CN109003267A - From the computer implemented method and system of the automatic detected target object of 3D rendering - Google Patents
From the computer implemented method and system of the automatic detected target object of 3D rendering Download PDFInfo
- Publication number
- CN109003267A CN109003267A CN201810789942.0A CN201810789942A CN109003267A CN 109003267 A CN109003267 A CN 109003267A CN 201810789942 A CN201810789942 A CN 201810789942A CN 109003267 A CN109003267 A CN 109003267A
- Authority
- CN
- China
- Prior art keywords
- target object
- rendering
- implemented method
- computer implemented
- learning network
- 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.)
- Granted
Links
Classifications
-
- 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
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- 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
- G06T2207/10081—Computed x-ray tomography [CT]
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20112—Image segmentation details
- G06T2207/20164—Salient point detection; Corner detection
-
- 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
- G06T2207/30061—Lung
- G06T2207/30064—Lung nodule
Abstract
This disclosure relates to a kind of computer implemented method and system from the automatic detected target object of 3D rendering.This method may include receiving the 3D rendering obtained by imaging device.This method, which may further include, uses multiple bounding boxes of the 3D learning network detection comprising target object by processor.Learning network can be trained to generate multiple Feature Mappings of different scale based on 3D rendering.This method can also include one group of parameter of the bounding box for determining that identification is each detected using 3D learning network by processor, and position target object based on one group of parameter by processor.This method quickly and accurately and can automatically detect target object from 3D rendering by means of 3D learning network.
Description
Cross reference to related applications
This application claims 62/542, No. 890 priority of the U.S. Provisional Application No. submitted for 9th in August in 2017, complete
Portion's content is incorporated herein by reference.
Technical field
The disclosure relates generally to image procossings and analysis.More specifically, this disclosure relates to for automatically fixed from 3D rendering
The method and system of position and detected target object.
Background technique
The accuracy and therapeutic effect of diagnosis depend on the quality of medical image analysis, especially target object (such as device
Official, tissue, target site etc.) detection.Compared with conventional two-dimensional imaging, volume (3D) imaging, such as volume CT can be captured more
Valuable medical information, to help more accurately to diagnose.But target object is usually by veteran healthcare givers
(such as radiologist) detection, rather than detected by machine, this makes its troublesome, time-consuming and easy error.
One example is that Lung neoplasm is detected from lung images.Fig. 1 shows the shaft-like face image from volume chest CT
An example.High density quality in white border frame corresponds to Lung neoplasm.In order to detect this Lung neoplasm, radiologist
It must screening be hundreds of and thousands of images from volume CT scan.Due to lacking 3d space information, only identified from 2D image
Tubercle is not a simple task.Differentiation lesser tubercle and blood vessel are very difficult in 2D image, because of 2D shaft-like view
In blood vessel be also round or ellipse, look like tubercle.In general, radiologist needs to check adjacent image with virtual
Ground (in brains) rebuilds 3d space relationship and/or checks that sagittal or coronal view (low resolution) are for reference.Therefore, it examines
Survey the experience that Lung neoplasm depends entirely on radiologist.
Although some bare metal learning methods are introduced into for detecting, usually artificially definition is special for these methods
Sign, and therefore accuracy in detection is low.In addition, this machine learning is typically limited to 2D image study, but empty due to lacking 3D
Between information and 3D study needed for considerable computing resource, can not directly detect target object in the 3 d image.
Present disclose provides a kind of method and system, can by means of 3D learning network from 3D rendering quickly and accurately
With automatically detect target object.It is this detection can include but is not limited to positioning target object, determine target object size,
And the type (such as blood vessel or Lung neoplasm) of identification target object.
Summary of the invention
In an arrangement, this disclosure relates to it is a kind of for from the computer realization side of the automatic detected target object of 3D rendering
Method.This method may include receiving the 3D rendering obtained by imaging device.This method may further include to be used by processor
The detection of 3D learning network includes multiple bounding boxes of target object.Learning network can be trained to different to be generated based on 3D rendering
Multiple Feature Mappings of scale.This method can also include determining that identification is each using 3D learning network by processor to detect
Bounding box one group of parameter, and target object is positioned based on one group of parameter by processor.
In some embodiments, one group of parameter includes the seat for identifying the position of each bounding box in the 3D rendering
Mark.
In some embodiments, one group of parameter includes identifying the size of the size of each bounding box.
In some embodiments, the 3D learning network is trained to execute recurrence to one group of parameter.
In some embodiments, the computer implemented method further comprises by multiple anchor frames and the 3D rendering phase
Association, wherein one group of parameter indicates offset of each bounding box relative to respective anchors frame.
In some embodiments, each anchor frame is associated with the grid cell of Feature Mapping.
In some embodiments, the anchor frame is according to the scaling of the Feature Mapping.
In some embodiments, wherein the modified picture size of the multiple Feature Mapping.
In some embodiments, wherein the multiple Feature Mapping uses the sliding window of variable size.
In some embodiments, the computer implemented method further includes creation initial boundary frame, wherein detection includes
Multiple bounding boxes of the target object include being classified as being associated with multiple labels by the initial boundary frame.
In some embodiments, the computer implemented method further includes to the bounding box detected using non-maximum suppression
System.
In some embodiments, the computer implemented method further include: divide the 3D rendering to obtain convex closure simultaneously
And the detection of the multiple bounding box is constrained using the convex closure.
In some embodiments, the learning network is further trained to divide institute in each bounding box detected
State target object.
In some embodiments, wherein the imaging device is Computerized tomographic imaging system.
In some embodiments, the target object is Lung neoplasm.
In some embodiments, the learning network is full convolutional neural networks.
In another program, present disclosure also relates to for the system from the automatic detected target object of 3D rendering.The system can be with
Including being configured as receiving the interface of the 3D rendering obtained by imaging device.The system can also include processor, the processor
It is configured with multiple bounding boxes that the detection of 3D learning network includes target object.Learning network can be trained to based on 3D
Multiple Feature Mappings of image generation different scale.Processor can be configured to determine knowledge using 3D learning network
One group of parameter of the bounding box not detected each, and target object is positioned based on one group of parameter.
In some embodiments, the processor includes graphics processing unit.
In yet another aspect, present disclosure also relates to be stored thereon with the non-transitory computer-readable medium of instruction.Instruction
It can be executed when executed by the processor for the method from the automatic detected target object of 3D rendering.This method may include receiving
The 3D rendering obtained by imaging device.This method, which may further include, includes the more of target object using the detection of 3D learning network
A bounding box.Learning network can be trained to generate multiple Feature Mappings of different scale based on 3D rendering.This method may be used also
To include determining one group of parameter for identifying each bounding box detected using 3D learning network, and be based on one group of ginseng
Number is to position target object.
It should be understood that foregoing general description and following detailed description are only exemplary and explanatory,
It is not the limitation to claimed invention.
