting each layer of planar image data into the trained deep learning network that fuses spatial features and context features, to obtain the prediction result of the hemorrhage area in each layer of planar image data, the depth The learning network is trained, and the loss function used for training is the Dice function.
  • the prediction result obtaining module 33 is set to:
  • the intracranial hemorrhage region detection device provided in the embodiment of the present application can execute the intracranial hemorrhage region detection method provided in any embodiment of the present application, and has corresponding functional modules and effects for executing the method.
  • the multiple units and modules included are only divided according to functional logic, but are not limited to the above-mentioned division, as long as the corresponding functions can be realized; in addition, multiple The names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the protection scope of the present application.
  • FIG. 4 is a schematic structural diagram of a computer device provided in Embodiment 3 of the present application, showing a block diagram of an exemplary computer device suitable for implementing the embodiment of the present application.
  • the computer device shown in FIG. 4 is only an example, and should not limit the functions and scope of use of this embodiment of the present application.
  • the computer equipment includes a processor 41, a memory 42, an input device 43 and an output device 44; the number of processors 41 in the computer equipment can be one or more, and one processor 41 is taken as an example in Figure 4 , the processor 41, the memory 42, the input device 43 and the output device 44 in the computer equipment may be connected via a bus or in other ways. In FIG. 4, the connection via a bus is taken as an example.
  • the memory 42 can be configured to store software programs, computer-executable programs and modules, such as program instructions/modules corresponding to the intracranial hemorrhage area detection method in the embodiment of the present application (for example, intracranial hemorrhage area detection Image data acquisition module 31, data sampling module 32, prediction result acquisition module 33 and bleeding area determination module 34) in the device.
  • the processor 41 executes various functional applications and data processing of the computer equipment by running the software programs, instructions and modules stored in the memory 42 , that is, realizes the above-mentioned intracranial hemorrhage area detection method.
  • the memory 42 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required by a function; the data storage area may store data created according to the use of the computer device, and the like.
  • the memory 42 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage devices.
  • memory 42 may include memory located remotely from processor 41 , and such remote memory may be connected to the computer device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
  • the input device 43 can be configured to acquire intracranial computerized tomography image data, and generate key signal input related to user settings and function control of the computer equipment.
  • the output device 44 includes devices such as a display screen, which can be set to display the prediction result of the final intracranial hemorrhage area to the user.
  • Embodiment 4 of the present application also provides a storage medium containing computer-executable instructions, the computer-executable instructions are used to execute a method for detecting an intracranial hemorrhage area when executed by a computer processor, the method comprising:
  • intracranial computerized tomography image data and converting the computerized tomography image data into three-dimensional image data; sampling the three-dimensional image data according to preset intervals and preset resolutions to obtain multi-layer planes Image data; each layer of plane image data is input into a trained deep learning network that fuses spatial features and context features to obtain the prediction result of the hemorrhage area in the plane image data of each layer; according to the hemorrhage area in each layer of plane image data The prediction results determine the intracranial three-dimensional hemorrhage area data.
  • a storage medium may be any of various types of memory devices or storage devices.
  • the term "storage medium” is intended to include: installation media such as Compact Disc Read-Only Memory (CD-ROM), floppy disks, or tape drives; computer system memory or random access memory, such as dynamic random access memory (Dynamic Random Access Memory, DRAM), Double Data Rate Random Access Memory (Double Data Rate, DDR RAM), Static Random Access Memory (Static RAM, SRAM), Extended Data Output Random Access Memory (Extended Data Output , EDO RAM), Rambus (Rambus) RAM, etc.; non-volatile memory, such as flash memory, magnetic media (such as hard disk or optical storage); registers or other similar types of memory elements, etc.
  • installation media such as Compact Disc Read-Only Memory (CD-ROM), floppy disks, or tape drives
  • computer system memory or random access memory such as dynamic random access memory (Dynamic Random Access Memory, DRAM), Double Data Rate Random Access Memory (Double Data Rate, DDR RAM),
  • the storage medium may also include other types of memory or combinations thereof. Also, the storage medium may be located in a computer system in which the program is executed, or may be located in a different second computer system connected to the computer system through a network such as the Internet. The second computer system may provide program instructions to the computer for execution.
  • the term "storage medium" may include two or more storage media that may reside in different locations, such as in different computer systems connected by a network.
  • a storage medium may store program instructions (eg, implemented as a computer program) that are executable by one or more processors.
  • a storage medium containing computer-executable instructions provided in an embodiment of the present application the computer-executable instructions are not limited to the method operations described above, and may also perform the steps in the intracranial hemorrhage area detection method provided in any embodiment of the present application. related operations.
  • a computer readable signal medium may include a data signal carrying computer readable program code in baseband or as part of a carrier wave. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. .
  • the program code contained on the computer readable medium can be transmitted by any appropriate medium, including but not limited to wireless, electric wire, optical cable, radio frequency (Radio Frequency, RF), etc., or any suitable combination of the above.
  • any appropriate medium including but not limited to wireless, electric wire, optical cable, radio frequency (Radio Frequency, RF), etc., or any suitable combination of the above.
  • the present application can be implemented by means of software and necessary general-purpose hardware, or can also be implemented by means of hardware.
  • the technical solution of the present application can be embodied in the form of a software product in essence, and the computer software product can be stored in a computer-readable storage medium, such as a computer's floppy disk, read-only memory (Read-Only Memory, ROM), RAM, flash memory (FLASH), hard disk or optical disk, etc., including a plurality of instructions to enable a computer device (which may be a personal computer, server, or network device, etc.) to execute the method described in the embodiment of the present application.
  • a computer-readable storage medium such as a computer's floppy disk, read-only memory (Read-Only Memory, ROM), RAM, flash memory (FLASH), hard disk or optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

