CN114202002A - Pulmonary nodule detection device based on improved FasterRCNN algorithm - Google Patents

Pulmonary nodule detection device based on improved FasterRCNN algorithm Download PDF

Info

Publication number
CN114202002A
CN114202002A CN202110811268.3A CN202110811268A CN114202002A CN 114202002 A CN114202002 A CN 114202002A CN 202110811268 A CN202110811268 A CN 202110811268A CN 114202002 A CN114202002 A CN 114202002A
Authority
CN
China
Prior art keywords
lung nodule
image
nodule
person
detected
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.)
Pending
Application number
CN202110811268.3A
Other languages
Chinese (zh)
Inventor
廖军
钱爽
刘礼
雍滋蕊
邓正巧
李小虎
张文彬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University
Original Assignee
Chongqing University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing University filed Critical Chongqing University
Priority to CN202110811268.3A priority Critical patent/CN114202002A/en
Publication of CN114202002A publication Critical patent/CN114202002A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30061—Lung
    • G06T2207/30064—Lung nodule

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a pulmonary nodule detection device based on an improved FasterRCNN algorithm, which comprises a data acquisition and labeling module, a data preprocessing module, a feature extraction module and a pulmonary nodule detection and classification module; the data acquisition and labeling module is mainly used for collecting and labeling related data so as to provide a real label for subsequent model learning; the data preprocessing module is mainly used for performing data enhancement and format conversion on the data set; the feature extraction module mainly extracts abstract features from the picture to obtain high-level semantic information; and the lung nodule detection and classification module is used for positioning and classifying the lung nodules in the picture according to the characteristics and outputting the position coordinates and the class probability of the lung nodules in the picture. The invention improves FasterRCNN, ensures accuracy and improves detection and classification speed.

