CN108257128B - Establishment method of pulmonary nodule detection device based on 3D convolutional neural network - Google Patents
Establishment method of pulmonary nodule detection device based on 3D convolutional neural network Download PDFInfo
- Publication number
- CN108257128B CN108257128B CN201810092248.3A CN201810092248A CN108257128B CN 108257128 B CN108257128 B CN 108257128B CN 201810092248 A CN201810092248 A CN 201810092248A CN 108257128 B CN108257128 B CN 108257128B
- Authority
- CN
- China
- Prior art keywords
- convolution
- residual
- network
- lung
- units
- 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.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 65
- 230000002685 pulmonary effect Effects 0.000 title claims abstract description 38
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 12
- 238000012549 training Methods 0.000 claims abstract description 39
- 230000011218 segmentation Effects 0.000 claims abstract description 7
- 206010056342 Pulmonary mass Diseases 0.000 claims description 59
- 230000006870 function Effects 0.000 claims description 34
- 210000004072 lung Anatomy 0.000 claims description 21
- 238000005070 sampling Methods 0.000 claims description 20
- 238000010586 diagram Methods 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 10
- 238000000605 extraction Methods 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 6
- 238000005065 mining Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
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
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30061—Lung
- G06T2207/30064—Lung nodule
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Apparatus For Radiation Diagnosis (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method for establishing a pulmonary nodule detection device based on a 3D convolutional neural network, which comprises the following steps: establishing a training sample; establishing a pulmonary nodule detection network: the pulmonary nodule segmentation network comprises convolution units, 64 × 64(32) residual convolution units A, 32 × 32(64) residual convolution units B, 16 × 16(64) residual convolution units C, 8 × 8(64) residual convolution units D and 16 × 16(64) residual convolution units E which are connected in sequence, an output feature graph of the residual convolution unit E and an output feature graph of the residual convolution unit C are spliced according to channels and then input to a residual convolution unit F, and an output feature graph of the residual convolution unit F and an output feature graph of the residual convolution unit B are spliced according to channels and then input to an RPN network to realize pulmonary nodule detection of the input graph; and training a pulmonary nodule detection network to obtain a pulmonary nodule detection device.
Description
Technical Field
The invention belongs to the field of image processing, and particularly relates to a method for establishing a pulmonary nodule detection device based on a 3D convolutional neural network.
Background
The existing methods for detecting lung nodules in a lung CT image by adopting a deep learning algorithm are many, but the detection accuracy is not high. The main reasons for the low precision are:
(1) the recall rate of the detection stage is lower than that of some special types of lung nodules, so that the condition of missed detection is caused, and the detection precision is low.
(2) Lung nodules are disproportionate in size and smaller lung nodules are easily overlooked.
Based on the two reasons, the lung nodules detected and segmented by the deep learning algorithm are insufficient in typicality and representativeness.
Therefore, improving the accuracy of pulmonary nodule detection and training the network to segment more representative nodules become problems to be solved urgently.
Disclosure of Invention
The invention aims to provide a method for establishing a pulmonary nodule detection device based on a 3D convolutional neural network. The device established by the method can more accurately and quickly detect and determine the three-dimensional image of the lung nodule in the lung CT.
In order to achieve the purpose, the invention provides the following technical scheme:
a method for establishing a pulmonary nodule detection apparatus based on a 3D convolutional neural network, the method comprising:
establishing a training sample: firstly, cutting an acquired three-dimensional lung CT image into a large number of cube small blocks, then processing the cube small blocks by adopting a data enhancement method, and finally processing the enhanced cube small blocks by adopting a difficultly-distinguished negative sample mining method to obtain 2n negative samples and n positive samples which are most difficultly distinguished to form a training sample set;
establishing a pulmonary nodule detection network: the lung nodule segmentation network comprises sequentially connected 128 × 128(24) convolution units, 64 × 64(32) residual convolution units A, 32 × 32(64) residual convolution units B, 16 × 16(64) residual convolution units C, 8 × 8(64) residual convolution units D, 16 × 16(64) residual convolution units E, 16 × 16(64) residual convolution units C and 16 × 16(64) residual convolution units C, output feature maps of the 32 × 32(64) residual convolution units F and 32 × 32(64) residual convolution units B are input into the RPN network according to channels after being spliced, and the detection of the lung nodule is realized by inputting the output feature maps of the 128 × 128(24) convolution units B, the RPN network after the output feature maps of the 16 × 16(64) residual convolution units C and the RPN network;
training a pulmonary nodule detection network: sampling small blocks with the sampling frequency of 2 times and the lung nodules larger than 30mm in a training sample set, sampling small blocks with the sampling frequency of 6 times and the lung nodules larger than 40mm in the training sample set, sampling the lung nodules with other sizes at the normal sampling frequency, inputting the sampled training samples into a lung nodule detection network, and training the lung nodule detection network to target the error convergence of the prediction output and the real output of the lung nodule detection network to obtain the lung nodule detection device.