本文公开了一种颅内出血区域检测方法、装置、计算机设备及存储介质。该颅内出血区域检测方法包括：获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据；根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据；将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果；根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。

Description

颅内出血区域检测方法、装置、计算机设备及存储介质
本申请要求在2021年12月08日提交中国专利局、申请号为202111491301.5的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本申请涉及医学检测技术领域，例如涉及一种颅内出血区域检测方法、装置、计算机设备及存储介质。
背景技术
颅内出血是发生在脑组织或脑室内的一种颅内出血，常导致高致残率和死亡率。电子计算机断层扫描(Computed Tomography，CT)因其快速、方便、对新鲜出血敏感等优点，成为诊断颅内出血的首选。通过对CT图像的分析，可以直接检查出血部位，估计颅内出血量，在制定手术方案或治疗方案时具有独特的优势。
颅内出血区域分割是对每个体素进行判断是否属于出血区域，相关技术中使用梯度类加权激活映射(Gradient-weighted Class Activation Mapping，Grad-CAM)技术来获得颅内出血区域的显著图，然而这种方法获得的显著图过于粗糙，仅仅可以用来作为一种解释卷积神经网络的指导性工具。一种更准确的颅内出血分割方法是通过有监督训练全卷机网络获得，即通过基于U-Net的方法来分割多种类型的颅内出血。通过基于U-Net的方法，可以结合测试时增强和模型融合等方式来在三个面(冠状面、矢状面和水平面)上分别对颅内出血区域进行分割。
但是由于在颅内出血的分割中边界不确定，病理形态与出血相似，使得正确识别出血区域非常困难。虽然出血区域是CT上Hounsfield单位(HU)值较高的区域，但出血区域与脑组织之间的过度区域往往是模糊的，它们之间的边界甚至很难被专家所分辨。此外，HU值高的部位不仅仅是因为出血所导致，比如松果体钙化和侧脑室脉络丛钙化等也会在CT影像中呈现较高的HU值，这也使得分割出血区域变得十分困难，相关技术中所使用的方法也很难达到较高的准确率。
发明内容
本申请提供一种颅内出血区域检测方法、装置、计算机设备及存储介质，以提高颅内出血检测结果的准确率。
本申请提供了一种颅内出血区域检测方法，该方法包括：
获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据；
根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据；
将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果；
根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。
本申请还提供了一种颅内出血区域检测装置，该装置包括：
影像数据获取模块，设置为获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据；
数据采样模块，设置为根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据；
预测结果获得模块，设置为将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果；
出血区域确定模块，设置为根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。
本申请还提供了一种计算机设备，该计算机设备包括：
一个或多个处理器；
存储器，设置为存储一个或多个程序；
当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现上述的颅内出血区域检测方法。
本申请还提供了一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现上述的颅内出血区域检测方法。
附图说明
图1为本申请实施例一提供的一种颅内出血区域检测方法的流程图；
图2为本申请实施例一提供的一种深度学习网络的示例性结构示意图；
图3为本申请实施例二提供的一种颅内出血区域检测装置的结构示意图；
图4为本申请实施例三提供的一种计算机设备的结构示意图。
具体实施方式
下面结合附图和实施例对本申请进行说明。此处所描述的具体实施例仅仅用于解释本申请。为了便于描述，附图中仅示出了与本申请相关的部分。
在讨论示例性实施例之前应当提到的是，一些示例性实施例被描述成作为流程图描绘的处理或方法。虽然流程图将多个步骤描述成顺序的处理，但是其中的许多步骤可以被并行地、并发地或者同时实施。此外，多个步骤的顺序可以被重新安排。当其操作完成时所述处理可以被终止，但是还可以具有未包括在附图中的附加步骤。所述处理可以对应于方法、函数、规程、子例程、子程序等等。
实施例一
图1为本申请实施例一提供的一种颅内出血区域检测方法的流程图。本实施例可适用于使用颅内CT扫描影像数据对颅内出血区域检测的情况，该方法可以由本申请实施例所提供的颅内出血区域检测装置来执行，该装置可以由硬件和/或软件的方式来实现，一般可集成于计算机设备中。如图1所示，包括如下步骤：
S11、获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据。
可以通过读取CT设备在完成扫描后所获得的元数据头数据(Meta Header Data，MHD)格式的数据文件，来获取颅内的电子计算机断层扫描影像数据，然后将获得的电子计算机断层扫描影像数据转换为三维影像数据，以便于后续的处理。对于数据的三维转换方法可以是相关技术中的任意方法，在本实施例不作限制。