Description

Pulmonary nodule detection device based on improved FasterRCNN algorithm
Technical Field
The invention relates to the technical field of computer medical images, in particular to a pulmonary nodule detection device based on an improved FasterRCNN algorithm.
Background
Lung cancer is one of the most common diseases and is the leading mortality. Causes of lung cancer include smoking, toxic particles in the air, aging, genes, sex, and the like. Unfortunately, it is felt that there is no effective way to control the incidence of cancer and, in the present case, it is almost impossible to find a cure when the patient is at an advanced stage of cancer. Confirmed cases of lung cancer are often accompanied by the continued appearance of nodules. Early detection of cancer can greatly improve the survival rate of patients. It can be seen that in the initial phase of a computer-aided diagnosis (CADx) protocol, the detection of lung nodules is significant and essential. Also, finding a large number of nodules due to fatigue is error prone for radiologists. To alleviate this problem, CAD schemes are urgently needed to share the detection work. The main functions of the CAD system are lesion detection, lesion segmentation and pathology analysis using image processing techniques and machine learning algorithms. However, the shape, size and density of lung nodules vary and are difficult to generalize into a particular class. Since the variability of nodules increases the difficulty of lung nodule detection, it is of great interest to researchers to propose some potential CADe schemes to achieve better detection performance.
Disclosure of Invention
The invention aims to provide a pulmonary nodule detection device based on an improved FasterRCNN algorithm, which comprises a data acquisition and labeling module, a data preprocessing module, a feature extraction module and a pulmonary nodule detection and classification module.
And the data acquisition and labeling module acquires a lung nodule CT image of the person to be detected and transmits the lung nodule CT image to the data preprocessing module.
The data preprocessing module is used for preprocessing the lung nodule CT image of the person to be detected to obtain a preprocessed lung nodule image of the person to be detected and transmitting the preprocessed lung nodule image to the feature extraction module.
The data preprocessing module preprocesses a lung nodule CT image of a person to be detected, and comprises the following steps: and (3) counting the HU value range of the lung region according to the linear attenuation coefficients of air, water and X-rays, converting the lung nodule CT image of the person to be detected into a lung nodule image of the person to be detected in a JPG format, and obtaining the preprocessed lung nodule image of the person to be detected.
The feature extraction module stores a feature extraction model.
The feature extraction model comprises a UNet network and a FasterRCNN neural network model.
The UNet network model is used for feature extraction. The UNet network model inputs the lung nodule image of the person to be detected and outputs the lung nodule image characteristic of the person to be detected.
The UNet network includes a downsampling layer, an upsampling layer, and a transverse connection layer. The downsampling layer is a convolutional layer. The transverse connection layer is a residual block.
The FasterRCNN neural network model is used for target position regression and category prediction of the lung nodule image; the input of the FasterRCNN neural Network is the lung nodule image characteristics of the person to be detected, the FasterRCNN neural Network determines the candidate Region of the lung nodule of the person to be detected by utilizing a Region pro-nodal Network, and the candidate Region is output as the accurate position, the category and the confidence coefficient of the lung nodule of the person to be detected;
the FasterRCNN neural network model generates a characteristic diagram with multiple receptive fields by using a scaled Encoder, and realizes multiplexing and fusion of characteristics.
Loss function L (p) of FasterRCNN neural network modeli，ti) As follows:
Figure BDA0003168321370000021
in the formula, piThe predicted classification probability for the ith target box. When the ith target box is a positive sample,
Figure BDA0003168321370000022
when the ith target box is a negative sample,
Figure BDA0003168321370000023
tiparameterized coordinates (x, y, w, h) of the predicted bounding box characterizing the ith target box.
Figure BDA0003168321370000024
Parameterized coordinates (x) of the real bounding box characterizing the ith target box*，y*，w*，h*). λ is a weight balance parameter. Class loss function
Figure BDA0003168321370000025
Function of regression loss
Figure BDA0003168321370000026
R represents the Smooth L1 Loss function.
And the feature extraction module inputs the preprocessed lung nodule image of the person to be detected into the feature extraction model, so that feature extraction and target region prediction are carried out on the preprocessed lung nodule image of the person to be detected, and a plurality of candidate lung nodule regions of the person to be detected are obtained.
The feature extraction module transmits the lung nodule candidate region of the person to be detected to the lung nodule detection and classification module.
The pulmonary nodule detection and classification module stores a pulmonary nodule detection classification model.