The pulmonary nodule detection device established by the invention has accurate pulmonary nodule detection effect and low calculation cost.
Wherein the cutting of the acquired three-dimensional lung CT image into a large number of cube small blocks comprises:
the three-dimensional lung CT image is cut according to the following conditions:
the first condition is as follows: 70% of the cube blocks contain at least one lung nodule target;
and a second condition: randomly selecting 30% cube small blocks from the whole lung;
if the cube contains regions beyond the lungs, filling the non-lung regions with meaningless values 170 in the CT image;
pixels of the lung nodule region are taken as positive samples, and pixels of other regions are taken as negative samples.
Processing the cube small blocks after the enhancement processing by adopting a difficult-to-distinguish negative sample mining method, and obtaining 2n negative samples and n positive samples which are difficult to distinguish to form a training sample set, wherein the training sample set comprises the following steps:
firstly, calculating the enhanced cube small blocks by adopting a lung nodule detection network, and outputting the classification confidence coefficient of each pixel, wherein the closer the classification confidence coefficient is to 1, the higher the probability that the cube small blocks contain lung nodules is, the closer the classification confidence coefficient is to 0, and the higher the probability that the cube small blocks do not contain lung nodules is;
then, calculating an absolute value of the difference between the classification confidence and the truth label, wherein the larger the absolute value is, the more difficult the negative sample is to be distinguished by the network;
finally, 2n negative samples and n positive samples which are most difficult to distinguish are selected to form a training sample set.
Specifically, the 64 × 64(32) residual convolution unit a includes three sequentially connected 32-dimensional input residual units I;
each residual error unit I comprises a convolution layer with 32-dimensional input, 8-dimensional output and 1 × 1 convolution kernel, a convolution layer with 8-dimensional input, 8-dimensional output and 3 × 3 convolution kernel and a convolution layer with 8-dimensional input, 32-dimensional output and 1 × 1 convolution kernel which are sequentially connected, each convolution layer is used for feature extraction, and after the output feature diagram and the input feature diagram of the cascade connection of the three convolution layers are added, the activation is carried out through a RELU function.
Specifically, each of the 32 × 32(64) residual convolution units B, 16 × 16(64) residual convolution units C, 8 × 8(64) residual convolution units D includes three sequentially connected 64-dimensional input residual units II;
each residual error unit II comprises a convolution layer with 64-dimensional input, 16-dimensional output and 1 × 1 convolution kernel, a convolution layer with 16-dimensional input, 16-dimensional output and 3 × 3 convolution kernel and a convolution layer with 16-dimensional input, 64-dimensional output and 1 × 1 convolution kernel which are sequentially connected, each convolution layer is used for feature extraction, and after the output feature diagram and the input feature diagram of the cascade connection of the three convolution layers are added, the activation is carried out through a RELU function.
Specifically, each of the 16 × 16(64) residual convolution units E and 32 × 32(64) residual convolution units F includes two sequentially connected 64-dimensional input residual units II.
Wherein, the loss function L of the RPN network is a cross entropy loss function C and a smoothL1The sum of the functions.
Wherein the cross entropy function C is:
where y is the desired output, i.e., the true label, a is the actual output of the network, a ═ σ (z), z ═ Σ Wj×Xj+b，WjAnd b is a network parameter;
the smooth L1 function is:
in which
wherein L islocA loss function value representing the target area and the current prediction; t is tuRepresenting the target region, v the current prediction, x, y the coordinates of the upper left corner of the region, w, h the width and height of the region, respectively.
Wherein, the convergence condition of the loss function L of the RPN network is as follows:
the loss functions L for 3 epochs in succession all have an average value which is lower than the value of the loss function for the preceding epoch.