S12、根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据。
在获得了三维影像数据之后，可以根据预设间隔和预设分辨率对三维影像数据进行采样，从而可以根据需要获得指定层数和指定点数的平面影像数据，而无需受到CT设备的精度限制。其中，预设间隔可以是10mm×1mm×1mm，预设分辨率可以是10×256×256。可选的，采样过程包括层内采样及层间采样，其中的层内采样可以采用双线性插值算法进行采样，而层间采样可以采用最近邻算法进行采样。在完成采样过程之后，还可以将获得的平面影像数据截断至0-90的范围内，并将数据归一化至-1到1之间，以便于后续的处理。
S13、将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果。
在获得多层平面影像数据之后，即可将每一层平面影像数据独立的通过融合空间特征和上下文特征的深度学习网络，从而通过该深度学习网络输出每层平面影像数据中更为精确的出血区域的预测结果。其中，深度学习网络可以是卷积神经网络。
可选的，所述深度学习网络包括空间模块、上下文模块以及注意力融合模块；所述空间模块用于提取每层平面影像数据中的空间特征，所述上下文模块用于捕获上下文特征，所述注意力融合模块用于对所述空间特征和所述上下文特征进行融合。
所述空间模块包括多个卷积核大小为3×3的卷积层。空间模块不需要很大的感受野，但需要对特定的空间特征更为敏感。浅层卷积网络具有较少的卷积层，能够很好地还原更精细的空间特征。因此，在本实施例中，可以选用轻量的卷积神经网络作为空间模块来提取空间细节特征，所提取的空间细节特征可以更好地用于确定出血区域与正常组织之间的边界。该轻量的卷积神经网络即可以是由多个卷积核大小为3×3的卷积层组成，并且，每个卷积层可以带有批量归一化层和线性整流函数(Rectified Linear Unit,ReLU)激活函数层，在相邻的卷积层之间还可以有一个最大池化层。
所述上下文模块包括深度残差网络(Deep Residual Network，ResNet)-18网络或EfficientNet网络。上下文模块用于区分出血区域和松果体钙化等解剖结构，出血区域和松果体钙化等解剖结构在空间细节上具有相似的局部表现。而深度卷积网络在提取上下文信息方面具有很强的能力，因此，在本实施例中，可以使用ResNet-18网络或EfficientNet网络等作为上下文模块，并且可以在其中的每个卷积层之后添加批量归一化层。
所述注意力融合模块为多层，且所述注意力融合模块中的当前层的输入包括当前层接收的上一层的输出或所述上下文模块输出的全局均值池化；所述对所述空间特征和所述上下文特征进行融合，包括：
f＝Concat(S t-1,C t-1,A t-1)
Map＝softmax((reshape(Conv(f))) T*reshape(Conv(f)))
A t＝Conv(reshape(reshape(Conv(f))*Map))
其中，S t-1表示所述空间模块的输出，C t-1表示所述上下文模块的输出，A t-1表示所述注意力融合模块中的当前层的上一层的输出，Concat表示特征融合，f 表示第一特征图，Conv(f)表示将第一特征图f通过卷积层，() T表示矩阵转置，reshape()表示矩阵重组，softmax表示归一化指数函数，Map表示第二特征图，A t表示所述注意力融合模块的当前层的输出，t为所述注意力融合模块的当前层的标识。
在提取出每个尺度上的空间特征和上下文特征后，可以采用注意力融合模块对其进行融合，以输出相应的预测特征图，则注意力融合模块可以采用上述公式进行融合。首先可以在通道维度上拼接S t-1、C t-1和A t-1以形成第一特征图f，然后可以形成三条分支，并分别使用卷积核大小为3×3的卷积层以及ReLU激活函数层作用在第一特征图f上，以得到特征图f 1、f 2和f 3。接着可以通过矩阵重组和矩阵转置将特征图f 1和f 2转换为相应的维度，并可通过矩阵乘法和softmax层获得第二特征图Map。第二特征图Map可以看成是全局注意力图，可以包括图像特征中任意两个位置之间的相关性，因此可以在第二特征图Map中捕捉到全局上下文信息，从而使用第二特征图Map可以引导第一特征图f去寻找空间细节的出血区域特征。然后特征图f 3通过矩阵重组转换为相应的维度，再通过矩阵乘法和第二特征图Map结合，最后，可以使用卷积核大小为1×1的卷积层、批量归一化层及ReLU激活函数层作用在结合后的输出上，以得到当前层注意力融合模块的输出。其中，卷积核大小为1×1的卷积层用于匹配通道数量，以确保输出的通道数能够匹配注意力融合模块中的下一层。
示例性的，本实施例中深度学习网络的结构可以如图2所示，空间模块包括五个卷积核大小为3×3的卷积层(Conv 3×3)。上下文模块包括一个卷积核大小为7×7的卷积层(Conv 7×7)和四个ResNet-18网络(Res-1、Res-2、Res-3和Res-4)，并分别对应接收空间模块的多个卷积层的输出，同时收空间模块的每层的输出又是收空间模块的下一层的输入，并最终输出全局均值池化。注意力融合模块包括的注意因子分解机(Attentional Factorization Machines，AFM)的数量为五层，并对应接收空间模块和上下文模块的多层输出，最后一层AFM接收上下文模块输出的全局均值池化，并将输出逐层向上传递。最后，最上层的AFM的输出再依次通过卷积核大小为3×3和卷积核大小为1×1的卷积层(Conv 3×3和Conv 1×1)，即可获得最后的预测结果。其中，空间模块的五个卷积层从上到下的滤波器数量分别可以是64、64、128、256和512。