And the lung nodule detection and classification module inputs the candidate region of the lung nodule of the person to be detected into a lung nodule detection classification model to obtain a prediction region and a prediction confidence probability of the lung nodule of the person to be detected.
The method for inputting the lung nodule candidate region of the person to be detected into the lung nodule detection classification model to obtain the predicted region and the predicted confidence probability of the lung nodule of the person to be detected comprises the following steps:
1) the pulmonary nodule detection classification model carries out RoI posing on a pulmonary nodule candidate region of a person to be detected, and converts regions with different sizes into regions with fixed length. And inputting the area with the fixed length into two connected full-connection layers to obtain the position and category output.
2) And performing class prediction by adopting a cross entropy loss function.
3) And (3) performing position regression by adopting a Smooth L1 Loss function to obtain a predicted region and a predicted confidence probability of the pulmonary nodule of the person to be detected.
Smooth L1 Loss function SmoothL1As follows:
Figure BDA0003168321370000031
in the formula, test
Figure BDA0003168321370000032
σ is a constant.
The step of establishing a pulmonary nodule detection classification model comprises the following steps:
1) the data acquisition and labeling module acquires a plurality of lung nodule CT images from different testers, and marks out a lung nodule region target frame in the lung nodule CT images.
And the data acquisition and labeling module transmits the labeled lung nodule CT image to the data preprocessing module.
2) The data preprocessing module preprocesses the lung nodule CT image to obtain a preprocessed lung nodule image, and divides the preprocessed lung nodule image into a training set and a testing set.
The data preprocessing module preprocesses the lung nodule CT image and comprises the following steps:
2.1) according to the linear attenuation coefficients of air, water and X-rays, counting the HU value range of the lung region, and converting the lung nodule CT image into a JPG picture.
2.2) expanding the JPG picture by using a turnover method, and converting the marked lung nodule area target frame into a VOC format.
3) The feature extraction module utilizes a K-means method to cluster the lung nodule region target frames of the lung nodule CT image to obtain a plurality of groups of lung nodule region target frames.
The objective function min E for clustering the pulmonary nodule region object box of the pulmonary nodule CT image is as follows:
Figure BDA0003168321370000033
in the formula, mujIs a cluster CjThe mean vector of (2); e is the square error; k is the number of clusters; j is the cluster number.
4) And establishing a feature extraction model.
5) And training the feature extraction model by using a training set to obtain the trained feature extraction model.
6) And (3) testing the trained feature extraction model by using the test set, finishing the establishment of the feature extraction model if the accuracy of the trained feature extraction model is greater than a preset threshold, and returning to the step 1) if the accuracy of the trained feature extraction model is not greater than the preset threshold.
It is worth explaining that the invention provides a pulmonary nodule detection device based on an improved fast RCNN algorithm, the device firstly takes a deep convolution neural network fast RCNN based on a regression target identification method as a basis, modifies the structure of the fast RCNN, adjusts parameters in a hierarchical structure, improves the structure into a neural network structure with dense convolution blocks, and mainly optimizes the problems of small target, complicated position, various characteristics, easy misdiagnosis and the like of pulmonary nodules, and finally the model greatly improves the precision ratio, the recall ratio and the efficiency of pulmonary nodule detection, and provides conditions for the real-time detection of pulmonary CT images and pulmonary nodules.
The technical effects of the present invention are undoubted, and the present invention has the following effects:
1) nodule detection and classification are directly carried out on the whole CT image, and preprocessing such as scaling and the like is not needed for the image, so that the loss of effective information in the image is avoided, and the detailed information in the image is fully utilized;
2) and clustering the sizes of the target frames of the data set by using K-means to obtain the length and width of 9 groups of target frames. Smaller errors are obtained when the boundary box is predicted, and the accuracy of target positioning and the recall rate of detection can be improved;
3) the common convolution is changed into a residual convolution block, gradient disappearance can be slowed down, the FPN is changed into a related Encoder, and time and memory consumption is reduced under the condition that multi-scale features are reserved.
Drawings
FIG. 1 is a schematic view of the apparatus of the present invention.
Fig. 2 is a structural diagram based on the improved fasterncnn algorithm.
FIG. 3 is a schematic diagram of a related Encode.
Fig. 4 is a flow chart illustrating the use of a pulmonary nodule detection apparatus based on the modified fasterncn algorithm.