Detailed description of the invention
In the attached drawing being not necessarily drawn to scale, identical appended drawing reference can describe similar in different views
Component.Like numeral with letter suffix or different letter suffix can indicate the different instances of similar component.Attached drawing is generally
By way of example rather than the mode of limitation shows various embodiments, and is used for together with specification and claims to institute
Disclosed embodiment is illustrated.In due course, all attached drawings make to be referred to identical reference everywhere it is same or
Similar part.Such embodiment is illustrative, and be not intended as this method, system or thereon have for realizing
The exhaustive or exclusive embodiment of the non-transitory computer-readable medium of the instruction of this method.
Fig. 1 shows the exemplary shaft-like image generated with chest volumetric computed tomography photography；
Fig. 2 shows exemplary nodule detection systems according to an embodiment of the present disclosure；
Fig. 3 shows according to an embodiment of the present disclosure from full articulamentum to the exemplary transformation of full convolutional layer；
Fig. 4 depicts the block diagram for illustrating exemplary medical image processing equipment according to an embodiment of the present disclosure；
Fig. 5 shows the schematic diagram of 3D learning network according to an embodiment of the present disclosure；
Fig. 6 shows the process of the example process according to an embodiment of the present disclosure for training convolutional neural networks model
Figure；
Fig. 7 shows the flow chart of the example process of target object for identification according to an embodiment of the present disclosure；
Fig. 8 shows the example process according to an embodiment of the present disclosure for from the automatic detected target object of 3D rendering；
Fig. 9 shows the exemplary nodule detection of the 3D learning network according to an embodiment of the present disclosure using n scale
Process；And
Figure 10 shows a kind of exemplary nodule segmentation of the 3D learning network according to an embodiment of the present disclosure using scale
Process.
Specific embodiment
Terms used herein " target object " can refer to any anatomical structure in subject's body, such as tissue, device
A part or target site of official.For example, target object can be Lung neoplasm.In the following embodiments, using Lung neoplasm as
The example of " target object " is illustrative and not limiting, but those skilled in the art can be by the lung knot in following each embodiment
Section is easily replaced other kinds of " target object ".
Fig. 2 shows according to an embodiment of the present disclosure for examining from the exemplary tubercle of the automatic detected target object of 3D rendering
Examining system 200.In this embodiment, Lung neoplasm is target object.Lung neoplasm can become the target area (target of the treatments such as radiotherapy
Area).As shown in Fig. 2, nodule detection system 200 includes: nodule detection model training unit 202, for training detection model；With
Nodule detection unit 204, for detecting position and the classification of tubercle object using trained detection model.Trained inspection
Nodule detection unit 204 can be transferred to from nodule detection model training unit 202 by surveying model, so that nodule detection unit 204
Trained detection model can be obtained and be applied to the 3D medicine figure for example obtained from 3D medical image databases 206
Picture.In some embodiments, detection model can be 3D learning network.
For example, the position of tubercle object can be identified by the center and its range of tubercle object.As needed, tubercle pair
The classification of elephant can identify by the label selected from (n+1) a tubercle label, such as, but not limited to non-nodules, the first ruler
Very little tubercle ..., the tubercle of the n-th size.In another example the position may include multiple bounding boxes comprising tubercle object.Optionally
Or additionally, the position of tubercle object may include one group of parameter for identifying each bounding box detected.One group of parameter
It may include the coordinate for identifying the position (for example, center) of each bounding box in 3D medical image.One group of parameter can be with
Size including the size of each bounding box in identification 3D medical image.Based on the bounding box detected and/or identify it one
Group parameter, can position tubercle object.
Training sample can store in training image database 201, and can be by nodule detection model training unit
202 obtain to train detection model.Each training sample includes the position of the tubercle object in medical image and corresponding medical image
It sets and classification information.
In some embodiments, the output result (position and classification including tubercle object) of nodule detection unit 204 can
It is visualized with using with primitive medicine 3D rendering (such as original volumetric CT image) Chong Die thermal map.In some embodiments
In, testing result can be transferred to training image database 201 by network 205, and with corresponding image together as attached
Training sample is added to be added.In this way it is possible to by including new testing result come continuous updating training image database
201.In some embodiments, nodule detection model training unit 202 can use the training sample updated periodically training inspection
Model is surveyed, to improve the accuracy in detection of trained detection model.
3D learning network can be realized by various neural networks.In some embodiments, 3D learning network can be
Feedover 3D convolutional neural networks.Multiple spies can be generated when being applied to Pulmonary volume CT image in this feedforward 3D convolutional neural networks
Sign mapping, each Feature Mapping correspond to corresponding a kind of tubercle object, such as tubercle of non-nodules, first size ..., the n-th ruler
Very little tubercle.In some embodiments, each grid cell of Feature Mapping can indicate in the correspondence area of Pulmonary volume CT image
The existence of the tubercle object of respective type in domain.Based on multiple Feature Mappings, multiple bounding boxes and these boundaries are generated
The scoring of the existence of tubercle object in frame.For example, score 1.0 can indicate the tubercle in bounding box there are respective type
Object, score 0.0 can indicate that the score being not present between the tubercle object and 0.0 and 1.0 of respective type in bounding box can
To show that there are the probability of the tubercle object of respective type in bounding box.In some embodiments, feedover 3D convolutional Neural net
It can be with right and wrong maximum suppression layer to generate final detection result after network.It alternatively, can be on 3D convolutional neural networks
It is additional to assist full articulamentum or the full convolutional layer of auxiliary as detection layers.In some embodiments, bounding box can be 3D's and by
One group of parameter identification.For example, it can pass through the frame on the coordinate (x, y, z) and x-axis, y-axis and z-axis at the center of bounding box respectively
Size (size_x, size_y, size_z) identifies.In some embodiments, 3D convolutional neural networks can be trained to come to institute
It states one group of parameter and executes recurrence, therefore the result of 3D convolutional neural networks may include that the classification results of tubercle object (detect
Tubercle object type) and the type for the tubercle object accordingly detected 6 regression parameter values.
In one embodiment, the feedforward 3D convolutional neural networks may include basic network, and can be from basic
Network obtains the Feature Mapping of scale 1 (the first scale).For example, basic network may include three convolution blocks and three detection layers
Fc1, fc2 and fc3, each convolution block are made of the maximum pond layer in two 3 × 3 × 3 convolutional layers, ReLU layers and 2 × 2 × 2.Convolution
Convolutional layer 1 and convolutional layer 2 in block 1 have 64 Feature Mappings, and the convolutional layer 1 and convolutional layer 2 in convolution block 2 have 128
Feature Mapping, and convolutional layer 1 in convolution block 3 and convolutional layer 2 have 256 Feature Mappings.In some embodiments, fc1,
Fc2 and fc3 can be the full articulamentum of auxiliary for classification task.In one embodiment, fc1 has after ReLU layers
512 neurons, fc2 has 128 neurons after ReLU layers, and the quantity of neuron possessed by fc3 depends on
Classification.For example, if by tubercle object be divided into 10 classes (such as non-nodules, the tubercle of the 1st size, the 2nd size tubercle ..., the 9th
The tubercle of size), then the neuronal quantity in fc3 layers is 10.