Compared with the prior art, the invention has the beneficial effects that:
the invention adopts the hard-to-divide negative sample mining method to mine the hard-to-divide negative sample of the cut CT image so as to improve the accuracy of a sample training network, and a residual convolution unit is introduced into the established pulmonary nodule detection network, thereby improving the capability of extracting features, the training speed and stability, obtaining more feature information, and improving the flexibility of pulmonary nodule detection by adding the RPN network. Through experimental verification, the lung nodule detection device obtained by training can more accurately and quickly detect and determine lung nodules in lung CT.
Drawings
Fig. 1 is a flowchart of a method for establishing a pulmonary nodule detection apparatus provided by an embodiment;
FIG. 2 is a schematic diagram of a lung nodule detection network according to an embodiment;
FIG. 3 is a schematic structural diagram of a residual error unit I provided by the embodiment;
fig. 4 is a schematic structural diagram of the residual error unit II provided in the embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
Fig. 1 is a flowchart of a method for establishing a lung nodule segmentation apparatus according to an embodiment. As shown in fig. 1, the method for establishing a pulmonary nodule segmentation apparatus provided in this embodiment includes the following steps:
and S101, establishing a training sample.
In general, the entire image is input as an object detection model. However, since the 3-dimensional CT image is too large, the conventional GPU cannot meet the requirement of the memory capacity, and the CT image cannot be directly input to the object detection model. In order to ensure good resolution, the present embodiment cannot compress the CT at will, so as to avoid losing many important information and being unfavorable to the detection result. Therefore, the acquired three-dimensional lung CT image is cut into a large number of cube patches. The cube tile is used as the input of the model. Specifically, a 128 × 128 × 128 (pixel) cube patch is cropped to the three-dimensional lung CT image according to the following conditions:
the first condition is as follows: 70% of the cube blocks contain at least one lung nodule target;
and a second condition: randomly selecting 30% cube small blocks from the whole lung;
if the cube contains regions beyond the lungs, filling the non-lung regions with meaningless values 170 in the CT image;
pixels of the lung nodule region are taken as positive samples, and pixels of other regions are taken as negative samples.
It should be noted that the lung nodule is not necessarily located in the center of the cube, and the pixel of the lung nodule region is a positive sample as long as it exists in the cube.
And after cutting, processing the cube small blocks by adopting a data enhancement method so as to increase the robustness of the training sample and relieve the over-fitting problem.
Since the number of negative samples far exceeds that of positive samples and the distribution is highly unbalanced, most of the samples appear to be easily distinguished, but there are some negative samples with suspected nodules having similar appearances. The embodiment adopts a hard-to-separate sample mining method to solve the problem. The specific process is as follows:
firstly, calculating the enhanced cube small blocks by using the lung nodule detection network provided by the invention, and outputting the classification confidence coefficient of each pixel, wherein the closer the classification confidence coefficient is to 1, the higher the probability that the cube small blocks contain lung nodules is, the closer the classification confidence coefficient is to 0, and the higher the probability that the cube small blocks do not contain lung nodules is;
then, calculating an absolute value of the difference between the classification confidence and the truth label, wherein the larger the absolute value is, the more difficult the negative sample is to be distinguished by the network;
finally, 2n negative samples and n positive samples which are most difficult to distinguish are selected to form a training sample set.
And S102, establishing a pulmonary nodule detection network.
As shown in fig. 2, the pulmonary nodule detection network includes a pulmonary nodule segmentation network including 128 × 128(24) convolution units 201, 64 × 64(32) residual convolution units 202, 32 × 32(64) residual convolution units 203 connected in sequence, the output feature maps of 16 × 16(64) residual convolution unit 204, 8 × 8(64) residual convolution unit 205, 16 × 16(64) residual convolution unit 206 and 16 × 16(64) residual convolution unit 206 are spliced according to channels and then input to 32 × 32(64) residual convolution unit 207, and the output feature maps of 32 × 32(64) residual convolution unit 207 and 32 × 32(64) residual convolution unit 203 are spliced according to channels and then input to RPN network 208 to realize pulmonary nodule detection on the input maps.
The 128 x 128(24) convolution unit 201 and all residual convolution units are used to perform feature extraction on the input map.
In fig. 2, R represents a splicing operation, that is, the feature map output from the 32 × 32(64) residual convolution unit 207 and the feature map output from the 32 × 32(64) residual convolution unit 203 are spliced according to channels.