FIG. 5(a) is a schematic diagram I of the classification result of lung nodule detection provided by the present invention; FIG. 5(b) is a schematic diagram II of the classification result of lung nodule detection provided by the present invention; fig. 5(c) is a schematic diagram III of the classification result of lung nodule detection provided by the present invention.
Detailed Description
The present invention is further illustrated by the following examples, but it should not be construed that the scope of the above-described subject matter is limited to the following examples. Various substitutions and alterations can be made without departing from the technical idea of the invention and the scope of the invention is covered by the present invention according to the common technical knowledge and the conventional means in the field.
Example 1:
referring to fig. 1 to 5, a pulmonary nodule detection apparatus based on an improved fasterncn algorithm includes a data acquisition and labeling module, a data preprocessing module, a feature extraction module, and a pulmonary nodule detection and classification module.
And the data acquisition and labeling module acquires a lung nodule CT image of the person to be detected and transmits the lung nodule CT image to the data preprocessing module.
The data preprocessing module is used for preprocessing the lung nodule CT image of the person to be detected to obtain a preprocessed lung nodule image of the person to be detected and transmitting the preprocessed lung nodule image to the feature extraction module.
The data preprocessing module preprocesses a lung nodule CT image of a person to be detected, and comprises the following steps: and (3) counting the HU value range of the lung region according to the linear attenuation coefficients of air, water and X-rays, converting the lung nodule CT image of the person to be detected into a lung nodule image of the person to be detected in a JPG format, and obtaining the preprocessed lung nodule image of the person to be detected.
The feature extraction module stores a feature extraction model.
The feature extraction model comprises a UNet network and a FasterRCNN neural network model.
The UNet network model is used for feature extraction. The UNet network model inputs the lung nodule image of the person to be detected and outputs the lung nodule image characteristic of the person to be detected.
The UNet network includes a downsampling layer, an upsampling layer, and a transverse connection layer. The downsampling layer is a convolutional layer. The transverse connection layer is a residual block.
The FasterRCNN neural network model is used for target position regression and category prediction of the lung nodule image; the FasterRCNN neural Network inputs the image characteristics of the pulmonary nodules of the person to be detected, determines a candidate Region of the pulmonary nodules of the person to be detected by utilizing a Region pro-nodal Network, and outputs the accurate position, category and confidence coefficient of the pulmonary nodules of the person to be detected;
the FasterRCNN neural network model utilizes a related Encoder to generate a feature map with multiple receptive fields, and multiplexing and fusion of features are realized.
Loss function L (p) of FasterRCNN neural network modeli，ti) As follows:
Figure BDA0003168321370000051
in the formula, piThe predicted classification probability for the ith target box. When the ith target box is a positive sample,
Figure BDA0003168321370000052
when the ith target box is a negative sample,
Figure BDA0003168321370000053
tiparameterized coordinates (x, y, w, h) of the predicted bounding box characterizing the ith target box.
Figure BDA0003168321370000054
Parameterized coordinates (x) of the real bounding box characterizing the ith target box*，y*，w*，h*). λ is a weight balance parameter. Class loss function
Figure BDA0003168321370000055
Function of regression loss
Figure BDA0003168321370000056
R represents the Smooth L1 Loss function.
And the feature extraction module inputs the preprocessed lung nodule image of the person to be detected into the feature extraction model, so that feature extraction and target region prediction are carried out on the preprocessed lung nodule image of the person to be detected, and a plurality of candidate lung nodule regions of the person to be detected are obtained.
The feature extraction module transmits the lung nodule candidate region of the person to be detected to the lung nodule detection and classification module.
The pulmonary nodule detection and classification module stores a pulmonary nodule detection classification model.
And the lung nodule detection and classification module inputs the candidate region of the lung nodule of the person to be detected into a lung nodule detection classification model to obtain a prediction region and a prediction confidence probability of the lung nodule of the person to be detected.
The method for inputting the lung nodule candidate region of the person to be detected into the lung nodule detection classification model to obtain the predicted region and the predicted confidence probability of the lung nodule of the person to be detected comprises the following steps:
1) the pulmonary nodule detection classification model carries out RoI posing on a pulmonary nodule candidate region of a person to be detected, and converts regions with different sizes into regions with fixed length. And inputting the area with the fixed length into two connected full-connection layers to obtain the position and category output.
2) And performing class prediction by adopting a cross entropy loss function.
3) And (3) performing position regression by adopting a Smooth L1 Loss function to obtain a predicted region and a predicted confidence probability of the pulmonary nodule of the person to be detected.