It in another embodiment, can be by the way that above-mentioned full articulamentum fc1, fc2 and fc3 be transformed to full convolutional layer respectively
Fc1-conv, Fc2-conv and Fc3-conv modify basic network.Therefore, because the acceleration for carrying out convolution algorithm to image is imitated
Fruit can be calculated acceleration based on modified basic network.Fig. 3 shows according to an embodiment of the present disclosure from full articulamentum
To the exemplary transformation of full convolutional layer.In some embodiments, the kernel size of Fc1-conv can be with the ruler of Feature Mapping
It is very little identical, if it is desired, Feature Mapping is, by pond, and Fc2-conv and Fc3-conv have after the output of convolution block 3
The kernel size of 1x1x1.In some embodiments, the quantity of the Feature Mapping of three full convolutional layers and corresponding full articulamentum
The quantity of Feature Mapping is identical.As shown in figure 3, weight w00, w01, w10 and w11 of convolution kernel are converted from corresponding full articulamentum
Respective weights w00, w01, w10 and w11.It in some embodiments, can be according to the quantity of convolution kernels size and Feature Mapping
To remold the weight of full articulamentum.
In one embodiment, above-mentioned basic network or its revision may be used as 3D learning network directly to detect side
Boundary's frame.Basic network can be applied to generate multiple Feature Mappings on the 3D rendering of input, and each Feature Mapping corresponds to pair
The particular category of elephant.In some embodiments, each grid cell of Feature Mapping corresponds to the relevant block in 3D rendering.Example
Such as, ith feature corresponding with the i-th class object is mapped, the value of grid cell can indicate that the i-th class object is present in 3D
Probability in the corresponding blocks of image.In some embodiments, can the value based on the correspondence grid cell of each Feature Mapping come
Classify to the object in relevant block.In addition, by by the coordinate of grid cell from Feature Mapping spatial alternation to 3D rendering
Initial boundary frame can be generated and using classification results as its label in space.In some embodiments, the volume for not filling
Product operation, can be used formula (1) to execute transformation.
Wherein, xfIt is the coordinate in predicted characteristics mapping, xoriIt is the coordinate in image space, s1、s2And s3Scale because
Number,It is lower rounding operation, ci_j(i=1,2,3, j=1,2) is the convolution kernels size of jth convolutional layer in the i-th convolution block,
c1_1=c1_2=c2_1=c2_2=c3_1=c3_2=3, c4=8, c5=1, c6=1.
In some embodiments, for using the convolution algorithm of filling, formula (2) can be used to execute transformation.
In some embodiments, 3D learning network can have several scales.It is, for example, possible to use be based on Analysis On Multi-scale Features
The convolutional network of mapping.The quantity of scale can be determined based on different Detection tasks.The Analysis On Multi-scale Features of convolutional network can
To realize in various ways.For example, multiple Feature Mappings can have identical scale size, but use different size of cunning
Dynamic window obtains.As another example, convolution filter or pond filter or both can be used will be under Feature Mapping
Different scales is sampled, while using the sliding window of identical size.As another example, down-sampling layer can also be used
It is different scales etc. by Feature Mapping down-sampling.It can be added using the disclosed convolutional network based on Analysis On Multi-scale Features mapping
Speed calculates, to make the detection based on 3D learning network that can also fit while being able to detect the object with extensive size range
For clinic.
In some embodiments, 3D learning network is (also referred to as complete using a series of full convolution filters on each scale
Articulamentum) generate fixed quantity testing result.In some embodiments, 3D learning network can return to multiple bounding boxes, often
A bounding box and two parts are associated: one group of parameter of object classification and the corresponding bounding box of identification.Object classification can have c
Classification.In one embodiment, c=3, wherein three object type be respectively non-lung background, non-nodules but for lung tissue and
Tubercle.For tubercle classification, bounding box surrounds corresponding tubercle object.It in some embodiments, can be each Feature Mapping
Grid cell introduces multiple anchor frames, to make the bounding box detected preferably track target object.For example, identifying corresponding boundary
One group of parameter of frame can be the returning on coordinate (centered on the grid cell assessed) and size relative to corresponding anchor frame
The offset returned.For tubercle classification, offset can be relative to respective anchors frame coordinate and size relative value (dx, dy,
Dz, dsize_x, dsize_y and dsize_z).For the Feature Mapping of each scale, if by k anchor frame and therein each
Grid cell is associated, then the corresponding grid cell acquisition of each of Feature Mapping of s scale k*s bounding box (and k*s in total
A anchor frame).Then, for each bounding box in k*s bounding box, the score of each classification in c classification can be calculated
And its 6 offsets relative to respective anchors frame.This will lead to (the c+ applied in Feature Mapping around each position in total
6) * k*s filter.In the case where the size of each Feature Mapping is m*n*d, it can produce (c+6) * k*s*m*n*d
Output.In some embodiments, anchor frame is 3D and associated with the Feature Mapping of different scale.In some embodiments,
The scaling that anchor frame can be mapped based on character pair.It is alternatively possible to be adjusted according to image information by regression algorithm
The positions and dimensions of anchor frame.
Fig. 4, which is depicted, illustrates according to an embodiment of the present disclosure be suitable for from the exemplary of the automatic detected target object of 3D rendering
The block diagram of Medical Image Processing equipment 300.Medical Image Processing equipment 300 may include network interface 328, be connect by means of network
Mouthfuls 328, Medical Image Processing equipment 300 may be coupled to network (not shown), such as, but not limited to the local area network in hospital or
Internet.Network can be by the external device (ED) of medical image processing devices 300 and such as image collecting device (not shown), medicine
Image data base 325, image data storage apparatus 326 connect.Image collecting device can be any of the image for obtaining object
Device, such as DSA imaging device, MRI imaging device, CT imaging device, PET imaging device, ultrasonic device, fluoroscopy are set
Other medical imaging devices of standby, SPECT imaging device or the medical image for obtaining patient.For example, imaging device can be with
It is lung CT imaging device etc..
In some embodiments, medical image processing devices 300 can be special intelligent device or universal intelligent device.Example
Such as, device 300 can be the service of the computer or cloud for image data acquiring and image data processing tasks customization
Device.For example, device 300 can be integrated into image collecting device.Optionally, list is rebuild the apparatus may include or with 3D
Member cooperation, 3D reconstruction unit rebuild 3D rendering for the 2D image based on being obtained by image acquisition equipment.
Medical image processing devices 300 may include image processor 321 and memory 322, and can additionally include
At least one of input/output 327 and image display 329.