64 × 64(32) residual convolution unit 202 includes three sequentially connected 32-dimensional input residual units I, as shown in fig. 3, each residual unit I includes sequentially connected 32-dimensional input, 8-dimensional output, convolution layer with convolution kernel of 1 × 1, convolution layer with 8-dimensional input, 8-dimensional output, convolution kernel of 3 × 3, convolution layer with 8-dimensional input, 32-dimensional output, convolution kernel of 1 × 1, each convolution layer is used for feature extraction, and after adding between output feature map and input feature map of three convolution layer cascade, it is activated by RELU function.
Each of the 32 × 32(64) residual convolution units 203, 16 × 16(64) residual convolution units 204, and 8 × 8(64) residual convolution units 205 includes three sequentially connected 64-dimensional input residual units II. As shown in fig. 4, each residual unit II includes sequentially connected convolutional layers with 64-dimensional input, 16-dimensional output, and 1 × 1 convolutional kernels, convolutional layers with 16-dimensional input, 16-dimensional output, and 3 × 3 convolutional kernels, and convolutional layers with 16-dimensional input, 64-dimensional output, and 1 × 1 convolutional kernels, each convolutional layer is used for feature extraction, and after adding the output feature maps and the input feature maps of the cascade of the three convolutional layers, the residual units are activated by the RELU function.
Each of the 16 × 16(64) residual convolution units 206 and 32 × 32(64) residual convolution units 207 includes two sequentially connected 64-dimensional input residual units II.
The shapes m (L) represent input and input picture sizes m, and the number of channels L. For example: the 128 x 128(24) convolution unit represents the input and output picture size as 128 x 128, and the number of channels as 24.
The RPN (Region pro-social Networks) network 208 is a regional Proposal network, consisting of 2 consecutive convolutional layers for extracting high confidence regions for the target problem. The RPN network 208 is configured to compute an input feature map and output a set of rectangular target suggestion boxes and a prediction score for each suggestion box. For the present embodiment, the RPN network 208 calculates the input lung nodule feature map, outputs the predicted lung nodule coordinates, diameter and probability, and implements classification and regression on the input image.
In this embodiment, in the pulmonary nodule detection network, the RELU function is an activation function, specifically, f (x) ═ max (0, x).
The loss function L of the RPN network is a cross entropy loss function C and smoothL1The sum of the functions.
Wherein the cross entropy function C is:
where y is the desired output, i.e., the true label, a is the actual output of the network, a ═ σ (z), z ═ Σ Wj×Xj+b，WjAnd b is a network parameter;
the smooth L1 function is:
in which
and S103, training a pulmonary nodule detection network to obtain the pulmonary nodule detection device.
For a set of training samples, although very small nodule objects are removed from the training samples in the set, there still remains a problem with nodule size and its imbalance, with the number of small nodules far exceeding large ones. If uniform sampling is used, learning of the network may be more inclined toward detection of nodules while sacrificing the accuracy of detection of nodules, which is contrary to the objects of the present invention. Therefore, in order to solve this problem, in this embodiment, the sampling frequency of the large nodule is increased for the training sample set. The method comprises the following specific steps:
sampling small blocks with the sampling frequency of 2 times and the lung nodules larger than 30mm in a training sample set, sampling small blocks with the sampling frequency of 6 times and the lung nodules larger than 40mm in the training sample set, sampling the lung nodules with other sizes at the normal sampling frequency, inputting the sampled training samples into a lung nodule detection network, and training the lung nodule detection network to target the error convergence of the prediction output and the real output of the lung nodule detection network to obtain the lung nodule detection device.
In the training process, the gradient of the network parameters is solved by adopting a chain rule. When the derivative value of the error to each parameter is obtained, the current gradient value is obtained. According to the gradient descent algorithm, the product of the gradient vector and the learning rate is subtracted from the parameter vector to form a parameter iteration process. And (3) taking the loss function value that the average value of the loss functions L of the continuous 3 epochs is lower than the previous epoch as a convergence target, and obtaining a final parameter vector through multiple parameter iterations, namely obtaining the lung nodule prediction device.
After obtaining the lung nodule segmentation device, preprocessing a sample to be detected (a lung CT image to be detected) according to the content in S101, inputting the preprocessed sample to the lung nodule detection device, and calculating to obtain the lung nodule probability, the coordinates and the diameter of the sample to be detected.