可选的，所述将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果，包括：通过所述深度学习网络输出所述每层平面影像数据中每个点的预测结果；将每个点的预测结果与预设阈值进行比较，并根据比较结果将所述每层平面影像数据二值化，以得到所述每层平面影像数据中出血区域的预测结果。
深度学习网络的输入可以是获得的每层平面影像数据，输出可以是该每层平面影像数据中每个点的预测结果，可以是每个点为出血点的概率。在获得深度学习网络输出的预测结果之后，可以将每个点对应的预测结果与预设阈值进行比较，预设阈值可以是预设概率阈值，当预测结果大于预设阈值时，即可将对应的点判定为出血点，并可将该点设置为1，当预测结果小于或等于预设阈值时，则可以将对应的点判定为正常背景点，并可将该点设置为0。在完成了一层平面影像数据中所有点的判定之后，即可得到该层平面影像数据中出血区域的预测结果。
S14、根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。
在得到每层平面影像数据中出血区域的预测结果之后，可以将多层预测得到的出血区域层叠起来，即可根据同样的预设间隔重新形成三维出血区域数据。
在上述技术方案的基础上，可选的，在所述将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果之前，还包括：对所述深度学习网络进行训练，训练所使用的损失函数为Dice函数。在对深度学习网络进行训练的过程中，可以使用Dice函数作为损失函数，公式为：
Figure PCTCN2021138657-appb-000001
其中，￡ Dice(p,y)表示损失函数，p表示深度学习网络的输出，y表示真实出血标注，|||| 2表示矩阵的2-范数，() T表示矩阵转置。由于颅内出血数据集存在类不平衡问题，即背景像素数量远大于出血像素数量，非出血的数量可以是出血数量的100多倍，如果使用交叉熵函数作为损失函数，则会使网络陷入局部最优解，使得网络总是预测为背景而没有任何出血像素。通过使用Dice函数作为损失函数，可以很好的解决这种类不平衡问题。
本申请实施例所提供的技术方案，首先获取颅内的电子计算机断层扫描影像数据，并将其转换为三维影像数据，然后根据预设间隔和预设分辨率对三维影像数据进行采样以得到多层平面影像数据，再将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到每层平面影像数据中出血区域的预测结果，最后即可根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。通过使用融合了空间特征的深度学习网络，可以区分出血区域与脑组织之间的模糊边界，通过使用融合了上下文全局特征的深度学习网络，可以更好地区分出血区域与正常组织，从而增强了颅内出血检测的性能，提高了颅内出血检测结果的准确率。
实施例二
图3为本申请实施例二提供的一种颅内出血区域检测装置的结构示意图，该装置可以由硬件和/或软件的方式来实现，一般可集成于计算机设备中，设置为执行本申请任意实施例所提供的颅内出血区域检测方法。如图3所示，该装置包括：
影像数据获取模块31，设置为获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据；数据采样模块32，设置为根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据；预测结果获得模块33，设置为将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果；出血区域确定模块34，设置为根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。
本申请实施例所提供的技术方案，首先获取颅内的电子计算机断层扫描影像数据，并将其转换为三维影像数据，然后根据预设间隔和预设分辨率对三维影像数据进行采样以得到多层平面影像数据，再将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到每层平面影像数据中出血区域的预测结果，最后即可根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。通过使用融合了空间特征的深度学习网络，可以区分出血区域与脑组织之间的模糊边界，通过使用融合了上下文全局特征的深度学习网络，可以更好地区分出血区域与正常组织，从而增强了颅内出血检测的性能，提高了颅内出血检测结果的准确率。
在上述技术方案的基础上，可选的，所述深度学习网络包括空间模块、上下文模块以及注意力融合模块；所述空间模块用于提取每层平面影像数据中的空间特征，所述上下文模块用于捕获上下文特征，所述注意力融合模块用于对所述空间特征和所述上下文特征进行融合。
在上述技术方案的基础上，可选的，所述空间模块包括多个卷积核大小为3×3的卷积层。
在上述技术方案的基础上，可选的，所述上下文模块包括ResNet-18网络或EfficientNet网络。
在上述技术方案的基础上，可选的，所述注意力融合模块为多层，且所述注意力融合模块中的当前层的输入包括当前层接收额上一层的输出或所述上下文模块输出的全局均值池化；注意力融合模块用于：
f＝Concat(S t-1,C t-1,A t-1)
Map＝softmax((reshape(Conv(f))) T*reshape(Conv(f)))
A t＝Conv(reshape(reshape(Conv(f))*Map))
其中，S t-1表示所述空间模块的输出，C t-1表示所述上下文模块的输出，A t-1表示所述注意力融合模块中的当前层的上一层的输出，Concat表示特征融合，f表示第一特征图，Conv(f)表示将第一特征图f通过卷积层，() T表示矩阵转置，reshape()表示矩阵重组，softmax表示归一化指数函数，Map表示第二特征图，A t表示所述注意力融合模块当前层的的输出，t为所述注意力融合模块的当前层的标识。
在上述技术方案的基础上，可选的，该颅内出血区域检测装置，还包括：
训练模块，设置为在所述将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果之前，对所述深度学习网络进行训练，训练所使用的损失函数为Dice函数。
在上述技术方案的基础上，可选的，预测结果获得模块33设置为：
通过所述深度学习网络输出所述每层平面影像数据中每个点的预测结果；将每个点的预测结果与预设阈值进行比较，并根据比较结果将所述每层平面影像数据二值化，以得到所述每层平面影像数据中出血区域的预测结果。
本申请实施例所提供的颅内出血区域检测装置可执行本申请任意实施例所提供的颅内出血区域检测方法，具备执行方法相应的功能模块和效果。
在上述颅内出血区域检测装置的实施例中，所包括的多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，多个功能单元的名称也只是为了便于相互区分，并不用于限制本申请的保护范围。
实施例三
图4为本申请实施例三提供的一种计算机设备的结构示意图，示出了适于用来实现本申请实施方式的示例性计算机设备的框图。图4显示的计算机设备仅仅是一个示例，不应对本申请实施例的功能和使用范围带来任何限制。如图4所示，该计算机设备包括处理器41、存储器42、输入装置43及输出装置44；计算机设备中处理器41的数量可以是一个或多个，图4中以一个处理器41为 例，计算机设备中的处理器41、存储器42、输入装置43及输出装置44可以通过总线或其他方式连接，图4中以通过总线连接为例。
存储器42作为一种计算机可读存储介质，可设置为存储软件程序、计算机可执行程序以及模块，如本申请实施例中的颅内出血区域检测方法对应的程序指令/模块(例如，颅内出血区域检测装置中的影像数据获取模块31、数据采样模块32、预测结果获得模块33及出血区域确定模块34)。处理器41通过运行存储在存储器42中的软件程序、指令以及模块，从而执行计算机设备的多种功能应用以及数据处理，即实现上述的颅内出血区域检测方法。
存储器42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计算机设备的使用所创建的数据等。此外，存储器42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42可包括相对于处理器41远程设置的存储器，这些远程存储器可以通过网络连接至计算机设备。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
输入装置43可设置为获取颅内的电子计算机断层扫描影像数据，以及产生与计算机设备的用户设置和功能控制有关的键信号输入等。输出装置44包括显示屏等设备，可设置为向用户展示最终颅内出血区域的预测结果。
实施例四
本申请实施例四还提供一种包含计算机可执行指令的存储介质，该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颅内出血区域检测方法，该方法包括：
获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据；根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据；将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果；根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。
存储介质可以是任何的多种类型的存储器设备或存储设备。术语“存储介质”旨在包括：安装介质，例如光盘只读存储器(Compact Disc Read-Only Memory，CD-ROM)、软盘或磁带装置；计算机系统存储器或随机存取存储器，诸如动态随机存取存储器(Dynamic Random Access Memory，DRAM)、双倍数据速 率随机存取存储器(双倍数据速率，DDR RAM)、静态随机存取存储器(Static RAM，SRAM)、扩展数据输出随机存取存储器(Extended Data Output，EDO RAM)，兰巴斯(Rambus)RAM等；非易失性存储器，诸如闪存、磁介质(例如硬盘或光存储)；寄存器或其它相似类型的存储器元件等。存储介质可以还包括其它类型的存储器或其组合。另外，存储介质可以位于程序在其中被执行的计算机系统中，或者可以位于不同的第二计算机系统中，第二计算机系统通过网络(诸如因特网)连接到计算机系统。第二计算机系统可以提供程序指令给计算机用于执行。术语“存储介质”可以包括可以驻留在不同位置中(例如在通过网络连接的不同计算机系统中)的两个或更多存储介质。存储介质可以存储可由一个或多个处理器执行的程序指令(例如实现为计算机程序)。
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不限于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颅内出血区域检测方法中的相关操作。
计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。
计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于无线、电线、光缆、射频(Radio Frequency，RF)等等，或者上述的任意合适的组合。
通过以上关于实施方式的描述，所属领域的技术人员可以了解到，本申请可借助软件及必需的通用硬件来实现，也可以通过硬件实现。本申请的技术方案本质上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，ROM)、RAM、闪存(FLASH)、硬盘或光盘等，包括多个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)执行本申请实施例所述的方法。