Image processor 321 can be including one or more general purpose processing devices (such as microprocessor, central processing list
First (CPU), graphics processing unit (GPU) etc.) processing equipment.More specifically, image processor 321 can be complicated order
Collection calculates (CISC) microprocessor, reduced instruction set computing (RISC) microprocessor, very long instruction word (VLIW) microprocessor, fortune
The processor of other instruction set of row or the combined processor of operating instruction collection.Image processor 321 is also possible to one or more
A dedicated treatment facility, such as specific integrated circuit (ASIC), field programmable gate array (FPGA), digital signal processor
(DSP), system on chip (SoC) etc..As it will appreciated by a person of ordinary skill, in some embodiments, image processor 321
It can be application specific processor, rather than general processor.Image processor 321 may include that one or more known treatments are set
Pentium that is standby, such as being manufactured by Intel companyTM、CoreTM、XeonTMOr ItaniumTMThe microprocessor of series, by AMD
The Turion of company's manufactureTM、AthlonTM、SempronTM、OpteronTM、FXTM、PhenomTMThe microprocessor or the sun of series
Micro-system (Sun Microsystems) manufacture various processors it is any.Image processor 321 can also include figure
Processing unit, such as from the manufacture of Nvidia company The GPU of series, by English spy
GMA, Iris of your TM manufactureTMThe GPU of the series or Radeon manufactured by Advanced Micro DevicesTMSerial GPU.Image processor 321
It can also include the processing unit accelerated, desktop A-4 (6, the 8) series of such as Advanced Micro Devices manufacture, Intel company's manufacture
Xeon PhiTMSeries.The disclosed embodiments are not limited to any kind of processor or processor circuit, these processors or place
Reason device circuit is otherwise configured to meet following calculating demand: identification, analysis, maintenance, generate and/or provide largely at
As data or the such imaging data of manipulation with consistently detect and position from 3D rendering with the disclosed embodiments target object or
Manipulate the data of any other type.In addition, term " processor " or " image processor " they may include more than one processor,
For example, multi core design or multiple processors, each processor in the multiple processor has multi core design.Image processor
321 can execute the sequence for the computer program instructions being stored in memory 322, with execute it is disclosed herein it is various operation,
Process, method.
Image processor 321 can be communicably coupled to memory 322 and be configured as executing the meter being stored therein
Calculation machine executable instruction.Memory 322 may include read-only memory (ROM), flash memory, random access memory (RAM), such as
The dynamic random access memory (DRAM) of synchronous dram (SDRAM) or Rambus DRAM, static memory (for example, flash memory,
Static random access memory) etc., store computer executable instructions in any format thereon.In some embodiments, it stores
Device 322 can store the computer executable instructions of one or more image processing programs 223.Computer program instructions can be by
Image processor 321 accesses, and reads from ROM or any other suitable storage location, and be loaded into RAM for image procossing
Device 321 executes.For example, memory 322 can store one or more software applications.It is stored in soft in memory 322
Part application program may include for example for the operating system (not shown) of general-purpose computing system and soft control equipment.This
Outside, memory 322 can store entire software application or only store a part (such as image procossing of software application
Program 223) can be executed by image processor 321.In addition, memory 322 can store multiple software modules, for realizing
With each step of the process of the consistent method or training 3D learning network from the automatic detected target object of 3D rendering of the disclosure.
In addition, memory 322 can store generation/caching data when executing computer program, such as medical image 324,
It includes the medical image from transmissions such as image collecting device, medical image databases 325, image data storage apparatus 326.This
The medical image 324 of sample may include the 3D medicine figure detected automatically that will carry out it target object received
Picture.In addition, medical image 324 can also include target object testing result of the 3D medical image together with it.
Image processor 321 can execute image processing program 223 to realize for detecting target pair automatically from 3D rendering
The method of elephant.In some embodiments, when executing image processing program 223, image processor 321 can scheme corresponding 3D
As with include that the testing result of object classification and the bounding box detected is associated, and by 3D rendering together with (such as with marking
Having) testing result storage is to memory 322.Optionally, memory 322 can be communicated with medical image databases 325 with therefrom
It obtains image (wherein there is object to be detected) or 3D rendering is sent jointly into medical image together with testing result
Library 325.
In some embodiments, 3D learning network can be stored in memory 322.Optionally, 3D learning network can
To be stored in remote equipment, discrete database (such as medical image databases 325), in distributed apparatus, and can be with
It is used by image processing program 223.3D rendering can be used as training sample together with testing result and be stored in medical image databases
In 325.
Input/output 327, which can be configured as, allows medical image processing devices 300 to receive and/or send data.It is defeated
Enter/export 327 may include allow equipment 300 communicate with user or other machines and equipment one or more digital and/or
Artificial traffic equipment.For example, input/output 327 may include that user is allowed to provide the keyboard and mouse of input.
Network interface 328 may include network adapter, cable connector, serial connector, USB connector, connect parallel
Connect device, the high speed data transfer adapter of such as optical fiber, USB 3.0, lightning, wireless network adapter such as WIFI adapter, electricity
Believe (3G, 4G/LTE etc.) adapter.Device 300 can be connected to network by network interface 328.Network can provide local area network
(LAN), wireless network, cloud computing environment are (for example, the software as service, the platform as service, set as the basis of service
Apply), the function of client-server, wide area network (WAN) etc..
Other than showing medical image, image display 329 can also show other information, such as classification results and inspection
The bounding box measured.For example, image display 329 can be LCD, CRT or light-emitting diode display.
There has been described various operations or functions, which can be implemented as software code or instruction or are defined as software generation
Code or instruction.Such content can be the source code or variance codes that can directly execute (" object " or " executable " form)
(" increment " or " block " code).Software code or instruction may be stored in a computer readable storage medium, and work as and be performed
When, machine can be made to execute described function or operation, and including for storing information in the form of machine-accessible
Any mechanism (for example, calculating equipment, electronic system etc.), such as recordable or non-recordable medium is (for example, read-only memory
(ROM), random access memory (RAM), magnetic disk storage medium, optical storage media, flash memory device etc.).
As described above, 3D learning network according to an embodiment of the present disclosure can be worked with end to end system, and directly pre-
Survey tubercle classification and bounding box.
In some embodiments, in order to reduce calculating and carrying cost, training program stage by stage can be used.Training program
Three phases can be divided into: (1) unsupervised learning；(2) the sorter network training based on small image block；(3) it is based on big image block
Detection network training.In some embodiments, stage (1) and (2) can be used for training sorter network (one of detection network
Point), all basic networks as disclosed herein, to generate good netinit for entire detection network.In some implementations
In example, the stage (3) can carry out end-to-end training on big image block.
In some embodiments, (such as modern times of image processor 321 cannot be encased in if original 3D rendering is too big
GPU) in memory, then it can be divided into multiple big image blocks according to the memory size of image processor 321 with civilian dress
Enter wherein.By the way that original 3D rendering is divided into small image block and big image block, and utilize the rank including unsupervised training
Segmentation training program, for the sorter network in small image block training 3D detection network, then in trained sorter network
On the basis of training 3D detect network, the required total calculating of training can be substantially reduced to be realized by modern GPU.
In one embodiment, 3D convolution autocoder can be used to generate, as shown in Figure 5 in initial network weight.
In some embodiments, encoder section is made of concatenated convolutional block (such as 3 convolution blocks), and decoder section is by corresponding to
In the cascade of 3 convolution blocks of encoder section deconvolute block composition.In the block that deconvolutes, the layer that deconvolutes is by up-sampling layer and it
Convolutional layer composition afterwards.As shown in figure 5, input picture block, executes convolution algorithm by encoder section, and by decoder section
De-convolution operation is executed, the image block of prediction is then exported.3D convolution autocoder can be trained so that output image block with
Input picture block (target image block) is identical.In one embodiment, the noise of such as Gaussian noise etc can be added to
It the image block of input but is not added in target output image block, to make study that there is robustness.In some embodiments,
Both input picture block and target output image block can be converted, rotation, deformation, amplification/diminution etc., so that
Study has higher robustness.Then, it has been initialized using the network weight of the encoder (that is, basic network) after training
The training process of supervision.