The experimental comparison shows that: a residual network is introduced into the pulmonary nodule detection network, that is, the recall rate of the pulmonary nodule detection network provided in this embodiment is 95%, the epoch number required for convergence is 60, and when the residual network is not introduced, the recall rate of the network is 85%, and the epoch number required for convergence is 70. Thus, the introduction of the residual network provides recall, training speed and stability of the pulmonary nodule detection network. In addition, when a sample to be tested is predicted, the method is obviously obtained, the lung nodule detection device obtained by introducing the residual error network has higher test speed, and the test result is more accurate.
The above-mentioned embodiments are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions, equivalents, etc. made within the scope of the principles of the present invention should be included in the scope of the present invention.
Claims (7)
1. A method for establishing a pulmonary nodule detection device based on a 3D convolutional neural network is characterized by comprising the following steps:
establishing a training sample: firstly, cutting an acquired three-dimensional lung CT image into a large number of cube small blocks, then processing the cube small blocks by adopting a data enhancement method, and finally processing the enhanced cube small blocks by adopting a difficultly-distinguished negative sample mining method to obtain 2n negative samples and n positive samples which are most difficultly distinguished to form a training sample set;
establishing a pulmonary nodule detection network: the lung nodule segmentation network comprises sequentially connected 128 × 128(24) convolution units, 64 × 64(32) residual convolution units A, 32 × 32(64) residual convolution units B, 16 × 16(64) residual convolution units C, 8 × 8(64) residual convolution units D, 16 × 16(64) residual convolution units E, 16 × 16(64) residual convolution units C and 16 × 16(64) residual convolution units C, output feature maps of the 32 × 32(64) residual convolution units F and 32 × 32(64) residual convolution units B are input into the RPN network according to channels after being spliced, and the detection of the lung nodule is realized by inputting the output feature maps of the 128 × 128(24) convolution units B, the RPN network after the output feature maps of the 16 × 16(64) residual convolution units C and the RPN network;
training a pulmonary nodule detection network: sampling small blocks with lung nodules larger than 30mm in a training sample set at 2 times of sampling frequency, sampling small blocks with lung nodules larger than 40mm in the training sample set at 6 times of sampling frequency, sampling lung nodules with other sizes at normal sampling frequency, inputting the sampled training samples into a lung nodule detection network, training the lung nodule detection network by taking error convergence of prediction output and real output of the lung nodule detection network as a target, and obtaining a lung nodule detection device;
processing the cube small blocks after the enhancement processing by adopting a difficult-to-distinguish negative sample mining method, and obtaining 2n negative samples and n positive samples which are difficult to distinguish to form a training sample set, wherein the training sample set comprises the following steps:
firstly, calculating the enhanced cube small blocks by adopting a lung nodule detection network, and outputting the classification confidence coefficient of each pixel, wherein the closer the classification confidence coefficient is to 1, the higher the probability that the cube small blocks contain lung nodules is, the closer the classification confidence coefficient is to 0, and the higher the probability that the cube small blocks do not contain lung nodules is;
then, calculating an absolute value of the difference between the classification confidence and the truth label, wherein the larger the absolute value is, the more difficult the negative sample is to be distinguished by the network;
finally, 2n negative samples and n positive samples which are most difficult to distinguish are selected to form a training sample set.
2. The method of claim 1, wherein the step of cropping the acquired three-dimensional lung CT image into a plurality of cube small blocks comprises:
the three-dimensional lung CT image is cut according to the following conditions:
the first condition is as follows: 70% of the cube blocks contain at least one lung nodule target;
and a second condition: randomly selecting 30% cube small blocks from the whole lung;
if the cube contains regions beyond the lungs, filling the non-lung regions with meaningless values 170 in the CT image;
pixels of the lung nodule region are taken as positive samples, and pixels of other regions are taken as negative samples.
3. The method for establishing a pulmonary nodule detection apparatus based on a 3D convolutional neural network as claimed in claim 1, wherein the 64 x 64(32) residual convolution unit a comprises three sequentially connected 32-dimensional input residual units I;
each residual error unit I comprises a convolution layer with 32-dimensional input, 8-dimensional output and 1 × 1 convolution kernel, a convolution layer with 8-dimensional input, 8-dimensional output and 3 × 3 convolution kernel and a convolution layer with 8-dimensional input, 32-dimensional output and 1 × 1 convolution kernel which are sequentially connected, each convolution layer is used for feature extraction, and after the output feature diagram and the input feature diagram of the cascade connection of the three convolution layers are added, the activation is carried out through a RELU function.
4. The method of establishing a 3D convolutional neural network-based pulmonary nodule detection apparatus as claimed in claim 1, wherein the 32 x 32(64) residual convolution units B, 16 x 16(64) residual convolution units C, 8 x 8(64) residual convolution units D each include three sequentially connected 64-dimensional input residual units II;
each residual error unit II comprises a convolution layer with 64-dimensional input, 16-dimensional output and 1 × 1 convolution kernel, a convolution layer with 16-dimensional input, 16-dimensional output and 3 × 3 convolution kernel and a convolution layer with 16-dimensional input, 64-dimensional output and 1 × 1 convolution kernel which are sequentially connected, each convolution layer is used for feature extraction, and after the output feature diagram and the input feature diagram of the cascade connection of the three convolution layers are added, the activation is carried out through a RELU function.
5. The method of establishing a 3D convolutional neural network-based pulmonary nodule detection apparatus as claimed in claim 1, wherein each of the 16 × 16(64) residual convolution units E and 32 × 32(64) residual convolution units F includes two sequentially connected 64-dimensional input residual units II;
each residual error unit II comprises a convolution layer with 64-dimensional input, 16-dimensional output and 1 × 1 convolution kernel, a convolution layer with 16-dimensional input, 16-dimensional output and 3 × 3 convolution kernel and a convolution layer with 16-dimensional input, 64-dimensional output and 1 × 1 convolution kernel which are sequentially connected, each convolution layer is used for feature extraction, and after the output feature diagram and the input feature diagram of the cascade connection of the three convolution layers are added, the activation is carried out through a RELU function.
6. The method for establishing a pulmonary nodule detection apparatus based on a 3D convolutional neural network as claimed in claim 1, wherein the loss function L of the RPN network is cross entropy loss function C and smoothL1The sum of the functions;
wherein the cross entropy function C is:
where y is the desired output, i.e., the true label, a is the actual output of the network, a ═ σ (z), z ═ Σ Wj×Xj+b，WjAnd b is a network parameter;
the smooth L1 function is:
in which
wherein L islocA loss function value representing the target area and the current prediction; t is tuRepresenting the target region, v the current prediction, x, y the coordinates of the upper left corner of the region, w, h the width and height of the region, respectively.
7. The method for establishing a pulmonary nodule detection apparatus based on a 3D convolutional neural network as claimed in claim 6, wherein the loss function L convergence condition of the RPN network is:
the loss functions L for 3 epochs in succession all have an average value which is lower than the value of the loss function for the preceding epoch.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810092248.3A CN108257128B (en) | 2018-01-30 | 2018-01-30 | Establishment method of pulmonary nodule detection device based on 3D convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810092248.3A CN108257128B (en) | 2018-01-30 | 2018-01-30 | Establishment method of pulmonary nodule detection device based on 3D convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108257128A CN108257128A (en) | 2018-07-06 |
| CN108257128B true CN108257128B (en) | 2020-09-18 |
Family
ID=62742961
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810092248.3A Active CN108257128B (en) | 2018-01-30 | 2018-01-30 | Establishment method of pulmonary nodule detection device based on 3D convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108257128B (en) |
Families Citing this family (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109003260B (en) * | 2018-06-28 | 2021-02-09 | 深圳视见医疗科技有限公司 | CT image pulmonary nodule detection method, device and equipment and readable storage medium |
| CN109034224B (en) * | 2018-07-16 | 2022-03-11 | 西安电子科技大学 | Hyperspectral classification method based on double branch network |
| CN109102502B (en) * | 2018-08-03 | 2021-07-23 | 西北工业大学 | Pulmonary nodule detection method based on three-dimensional convolutional neural network |
| CN111160346A (en) * | 2018-11-07 | 2020-05-15 | 电子科技大学 | Ischemic stroke segmentation system based on three-dimensional convolution |
| CN109559297B (en) * | 2018-11-15 | 2023-04-11 | 大连大学 | Pulmonary nodule detection method based on three-dimensional region generation network |
| CN109685765B (en) * | 2018-11-21 | 2023-03-14 | 浙江大学山东工业技术研究院 | X-ray film pneumonia result prediction device based on convolutional neural network |