Claims (10)

  1. 一种颅内出血区域检测方法，包括：
    获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据；
    根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据；
    将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果；
    根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。
  2. 根据权利要求1所述的颅内出血区域检测方法，其中，所述深度学习网络包括空间模块、上下文模块以及注意力融合模块；所述空间模块用于提取每层平面影像数据中的空间特征，所述上下文模块用于捕获上下文特征，所述注意力融合模块用于对所述空间特征和所述上下文特征进行融合。
  3. 根据权利要求2所述的颅内出血区域检测方法，其中，所述空间模块包括多个卷积核大小为3×3的卷积层。
  4. 根据权利要求2所述的颅内出血区域检测方法，其中，所述上下文模块包括深度残差网络ResNet-18网络或EfficientNet网络。
  5. 根据权利要求2所述的颅内出血区域检测方法，其中，所述注意力融合模块为多层，且所述注意力融合模块中的当前层的输入包括当前层接收的上一层的输出或所述上下文模块输出的全局均值池化；
    所述对所述空间特征和所述上下文特征进行融合，包括：
    f＝Concat(S t-1,C t-1,A t-1)；
    Map＝softmax((reshape(Conv(f))) T*reshape(Conv(f)))；
    A t＝Conv(reshape(reshape(Conv(f))*Map))；
    其中，S t-1表示所述空间模块的输出，C t-1表示所述上下文模块的输出，A t-1表示所述注意力融合模块中的当前层的上一层的输出，Concat表示特征融合，f表示第一特征图，Conv(f)表示将第一特征图f通过卷积层，() T表示矩阵转置，reshape()表示矩阵重组，softmax表示归一化指数函数，Map表示第二特征图，A t表示所述注意力融合模块的当前层的输出，t为所述注意力融合模块的当前层的标识。
  6. 根据权利要求1所述的颅内出血区域检测方法，在所述将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果之前，还包括：
    对所述深度学习网络进行训练，其中，训练所使用的损失函数为Dice函数。
  7. 根据权利要求1所述的颅内出血区域检测方法，其中，所述将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果，包括：
    通过所述深度学习网络输出所述每层平面影像数据中每个点的预测结果；
    将每个点的预测结果与预设阈值进行比较，并根据比较结果将所述每层平面影像数据二值化，以得到所述每层平面影像数据中出血区域的预测结果。
  8. 一种颅内出血区域检测装置，包括：
    影像数据获取模块，设置为获取颅内的电子计算机断层扫描影像数据，并将所述电子计算机断层扫描影像数据转换为三维影像数据；
    数据采样模块，设置为根据预设间隔和预设分辨率对所述三维影像数据进行采样，以得到多层平面影像数据；
    预测结果获得模块，设置为将每层平面影像数据输入训练后的融合空间特征和上下文特征的深度学习网络，以得到所述每层平面影像数据中出血区域的预测结果；
    出血区域确定模块，设置为根据每层平面影像数据中出血区域的预测结果确定颅内的三维出血区域数据。
  9. 一种计算机设备，包括：
    至少一个处理器；
    存储器，设置为存储至少一个程序；
    当所述至少一个程序被所述至少一个处理器执行，使得所述至少一个处理器实现如权利要求1-7中任一项所述的颅内出血区域检测方法。
  10. 一种计算机可读存储介质，存储有计算机程序，其中，所述程序被处理器执行时实现如权利要求1-7中任一项所述的颅内出血区域检测方法。
PCT/CN2021/138657 2021-12-08 2021-12-16 颅内出血区域检测方法、装置、计算机设备及存储介质 Ceased WO2023102995A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111491301.5A CN114155232A (zh) 2021-12-08 2021-12-08 颅内出血区域检测方法、装置、计算机设备及存储介质
CN202111491301.5 2021-12-08

Publications (1)

Publication Number Publication Date
WO2023102995A1 true WO2023102995A1 (zh) 2023-06-15

Family

ID=80454015

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/138657 Ceased WO2023102995A1 (zh) 2021-12-08 2021-12-16 颅内出血区域检测方法、装置、计算机设备及存储介质

Country Status (2)

Country Link
CN (1) CN114155232A (zh)
WO (1) WO2023102995A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118015012A (zh) * 2024-01-15 2024-05-10 宁波大学 一种脑部mri图像脑微出血自动分割方法
CN118154591A (zh) * 2024-05-09 2024-06-07 杭州脉流科技有限公司 颅内大血管闭塞点检测方法、装置、设备和存储介质
CN118608544A (zh) * 2024-06-20 2024-09-06 首都医科大学附属北京天坛医院 一种用于颅脑的出血灶分割的方法及相关产品
CN118657733A (zh) * 2024-06-20 2024-09-17 首都医科大学附属北京天坛医院 一种用于颅脑出血影像的脑分区的方法及相关产品
CN120932045A (zh) * 2025-07-04 2025-11-11 强联智创（北京）科技有限公司 用于缺血性脑卒中预后预测的装置及存储介质

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114943717B (zh) * 2022-05-31 2023-04-07 北京医准智能科技有限公司 一种乳腺病灶检测方法、装置、电子设备及可读存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110002523A1 (en) * 2008-03-03 2011-01-06 Prakash Bhanu K N Method and System of Segmenting CT Scan Data
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
CN112614145A (zh) * 2020-12-31 2021-04-06 湘潭大学 一种基于深度学习的颅内出血ct图像分割方法
US20210125074A1 (en) * 2019-10-28 2021-04-29 Lunit Inc. Method for training neural network and device thereof

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111915556B (zh) * 2020-06-22 2024-05-14 杭州深睿博联科技有限公司 一种基于双分支网络的ct图像病变检测方法、系统、终端及存储介质
CN112215850A (zh) * 2020-08-21 2021-01-12 天津大学 一种带注意力机制的级联空洞卷积网络脑肿瘤分割方法
CN112037237B (zh) * 2020-09-01 2023-04-07 腾讯科技（深圳）有限公司 一种图像处理方法、装置、计算机设备及介质
CN112288708B (zh) * 2020-10-27 2021-10-15 推想医疗科技股份有限公司 Ct图像中淋巴结的检测方法、装置、介质及电子设备
CN113724187B (zh) * 2021-03-16 2025-06-27 腾讯科技（深圳）有限公司 医疗影像处理方法、装置、电子设备和存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110002523A1 (en) * 2008-03-03 2011-01-06 Prakash Bhanu K N Method and System of Segmenting CT Scan Data
US20210125074A1 (en) * 2019-10-28 2021-04-29 Lunit Inc. Method for training neural network and device thereof
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
CN112614145A (zh) * 2020-12-31 2021-04-06 湘潭大学 一种基于深度学习的颅内出血ct图像分割方法

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118015012A (zh) * 2024-01-15 2024-05-10 宁波大学 一种脑部mri图像脑微出血自动分割方法
CN118154591A (zh) * 2024-05-09 2024-06-07 杭州脉流科技有限公司 颅内大血管闭塞点检测方法、装置、设备和存储介质
CN118608544A (zh) * 2024-06-20 2024-09-06 首都医科大学附属北京天坛医院 一种用于颅脑的出血灶分割的方法及相关产品
CN118657733A (zh) * 2024-06-20 2024-09-17 首都医科大学附属北京天坛医院 一种用于颅脑出血影像的脑分区的方法及相关产品
CN120932045A (zh) * 2025-07-04 2025-11-11 强联智创（北京）科技有限公司 用于缺血性脑卒中预后预测的装置及存储介质

Also Published As

Publication number Publication date
CN114155232A (zh) 2022-03-08

Similar Documents

Publication Publication Date Title
WO2023102995A1 (zh) 颅内出血区域检测方法、装置、计算机设备及存储介质
US20230177678A1 (en) Method and system for domain knowledge augmented multi-head attention based robust universal lesion detection
JP6623265B2 (ja) 偽陽性低減での小結節検出
Nazir et al. Machine Learning‐Based Lung Cancer Detection Using Multiview Image Registration and Fusion
US20170004619A1 (en) System and method for automatic pulmonary embolism detection
CN113724185B (zh) 用于图像分类的模型处理方法、装置及存储介质
CN109754007A (zh) 前列腺手术中外包膜智能检测和预警方法及系统
CN113592769B (zh) 异常图像的检测、模型的训练方法、装置、设备及介质
Elhanashi et al. Classification and localization of multi-type abnormalities on chest X-rays images
CN113313699B (zh) 基于弱监督学习的x光胸部疾病分类及定位方法、电子设备
US12243232B1 (en) Racially unbiased deep learning-based mammogram analyzer
CN114897764A (zh) 基于标准化通道注意力的肺结节假阳性排除方法及装置
EP4327333A1 (en) Methods and systems for automated follow-up reading of medical image data
US11521323B2 (en) Systems and methods for generating bullseye plots
Hu et al. A lung nodule segmentation model based on the transformer with multiple thresholds and coordinate attention
Al-Antari et al. A hybrid segmentation and classification CAD framework for automated myocardial infarction prediction from MRI images
CN116910295B (zh) 基于抗混淆因子的哈希图像检索方法
Dahmani et al. Vision transformers (vit) for enhanced skin cancer classification
CN116740609B (zh) 肾脏肿瘤超声视频的分类方法、装置、设备及存储介质
CN113192085A (zh) 三维器官图像分割方法、装置及计算机设备
US12482097B2 (en) Systems and methods for identifying cancer in pets
Krishna et al. Deep learning-based robust hybrid approaches for brain tumor classification in magnetic resonance images
JP7130107B2 (ja) 領域特定装置、方法およびプログラム、学習装置、方法およびプログラム、並びに識別器
Babu et al. Diagnosis of lung cancer detection using various deep learning models: A survey
KR20230151146A (ko) 딥러닝 기반 이미지 변환 방법 및 그 시스템

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: 21966911

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: 21966911

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 21966911

Country of ref document: EP

Kind code of ref document: A1