Detecting network includes the sorter network.It in one embodiment, can the training point before training detects network
Class network.Fig. 6 shows the flow chart for training the convolutional neural networks training process 400 of sorter network therein.Process
400 start from receiving the step 450 of 3D training image and relevant classification result.It in some embodiments, can be in step 452
3D training image is for example divided into image block using sliding window.Then, in step 454, by single image block together with its point
Class result is input in sorter network together as training data.In some embodiments, the weight of sorter network may be
It is initialised.In step 456, the classifier parameters of sorter network can be determined based on training data.In step 458, classification
The determination of device parameter may include being verified for loss function.In some embodiments, step 456 and 458 can also be by
It is integrated in identical step, wherein each image block can be based on, joins for the classifier of loss function Optimum Classification network
Number.In some embodiments, optimization process can be executed by any algorithms most in use, and including but not limited to gradient decline is calculated
Method, Newton method, conjugate gradient algorithms, quasi-Newton method and Levenberg Marquardt algorithm etc..In step 460, it determine that being
It is no to have been processed by all image blocks, if it is, will there is training for the classifier parameters currently optimized in step 462
Sorter network output be used as trained model.Otherwise, the process back to step 454 to handle subsequent image block, until
All image blocks are handled.In step 452, sliding window can be used by 3D training image and be divided into image block.In a reality
Apply in example, as above by reference to explaining Fig. 3, convolutional network as sorter network (or detection network) it is last it is several entirely
Articulamentum can be converted into full convolutional layer.The striding just as sliding window that stride of convolution algorithm, these operations are equivalent.
Since the fast convolution on GPU calculates, huge acceleration is obtained.
In some embodiments, it is based on trained sorter network, can be constructed on big image block and training detects net
Network.For example, the training process 400 of sorter network can be adapted for training detection network.It distinguishes as follows: being inputted in step 454
Training data is big image block and the information about the bounding box comprising target object wherein detected, and such as identification is corresponding
One group of parameter of the bounding box detected.Such one group of parameter can include but is not limited to label (with the mesh in big image block
Mark object is associated) and one group of location parameter of bounding box for detecting.As an example, label 0 can indicate 3D rendering block not
Comprising tubercle (that is, the bounding box wherein detected includes non-nodules target object), and label 1-9 can respectively indicate 3D figure
(that is, the bounding box wherein detected includes the tubercle target object of size n, n is 1-9 to tubercle as block comprising different size 1-9
Integer in range).There is also differences in step 456 and 458, wherein to be to belong to inspection for the parameter of loss function optimization
The parameter of survey grid network.
In some embodiments, in order to train detection network, label can be distributed to each anchor frame.In one embodiment
In, if friendship that anchor frame and any ground truth frame overlap and being higher than specific threshold (such as 0.7) than (IoU), correspond to ground
The label of face true value frame can be assigned to anchor frame.It note that single ground truth frame (such as tight bounding box comprising tubercle)
Respective labels can be distributed to several anchor frames.For example, for non-nodules anchor frame for the IoU ratio of ground truth frame outside all lungs
Rate is lower than 0.3, then intrapulmonary non-nodules label can be distributed to the non-nodules anchor frame.On the contrary, if non-nodules anchor frame is for institute
There is the IoU ratio of intrapulmonary ground truth frame lower than 0.3, non-nodules label outside lung can be distributed to the non-nodules anchor frame.
In one embodiment, the loss function for training detection network can be multitask loss function, with covering
Both classification task and bounding box prediction task.For example, multitask loss function can be defined by formula (3).
Wherein i is the index of anchor frame in trained small lot, piIt is the prediction probability that anchor frame i is tubercle.Ground truth label
It is pi*, tiIndicate 6 parametrization coordinates of the bounding box of prediction, tiIt * is the ginseng of ground truth frame associated with tubercle anchor frame
Numberization coordinate.LclsIt is to intersect entropy loss, LregIt is robust loss function.In some embodiments, NclsAnd NregCan be respectively
The quantity that frame is corresponded in small lot, for normalizing.λ is the weighting parameters between classification task and recurrence task.
As an example, bounding box can be returned and uses 6 parametric regressions, and this 6 ginsengs can be defined by formula (4)
Number.
Wherein x, y, z, w, h and d indicate bounding box centre coordinate and its width, height and depth.Variable x, xaAnd x*Point
Bounding box, anchor frame and the ground truth frame (being equally applicable to y, z, w, h and d) that Yong Yu do not predict.
Fig. 7 shows the flow chart of the example process 500 of the target object in 3D rendering scanning for identification.Target pair
As identification process 500 step 512 start, receive trained nodule detection model in this step.In step 514, receive
It may include the 3D medical image of target object.Then, in step 452, sliding window can be used and divide 3D medical image
For image block.In step 516, multiple bounding boxes are detected from image block using detection network.In step 518, determine that identification is each
The label of a bounding box and one group of location parameter.Alternatively, step 516 and 518 can integrate in same step.Then,
It can be classified using label to bounding box in step 520, and classified bounding box and its location parameter can be used
To position the target object of such as tubercle.In step 460, it is determined whether all image blocks have been handled, if it is, in step
462, the nodule detection result of each image block is integrated, to obtain and export complete nodule detection result.If it is not, then place
Reason 500 returns to step 516.Although recongnition of objects process 500 as shown in Figure 7 be based on sliding window, it is unlimited
In this.In one embodiment, as above by reference to described in Fig. 3, convolutional network as sorter network (or detection network)
Last several full articulamentums can be converted into full convolutional layer.The striding of convolution algorithm is played similar with striding for sliding window
Effect, these operations are equivalent.Since the fast convolution on GPU calculates, huge acceleration is obtained.
The exemplary mistake being used for from the automatic detected target object of 3D rendering of Fig. 8 diagram according to another embodiment of the present disclosure
Journey.As shown in figure 8,3D Pulmonary volume CT image input detection systems, the system can be detected it using trained model
In tubercle, and identify include each tubercle corresponding bounding box, the positions and dimensions including corresponding bounding box.
In some embodiments, 3D rendering can be divided into along various directions (such as, but not limited to direction z) smaller
Chunking, then can will test network and its related algorithm and be applied to each chunking to obtain the corresponding detection of target object and tie
Fruit.The testing result of each chunking can be gathered, and the chunking with corresponding testing result can be integrated, to generate for whole
The complete testing result of a 3D rendering.
