WO2020224123A1 - 一种基于深度学习的致痫灶三维自动定位系统 - Google Patents

一种基于深度学习的致痫灶三维自动定位系统 Download PDF

Info

Publication number
WO2020224123A1
WO2020224123A1 PCT/CN2019/103530 CN2019103530W WO2020224123A1 WO 2020224123 A1 WO2020224123 A1 WO 2020224123A1 CN 2019103530 W CN2019103530 W CN 2019103530W WO 2020224123 A1 WO2020224123 A1 WO 2020224123A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
pet
layer
data
epileptic
Prior art date
Application number
PCT/CN2019/103530
Other languages
English (en)
French (fr)
Inventor
卓成
张沁茗
张腾
廖懿
王夏婉
冯建华
张宏
田梅
Original Assignee
浙江大学
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 浙江大学 filed Critical 浙江大学
Priority to US17/047,392 priority Critical patent/US11645748B2/en
Publication of WO2020224123A1 publication Critical patent/WO2020224123A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/02—Arrangements for diagnosis sequentially in different planes; Stereoscopic radiation diagnosis
    • A61B6/03—Computed tomography [CT]
    • A61B6/037—Emission tomography
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/02—Arrangements for diagnosis sequentially in different planes; Stereoscopic radiation diagnosis
    • A61B6/03—Computed tomography [CT]
    • A61B6/032—Transmission computed tomography [CT]
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/50—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment specially adapted for specific body parts; specially adapted for specific clinical applications
    • A61B6/501—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment specially adapted for specific body parts; specially adapted for specific clinical applications for diagnosis of the head, e.g. neuroimaging or craniography
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/52—Devices using data or image processing specially adapted for radiation diagnosis
    • A61B6/5211—Devices using data or image processing specially adapted for radiation diagnosis involving processing of medical diagnostic data
    • A61B6/5217—Devices using data or image processing specially adapted for radiation diagnosis involving processing of medical diagnostic data extracting a diagnostic or physiological parameter from medical diagnostic data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/255—Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/52—Devices using data or image processing specially adapted for radiation diagnosis
    • A61B6/5211—Devices using data or image processing specially adapted for radiation diagnosis involving processing of medical diagnostic data
    • 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
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10104—Positron emission tomography [PET]
    • 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
    • G06T2207/30016—Brain
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images