Smooth L1 Loss function SmoothL1As follows:
Figure BDA0003168321370000061
in the formula, test
Figure BDA0003168321370000062
σ is a constant.
The step of establishing a pulmonary nodule detection classification model comprises the following steps:
1) the data acquisition and labeling module acquires a plurality of lung nodule CT images from different testers, and marks out a lung nodule region target frame in the lung nodule CT images.
And the data acquisition and labeling module transmits the labeled lung nodule CT image to the data preprocessing module.
2) The data preprocessing module preprocesses the lung nodule CT image to obtain a preprocessed lung nodule image, and divides the preprocessed lung nodule image into a training set and a testing set.
The data preprocessing module preprocesses the lung nodule CT image and comprises the following steps:
2.1) according to the linear attenuation coefficients of air, water and X-rays, counting the HU value range of the lung region, and converting the lung nodule CT image into a JPG picture.
2.2) expanding the JPG picture by using a turnover method, and converting the marked lung nodule area target frame into a VOC format.
3) The feature extraction module utilizes a K-means method to cluster the lung nodule region target frames of the lung nodule CT image to obtain a plurality of groups of lung nodule region target frames.
The objective function min E for clustering the pulmonary nodule region object box of the pulmonary nodule CT image is as follows:
Figure BDA0003168321370000071
in the formula, mujIs a cluster CjThe mean vector of (2); e is the square error; k is the number of clusters; j is the cluster number.
4) And establishing a feature extraction model.
5) And training the feature extraction model by using a training set to obtain the trained feature extraction model.
6) And (3) testing the trained feature extraction model by using the test set, finishing the establishment of the feature extraction model if the accuracy of the trained feature extraction model is greater than a preset threshold, and returning to the step 1) if the accuracy of the trained feature extraction model is not greater than the preset threshold.
Example 2:
the pulmonary nodule detection device based on the improved FasterRCNN algorithm is used as follows:
1) collecting and marking lung nodule data, and preprocessing an original CT picture, wherein the method comprises the following specific steps:
1.1) acquiring a pulmonary nodule data set and labeling data;
1.2) calculating the HU value range of the lung region according to the linear attenuation coefficients of air, water and X-rays, and converting the DCM file into a corresponding JPG picture;
1.3) expanding the data set by turning the JPG picture by 90 degrees, 180 degrees and 270 degrees, converting all data into the data set in the VOC format, and simultaneously dividing the data set into a training set, a verification set and a test set.
2) Clustering the real boxes in the training set by using K-means to obtain 9 groups of Anchor boxes with different lengths and widths, and specifically comprising the following steps:
2.1) after data preprocessing operation, clustering the size of a target frame of a data set by adopting a K-means algorithm, wherein the objective function optimized by the K-means algorithm is a minimized square error E:
Figure BDA0003168321370000081
wherein muiIs a cluster CiA mean vector of (a), sometimes referred to as a centroid;
2.2) sorting the clustering results to sequentially obtain 9 groups of Anchor Box with different sizes.
3) The method comprises the following steps of utilizing a UNet network to carry out feature extraction on a CT picture, and selecting a candidate region through a region suggestion network, wherein the specific steps are as follows:
3.1) replacing the convolution layer in the UNet network with a residual block, which is beneficial to preventing gradient from disappearing; the pooling layer in the UNet network is replaced by a convolution layer with the step length of 2, so that the detail information can be kept;
3.2) FasterRCNN area suggest that the network uses FPN for area selection at different scales. Consider that lung nodule size differences are not significant and FPN networks can greatly increase temporal memory consumption. Detecting by adopting a layer of characteristics, adding a related Encoder, generating a characteristic diagram with multiple receptive fields, and realizing characteristic multiplexing and fusion;
3.3) carrying out regional Proposal on the layer feature map, and predicting the Anchor Box offset value and the confidence coefficient through the Region Proposal Network.
4) Performing position regression and nodule classification on the candidate region recommended in the step 3 to obtain a prediction box and a prediction confidence probability of the nodule, and specifically comprising the following steps of:
4.1) performing RoI poling on candidate regions output by the Region Proposal Networks, converting the regions with different sizes into output with fixed length, and obtaining position and category output through two full connection layers;
4.2) adopting a cross entropy loss function to carry out category prediction;
4.3) position regression with Smooth L1 Loss function.
The loss function form of fast RCNN is as follows:
Figure BDA0003168321370000082
wherein p isiFor the predicted classification probability of the ith Anchor Box, if the ith Anchor Box is a positive sample,
Figure BDA0003168321370000083
if the ith Anchor Box is a negative sample,
Figure BDA0003168321370000084
tifor the parameterized coordinates (x, y, w, h) of the predicted bounding Box of the ith Anchor Box,
Figure BDA0003168321370000085
parameterized coordinates (x) of the real bounding Box for the ith Anchor Box*，y*，w*，h*) (ii) a λ is a weight balance parameter; class loss function
Figure BDA0003168321370000086
Figure BDA0003168321370000087
Function of regression loss
Figure BDA0003168321370000088
Where R is the Smooth L1 function, defined as follows:
Figure BDA0003168321370000091
wherein
Figure BDA0003168321370000092
σ＝3。
Example 3:
referring to fig. 3, a pulmonary nodule detection apparatus based on the modified fasterncn algorithm is used as follows:
firstly, acquiring a lung nodule CT image, performing data enhancement on the lung nodule image to expand a data set, and simultaneously dividing the data set according to a patient; then, clustering the size of a target frame of the data set by using a K-means method, and sequentially obtaining the first 9 groups of values as the length and the width of the Anchor Box according to descending order; inputting the training set into a network to extract basic features of the lung nodule image, classifying and regressing Anchor Box generated by the network by using a region suggestion network RPN, recommending and generating candidate regions according to the confidence level and the cross-over ratio ordering, and performing ROI pooling on the candidate regions to obtain corresponding feature mapping maps; and finally, flattening the feature mapping image output by the RPN network, and outputting the position and category information of the lung nodule bounding box in the image through two full-connection layers.
Example 4:
the pulmonary nodule detection device based on the improved FasterRCNN algorithm is used as follows:
1) lung nodule CT image preprocessing
The data set used by the invention contains 5559 lung nodule pictures of 300 patients, the data set is expanded by turning the lung nodule pictures by 90 degrees and 180 degrees, all data labels are converted into VOC formats, 16679 lung nodule pictures and label information corresponding to the VOC formats are obtained, and the data set is divided into a training set and a test set according to the patients.
2) Extracting image feature information
The network used in the network structure of the present invention uses UNet as a base network and is improved. The original UNet network extracts features in a convolution mode, maximum pooling is adopted for down-sampling, and a residual error structure of ResNet is introduced into the new network, so that when the network becomes deeper, the gradient is slowed down to disappear, and the model is converged more quickly; meanwhile, the maximum pooling is changed into a convolution kernel with the step length of 2 to carry out downsampling so as to keep picture detail information; the FPN network can multiplex and fuse features of feature maps of different scales, so that the detection capability of the network on targets of different sizes is enhanced, but huge time memory overhead is brought, a novel network replaces the FPN with a scaled Encoder, the time memory overhead can be effectively reduced while multi-scale feature information is kept, and therefore the time memory overhead is effectively reduced, and the method is suitable for the FPN network
Replacing the convolutional layer in the original UNet network with ResBlock {4,6, 8, 4} containing a residual error structure, wherein each Battleeck in the ResBlock uses a convolutional layer with a combination of 1 multiplied by 1+3 multiplied by 3+1 multiplied by 1, and the original maximum pooling downsampling is replaced by a convolution kernel with the step size of 2;
the feature pyramid network FPN is replaced by a Dilated Encoder, and more abundant features can be obtained by convolving the Residual Block of the hole first by 1x1 and 3x3 containing BN layers, and then using the Residual Block with different scaled rates to generate a feature map with multiple sense fields. The scaled Rate used here is {2, 4,6, 8}, and the scaled Encoder is schematically shown in FIG. 2.
3) Classification and regression of Anchor Box Using RPN
Different from the selection of the fixed Anchor Box, the invention uses a K-means method to cluster the size of the target frame of the data set, and sequentially obtains the first 9 groups of values as the length and the width of the Anchor Box according to descending order. And then, carrying out convolution operation on the RPN on the basis of the extracted feature map, wherein the RPN is responsible for generating an Anchor frame and classifying and regressing the Anchor Box, and the RPN is mainly used for generating a target candidate region. And (3) carrying out category judgment on the Anchor Box as a foreground or a background through BCE Loss, correcting the length and the width of the Anchor Box by utilizing Smooth L1 Loss regression, and finally recommending and generating a candidate region according to the confidence level and the cross-over ratio.
4) Performing ROI Pooling on the candidate region to obtain a corresponding region feature map, performing type judgment and position refinement RPN to generate a target candidate region, generating a feature map with a fixed size (the size of the feature map is 7 x 7) through ROI Pooling, sending the feature map into a subsequent full-connection layer to obtain ROI feature vectors, and performing judgment on the type of the target region and refinement on the position of the target region.

Claims (9)

1. A pulmonary nodule detection device based on an improved FasterRCNN algorithm is characterized by comprising a data acquisition and labeling module, a data preprocessing module, a feature extraction module and a pulmonary nodule detection and classification module.
The data acquisition and labeling module acquires a lung nodule CT image of a person to be detected and transmits the lung nodule CT image to the data preprocessing module;
the data preprocessing module is used for preprocessing a lung nodule CT image of a person to be detected to obtain a preprocessed lung nodule image of the person to be detected and transmitting the preprocessed lung nodule image to the feature extraction module;
the feature extraction module stores a feature extraction model;
the feature extraction module inputs the preprocessed lung nodule images of the testees into a feature extraction model, so that feature extraction and target region prediction are carried out on the preprocessed lung nodule images of the testees to obtain a plurality of candidate lung nodule regions of the testees;
the feature extraction module transmits the lung nodule candidate region of the person to be detected to the lung nodule detection and classification module;
the pulmonary nodule detection and classification module stores a pulmonary nodule detection classification model;
and the lung nodule detection and classification module inputs the candidate region of the lung nodule of the person to be detected into a lung nodule detection classification model to obtain a prediction region and a prediction confidence probability of the lung nodule of the person to be detected.
2. The pulmonary nodule detection apparatus based on the improved fasterncnn algorithm as claimed in claim 1, wherein: the data preprocessing module preprocesses a lung nodule CT image of a person to be detected, and comprises the following steps: and (3) counting the HU value range of the lung region according to the linear attenuation coefficients of air, water and X-rays, converting the lung nodule CT image of the person to be detected into a lung nodule image of the person to be detected in a JPG format, and obtaining the preprocessed lung nodule image of the person to be detected.
3. The pulmonary nodule detection apparatus based on the improved fasterncnn algorithm as claimed in claim 1, wherein: the feature extraction model comprises a UNet network model and a FasterRCNN neural network model;
the UNet network model is used for feature extraction; the input of the UNet network model is a lung nodule image of the person to be detected, and the output is the lung nodule image characteristic of the person to be detected;
the UNet network comprises a down-sampling layer, an up-sampling layer and a transverse connection layer; the down-sampling layer is a convolution layer; the transverse connection layer is a residual block.
The FasterRCNN neural network model is used for target position regression and category prediction of the lung nodule image; the input of the FasterRCNN neural Network is the lung nodule image characteristics of the person to be detected, the FasterRCNN neural Network determines the candidate Region of the lung nodule of the person to be detected by utilizing a Region pro-nodal Network, and the candidate Region is output as the accurate position, the category and the confidence coefficient of the lung nodule of the person to be detected;
the FasterRCNN neural network model generates a characteristic diagram with multiple receptive fields by using a scaled Encoder, and realizes multiplexing and fusion of characteristics.
4. The device according to claim 3, wherein the lung nodule detection device based on the improved FasterRCNN algorithm comprises: loss function L (p) of FasterRCNN neural network modeli，ti) As follows:
Figure FDA0003168321360000021
in the formula, piThe predicted classification probability of the ith target frame; when the ith target box is a positive sample,
Figure FDA0003168321360000022
when the ith target box is a negative sample,
Figure FDA0003168321360000023
tiparameterized coordinates (x, y, w, h) of a predicted bounding box characterizing the ith target box;
Figure FDA0003168321360000024
parameterized coordinates (x) of the real bounding box characterizing the ith target box*，y*，w*，h*) (ii) a λ is a weight balance parameter; class loss function
Figure FDA0003168321360000025
Function of regression loss
Figure FDA0003168321360000026
R represents the Smooth L1 Loss function.
5. The pulmonary nodule detection apparatus based on the improved fasterncnn algorithm as claimed in claim 1, wherein: the method for inputting the lung nodule candidate region of the person to be detected into the lung nodule detection classification model to obtain the predicted region and the predicted confidence probability of the lung nodule of the person to be detected comprises the following steps:
1) the pulmonary nodule detection classification model carries out RoI posing on a pulmonary nodule candidate region of a person to be detected, and converts regions with different sizes into regions with fixed length; inputting the area with fixed length into two connected full connection layers to obtain position and category output;
2) performing category prediction by adopting a cross entropy loss function;
3) and (3) performing position regression by adopting a Smooth L1 Loss function to obtain a predicted region and a predicted confidence probability of the pulmonary nodule of the person to be detected.
6. The device according to claim 5, wherein the lung nodule detection device based on the improved FasterRCNN algorithm comprises: smooth L1 Loss function SmoothL1As follows:
Figure FDA0003168321360000027
in the formula, test
Figure FDA0003168321360000028
σ is a constant.
7. The device according to claim 1, wherein the step of establishing a pulmonary nodule detection classification model comprises:
1) the data acquisition and labeling module acquires a plurality of lung nodule CT images from different testers, and marks a lung nodule region target frame in the lung nodule CT images;
the data acquisition and labeling module transmits the labeled lung nodule CT image to the data preprocessing module;
2) the data preprocessing module is used for preprocessing the lung nodule CT image to obtain a preprocessed lung nodule image, and dividing the preprocessed lung nodule image into a training set and a test set;
3) the feature extraction module clusters the lung nodule region target frames of the lung nodule CT image by using a K-means method to obtain a plurality of groups of lung nodule region target frames;
4) establishing a feature extraction model;
5) training the feature extraction model by using a training set to obtain a trained feature extraction model;
6) and (3) testing the trained feature extraction model by using the test set, finishing the establishment of the feature extraction model if the accuracy of the trained feature extraction model is greater than a preset threshold, and returning to the step 1) if the accuracy of the trained feature extraction model is not greater than the preset threshold.
8. The device according to claim 7, wherein the data preprocessing module preprocesses the CT image of the lung nodule to include:
1) according to the linear attenuation coefficients of air, water and X-rays, counting the HU value range of the lung region, and converting the lung nodule CT image into a JPG image;
2) and expanding the JPG picture by using a turning method, and converting the marked target frame of the lung nodule area into a VOC format.
9. The device according to claim 7, wherein the lung nodule detection device based on the improved fasterncnn algorithm comprises: the objective function min E for clustering the pulmonary nodule region object box of the pulmonary nodule CT image is as follows:
Figure FDA0003168321360000031
in the formula, mujIs a cluster CjThe mean vector of (2); e is the square error; k is the number of clusters; j is the cluster number.
CN202110811268.3A 2021-07-19 2021-07-19 Pulmonary nodule detection device based on improved FasterRCNN algorithm Pending CN114202002A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110811268.3A CN114202002A (en) 2021-07-19 2021-07-19 Pulmonary nodule detection device based on improved FasterRCNN algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110811268.3A CN114202002A (en) 2021-07-19 2021-07-19 Pulmonary nodule detection device based on improved FasterRCNN algorithm

Publications (1)

Publication Number Publication Date
CN114202002A true CN114202002A (en) 2022-03-18

Family

ID=80645857

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110811268.3A Pending CN114202002A (en) 2021-07-19 2021-07-19 Pulmonary nodule detection device based on improved FasterRCNN algorithm

Country Status (1)

Country Link
CN (1) CN114202002A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115082426A (en) * 2022-07-20 2022-09-20 湖北经济学院 Follicle detection method and device based on deep learning model
CN116563237A (en) * 2023-05-06 2023-08-08 大连工业大学 Deep learning-based chicken carcass defect hyperspectral image detection method

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115082426A (en) * 2022-07-20 2022-09-20 湖北经济学院 Follicle detection method and device based on deep learning model
CN115082426B (en) * 2022-07-20 2022-11-04 湖北经济学院 Follicle detection method and device based on deep learning model
CN116563237A (en) * 2023-05-06 2023-08-08 大连工业大学 Deep learning-based chicken carcass defect hyperspectral image detection method
CN116563237B (en) * 2023-05-06 2023-10-20 大连工业大学 Deep learning-based chicken carcass defect hyperspectral image detection method

Similar Documents

Publication Publication Date Title
CN109685776B (en) Pulmonary nodule detection method and system based on CT image
CN110287849B (en) Lightweight depth network image target detection method suitable for raspberry pi
EP3432263B1 (en) Semantic segmentation for cancer detection in digital breast tomosynthesis
CN108171232B (en) Deep learning algorithm-based bacterial and viral pneumonia classification method for children
CN111415316B (en) Defect data synthesis method based on generation countermeasure network
CN109919230B (en) Medical image pulmonary nodule detection method based on cyclic feature pyramid
CN108830326B (en) Automatic segmentation method and device for MRI (magnetic resonance imaging) image
CN112488210A (en) Three-dimensional point cloud automatic classification method based on graph convolution neural network
WO2020133636A1 (en) Method and system for intelligent envelope detection and warning in prostate surgery
CN110969245B (en) Target detection model training method and device for medical image
CN114220035A (en) Rapid pest detection method based on improved YOLO V4
CN112085714B (en) Pulmonary nodule detection method, model training method, device, equipment and medium
CN110210534B (en) Multi-packet fusion-based high-resolution remote sensing image scene multi-label classification method
CN107767416B (en) Method for identifying pedestrian orientation in low-resolution image
CN111310622A (en) Fish swarm target identification method for intelligent operation of underwater robot
CN115063573A (en) Multi-scale target detection method based on attention mechanism
CN114202002A (en) Pulmonary nodule detection device based on improved FasterRCNN algorithm
CN112700461B (en) System for pulmonary nodule detection and characterization class identification
CN113393416A (en) Method and medium for identifying and locating images using triple loss and prediction regions
CN112598613A (en) Determination method based on depth image segmentation and recognition for intelligent lung cancer diagnosis
CN108734200A (en) Human body target visible detection method and device based on BING features
CN114332572B (en) Method for extracting breast lesion ultrasonic image multi-scale fusion characteristic parameters based on saliency map-guided hierarchical dense characteristic fusion network
CN116091490A (en) Lung nodule detection method based on YOLOv4-CA-CBAM-K-means++ -SIOU
CN114283326A (en) Underwater target re-identification method combining local perception and high-order feature reconstruction
CN109902692A (en) A kind of image classification method based on regional area depth characteristic coding

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