Fig. 9 shows the exemplary nodule detection process of the 3D learning network using n scale.As shown in figure 9, the 3D of input
Medical image is W*H*Z volume CT scan.It is fitted into GPU memory to will test network and Feature Mapping, inputs CT scan quilt
It is divided into smaller chunking, it is noted that two chunking W*H*Z have been illustrated in Fig. 91CT scan and W*H*Z2CT scan, as illustrating
An example with detection process is illustrated, can actually select the quantity of chunking, as needed to be suitble to the property of GPU
Energy.For example, basic network corresponds to scale 1, and the learning network for corresponding to other scales including 2~n of scale can lead to
It crosses and scaling (rescale) operation of basic network is realized, including but not limited to convolution sum maximum pond.Each chunking can
To there is 3 class labels for bounding box.The Feature Mapping of 3 classifications of various scales can be used to detect each chunking
Interior bounding box.The testing result may include 6 offset parameters of the class label of bounding box and the recurrence relative to anchor frame
(dx, dy, dz, dsize_x, dsize_y and dsize_z).The bounding box of all chunkings detected, i.e., more frames shown in Fig. 9
Prediction, can organize merging and be transformed into coordinates of original image coordinates system, carry out the non-maximum suppression of 3D, then to obtain final detection knot
Fruit.By the non-maximum suppression of 3D, redundancy bounding box can be cancelled, to simplify and clarify the inspection of target object in 3D medical image
Survey/positioning result.For example, as the non-maximum suppression of 3D as a result, can determine the boundary detected for a tubercle
Frame.
It is alternatively possible to execute segmentation before running detection algorithm, 3D medical image is tied to will test algorithm
Potential region rather than entire 3D medical image.Thus, it is possible to improve detection accuracy, while can reduce needed for detection network
Calculation amount.
Using Lung neoplasm as the example of target object, it is known that Lung neoplasm is always in intrapulmonary.It in one embodiment, can be pre-
Lung segmentation is first carried out further to remove the false alarm of intrapulmonary.Particularly, can to first carry out in advance lung segmentation convex to generate lung
Packet, then constrains nodule detection using convex closure.Lung segmentation can be executed by various means, including but not limited to convolution
Network, movable contour model, fractional spins etc..In some embodiments, it can be executed in low resolution scanning
This lung segmentation, and original resolution can be upsampled to result, 3D learning network and Feature Mapping are packed into GPU
Memory, while accelerating cutting procedure.
In clinical setting, radiologist usually requires to carry out quantitative analysis to Lung neoplasm.For example, in addition to detecting
Except bounding box, they also need the boundary of tubercle, and the accurate size etc. of the detected tubercle depending on nodule segmentation.?
In some embodiments, segmentation can be executed based on the bounding box detected.It is, for example, possible to use 3D convolutional networks to detect
Lung neoplasm bounding box in execute segmentation.Parted pattern/learning network can be instructed on lesser nodule image block as a result,
The image-region practiced and be applied in detected bounding box.
In one embodiment, as shown in Figure 10, nodule segmentation and detection can be integrated into during nodule segmentation, with
Realize detection and cutting procedure end to end.It, can be with although input W*H*Z 3D scanning is divided into two CT scan in Figure 10
Any appropriate number of CT scan can be divided by imagining input 3D scanning.It, can be basic for the CT scan of each division
Various scaling operations, including the operation of convolution sum maximum pondization are executed on network, to obtain the testing result of corresponding scale.One
Denier detects bounding box, bounding box can be scaled back to Feature Mapping space (for example, realigning characteristic layer to the end), so
ROI (area-of-interest) pond is applied on it afterwards, to generate ROI region.Partitioning algorithm can be executed to each ROI with straight
Deliver a child into nodule segmentation.As an example, this partitioning algorithm can be realized by full convolutional network dividing layer.It is such
Partitioning algorithm can also be realized by a series of warp laminations after convolutional layer or up-sampling layer.In some embodiments,
The nodule segmentation result of each segmentation CT scan can be integrated, to obtain the entire tubercle point in the original CT scan inputted
Cut result.In one embodiment, pond is of the same size ROI using bilinear interpolation and resampling, to accelerate
GPU is calculated.
It is envisioned that nodule segmentation process as shown in Figure 10, can extend from previous detection network, wherein examining
It surveys and the segmentation stage can share identical detection network, such as basic network.
In some embodiments, the training of network used in Figure 10 can be proceed as follows.Firstly, training detection net
Network detects the weight of network portion to obtain.Then, give detection network portion weight, using ground truth bounding box to point
Network portion is cut to be trained.Segmentation network portion can be trained using several loss functions, before being including but not limited to based on
The normalization cross entropy of scape and background voxels quantity.Above two network can be combined to obtain nodule segmentation result.
Detection network portion and segmentation network portion can respectively, be sequentially or simultaneously trained to.Implement at one
In example, during the training stage, detection network portion and segmentation both network portion can simultaneously for associated losses function into
Row training, wherein carrying out supervised segmentation using ground truth bounding box and segmentation.Associated losses function can be defined by formula (5).
Wherein, item is identical those of in the item with formula (3) of front and therefore omits its definition.Last last item is segmentation
Lose component.NsegIt is the quantity of cut zone in a small lot.LsegIt is the voxel losses function in a region, j is instruction
Practice the index of area-of-interest in small lot, SjIt is the prediction probability of area-of-interest, SjIt * is ground truth segmentation.
The description of front is for illustrative purposes and presents.This is not exhausted, and be not limited to it is disclosed really
Cut form or embodiment.In view of the explanation and practice of the disclosed embodiments, the modifications and changes of embodiment will become it is aobvious and
It is clear to.
It in the document, include one or more using term " one " or " a " as common in the patent literature
In one, independently of "at least one" or any other example or usage of " one or more ".Herein, unless in addition referring to
Out, term "or" is for referring to nonexcludability, or " A or B " is made to include " A but do not include B ", " B but do not include A " and " A with
B".In the document, term " including (including) " and " wherein (in which) " be used as corresponding term " including
And the plain English equivalent of " wherein (wherein) " (comprising) ".Moreover, in the following claims, term
" including (including) " and " including (comprising) " is open, that is, including in addition to those exist in the claims
Equipment, system, equipment, product, composition, formula or the process of element other than the element listed after the term, are also regarded as and fall
Enter in the scope of protection of the claims.In addition, in the following claims, term " first ", " second " and " third " etc.
It is merely used as label, it is no intended to which requirement numerically is applied to its object.
Illustrative methods described herein can be at least partly machine or computer implemented.Some examples can wrap
The computer-readable medium or machine readable media with instruction encoding are included, described instruction can be operated to configure electronic equipment and execute such as
Method described in above example.The realization of this method may include software code, such as microcode, assembler language code,
More advanced language codes etc..Various software programming techniques can be used to create in various programs or program module.For example, can be with
Program segment or program module are designed using Java, Python, C, C++, assembler language or any of programming language.One
Or multiple such software sections or module can be integrated into computer system and/or computer-readable medium.It is this soft
Part code may include the computer-readable instruction for executing various methods.Software code can form computer program product
Or a part of computer program module.In addition, in one example, software code can such as during execution or other when
Between be tangibly stored in one or more volatibility, non-transitory or non-volatile visible computer readable medium.These have
The example of the computer-readable medium of shape can include but is not limited to hard disk, moveable magnetic disc, removable CD (for example, CD
And digital video disc), cassette, storage card or stick, random access memory (RAM), read-only memory (ROM) etc..