| CN109685768B (en) * | 2018-11-28 | 2020-11-20 | 心医国际数字医疗系统(大连)有限公司 | Pulmonary nodule automatic detection method and system based on pulmonary CT sequence |
| CN110009599A (en) * | 2019-02-01 | 2019-07-12 | 腾讯科技（深圳）有限公司 | Liver masses detection method, device, equipment and storage medium |
| CN110533637B (en) * | 2019-08-02 | 2022-02-11 | 杭州依图医疗技术有限公司 | Method and device for detecting object |
| CN110533120B (en) * | 2019-09-05 | 2023-12-12 | 腾讯科技（深圳）有限公司 | Image classification method, device, terminal and storage medium for organ nodule |
| CN111179247A (en) * | 2019-12-27 | 2020-05-19 | 上海商汤智能科技有限公司 | Three-dimensional target detection method, training method of model thereof, and related device and equipment |
| CN112132815B (en) * | 2020-09-25 | 2023-11-17 | 广州视源电子科技股份有限公司 | Pulmonary nodule detection model training method, detection method and device |
| CN112561060B (en) * | 2020-12-15 | 2022-03-22 | 北京百度网讯科技有限公司 | Neural network training method and device, image recognition method and device and equipment |
| CN112614133B (en) * | 2021-03-05 | 2021-07-06 | 北京小白世纪网络科技有限公司 | Three-dimensional pulmonary nodule detection model training method and device without anchor point frame |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
| CN107016665A (en) * | 2017-02-16 | 2017-08-04 | 浙江大学 | A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks |
| CN107590797A (en) * | 2017-07-26 | 2018-01-16 | 浙江工业大学 | CT image pulmonary nodule detection method based on three-dimensional residual error neural network |
-
2018
- 2018-01-30 CN CN201810092248.3A patent/CN108257128B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
| CN107016665A (en) * | 2017-02-16 | 2017-08-04 | 浙江大学 | A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks |
| CN107590797A (en) * | 2017-07-26 | 2018-01-16 | 浙江工业大学 | CT image pulmonary nodule detection method based on three-dimensional residual error neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108257128A (en) | 2018-07-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108257128B (en) | Establishment method of pulmonary nodule detection device based on 3D convolutional neural network | |
| CN108346154B (en) | Method for establishing lung nodule segmentation device based on Mask-RCNN neural network | |
| CN108230323B (en) | Pulmonary nodule false positive screening method based on convolutional neural network | |
| CN109118479B (en) | Capsule network-based insulator defect identification and positioning device and method | |
| CN108960135B (en) | Dense ship target accurate detection method based on high-resolution remote sensing image | |
| CN109840556B (en) | Image classification and identification method based on twin network | |
| CN114581782B (en) | Fine defect detection method based on coarse-to-fine detection strategy | |
| CN111046950B (en) | Image processing method and device, storage medium and electronic device | |
| CN108171119B (en) | SAR image change detection method based on residual error network | |
| CN112800934B (en) | Behavior recognition method and device for multi-class engineering vehicle | |
| CN111369623A (en) | Lung CT image identification method based on deep learning 3D target detection | |
| CN117152746B (en) | Method for acquiring cervical cell classification parameters based on YOLOV5 network | |
| CN108388901B (en) | Collaborative significant target detection method based on space-semantic channel | |
| CN115272826A (en) | Image identification method, device and system based on convolutional neural network | |
| CN114092814A (en) | Unmanned plane navel orange tree image target identification and statistics method based on deep learning | |
| CN111179278B (en) | Image detection method, device, equipment and storage medium | |
| CN116228685B (en) | Deep learning-based lung nodule detection and rejection method | |
| CN114119577B (en) | High-speed rail tunnel leakage cable buckle detection method | |
| CN113592829B (en) | Deep learning sediment particle identification method and device based on segmentation and recombination | |
| CN115965796A (en) | Metal corrosion detection method and system based on image recognition | |
| CN109753896A (en) | A kind of unsupervised heterologous method for detecting change of remote sensing image based on general character autocoder | |
| CN112199984B (en) | Target rapid detection method for large-scale remote sensing image | |
| CN104616302A (en) | Real-time object identification method | |
| CN115406359A (en) | Image-based concrete crack measuring method | |
| CN106340038B (en) | A kind of depth order inference method based on object in monocular image |
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 |