Definitions

  • the invention relates to the technical field of medical imaging engineering, in particular to a three-dimensional automatic positioning system for epileptic foci based on deep learning.
  • the detection systems for epilepsy diseases include positron emission computed tomography (PET), magnetic resonance imaging (MRI), single photon emission computed tomography (SPECT) and electroencephalogram (EEG).
  • PET is used for the detection of epilepsy diseases.
  • the prognosis has a higher sensitivity.
  • PET is used for the detection of epilepsy diseases.
  • the prognosis has a higher sensitivity.
  • the traditional routine clinical diagnosis of visual evaluation of 3D PET images is very time-consuming and is affected by the clinical experience of doctors. Therefore, it is very important to propose an accurate and rapid epileptogenic focus location system.
  • the imaging technology usually judges abnormalities based on the statistical inference of the standard uptake value (SUV) and/or the asymmetry index (AI) of the region or voxel.
  • Regional statistical methods usually divide the brain into larger regions of interest (ROI), and then compare the average values of SUVs or AIs in the regions. Because the area is often much larger than the lesion area, this method will ignore subtle changes, which will reduce its detection sensitivity.
  • Voxel statistical methods usually use statistical parameter mapping (SPM) software to compare the data of a single case and a control group.
  • SPM statistical parameter mapping
  • the purpose of the present invention is to provide a three-dimensional positioning system for epileptic foci in the brain region based on deep learning in view of the shortcomings of current medical image focus positioning technology, which is used to automatically locate the location of the epileptogenic foci in the brain with high accuracy of positioning results ,
  • the model has high robustness.
  • a three-dimensional automatic positioning system for epileptogenic foci based on deep learning which includes the following modules:
  • PET image acquisition and marking module including image acquisition and epileptic focus area marking:
  • a 3D PET/CT scanner is used to acquire a PET image of the brain.
  • the subject maintains the same posture during the acquisition process to acquire the PET image.
  • the image format is converted, that is, the original collected image sequence in the DICOM format is converted into an easy-to-process NIFTI format image.
  • PET image registration module Taking cross-correlation as the similarity measure between images, using the Symmetric Differential Homeomorphism (SyN) algorithm to deform all PET images and their marked images into the same symmetrical standard space to achieve PET image acquisition, The registration of the marked image with the standard symmetrical brain template. After registration, the Gaussian smoothing algorithm is used to reduce the registration error caused by individual differences, and the Gaussian smoothing process selects the FWHM of the Gaussian function to be 5-15mm. Perform z-score normalization on the smoothed image.
  • SyN Symmetric Differential Homeomorphism
  • Adopt a deep learning system based on symmetry including the following modules:
  • Radial distortion and image intensity enhancement are performed on the registered images and labels to obtain newly generated images and labels.
  • Radial distortion refers to the deviation of image pixel points along the radial direction with the distortion center as the center point. The calculation process of radial distortion is:
  • P u is a pixel of the original image
  • P d is a pixel of the distorted image
  • P c is the distortion center
  • Image intensity enhancement includes filtering processing, image noise processing and multiplicative and additive transformation of image gray values in space.
  • the formula for image intensity enhancement is:
  • P a is the image pixel after the image intensity has been enhanced
  • g_mult is the image pixel of the multiplicative Gaussian bias field
  • g_add is the image pixel of the additive Gaussian bias field.
  • Image block division divide the enhanced image data into image blocks, use a three-dimensional sliding window to divide the left and right hemispheres L and R of the PET image into mirror pairs of image blocks, and scale the mirror pair data of the image blocks Divided into training set, validation set and test set.
  • the training set, validation set, and test set all contain two types of PET image block data that carry epileptic foci and normal.
  • the resolution of each PET image data is X ⁇ Y ⁇ Z pixels
  • the size of the sliding scanning window block is set to m ⁇ m ⁇ m
  • the sliding step size is t. Then the size of each image block is m ⁇ m ⁇ m, and the left and right hemispheres of a PET image can be divided into To the image block.
  • Network building module build a deep twin network SiameseNet.
  • the network consists of two identical convolutional neural networks, a fully connected layer and an output layer.
  • Each convolutional neural network has a ten-layer structure.
  • the first layer includes a convolutional layer (conv), a batch normalization unit (batch normalization), a Relu function and a pooling layer (pool) connected in sequence;
  • the -9 layer is eight ResBlocks, each ResBlock contains two convolutional layers, two normalization operations, and a Relu function connected in sequence;
  • the tenth layer is a convolutional layer, the first of the two convolutional neural networks
  • the 10-layer output is connected to a fully connected layer (fc) for nonlinear transformation. Finally connect an output layer.
  • the two convolutional neural networks of SiameseNet share the same weight parameter ⁇ in each layer, input a pair of mirrored image pairs into the network, obtain the feature L_feature and feature R_feature of the two high-dimensional images, and calculate the two high-dimensional image features
  • the absolute difference d
  • is passed into the multi-layer perceptron (MLP) of the fully connected layer for probabilistic regression.
  • MLP multi-layer perceptron
  • the dimensions of the fully connected layer vector are 2048, 1024, 512, and 2 in order.
  • the output layer uses the classification probability of the softmax regression function, that is, the probability that the image block carries the epileptic focus or is normal.
  • cross entropy function is used as the loss function of the network in model training.
  • the calculation method of cross entropy Loss(a,b) is:
  • n the number of samples
  • a the correct probability distribution
  • b the probability distribution predicted by the network model.
  • Standard stochastic gradient descent (SGD) is used to update the weight parameter ⁇ , the formula is:
  • ⁇ is the learning rate
  • ⁇ k is the k-th weight parameter
  • Image classification Use the trained model to calculate the probability heat map of the PET image in the test set.
  • the probability heat map is a probability map formed by stitching the corresponding probabilities of different image blocks on the PET image, and the size is After that, the logistic regression algorithm is used to classify the probability heat map corresponding to each PET image, and the classification result is obtained, which is a normal PET image or a PET image with epileptic foci.
  • Epileptic focus location Bilinear interpolation is performed on the probability heat map of the PET image identified as carrying the epileptic focus, the probability heat map is changed to the original image size, and the area larger than the probability threshold is predicted as the epileptic focus area.
  • the system proposed by the present invention can accurately detect the images of patients with metabolic abnormalities, and compared with the existing SPM software, the epileptic focus area predicted by the system is more consistent with the doctor's visual evaluation, and maintains a higher accuracy and efficiency. Therefore, it has high value in helping doctors locate the epileptic area and follow-up surgical treatment.
  • the system proposed by the present invention is effective for detecting epileptic foci in different brain regions of the whole brain, and is suitable for epileptic patients with epileptic foci in different brain regions.
  • the present invention utilizes image enhancement and division of mirror pairs of image blocks to increase the sample size, training models and test data are performed on this basis, avoiding over-fitting of network training, and improving the robustness of network training.
  • the present invention uses sample weighting as data enhancement, and sets larger weights for a small number of samples, so as to balance the sample ratio of each batch of normal area samples with the sample ratio of epileptic regions during the training process.
  • FIG. 1 is a structural block diagram of a three-dimensional positioning system for epileptogenic foci based on deep learning according to an embodiment of the present invention
  • FIG. 2 is an implementation flowchart of a three-dimensional positioning system for epileptogenic foci based on deep learning according to an embodiment of the present invention
  • Figure 3 is a schematic diagram of the construction of a deep twin network SiameseNet according to an embodiment of the present invention
  • Figure 4 is a schematic diagram of a single residual neural network structure of SiameseNet of the present invention.
  • Fig. 5 is a probability heat map corresponding to a PET image according to an embodiment of the present invention.
  • the three-dimensional automatic positioning system for epileptogenic foci includes the following modules:
  • PET image acquisition and marking module including image acquisition and epileptic focus area marking:
  • a 3D PET/CT scanner is used to acquire a PET image of the brain.
  • the subject maintains the same posture during the acquisition process to acquire the PET image.
  • the image format is converted, that is, the original collected image sequence in the DICOM format is converted into an easy-to-process NIFTI format image.
  • PET image registration module Taking cross-correlation as the similarity measure between images, using the Symmetric Differential Homeomorphism (SyN) algorithm to deform all PET images and their marked images into the same symmetrical standard space to achieve PET image acquisition, The registration of the marked image with the standard symmetrical brain template. For deforming the original image I to the image J, minimize the following objective function:
  • the first term is the smoothing term, where L is the smoothing operator and v is the velocity field.
  • the ⁇ in the second term controls the accuracy of matching.
  • C(I,J) is the similarity measure, where C(I,J) can be expressed as:
  • Gaussian smoothing algorithm After registration, Gaussian smoothing algorithm is used to reduce the registration error caused by individual differences.
  • the Gaussian smoothing process selects the FWHM of the Gaussian function to be 5-15mm to eliminate registration errors caused by individual differences. Perform z-score standardization on the smoothed image:
  • ⁇ is the mean value of a registered image J
  • ⁇ is the variance of an image
  • Adopt a deep learning system based on symmetry including the following modules
  • Radial distortion and image intensity enhancement are performed on the registered images and labels to obtain newly generated images and labels.
  • Radial distortion refers to the deviation of the image pixel points along the radial direction with the distortion center as the center point. The calculation process of radial distortion is:
  • P u is a pixel of the original image
  • P d is a pixel of the distorted image
  • P c is the distortion center
  • Image intensity enhancement includes filtering processing, image noise processing and multiplicative and additive transformation of image gray values in space.
  • the formula for image intensity enhancement is:
  • P a is the image pixel after the image intensity has been enhanced
  • g_mult is the image pixel of the multiplicative Gaussian bias field
  • g_add is the image pixel of the additive Gaussian bias field.
  • Image block division divide the enhanced image data into image blocks, use a three-dimensional sliding window to divide the left and right hemispheres L and R of the PET image into mirror pairs of image blocks, and scale the mirror pair data of the image blocks It is divided into a training set, a verification set and a test set; the training set, the verification set and the test set all contain two types of PET image block data carrying epileptic foci and normal.
  • the resolution of each PET image data is X ⁇ Y ⁇ Z pixels
  • the size of the sliding scanning window block is set to m ⁇ m ⁇ m
  • the sliding step size is t.
  • the size of each image block is m ⁇ m ⁇ m, and the left and right hemispheres of a PET image can be divided into To the image block.
  • Network building module build a deep twin network SiameseNet.
  • the network consists of two identical convolutional neural networks, a fully connected layer and an output layer.
  • Each convolutional neural network has a ten-layer structure.
  • the first layer includes a convolutional layer (conv), a batch normalization unit (batch normalization), a Relu function and a pooling layer (pool) connected in sequence;
  • the -9 layer is eight ResBlocks, each ResBlock contains two convolutional layers, two normalization operations, and a Relu function connected in sequence;
  • the tenth layer is a convolutional layer, the first of the two convolutional neural networks
  • the 10-layer output is connected to a fully connected layer (fc) for nonlinear transformation. Finally connect an output layer.
  • the parameter setting of a random dropout can be 0.5.
  • output conv is the three-dimensional size of the output image data of each convolutional layer (the length, width and depth of the image)
  • input conv is the three-dimensional size of the input image
  • kernal is the three-dimensional convolution kernel
  • the size, stride is the step size of the convolution kernel.
  • batch standardization operation For each convolutional layer, use batch standardization operation to accelerate the convergence speed and stability of the network.
  • the formula for batch standardization operation is:
  • input norm is each batch of input data
  • ⁇ and ⁇ are the mean and variance of each batch of data
  • ⁇ and ⁇ are scaling and translation variables, respectively
  • ⁇ is added to increase training stability Smaller constant data
  • the activation function connected to each convolutional layer selects the Relu function to shorten the training period.
  • the calculation method of the Relu function is:
  • input relu is the input data of the Relu function
  • output relu is the output data of the Relu function
  • the two convolutional neural networks of SiameseNet share the same weight parameter ⁇ in each layer, and input the mirror pair of a pair of image blocks into the network.
  • the size of the input image block is 48 ⁇ 48 ⁇ 48 ⁇ 1, where 48 ⁇ 48 ⁇ 48 represents the length, width, and height of the image block, and 1 represents the number of channels of the image block.
  • the feature size obtained is 24 ⁇ 24 ⁇ 24 ⁇ 64
  • the feature size obtained through ResBlocks is 12 ⁇ 12 ⁇ 12 ⁇ 64, 12 ⁇ 12 ⁇ 12 ⁇ 64, 6 ⁇ 6 ⁇ 6 ⁇ 128, 6 ⁇ 6 ⁇ 6 ⁇ 128, 3 ⁇ 3 ⁇ 3 ⁇ 256, 3 ⁇ 3 ⁇ 3 ⁇ 256, 3 ⁇ 3 ⁇ 3 ⁇ 512 and 3 ⁇ 3 ⁇ 3 ⁇ 512
  • MLP layer perceptron
  • the output layer uses the classification probability of the softmax regression function, that is, the probability that the image block carries the epileptogenic focus or is normal.
  • the formula of softmax is:
  • d j represents the output of different categories
  • g represents the number of categories
  • j 1, 2, ... g.
  • cross entropy function is used as the loss function of the network in model training.
  • the calculation method of cross entropy Loss(a,b) is:
  • n the number of samples
  • a the correct probability distribution
  • b the probability distribution predicted by the network model.
  • Standard stochastic gradient descent (SGD) is used to update the weight parameter ⁇ , and the formula is:
  • ⁇ is the learning rate
  • ⁇ k is the k-th weight parameter
  • the training phase and the test phase flow chart are shown in Figure 4.
  • the basic network framework adopted by SiameseNet is ResNet18, and the two ResNets share the same network weight parameter ⁇ , using PET images and
  • the training set of normal images trains the network, and the network model is obtained through the training process.
  • a small number of mirror image pairs of image background blocks are added to the normal samples of the training set to reduce the influence of the image background on the model.
  • Image classification Use the trained model to calculate the probability heat map of the PET image in the test set.
  • the probability heat map is a probability map formed by stitching the corresponding probabilities of different image blocks on the PET image.
  • the size is After that, the logistic regression algorithm is used to classify the probability heat map corresponding to each PET image, and the classification result is obtained, which is a normal PET image or a PET image with epileptic foci.
  • f(m+u,n+v) is the newly calculated pixel value
  • f(m,n), f(m+1,n), f(m,n+1) and f(m+1, n+1) are the four original pixel values around the new pixel value
  • u and v are the distance between the original pixel and the new pixel.
  • the collected PET data set is first divided into a training set, a validation set and a test set, and the twin network learning system is used to extract two left and right brain image blocks.
  • Feature vector calculate the absolute difference between two feature vectors, and add a multilayer perceptron to perform probability regression.
  • a sliding window block is used to scan and test each entire image.
  • the probability heat map is output, and the detection result map is finally obtained, so as to realize the classification and positioning of the epilepsy focus in the PET image, and the final classification result of the entire image
  • the AUC is 94%, and compared with the existing SPM software, the epileptic focus area predicted by the system is more consistent with the doctor’s visual assessment, maintaining a higher accuracy and efficiency.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Medical Informatics (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • General Physics & Mathematics (AREA)
  • Public Health (AREA)
  • Optics & Photonics (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Pathology (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Quality & Reliability (AREA)
  • Databases & Information Systems (AREA)
  • Pulmonology (AREA)
  • Neurosurgery (AREA)
  • Neurology (AREA)
  • Dentistry (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Primary Health Care (AREA)
  • Physiology (AREA)

Abstract

一种基于深度学习的致痫灶三维自动定位系统，该系统包括：PET图像采集和标记模块；PET图像与标准对称脑模版的配准模块；PET图像数据预处理模块，生成左右脑图像块的镜像对；孪生网络SiameseNet训练模块，包含两个共享权重参数的深度残差卷积神经网络，输出层连接多层感知机和softmax层，利用携带致痫灶的图像和正常图像的训练集对所述网络进行训练获得网络模型；分类模块和致痫灶定位模块，利用训练好的网络模型对新输入的PET图像生成概率热图，先通过分类器判断图像为正常或者携带致痫灶样本，再预测致痫灶区域的位置。该系统通过引入图像块的镜像对和孪生网络SiameseNet，来自动定位PET图像的致痫灶，能有效地提高致痫灶定位的准确度和效率，并具有较高的鲁棒性。

Description

一种基于深度学习的致痫灶三维自动定位系统 技术领域
本发明涉及医疗影像工程技术领域，特别涉及一种基于深度学习的致痫灶三维自动定位系统。
背景技术
随着医学成像技术和人工智能技术的发展，自动和半自动的计算机辅助诊断系统被广泛应用于精准诊断和治疗中，以提高诊断准确率及预后。目前，针对癫痫疾病的检测系统包括正电子发射计算机断层扫描(PET)、核磁共振成像(MRI)、单光子发射计算机断层扫描(SPECT)和脑电图(EEG)，其中PET对于癫痫疾病的检测和预后具有更高的灵敏度。对癫痫疾病类型的确定、难治性癫痫的手术治疗等中，均需要用诊断系统精确地定位致痫灶的位置。然而，传统的对三维PET图像进行视觉评估的常规临床诊断非常耗时，且受到医生临床经验的影响。因此，提出一种准确、快速的致痫灶定位系统十分重要。
现有的技术面临的挑战主要有：1.影像学技术通常基于区域或体素的标准摄取值(SUV)和/或不对称指数(AI)的统计推断判断异常。区域统计方法通常将大脑分割成较大的感兴趣区域(ROI)，然后比较区域内的SUV或AI的平均值。由于区域常远大于病灶区，导致该方法会忽略细微变化，从而导致其检测灵敏度降低。体素统计方法通常使用统计参数映射(SPM)软件来比较单个病例和控制组的数据，然而，体素统计方法对配准误差高度敏感，容易在错位区域产生假阳性。2.现有的算法大多仅适用于二维自然图像处理，而由于PET成像是一个由平行扫描图像帧组成的三维结构，二维定位算法会忽略重要的帧间信息。3.由于医学图像数据量少、缺少高质量的标注数据和训练样本、正负样本数量差异较大导致样本的不平衡等问题，训练出来的模型可能是过拟合或者模型泛化能力不高。
综上，提供一种致痫灶区域的三维自动定位系统，利用PET图像帧内和帧间的对称性信息，以提高致痫灶定位的准确度和效率，成为目前亟待解决的重要技术问题。
发明内容
本发明的目的在于针对目前医学图像病灶定位技术的不足，提供了一种基于深度学习的脑部区域致痫灶三维定位系统，用于自动定位脑部致痫灶的位置，定位结果准确率高，模型具有较高的鲁棒性。
本发明的目的是通过以下技术方案来实现的：一种基于深度学习的致痫灶三维自动定位 系统，所述系统包括以下模块：
(1)PET图像采集和标记模块，包括图像采集和致痫灶区域标记：
1.1)采集图像：使用3D PET/CT扫描仪采集脑部PET图像，受试者在采集过程中保持相同体位，获取PET图像。图像采集后进行图像格式转换，即将DICOM格式的原始采集图像序列转换成易处理的NIFTI格式图像。
1.2)标记样本：将PET图像分为正常样本集和携带致痫灶的样本集，并对携带致痫灶的样本集手动标记致痫灶区域，其中，致痫灶区域标记为1，其余区域标记为0。
(2)PET图像配准模块：以互相关作为图像间的相似性度量，运用对称微分同胚(SyN)算法将所有PET图像及其标记图像形变到同一对称标准空间，以实现PET采集图像、标记图像与标准对称脑模版的配准。配准后，采用高斯平滑算法减少个体差异带来的配准误差，高斯平滑处理选择高斯函数的半峰全宽FWHM为5～15mm。对平滑后的图像进行z-score标准化。
(3)采用基于对称性的深度学习系统，包含以下模块：
3.1)数据预处理模块：
3.1.1)数据增强：对配准后的图像和标签进径向畸变和图像强度增强，得到新生成的图像和标签。径向畸变是图像像素点以畸变中心为中心点，沿着径向的位置产生偏差，径向畸变的计算过程为：
P u＝P d+(P d-P c)(k 1r 2+k 2r 4+k 3r 6+…)
其中，P u是原图像的一个像素点，P d是畸变后图像的一个像素点，P c是畸变中心，k i(i＝1,2,3…)是径向畸变的畸变系数，r是P d和P c在矢量空间上的距离。
图像强度增强包括滤波处理、图像加噪处理和图像灰度值在空间的乘性、加性变换，图像强度增强的公式为：
P a＝g_mult×P u+g_add
其中P a是图像强度增强后的图像像素点，g_mult是乘性高斯偏置场的图像像素点，g_add是加性高斯偏置场的图像像素点。
3.1.2)图像块划分：对增强后的图像数据进行图像块划分，用三维滑动窗口将PET图像的左右半脑L和R划分为图像块的镜像对，将图像块的镜像对数据按比例分为训练集、验证集和测试集。所述训练集、验证集和测试集中均包含有携带致痫灶和正常两种类型的PET图像块数据。图像数据集中，每一张PET图像数据的分辨率为X×Y×Z像素，设置滑动扫描窗口块的大小为m×m×m，滑动步长为t。则每个图像块的大小为m×m×m，对于一张PET图像的左右半脑，可划分为
Figure PCTCN2019103530-appb-000001
对图像块。
3.2)网络构建模块：构建深度孪生网络SiameseNet。该网络包含两个相同的卷积神经网络、一个全连接层及一个输出层。每个卷积神经网络有十层结构，第1层包括依次连接的一个卷积层(conv)、一个批标准化操作单元(batch normalization)，一个Relu函数和一个池化层(pool)；第2-9层是八个ResBlock，每个ResBlock均包含依次连接的两个卷积层、两次归一化操作和一个Relu函数；第10层为一个卷积层，两个卷积神经网络的第10层输出连接一个全连接层(fc)，进行非线性变换。最后连接一个输出层。
SiameseNet的两个卷积神经网络在每层共享相同的权重参数θ，将一对图像块的镜像对输入网络，获取两个高维图像的特征L_feature和特征R_feature，计算两个高维图像特征的绝对差值d＝|L_feature-R_feature|，并将其传入到全连接层的多层感知机(MLP)中进行概率回归，全连接层向量的维度依次为2048、1024、512和2。输出层采用softmax回归函数的分类概率，即图像块携带致痫灶或正常的概率。
在模型训练中采用交叉熵函数作为网络的损失函数。交叉熵Loss(a,b)的计算方式为：
Figure PCTCN2019103530-appb-000002
其中，n表示样本数量，a是正确的概率分布，b是网络模型预测的概率分布。采用标准随机梯度下降(SGD)更新权重参数θ，其公式为：
Figure PCTCN2019103530-appb-000003
其中，η是学习速率，θ k是第k次的权重参数。
3.3)测试图像检测模块：
图像分类：利用训练好的模型计算测试集PET图像的概率热图，概率热图是一张PET图像上不同图像块对应概率拼接而成的概率图，大小为
Figure PCTCN2019103530-appb-000004
之后采用逻辑回归算法对每一张PET图像对应的概率热图进行分类，获得分类结果，即为正常PET图像或携带致痫灶PET图像。
致痫灶定位：对识别为携带致痫灶PET图像的概率热图进行双线性插值，将概率热图改变为原始图像尺寸，将大于概率阈值的区域预测为致痫灶区域。
本发明的有益效果如下：
1)能够自动学习PET图像数据中的深度特征。传统的视觉评估需要医生一帧帧地观察和判断，极度依赖医生的经验及技术水平，且消耗大量时间。孪生网络能够自动地学习PET图像中的高维不对称特征特征，以发现PET图像和致痫灶之间的内在联系。与传统的致痫灶 定位系统相比，本发明提出的系统能够学习到人眼难以识别的高阶特征，且兼顾了单侧癫痫患者代谢分布不对称这一先验知识。
2)能够实现对病灶区的精准定位。本发明提出的系统能准确检测代谢异常的患者图像，并且与现有的SPM软件相比，该系统预测出的致痫灶区域和医师视觉评估更一致，并保持较高的准确率和效率。因此，在帮助医生定位致痫区和后续手术治疗方面有较高的价值。
3)能够适用于不同脑区的致痫灶检测。本发明提出的系统对全脑不同脑区的致痫灶检测均有效，适用于致痫灶在不同脑区的癫痫患者。
4)能够实现小数据量的网络训练。本发明利用图像增强和划分图像块的镜像对，以增加样本量，在此基础上进行训练模型和测试数据，避免了网络训练的过拟合，提高了网络训练的鲁棒性。此外，为了平衡正常和病患数据样本，本发明采用样本加权作为数据增强，为少数样本设置了较大的权重，以在训练过程中使每批正常面积样本比例与致痫区样本比例均衡。
附图说明
图1是本发明一个实施例的基于深度学习的致痫灶三维定位系统的结构框图；
图2是本发明一个实施例的基于深度学习的致痫灶三维定位系统的实现流程图；
图3是本发明一个实施例的深度孪生网络SiameseNet构建示意图；
图4是本发明SiameseNet的单个残差神经网络结构示意图；
图5是本发明一个实施例的PET图像对应的概率热图。
具体实施方式
下面结合附图和具体实施例对本发明作进一步详细说明。
如图1所示，本发明一个实施例的致痫灶三维自动定位系统包括以下模块：
(1)PET图像采集和标记模块，包括图像采集和致痫灶区域标记：
1.1)采集图像：使用3D PET/CT扫描仪采集脑部PET图像，受试者在采集过程中保持相同体位，获取PET图像。图像采集后进行图像格式转换，即DICOM格式的原始采集图像序列转换成易处理的NIFTI格式图像。
1.2)标记样本：将PET图像分为正常样本集和携带致痫灶的样本集，并对携带致痫灶的样本集手动标记致痫灶区域，其中，致痫灶区域标记为1，其余区域标记为0。
(2)PET图像配准模块：以互相关作为图像间的相似性度量，运用对称微分同胚(SyN)算法将所有PET图像及其标记图像形变到同一对称标准空间，以实现PET采集图像、标记图像与标准对称脑模版的配准。对于将原始图像I形变到图像J，最小化以下目标函数：
Figure PCTCN2019103530-appb-000005
第一项是平滑项，式中，L是平滑算符，v是速度场。第二项中的λ控制匹配的精确性。C(I,J)是相似性度量，其中C(I,J)可表示为：
Figure PCTCN2019103530-appb-000006
配准后，采用高斯平滑算法减少个体差异带来的配准误差。高斯平滑处理选择高斯函数的半峰全宽FWHM为5～15mm，以消除个体差异带来配准误差。对平滑后的图像进行z-score标准化：
Figure PCTCN2019103530-appb-000007
其中μ为一张配准后图像J的均值，σ为一张图像的方差。
(3)采用基于对称性的深度学习系统，包含以下模块
3.1)数据预处理模块：
3.1.1)数据增强：对配准后的图像和标签进行径向畸变和图像强度增强，得到新生成的图像和标签。径向畸变是图像像素点以畸变中心为中心点,沿着径向的位置产生偏差，径向畸变的计算过程为：
P u＝P d+(P d-P c)(k 1r 2+k 2r 4+k 3r 6+…)
其中，P u是原图像的一个像素点，P d是畸变后图像的一个像素点，P c是畸变中心，k i(i＝1,2,3…)是径向畸变的畸变系数，r是P d和P c在矢量空间上的距离。
图像强度增强包括滤波处理、图像加噪处理和图像灰度值在空间的乘性、加性变换，图像强度增强的公式为：
P a＝g_mult×P u+g_add
其中P a是图像强度增强后的图像像素点，g_mult是乘性高斯偏置场的图像像素点，g_add是加性高斯偏置场的图像像素点。
3.1.2)图像块划分：对增强后的图像数据进行图像块划分，用三维滑动窗口将PET图像的左右半脑L和R划分为图像块的镜像对，将图像块的镜像对数据按比例分为训练集、验证集和测试集；所述训练集、验证集和测试集中均包含有携带致痫灶和正常两种类型的PET图像块数据。图像数据集中，每一张PET图像数据的分辨率为X×Y×Z像素，设置滑动扫描窗口块的大小为m×m×m，滑动步长为t。则每个图像块的大小为m×m×m，对于一张PET图像的左右半脑，可划分为
Figure PCTCN2019103530-appb-000008
对图像块。
3.2)网络构建模块：构建深度孪生网络SiameseNet。该网络包含两个相同的卷积神经网络、一个全连接层及一个输出层。每个卷积神经网络有十层结构，第1层包括依次连接的一个卷积层(conv)、一个批标准化操作单元(batch normalization)，一个Relu函数和一个池化 层(pool)；第2-9层是八个ResBlock，每个ResBlock均包含依次连接的两个卷积层、两次归一化操作和一个Relu函数；第10层为一个卷积层，两个卷积神经网络的第10层输出连接一个全连接层(fc)，进行非线性变换。最后连接一个输出层。一次随机失活(dropout)其参数设置可以为0.5。
SiameseNet网络模型中，卷积层操作的计算过程为：
Figure PCTCN2019103530-appb-000009
其中，output conv是每一卷积层输出图像数据的三维大小(图像的长度、宽度和深度)，input conv是输入图像的三维大小，pad表示在图像周围填充像素，kernal为卷积核的三维大小，stride为卷积核的步长。
对于每一个卷积层使用批标准化操作，加速网络的收敛速度及稳定性，批标准化操作的公式为：
Figure PCTCN2019103530-appb-000010
Figure PCTCN2019103530-appb-000011
其中，input norm是输入的每个批数据，
Figure PCTCN2019103530-appb-000012
是归一化数据，output norm批标准化操作输出的批数据，μ和σ分别是每个批数据的均值和方差，γ和β分别是缩放和平移变量，ε是为了增加训练稳定性而加入的较小的常量数据；
与每个卷积层相连的激活函数选用Relu函数，可缩短训练周期，Relu函数的计算方式为：
output relu＝max(input relu,0)
其中，input relu是Relu函数的输入数据，output relu是Relu函数的输出数据。
SiameseNet的两个卷积神经网络在每层共享相同的权重参数θ，将一对图像块的镜像对输入网络，如图3所示，输入图像块的大小为48×48×48×1，其中48×48×48代表图像块的长、宽、高，1代表图像块的通道数量。经过第1层卷积后，得到的特征大小为24×24×24×64，经过ResBlocks分别得到的特征大小为12×12×12×64、12×12×12×64、6×6×6×128、6×6×6×128、3×3×3×256、3×3×3×256、3×3×3×512和3×3×3×512，经过第10层卷积层得到大小为1×1×1×2048的两个高维特征L_feature和R_feature，计算两个高维图像特征的绝对差值d＝|L_feature-R_feature|，并将其传入到全连接层的多层感知机(MLP)中进行概率回归，全连接层向量的维度依次为1×1×1×1024、1×1×1×512和1×1×1×2，全连接层中间采用dropout层并设置p＝0.5，减少网络参数，防止过拟合。输出层采用softmax回归函数的分类概率，即图像块携带致痫灶或正常的概率，softmax的公式为：
Figure PCTCN2019103530-appb-000013
其中，d j代表不同类别的输出，g代表分类数，j＝1,2,…g。
在模型训练中采用交叉熵函数作为网络的损失函数。交叉熵Loss(a,b)的计算方式为：
Figure PCTCN2019103530-appb-000014
其中，n表示样本数量，a是正确的概率分布，b是网络模型预测的概率分布。采用标准随机梯度下降(SGD)更新权重权重参数θ，其公式为：
Figure PCTCN2019103530-appb-000015
其中，η是学习速率，θ k是第k次的权重参数。
在本发明实例中，所述训练阶段和测试阶段流程图如图4所示，SiameseNet所采用的基本网络框架是ResNet18，两个ResNet共享相同的网络权重参数θ，利用携带致痫灶PET图像和正常图像的训练集对所述网络进行训练，通过训练过程得到网络模型。此外，在训练集的正常样本中加入少量图像背景块的镜像对，以减少图像背景对模型产生的影响。
3.3)测试图像检测模块：
图像分类：利用训练好的模型计算测试集PET图像的概率热图，如图5所示，概率热图是一张PET图像上不同图像块对应概率拼接而成的概率图，大小为
Figure PCTCN2019103530-appb-000016
之后采用逻辑回归算法对每一张PET图像对应的概率热图进行分类，获得分类结果，即为正常PET图像或携带致痫灶PET图像。
致痫灶定位：对识别为携带致痫灶PET图像的概率热图进行双线性插值，将概率热图改变为与原始图像尺寸相同的概率图heatmap，将大于概率阈值的区域预测为致痫灶区域。双线性插值的计算公式为；
f(m+u,n+v)＝(1-u)(1-v)f(m,n)+u(1-v)f(m+1,n)
+(1-u)vf(m,n+1)+uvf(m+1,n+1)
其中，f(m+u,n+v)为新计算的像素值，f(m,n)，f(m+1,n)，f(m,n+1)和f(m+1,n+1)分别是新像素值周围的四个原像素值，u和v为原像素点和新像素点之间的距离。通过设置阈值k(heatmap≥heatmap_max×k)，其中heatmap_max是heatmap的最大值，最终获得预测的致痫灶区域。
在一个应用本实施例系统的具体案例中，如图4所示，首先将采集的PET数据集分为训练集、验证集和测试集，利用孪生网络学习系统，提取左右脑图像块的两个特征向量，计算两个特征向量的绝对差值，在其后添加多层感知机进行概率回归。最后在每一整张图像上用 滑动窗口块进行扫描测试，扫描完后输出概率热图，最终得到检测结果图，从而实现对PET图像中癫痫灶的分类和定位，最终整张图像分类结果的AUC为94％，并且与现有的SPM软件相比，该系统预测出的致痫灶区域和医师视觉评估更一致，保持较高的准确率和效率。
本专利不局限于上述最佳实施方式。任何人在本专利的启示下都可以得出其他各种形式的基于深度学习的致痫灶定位系统，凡依照本发明申请专利范围所做的均等变化与修饰，皆应属本专利的涵盖范围。

Claims (9)

  1. 一种基于深度学习的致痫灶三维自动定位系统，其特征在于，该系统包括以下模块：
    (1)PET图像采集和标记模块，用于图像采集和致痫灶区域标记：
    1.1)采集图像：受试者在PET扫描仪上使用3D脑部图像采集，在相同体位状态下获取PET脑图像。
    1.2)标记样本：将PET图像分为正常样本集和携带致痫灶的样本集，并对携带致痫灶的样本集手动标记致痫灶区域，其中，致痫灶区域标记为1，其余区域标记为0。
    (2)PET图像配准模块：以互相关作为原始图像与配准图像的相似性度量，运用对称微分同胚(SyN)算法将所有PET图像及其标记图像配准到同一对称标准空间，实现PET采集图像、标记图像与标准对称脑模版的配准。
    (3)采用基于对称性的深度学习系统，包含以下模块：
    3.1)数据预处理模块：
    3.1.1)数据增强：对配准后的图像和标签进径向畸变和图像强度增强，得到新生成的图像和标签。
    3.1.2)图像块划分：对增强后的图像数据进行图像块划分，用三维滑动窗口将PET图像的左右半脑L和R划分为图像块的镜像对，将图像块的镜像对数据按比例分为训练集和测试集；所述训练集和测试集中均包含有携带致痫灶和正常两种类型的PET图像块数据。
    3.2)网络构建模块：构建深度孪生网络SiameseNet，该网络包含两个相同的卷积神经网络、一个全连接层以及一个输出层。SiameseNet将一对图像块的镜像对输入每层共享权重参数θ的两个卷积神经网络，以获取两个高维图像的特征L_feature和特征R_feature，计算两个高维图像特征的绝对差值d＝|L_feature-R_feature|，并将其传入到全连接层的多层感知机中进行概率回归，输出层采用softmax回归函数的分类概率，即图像块携带致痫灶或正常的概率。
    3.3)测试图像检测模块：
    图像分类：利用训练好的网络计算测试集PET图像的概率热图，采用逻辑回归算法对每一张PET图像对应的概率热图进行分类，获得分类结果，即为正常PET图像或携带致痫灶PET图像。
    致痫灶定位：对识别为携带致痫灶PET图像的概率热图进行双线性插值，将概率热图改变为原始图像尺寸，将大于概率阈值的区域预测为致痫灶区域。
  2. 根据权利要求1所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，1.1)采集图像过程中，将获取的PET脑图像进行格式转换，即DICOM格式的原始采集图像转换成NIFTI格式图像。
  3. 根据权利要求1所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，(2)PET图像配准模块中，采用高斯平滑算法减少配准误差，高斯平滑处理选择高斯函数的半峰全宽FWHM为5～15mm，并对平滑后的图像进行z-score标准化。
  4. 根据权利要求1所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，3.1.1)数据增强过程中的径向畸变具体为：径向畸变是图像像素点以畸变中心为中心点，沿着径向的位置产生偏差，径向畸变的计算过程为：
    P u＝P d+(P d-P c)(k 1r 2+k 2r 4+k 3r 6+…)
    其中，P u是原图像的一个像素点，P d是畸变后图像的一个像素点，P c是畸变中心，k i(i＝1,2,3…)是径向畸变的畸变系数，r是P d和P c在矢量空间上的距离。
  5. 根据权利要求1所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，3.1.1)数据增强过程中的图像强度增强包括滤波处理、图像加噪处理和图像灰度值在空间的乘性、加性变换，图像强度增强的公式为：
    P a＝g_mult×P u+g_add
    其中P a是图像强度增强后的图像像素点，g_mult是乘性高斯偏置场的图像像素点，g_add是加性高斯偏置场的图像像素点。
  6. 根据权利要求1所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，3.1.2)图像块划分中，图像数据集中的每一张PET图像数据的分辨率为X×Y×Z像素，设置滑动扫描窗口块的大小为m×m×m，滑动步长为t，则每个图像块的大小为m×m×m，对于一张PET图像的左右半脑，可划分为
    Figure PCTCN2019103530-appb-100001
    对图像块。
  7. 根据权利要求1所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，3.2)网络构建模块中，SiameseNet的每个卷积神经网络有十层结构，第1层包括依次连接的一个卷积层、一个批标准化操作单元、一个Relu函数和一个池化层；第2-9层是八个ResBlock，每个ResBlock均包含依次连接的两个卷积层、两次归一化操作和一个Relu函数；第10层为一个卷积层，两个卷积神经网络的第10层输出连接一个全连接层进行非线性变换，全连接层向量的维度依次为2048、1024、512和2；最后连接一个输出层。
  8. 根据权利要求7所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，3.2)网络构建模块，在模型训练中采用交叉熵函数作为网络的损失函数，交叉熵Loss(a,b)的计算公式为：
    Figure PCTCN2019103530-appb-100002
    其中，n表示样本数量，a是正确的概率分布，b是网络模型预测的概率分布；
    采用标准随机梯度下降更新权重参数θ，其公式为：
    Figure PCTCN2019103530-appb-100003
    其中，η是学习速率，θ k是第k次的权重参数。
  9. 根据权利要求1所述的一种基于深度学习的致痫灶三维自动定位系统，其特征在于，3.2)网络构建模块，SiameseNet网络模型中卷积层操作的计算过程为：
    Figure PCTCN2019103530-appb-100004
    其中，output conv是每一卷积层输出图像数据的三维大小，input conv是输入图像的三维大小，pad表示在图像周围填充像素，kernal为卷积核的三维大小，stride为卷积核的步长；
    对于每一个卷积层使用批标准化操作，批标准化操作的公式为：
    Figure PCTCN2019103530-appb-100005
    Figure PCTCN2019103530-appb-100006
    其中，input norm是输入的每个批数据，
    Figure PCTCN2019103530-appb-100007
    是归一化数据，output norm批标准化操作输出的批数据，μ和σ分别是每个批数据的均值和方差，γ和β分别是缩放和平移变量，ε是为了增加训练稳定性而加入的较小的常量数据。
PCT/CN2019/103530 2019-06-24 2019-08-30 一种基于深度学习的致痫灶三维自动定位系统 WO2020224123A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/047,392 US11645748B2 (en) 2019-06-24 2019-08-30 Three-dimensional automatic location system for epileptogenic focus based on deep learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910549416.1A CN110390351B (zh) 2019-06-24 2019-06-24 一种基于深度学习的致痫灶三维自动定位系统
CN201910549416.1 2019-06-24

Publications (1)

Publication Number Publication Date
WO2020224123A1 true WO2020224123A1 (zh) 2020-11-12

Family

ID=68285820

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/103530 WO2020224123A1 (zh) 2019-06-24 2019-08-30 一种基于深度学习的致痫灶三维自动定位系统

Country Status (3)

Country Link
US (1) US11645748B2 (zh)
CN (1) CN110390351B (zh)
WO (1) WO2020224123A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112419278A (zh) * 2020-11-25 2021-02-26 上海应用技术大学 一种基于深度学习的实木地板分类方法
CN113269711A (zh) * 2021-04-06 2021-08-17 东北大学 脑图像的处理方法及装置、电子设备和存储介质
CN113506233A (zh) * 2021-07-08 2021-10-15 西安电子科技大学 基于深度学习的sar自聚焦方法
CN113643336A (zh) * 2021-07-26 2021-11-12 之江实验室 基于球极坐标系深度神经网络的三维图像刚性匹配方法
CN113688942A (zh) * 2021-02-09 2021-11-23 四川大学 基于深度学习的头侧片腺样体图像自动评估方法和装置
CN113808097A (zh) * 2021-09-14 2021-12-17 北京主导时代科技有限公司 一种列车的关键部件丢失检测方法及其系统
CN114376522A (zh) * 2021-12-29 2022-04-22 四川大学华西医院 构建用于识别青少年肌阵挛癫痫的计算机识别模型的方法
CN114820535A (zh) * 2022-05-05 2022-07-29 深圳市铱硙医疗科技有限公司 动脉瘤的图像检测方法、装置、计算机设备及存储介质
CN115661680A (zh) * 2022-11-15 2023-01-31 北京轨道未来空间科技有限公司 卫星遥感图像处理方法
CN116051810A (zh) * 2023-03-30 2023-05-02 武汉纺织大学 一种基于深度学习的智慧服装定位方法
EP4235685A1 (en) * 2022-02-23 2023-08-30 Siemens Healthcare GmbH Method, system and computer program for detection of a disease information
CN116958128A (zh) * 2023-09-18 2023-10-27 中南大学 基于深度学习的医学图像自动定位方法

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111127532B (zh) * 2019-12-31 2020-12-22 成都信息工程大学 基于深度学习特征光流的医学图像形变配准方法及系统
US11887298B2 (en) * 2020-01-07 2024-01-30 Rensselaer Polytechnic Institute Fluorescence lifetime imaging using deep learning
CN111460991A (zh) * 2020-03-31 2020-07-28 科大讯飞股份有限公司 异常检测方法、相关设备及可读存储介质
CN111369558A (zh) * 2020-04-08 2020-07-03 哈尔滨理工大学 一种基于多模态脑影像的儿童癫痫定位方法
CN111584066B (zh) * 2020-04-13 2022-09-09 清华大学 基于卷积神经网络与对称信息的脑部医学影像诊断方法
CN111640107B (zh) * 2020-06-02 2024-02-06 无锡北邮感知技术产业研究院有限公司 一种致痫灶位置检测方法及装置
CN111951228B (zh) * 2020-07-22 2022-03-15 电子科技大学 一种融合梯度激活映射和深度学习模型的癫痫灶定位系统
CN112185523B (zh) * 2020-09-30 2023-09-08 南京大学 基于多尺度卷积神经网络的糖尿病视网膜病变分类方法
CN113112476B (zh) * 2021-04-14 2023-08-29 中国人民解放军北部战区总医院 一种识别致痫灶和/或预测其病理分型的方法和系统
CN113724307B (zh) * 2021-09-02 2023-04-28 深圳大学 基于特征自校准网络的图像配准方法、装置及相关组件
CN115081486B (zh) * 2022-07-05 2023-07-04 华南师范大学 一种癫痫发作前期颅内脑电网络定位癫痫灶系统及方法
CN116071350B (zh) * 2023-03-06 2023-07-04 同心智医科技(北京)有限公司 基于深度学习的大脑微出血识别方法、装置及存储介质
CN116168352B (zh) * 2023-04-26 2023-06-27 成都睿瞳科技有限责任公司 基于图像处理的电网障碍物识别处理方法及系统
CN117690134B (zh) * 2024-02-02 2024-04-12 苏州凌影云诺医疗科技有限公司 一种esd手术中电刀目标位置的标记辅助方法和装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107403201A (zh) * 2017-08-11 2017-11-28 强深智能医疗科技（昆山）有限公司 肿瘤放射治疗靶区和危及器官智能化、自动化勾画方法
CN108629784A (zh) * 2018-05-08 2018-10-09 上海嘉奥信息科技发展有限公司 一种基于深度学习的ct图像颅内血管分割方法及系统
CN109447966A (zh) * 2018-10-26 2019-03-08 科大讯飞股份有限公司 医学图像的病灶定位识别方法、装置、设备及存储介质
CN109523521A (zh) * 2018-10-26 2019-03-26 复旦大学 基于多切片ct图像的肺结节分类和病灶定位方法和系统
US20190130569A1 (en) * 2017-10-26 2019-05-02 Wisconsin Alumni Research Foundation Deep learning based data-driven approach for attenuation correction of pet data
CN109754387A (zh) * 2018-11-23 2019-05-14 北京永新医疗设备有限公司 医学图像病灶检测定位方法、装置、电子设备及存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TW201615152A (zh) * 2014-10-23 2016-05-01 Univ Nat Yang Ming 用於正子斷層影像之衰減修正方法
US10586335B2 (en) 2017-08-28 2020-03-10 Intel Corporation Hand segmentation in a 3-dimensional image
US11717686B2 (en) * 2017-12-04 2023-08-08 Neuroenhancement Lab, LLC Method and apparatus for neuroenhancement to facilitate learning and performance

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107403201A (zh) * 2017-08-11 2017-11-28 强深智能医疗科技（昆山）有限公司 肿瘤放射治疗靶区和危及器官智能化、自动化勾画方法
US20190130569A1 (en) * 2017-10-26 2019-05-02 Wisconsin Alumni Research Foundation Deep learning based data-driven approach for attenuation correction of pet data
CN108629784A (zh) * 2018-05-08 2018-10-09 上海嘉奥信息科技发展有限公司 一种基于深度学习的ct图像颅内血管分割方法及系统
CN109447966A (zh) * 2018-10-26 2019-03-08 科大讯飞股份有限公司 医学图像的病灶定位识别方法、装置、设备及存储介质
CN109523521A (zh) * 2018-10-26 2019-03-26 复旦大学 基于多切片ct图像的肺结节分类和病灶定位方法和系统
CN109754387A (zh) * 2018-11-23 2019-05-14 北京永新医疗设备有限公司 医学图像病灶检测定位方法、装置、电子设备及存储介质

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112419278A (zh) * 2020-11-25 2021-02-26 上海应用技术大学 一种基于深度学习的实木地板分类方法
CN112419278B (zh) * 2020-11-25 2024-04-19 上海应用技术大学 一种基于深度学习的实木地板分类方法
CN113688942A (zh) * 2021-02-09 2021-11-23 四川大学 基于深度学习的头侧片腺样体图像自动评估方法和装置
CN113269711B (zh) * 2021-04-06 2024-01-09 东北大学 脑图像的处理方法及装置、电子设备和存储介质
CN113269711A (zh) * 2021-04-06 2021-08-17 东北大学 脑图像的处理方法及装置、电子设备和存储介质
CN113506233A (zh) * 2021-07-08 2021-10-15 西安电子科技大学 基于深度学习的sar自聚焦方法
CN113506233B (zh) * 2021-07-08 2024-04-19 西安电子科技大学 基于深度学习的sar自聚焦方法
CN113643336A (zh) * 2021-07-26 2021-11-12 之江实验室 基于球极坐标系深度神经网络的三维图像刚性匹配方法
CN113643336B (zh) * 2021-07-26 2024-03-15 之江实验室 基于球极坐标系深度神经网络的三维图像刚性匹配方法
CN113808097A (zh) * 2021-09-14 2021-12-17 北京主导时代科技有限公司 一种列车的关键部件丢失检测方法及其系统
CN113808097B (zh) * 2021-09-14 2024-04-12 北京主导时代科技有限公司 一种列车的关键部件丢失检测方法及其系统
CN114376522B (zh) * 2021-12-29 2023-09-05 四川大学华西医院 构建用于识别青少年肌阵挛癫痫的计算机识别模型的方法
CN114376522A (zh) * 2021-12-29 2022-04-22 四川大学华西医院 构建用于识别青少年肌阵挛癫痫的计算机识别模型的方法
EP4235685A1 (en) * 2022-02-23 2023-08-30 Siemens Healthcare GmbH Method, system and computer program for detection of a disease information
CN114820535B (zh) * 2022-05-05 2023-09-12 深圳市铱硙医疗科技有限公司 动脉瘤的图像检测方法、装置、计算机设备及存储介质
CN114820535A (zh) * 2022-05-05 2022-07-29 深圳市铱硙医疗科技有限公司 动脉瘤的图像检测方法、装置、计算机设备及存储介质
CN115661680A (zh) * 2022-11-15 2023-01-31 北京轨道未来空间科技有限公司 卫星遥感图像处理方法
CN116051810A (zh) * 2023-03-30 2023-05-02 武汉纺织大学 一种基于深度学习的智慧服装定位方法
CN116958128A (zh) * 2023-09-18 2023-10-27 中南大学 基于深度学习的医学图像自动定位方法
CN116958128B (zh) * 2023-09-18 2023-12-26 中南大学 基于深度学习的医学图像自动定位方法

Also Published As

Publication number Publication date
CN110390351B (zh) 2020-07-24
US20220230302A1 (en) 2022-07-21
CN110390351A (zh) 2019-10-29
US11645748B2 (en) 2023-05-09

Similar Documents

Publication Publication Date Title
WO2020224123A1 (zh) 一种基于深度学习的致痫灶三维自动定位系统
US11151721B2 (en) System and method for automatic detection, localization, and semantic segmentation of anatomical objects
Candemir et al. A review on lung boundary detection in chest X-rays
WO2018120942A1 (zh) 一种多模型融合自动检测医学图像中病变的系统及方法
CN113040715B (zh) 一种基于卷积神经网络的人脑功能网络分类方法
CN109859184B (zh) 一种连续扫描乳腺超声图像的实时检测与决策融合方法
CN114694236B (zh) 一种基于循环残差卷积神经网络的眼球运动分割定位方法
Yoon et al. Medical image analysis using artificial intelligence
Wu et al. Registration of longitudinal brain image sequences with implicit template and spatial–temporal heuristics
CN112365464A (zh) 一种基于gan的医学图像病变区域弱监督定位方法
Saidy et al. Chest X-ray image segmentation using encoder-decoder convolutional network
CN113658721B (zh) 阿尔兹海默疾病进程预测方法
Tripathi et al. Optimal pneumonia detection using convolutional neural networks from x-ray images
Manikandan et al. Segmentation and Detection of Pneumothorax using Deep Learning
Zeng et al. TUSPM-NET: A multi-task model for thyroid ultrasound standard plane recognition and detection of key anatomical structures of the thyroid
CN115210755A (zh) 解决训练数据中遗漏注释的类别不同损失函数
Elkhill et al. Geometric learning and statistical modeling for surgical outcomes evaluation in craniosynostosis using 3D photogrammetry
Dandan et al. A multi-model organ segmentation method based on abdominal ultrasound image
CN113298754A (zh) 一种前列腺组织轮廓线控制点的检测方法
Singh et al. Preprocessing of Medical Images using Deep Learning: A Comprehensive Review
Oliveira et al. Automatic segmentation of posterior fossa structures in pediatric brain mris
Rachmawati et al. Bone Scan Image Segmentation based on Active Shape Model for Cancer Metastasis Detection
Vieira et al. Using a Siamese Network to Accurately Detect Ischemic Stroke in Computed Tomography Scans
Vetrithangam et al. prediction of pneumonia disease from x-ray images using a modified resnet152v2 deep learning model
Ara et al. Novel approach of brain tumor segmentation using convolutional neural network hybridized with water cycle algorithm

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: 19928172

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: 19928172

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 19928172

Country of ref document: EP

Kind code of ref document: A1