In addition, although there is described herein illustrative embodiments, range include with based on the disclosure equivalent elements,
Modification is omitted, any and all embodiments of combination (for example, combination of the scheme across various embodiments), adjustment or change.Power
Benefit require in element will be construed broadly as based on language used in claim, and be not limited in this specification or
The example of the duration description of the application, these examples are to be interpreted as nonexcludability.In addition, the step of disclosed method
Suddenly it can modify in any way, including by rearrangement step or insertion or delete step.Therefore, it is intended that only will
Description is considered as example, and real range is indicated by following following claims and its whole equivalency ranges.
Above description is intended to illustrative rather than restrictive.For example, above-mentioned example (or one or more side
Case) it can be in combination with one another.Other embodiments can be used when checking above description by those of ordinary skill in the art.And
And in detailed description above, various features can be grouped together to simplify the disclosure.This is not construed as being intended to
Not claimed open feature is allowed all to be essential for any claim.But subject matter can be
It is combined in the feature fewer than all features of a disclosed embodiment.Therefore, thus following following claims is used as example or reality
Example is applied to be incorporated into specific embodiment, wherein each claim is independently used as individual embodiment, and be contemplated that
It is that these embodiments can be combined with each other with various combinations or displacement.The scope of the present invention should refer to appended claims
And it assigns the full scope of the equivalent of these claims and determines.
Claims (20)
1. a kind of computer implemented method from the automatic detected target object of 3D rendering, comprising:
Receive the 3D rendering obtained by imaging device；
Multiple bounding boxes of the 3D learning network detection comprising the target object are used by processor, wherein the learning network
It is trained to generate multiple Feature Mappings with variation scale based on the 3D rendering；
One group of parameter for identifying each bounding box detected is determined using the 3D learning network by processor；And
One group of parameter is based on by processor and positions the target object.
2. computer implemented method according to claim 1, wherein one group of parameter includes identifying in the 3D rendering
Each bounding box position coordinate.
3. computer implemented method as described in claim 1, wherein one group of parameter includes identify each bounding box big
Small size.
4. computer implemented method according to claim 1, wherein the 3D learning network is trained to described one group
Parameter executes recurrence.
5. computer implemented method according to claim 1 further comprises that multiple anchor frames are related to the 3D rendering
Connection, wherein one group of parameter indicates offset of each bounding box relative to respective anchors frame.
6. computer implemented method as claimed in claim 5, wherein each anchor frame is related to the grid cell of Feature Mapping
Connection.
7. computer implemented method according to claim 6, wherein the anchor frame contracts according to the scale of the Feature Mapping
It puts.
8. computer implemented method according to claim 1, wherein the modified image ruler of the multiple Feature Mapping
It is very little.
9. computer implemented method according to claim 1, wherein the multiple Feature Mapping uses the cunning of variable size
Dynamic window.
10. computer implemented method according to claim 1 further includes creation initial boundary frame, wherein detection includes institute
The multiple bounding boxes for stating target object include being classified as being associated with multiple labels by the initial boundary frame.
11. computer implemented method according to claim 1 further includes to the bounding box detected using non-maximum suppression
System.
12. computer implemented method according to claim 1, further includes: divide the 3D rendering with obtain convex closure and
The detection of the multiple bounding box is constrained using the convex closure.
13. computer implemented method according to claim 1, wherein the learning network is further trained each
The segmentation target object in the bounding box detected.
14. computer implemented method according to claim 1, wherein the imaging device is computed tomography imaging
System.
15. computer implemented method according to claim 1, wherein the target object is Lung neoplasm.
16. computer implemented method according to claim 1, wherein the learning network is full convolutional neural networks.
17. a kind of system from the automatic detected target object of 3D rendering, comprising:
Interface is configured as receiving the 3D rendering obtained by imaging device；And
Processor is configured as:
Multiple bounding boxes comprising the target object are detected using 3D learning network, wherein the learning network be trained to
Multiple Feature Mappings with variation scale are generated based on the 3D rendering；
One group of parameter for identifying each bounding box detected is determined using 3D learning network；And
The target object is positioned based on one group of parameter.
18. system as claimed in claim 17, wherein the processor includes graphics processing unit.
19. system according to claim 17, wherein imaging device is Computerized tomographic imaging system.
20. a kind of non-transitory computer-readable medium for being stored thereon with instruction, wherein described instruction is executed by processor
Method of the Shi Zhihang from the automatic detected target object of 3D rendering, which comprises
Receive the 3D rendering obtained by imaging device；
Multiple bounding boxes comprising the target object are detected using 3D learning network, wherein the learning network be trained to
Multiple Feature Mappings with variation scale are generated based on the 3D rendering；
One group of parameter of the bounding box that identification each detects is determined using the 3D learning network；And
Target object is positioned based on one group of parameter.
Applications Claiming Priority (4)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US201762542890P | 2017-08-09 | 2017-08-09 | |
| US62/542,890 | 2017-08-09 | ||
| US15/996,434 US10867384B2 (en) | 2017-08-09 | 2018-06-02 | System and method for automatically detecting a target object from a 3D image |
| US15/996,434 | 2018-06-02 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109003267A true CN109003267A (en) | 2018-12-14 |
| CN109003267B CN109003267B (en) | 2021-07-30 |
Family
ID=64599870
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810789942.0A Active CN109003267B (en) | 2017-08-09 | 2018-07-18 | Computer-implemented method and system for automatically detecting target object from 3D image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109003267B (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109766887A (en) * | 2019-01-16 | 2019-05-17 | 中国科学院光电技术研究所 | A kind of multi-target detection method based on cascade hourglass neural network |
| CN110059548A (en) * | 2019-03-08 | 2019-07-26 | 北京旷视科技有限公司 | Object detection method and device |
| CN110852314A (en) * | 2020-01-16 | 2020-02-28 | 江西高创保安服务技术有限公司 | Article detection network method based on camera projection model |
| CN110930454A (en) * | 2019-11-01 | 2020-03-27 | 北京航空航天大学 | Six-degree-of-freedom pose estimation algorithm based on boundary box outer key point positioning |
| CN111523452A (en) * | 2020-04-22 | 2020-08-11 | 北京百度网讯科技有限公司 | Method and device for detecting human body position in image |
| CN111563496A (en) * | 2019-02-13 | 2020-08-21 | 西门子医疗有限公司 | Continuous learning for automatic view planning for image acquisition |
| CN112348049A (en) * | 2020-09-28 | 2021-02-09 | 北京师范大学 | Image recognition model training method and device based on automatic coding |
| CN112785565A (en) * | 2021-01-15 | 2021-05-11 | 上海商汤智能科技有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112801068A (en) * | 2021-04-14 | 2021-05-14 | 广东众聚人工智能科技有限公司 | Video multi-target tracking and segmenting system and method |
| CN112885453A (en) * | 2019-11-29 | 2021-06-01 | 西门子医疗有限公司 | Method and system for identifying pathological changes in subsequent medical images |
| CN113538470A (en) * | 2021-06-16 | 2021-10-22 | 唯智医疗科技(佛山)有限公司 | Image interlayer boundary determining method and device based on neural network |
| CN113711234A (en) * | 2019-03-15 | 2021-11-26 | 英威达纺织（英国）有限公司 | Yarn quality control |
| CN114730486A (en) * | 2019-11-22 | 2022-07-08 | 国际商业机器公司 | Generating training data for object detection |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
| CN106815579A (en) * | 2017-01-22 | 2017-06-09 | 深圳市唯特视科技有限公司 | A kind of motion detection method based on multizone double fluid convolutional neural networks model |
-
2018
- 2018-07-18 CN CN201810789942.0A patent/CN109003267B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
| CN106815579A (en) * | 2017-01-22 | 2017-06-09 | 深圳市唯特视科技有限公司 | A kind of motion detection method based on multizone double fluid convolutional neural networks model |
Non-Patent Citations (2)
| Title |
|---|
| SHURAN SONG AND JIANXIONG XIAO: "Deep Sliding Shapes for Amodal 3D Object Detection in RGB-D Images", 《2016IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| TSUNG-YI LIN ET AL.: "Feature Pyramid Network for Object Detection", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109766887B (en) * | 2019-01-16 | 2022-11-11 | 中国科学院光电技术研究所 | Multi-target detection method based on cascaded hourglass neural network |
| CN109766887A (en) * | 2019-01-16 | 2019-05-17 | 中国科学院光电技术研究所 | A kind of multi-target detection method based on cascade hourglass neural network |
| CN111563496A (en) * | 2019-02-13 | 2020-08-21 | 西门子医疗有限公司 | Continuous learning for automatic view planning for image acquisition |
| CN110059548A (en) * | 2019-03-08 | 2019-07-26 | 北京旷视科技有限公司 | Object detection method and device |
| CN113711234A (en) * | 2019-03-15 | 2021-11-26 | 英威达纺织（英国）有限公司 | Yarn quality control |
| CN113711234B (en) * | 2019-03-15 | 2024-01-23 | 英威达纺织（英国）有限公司 | Yarn quality control |
| CN110930454A (en) * | 2019-11-01 | 2020-03-27 | 北京航空航天大学 | Six-degree-of-freedom pose estimation algorithm based on boundary box outer key point positioning |
| CN110930454B (en) * | 2019-11-01 | 2022-11-22 | 北京航空航天大学 | Six-degree-of-freedom pose estimation algorithm based on boundary box outer key point positioning |
| CN114730486B (en) * | 2019-11-22 | 2023-04-04 | 国际商业机器公司 | Method and system for generating training data for object detection |
| CN114730486A (en) * | 2019-11-22 | 2022-07-08 | 国际商业机器公司 | Generating training data for object detection |
| CN112885453A (en) * | 2019-11-29 | 2021-06-01 | 西门子医疗有限公司 | Method and system for identifying pathological changes in subsequent medical images |
| CN110852314B (en) * | 2020-01-16 | 2020-05-22 | 江西高创保安服务技术有限公司 | Article detection network method based on camera projection model |
| CN110852314A (en) * | 2020-01-16 | 2020-02-28 | 江西高创保安服务技术有限公司 | Article detection network method based on camera projection model |
| CN111523452A (en) * | 2020-04-22 | 2020-08-11 | 北京百度网讯科技有限公司 | Method and device for detecting human body position in image |
| CN111523452B (en) * | 2020-04-22 | 2023-08-25 | 北京百度网讯科技有限公司 | Method and device for detecting human body position in image |
| CN112348049A (en) * | 2020-09-28 | 2021-02-09 | 北京师范大学 | Image recognition model training method and device based on automatic coding |
| WO2022151755A1 (en) * | 2021-01-15 | 2022-07-21 | 上海商汤智能科技有限公司 | Target detection method and apparatus, and electronic device, storage medium, computer program product and computer program |
| CN112785565A (en) * | 2021-01-15 | 2021-05-11 | 上海商汤智能科技有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112785565B (en) * | 2021-01-15 | 2024-01-05 | 上海商汤智能科技有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112801068B (en) * | 2021-04-14 | 2021-07-16 | 广东众聚人工智能科技有限公司 | Video multi-target tracking and segmenting system and method |
| CN112801068A (en) * | 2021-04-14 | 2021-05-14 | 广东众聚人工智能科技有限公司 | Video multi-target tracking and segmenting system and method |
| CN113538470A (en) * | 2021-06-16 | 2021-10-22 | 唯智医疗科技(佛山)有限公司 | Image interlayer boundary determining method and device based on neural network |
| CN113538470B (en) * | 2021-06-16 | 2024-02-23 | 唯智医疗科技(佛山)有限公司 | Image interlayer boundary determining method and device based on neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109003267B (en) | 2021-07-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109003267A (en) | From the computer implemented method and system of the automatic detected target object of 3D rendering | |
| US10867384B2 (en) | System and method for automatically detecting a target object from a 3D image | |
| US11386557B2 (en) | Systems and methods for segmentation of intra-patient medical images | |
| CN109906470B (en) | Image segmentation using neural network approach | |
| CN109493347B (en) | Method and system for segmenting sparsely distributed objects in an image | |
| US9965863B2 (en) | System and methods for image segmentation using convolutional neural network | |
| EP3195257B1 (en) | Systems and methods for segmenting medical images based on anatomical landmark-based features | |
| EP2598034B1 (en) | Adaptive visualization for direct physician use | |
| EP3591616A1 (en) | Automated determination of a canonical pose of a 3d dental structure and superimposition of 3d dental structures using deep learning | |
| CN109147940A (en) | From the device and system of the medical image automatic Prediction physiological status of patient | |
| CN109313940A (en) | The virtual assessment in medical device implantation path | |
| CN108986891A (en) | Medical imaging processing method and processing device, electronic equipment and storage medium | |
| CN110458833A (en) | Medical image processing method, medical supply and storage medium based on artificial intelligence | |
| CN109410188A (en) | System and method for being split to medical image | |
| Yerukalareddy et al. | Brain tumor classification based on mr images using GAN as a pre-trained model | |
| Svoboda et al. | Generation of 3D digital phantoms of colon tissue | |
| US11704796B2 (en) | Estimating bone mineral density from plain radiograph by assessing bone texture with deep learning | |
| Heeneman et al. | Lung nodule detection by using Deep Learning | |
| Foysal et al. | Detection of COVID-19 case from chest CT images using deformable deep convolutional neural network | |
| AU2019204365B1 (en) | Method and System for Image Segmentation and Identification | |
| CN112884706B (en) | Image evaluation system based on neural network model and related product | |
| JP2023545570A (en) | Detecting anatomical abnormalities by segmentation results with and without shape priors | |
| Zhou et al. | Learning stochastic object models from medical imaging measurements by use of advanced ambientgans | |
| Alshamrani et al. | Automation of Cephalometrics Using Machine Learning Methods | |
| US20220414882A1 (en) | Method for automatic segmentation of coronary sinus |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |