CN111798455A - Thyroid nodule real-time segmentation method based on full convolution dense cavity network - Google Patents
Thyroid nodule real-time segmentation method based on full convolution dense cavity network Download PDFInfo
- Publication number
- CN111798455A CN111798455A CN201910914157.8A CN201910914157A CN111798455A CN 111798455 A CN111798455 A CN 111798455A CN 201910914157 A CN201910914157 A CN 201910914157A CN 111798455 A CN111798455 A CN 111798455A
- Authority
- CN
- China
- Prior art keywords
- convolution
- full
- dense
- thyroid
- data
- 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
Images
Classifications
-
- 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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10132—Ultrasound image
-
- 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
- G06T2207/30096—Tumor; Lesion
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a thyroid nodule real-time segmentation method based on a full-convolution dense cavity network, which comprises the following steps of: step one, thyroid data is obtained and preprocessed; marking the obtained data to be used as a data set for training a full-convolution dense void network model; constructing a full-convolution dense cavity network model based on dense connection, and performing parameter training; replacing the convolution kernels in the convolution layer with cavity convolution and decomposing by using convolution kernel decomposition; fifthly, carrying out data standardization and nonlinear activation processing on the input of the convolutional layer; and step six, analyzing and comparing the segmentation effect and efficiency of the full-convolution dense cavity network model.
Description
Technical Field
The invention belongs to the field of deep learning and image processing, relates to a convolutional neural network and an image semantic segmentation technology, and particularly relates to a thyroid nodule real-time segmentation method based on a full-convolution dense void network.
Background
Thyroid nodules are the most common abnormalities in the endocrine system, and their potential malignancy makes them clinically important. Ultrasound examination is the method of imaging of choice for diagnosing thyroid nodules. In clinical practice, a radiologist diagnoses thyroid gland malignancy and malignancy according to macroscopic evaluation criteria such as the aspect ratio of nodes, the existence of calcification, structures (diffuse, single-shot or multiple-shot), boundaries, echoic characteristics (hyperechoic, isoechoic, and hypoechoic), and the like in an ultrasonic image, but the radiologist may generate different diagnosis results for one ultrasonic image of the thyroid due to the influences of cognitive ability, subjective experience, fatigue degree, and the like. In addition, low contrast and speckle noise in the ultrasound images can also affect the diagnosis of the physician.
In recent years, researchers have paid more attention to deep learning-based computer-aided diagnosis, and research on thyroid ultrasound image-aided diagnosis has been actively developed. The existing deep learning method [1] [2] achieves high precision, however, models obtained by training the existing deep learning method [1] [2] have a large number of weight parameters, a large number of computing resources are needed, and the storage space and the processor performance of medical equipment are limited, so that the practical application of the deep learning method is limited to a certain extent. On the other hand, the purpose of computer-aided diagnosis is to shorten the diagnosis time and improve the diagnosis efficiency and accuracy, which requires that the depth model has high accuracy and real-time performance.
In recent years, the design of deeper neural networks achieves higher precision than many traditional computer vision algorithms in the tasks of image classification, semantic segmentation, target detection and the like, however, a large amount of computing resources and longer reasoning time are also required, so that the depth model cannot run on some resource-limited platforms. In order to solve the balance problem between high precision and computing resources, most of the existing methods focus on network pruning [3], low-bit qualification [4] and design of efficient network architecture. However, whether network pruning or low-bit quantification is performed on the trained model, the accuracy of the model is inevitably affected. In contrast, designing an efficient network architecture can reduce the computational resources required by the model without losing accuracy. The existing deep network has a lot of parameters, however, some parameters have very little or no effect when the network runs, and the network designed in the document [5] has a clear performance improvement and unchanged parameter number compared with the previous network, so that the network parameters are reduced, and the effect of improving the residual parameters is very useful for balancing the accuracy and the operation resources of the network.
Disclosure of Invention
The invention aims to solve the problem that the existing semantic segmentation model has too many parameters and cannot be efficiently operated on computer-aided diagnosis equipment with limited computing resources.
The purpose of the invention is realized by the following technical scheme:
a thyroid nodule real-time segmentation method based on a full convolution dense cavity network comprises the following steps:
the method comprises the following steps: obtaining thyroid data and preprocessing the thyroid data;
step two: marking the obtained data to be used as a data set for training a full-convolution dense void network model;
step three: constructing a full-convolution dense network model based on dense connection, and performing parameter training;
step four: replacing the convolution kernel in the convolution layer with a hole convolution and decomposing by using convolution kernel decomposition;
step five: carrying out data standardization and nonlinear activation processing on the input of the convolutional layer;
step six: and analyzing and comparing the precision index and the efficiency of the segmentation effect and the efficiency of the full convolution dense cavity network model.
Further, marking the nodule edge of the obtained thyroid gland ultrasonic image in the step two, and normalizing the size of the thyroid gland ultrasonic image; the formed data set includes a training set and a test set.
Further, in the third step, the data set obtained in the second step is used for training full-convolution dense void network model parameters; the full convolutional dense hole network model is an end-to-end model for thyroid nodule segmentation, follows the architecture of an automatic encoder-decoder, and uses dense connections to perform cross-layer transmission on features extracted from convolutional layers.
Further, in the fourth step, the full convolution dense void network model is composed of convolution layers, convolution kernels in different convolution layers are replaced by void convolutions with different void rates, and a convolution kernel decomposition is used for decomposing a two-dimensional convolution kernel into a one-dimensional convolution kernel.
Further, step five, on the basis of step four, standardizing and activating the input of the convolutional layer by using batch standardization (BN) and a linear correction unit activation function (ReLU), and processing the parameters of the full-convolutional dense void network by using random inactivation (dorpout); specifically, the method includes performing data normalization processing on the input of the convolutional layer by using Batch Normalization (BN), and increasing the nonlinearity of data by using a linear modified unit activation function (ReLU).
Compared with the prior art, the technical scheme of the invention has the following beneficial effects:
1. the model constructed by the method uses the combination of dense connection, hollow convolution and convolution kernel decomposition, so that the model is more efficient while keeping similar learning performance. Dense connection can greatly improve feature reuse, so that a model can generate a large number of features by using a small number of convolution kernels, and the size of the final model is smaller. Convolution kernel decomposition decomposes one convolution kernel of 3 × 3 size into two convolution kernels of 3 × 1 and 1 × 3 size, the model parameters decrease, but the nonlinearity of the model increases, so the accuracy does not decrease significantly. The hole convolution can obtain more context information while keeping the same parameter quantity, thereby improving the accuracy of the model.
2. The thyroid gland data set constructed by the method achieves similar segmentation precision to the high-precision model, the IOU is 0.57% higher than that of the FCDenseNet model with the best overall effect and is 70.72%, the running time of the model on a single NVIDATITAN XpGPU is less than 1/6 of the high-precision model and is only 7.74ms, the efficiency is competitive with that of the high-efficiency model, and the running time of the model is only 2.24ms more than that of the fastest ENet.
3. The model trained by the method of the invention has good balance between the segmentation precision and the speed, the size of the model is only 3.9Mb, and the method is suitable for equipment which needs robustness and efficiency.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention.
Fig. 2a is a screened thyroid ultrasound image, and fig. 2b shows a label (ground route) of the image.
FIG. 3 is a schematic diagram of the structure of a single convolutional layer. In fig. 3, c represents the number of channels of the input features, 3 × 1 and 1 × 3 represent the sizes of convolution kernels, k represents the number of convolution kernels, L represents the void rate of the convolution kernels,
representing the connection of the input to the output.
Fig. 4a shows an image input to the full-convolution dense hole network model, and fig. 4b shows a division result of the FCDDN model. Fig. 4c shows the segmentation result of the FC DenseNet model; FIG. 4d shows the segmentation result of the U-Net model.
Detailed Description
The invention is described in further detail below with reference to the figures and specific examples. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The invention provides a thyroid nodule segmentation method based on a full-convolution dense cavity network, which is an overall flow schematic diagram of a specific embodiment of the thyroid nodule segmentation method as shown in fig. 1 and comprises the following steps:
the method comprises the following steps: obtaining thyroid data and preprocessing the thyroid data;
step 101: obtaining the thyroid ultrasound image data after pathological verification from a hospital, taking the image out of a folder with a medical record, modifying the name of the image and making a backup, and then screening the data with clear images and removed nodule physiological structures.
Step two: marking the obtained data to be used as a data set for training a full-convolution dense void network model;
step 201: with the help of the ultrasound department and the radiologist, the node edge marking is carried out on the screened ultrasound images, and a group channel is obtained according to the marking, so as to form a data set of the invention, the images are randomly divided into a training set and a testing set, and then the sizes of the images and the group channel are uniformly scaled to 256 × 256 (the images and the group channel are shown in fig. 2a and 2 b).
Step three: constructing a full-convolution dense network model based on dense connection, and performing parameter training;
step 301: the full-volume Dense hole network model is built according to the framework of an automatic encoder-decoder, and can be called as a full volumetric specific detect DilatedNet model; referred to as the FCDDN model; wherein layers 1 to 8 constitute an encoder and layers 9 to 15 constitute a decoder. Layer 1 is a convolution layer with convolution kernel size of 3 × 3 and number of 48; the 2 nd to 8 th layers are 1 convolutional layer, a pooling layer, 4 convolutional layers and a pooling layer in sequence, wherein the convolutional cores of the convolutional layers are 3 multiplied by 3, the number of the convolutional layers is 16, the pooling layer is 2 multiplied by 2, and the step length is 2; the 9 th layer to the 14 th layer are sequentially an deconvolution layer, 1 convolution layer, an deconvolution layer and 1 convolution layer, wherein the deconvolution size is 3 x 3, the step length is 2, the convolution kernel size of the convolution layer is 3 x 3, and the number is 16; layer 15 is a convolutional layer with convolutional kernel size of 1 × 1 and convolutional kernel number of 2, and is used to classify features.
Step 302: parameters of the full convolution dense void network model are trained using a training set in the thyroid dataset. The initial learning rate is 0.0001, each training round is divided by 2, 240 training rounds are performed, the final learning rate is 0.00005, dropout is 0.8, and the random gradient descent optimizer the adam optimization is used for training.
Step four: replacing the convolution kernel in the convolution layer with a hole convolution and decomposing by using convolution kernel decomposition;
step 401: the full convolution dense hole network model comprises a plurality of convolution layers, convolution kernels in different convolution layers are replaced by hole convolutions with different hole rates, the hole rates are set according to the number of the convolution layers in the layers of the full convolution dense hole network model, the hole rates comprise one convolution layer, the hole rate of the hole convolution is 2, the hole rates comprise 4 convolution layers, and the hole rates of the convolution layers are sequentially 2, 4, 8 and 16.
Step 402: it is proposed to decompose the 2-dimensional convolution kernels in layers 2 to 14 into one-dimensional convolution kernels, all 3 x 3 in size, using convolution kernel decomposition, which can be decomposed into two successive convolution kernels of 3 x 1 and 1 x 3 in size. For convolution kernels of size 3 x 3, the convolution kernel decomposition may reduce the model parameters to 2/3 as they are.
Step five: carrying out data batch standardization and nonlinear activation processing on the input of the convolutional layer;
step 501: the method comprises the steps of firstly calculating a data mean value, then calculating a data variance, then standardizing the data by using a standard deviation formula, and finally introducing two parameters of gamma and beta to carry out translation and scaling processing on the data.
Step 502: and carrying out nonlinear activation processing on the data after batch standardization processing by using a Relu function. By this time, the building of the structure of the convolutional layer is finished, a complete schematic diagram is shown in fig. 3, data input into the convolutional layer is subjected to batch standardization processing, nonlinear activation processing and convolutional kernels with the sizes of 3 × 1 and 1 × 3, finally, the data input into the convolutional layer and the data output are connected together, Dropout only randomly deactivates part of neurons during network training, and the trained network does not deactivate the neurons any more.
Step six: and analyzing and comparing the segmentation effect and efficiency of the full convolution dense cavity network model.
The trained full-convolution dense hole network model is used for nodule segmentation of a test set, accuracy is an important aspect for measuring segmentation effect, main indexes for evaluating the accuracy are IOU, TPF and FPF, segmentation efficiency is an important aspect for measuring the quality of the model, and the evaluation efficiency is only the index of running time and storage size. Table (1) quantitatively compares and analyzes the segmentation effect and efficiency of the FCDDN model with other models, and from the IOU, TPF and FPF of each model, the IOU, TPF and FPF of the FCDDN model proposed by the present invention are 72.72%, 96.10% and 0.677% respectively, and have a small overall difference compared with the U-Net and fcdensnet models with the best accuracy, although the IOU of U-Net is 3.76% higher than that of the model of the present invention, and the running time is nearly twice as much, and the parameters are two orders of magnitude more, and from fig. 4a to fig. 4d, the segmentation result of U-Net is often smaller than groudtuth, which may have a certain impact on the diagnosis of thyroid gland by doctors. From the view of running time and Model size (Model size), the full-convolution dense network Model provided by the invention has the same high efficiency as the high-efficiency models (ENet and ERFNet), the Model size is even much smaller, and the segmentation accuracy is higher than that of the high-efficiency models. By integrating a plurality of evaluation indexes, the full convolution dense network model provided by the invention is optimal in all comparison models.
TABLE 1 results of model segmentation and comparison
Specifically, the experimental results are evaluated by three evaluation indexes, namely, interaction-over-Interaction (IOU), True Positive Fraction (TPF) and False Positive Fraction (FPF). The calculation modes of the three evaluation indexes are shown as formulas (1), (2) and (3).
IOU＝area(A∩B)/area(A∪B) (1)
TPF＝area(A∩B)/area(A) (2)
FPF＝(area(A)-area(A∩B))/(area(C)-area(A)) (3)
Wherein, A is the node area in the ground route, B is the node area in the model segmentation result, and C is the ground route. The larger the IOU and TPF are, the smaller the FPF is, and the better the segmentation effect is. In addition, the running time of the model is tested and compared, and the shorter the time is, the higher the model segmentation efficiency is.
The experimental result shows that compared with a high-precision network, the full-volume dense network model provided by the invention achieves similar precision on an IOU (input output Unit), but the quantity of parameters is obviously less, and the running time is obviously shorter. U-Net [6] is less than 4% higher than our method in IOU, but 11% lower in TPF, as can be seen from the effect graph of segmentation, since the result of segmentation by the method of the present invention can contain the group channel as a whole, but U-Net is not. The parameters of the high-precision network are much more than those of the network provided by the invention, and the parameters of the depth model can be illustrated to have a large amount of redundancy. It can be seen from the comparison of FCDDN and FC DenseNet [7], designing an efficient network architecture can improve efficiency without loss of accuracy. Compared with an efficient network, the method is only slightly longer than the fastest ENet [8] in operation time, but has the precision 2.49% higher than the highest ERFNet [9 ]. The final experiment result shows that the efficiency of the model is greatly improved while the high precision is kept.
The invention provides a semantic segmentation network structure with high precision and high efficiency, and a trained model is called as an FCDDN model. The combination of dense connections, hole convolutions and convolution kernels are used in the model, making it more efficient while maintaining similar learning performance. The FCDDN model achieves similar segmentation accuracy on thyroid datasets as the high accuracy model and runs less than 1/6 of the high accuracy model on a single NVIDATITAN Xp GPU, with efficiency as competitive as the high efficiency model. Finally, the model trained by the method of the invention has good balance between the segmentation precision and the segmentation speed, so that the method is suitable for equipment which needs both robustness and efficiency.
The present invention is not limited to the above-described embodiments. The foregoing description of the specific embodiments is intended to describe and illustrate the technical solutions of the present invention, and the above specific embodiments are merely illustrative and not restrictive. Those skilled in the art can make many changes and modifications to the invention without departing from the spirit and scope of the invention as defined in the appended claims.
Reference documents:
[1]Liu Shu,Qi Lu,Haifang Qin,Jianping Shi,and Jiaya Jia.Pathaggregation network for instance segmentation.2018.
[2]Liang Chieh Chen,Yukun Zhu,George Papandreou,Florian Schroff,andHartwig Adam.Encoder-decoder with atrous separable convolution for semanticimage segmentation.2018
[3]Yihui He,Xiangyu Zhang,and Jian Sun.Channel pruning foraccelerating very deep neural networks.2017
[4]Bohan Zhuang,Chunhua Shen,Mingkui Tan,Lingqiao Liu,and IanReid.Towards effective low-bitwidth convolutional neural networks.2017.
[5]Francois Chollet.Xception:Deep learning with depthwise separableconvolutions.In IEEE Conference on Computer Vision&Pattern Recognition,2016.
[6]Olaf Ronneberger,Philipp Fischer,and Thomas Brox.U-net:Convolutional networks for biomedical image segmentation.In InternationalConference on Medical Image Computing&Computer-assisted Intervention,2015.
[7]Simon Jegou,Michal Drozdzal,David′Vazquez,Adriana Romero,andYoshua Bengio.The one hundred layers tiramisu:Fully convolutional densenetsfor semantic segmentation.2016.
[8]Adam Paszke,Abhishek Chaurasia,
Sangpil Kim,and Eugenio Culurciello.Enet:A deep neural networkarchitecture for real-time semantic segmentation.2016.
[9]Eduardo Romera,Jose M.Alvarez,Luis M.Bergasa,Roberto Arroyo,Eduardo Romera,Jose M.Alvarez,Luis M.Bergasa,Roberto Arroyo,Eduardo Romera,and Jose M.Alvarez.Erfnet:Efficient residual factorized convnet for real-timesemantic segmentation.IEEE Transactions on Intelligent TransportationSystems,PP(99):1–10,2017.
Claims (5)
1. a thyroid nodule real-time segmentation method based on a full convolution dense cavity network is characterized by comprising the following steps:
the method comprises the following steps: obtaining thyroid data and preprocessing the thyroid data;
step two: marking the obtained data to be used as a data set for training a full-convolution dense void network model;
step three: constructing a full-convolution dense cavity network model based on dense connection, and performing parameter training;
step four: replacing the convolution kernel in the convolution layer with a hole convolution and decomposing by using convolution kernel decomposition;
step five: carrying out data standardization and nonlinear activation processing on the input of the convolutional layer;
step six: and analyzing and comparing the precision index and the efficiency of the segmentation effect and the efficiency of the full convolution dense cavity network model.
2. The thyroid nodule real-time segmentation method based on the full-convolution dense void network according to claim 1, wherein in the second step, the obtained thyroid ultrasound image is marked with a nodule edge, and the size of the thyroid ultrasound image is normalized; the formed data set includes a training set and a test set.
3. The thyroid nodule real-time segmentation method based on the full convolution dense cavity network as claimed in claim 1, wherein the data set obtained in step two is used in step three to train the model parameters of the full convolution dense cavity network; the full convolutional dense hole network model is an end-to-end model for thyroid nodule segmentation, follows the architecture of an automatic encoder-decoder, and uses dense connections to perform cross-layer transmission on features extracted from convolutional layers.
4. The method for real-time thyroid nodule segmentation based on full convolution dense hole network as claimed in claim 1, wherein in step four, the full convolution dense hole network model is composed of convolution layers, convolution kernels in different convolution layers are replaced by hole convolutions with different hole rates, and convolution kernel decomposition is used to decompose two-dimensional convolution kernels into one-dimensional convolution kernels.
5. The method for real-time thyroid nodule segmentation based on full-convolution dense void network as claimed in claim 1, wherein step five is to normalize and activate the input of the convolution layer on the basis of step four by using Batch Normalization (BN) and linear modified unit activation function (ReLU), and process the full-convolution dense void network model parameters by using random deactivation (dorpout); specifically, the method includes performing data normalization processing on the input of the convolutional layer by using Batch Normalization (BN), and increasing the nonlinearity of data by using a linear modified unit activation function (ReLU).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910914157.8A CN111798455B (en) | 2019-09-25 | 2019-09-25 | Thyroid nodule real-time segmentation method based on full convolution dense cavity network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910914157.8A CN111798455B (en) | 2019-09-25 | 2019-09-25 | Thyroid nodule real-time segmentation method based on full convolution dense cavity network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111798455A true CN111798455A (en) | 2020-10-20 |
| CN111798455B CN111798455B (en) | 2023-07-04 |
Family
ID=72805521
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910914157.8A Active CN111798455B (en) | 2019-09-25 | 2019-09-25 | Thyroid nodule real-time segmentation method based on full convolution dense cavity network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111798455B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112381846A (en) * | 2020-12-11 | 2021-02-19 | 江南大学 | Ultrasonic thyroid nodule segmentation method based on asymmetric network |
| CN113610859A (en) * | 2021-06-07 | 2021-11-05 | 东北大学 | Automatic thyroid nodule segmentation method based on ultrasonic image |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056595A (en) * | 2015-11-30 | 2016-10-26 | 浙江德尚韵兴图像科技有限公司 | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network |
| CN108776969A (en) * | 2018-05-24 | 2018-11-09 | 复旦大学 | Breast ultrasound image lesion segmentation approach based on full convolutional network |
| CN109087327A (en) * | 2018-07-13 | 2018-12-25 | 天津大学 | A kind of thyroid nodule ultrasonic image division method cascading full convolutional neural networks |
| CN109344883A (en) * | 2018-09-13 | 2019-02-15 | 西京学院 | Fruit tree diseases and pests recognition methods under a kind of complex background based on empty convolution |
-
2019
- 2019-09-25 CN CN201910914157.8A patent/CN111798455B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056595A (en) * | 2015-11-30 | 2016-10-26 | 浙江德尚韵兴图像科技有限公司 | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network |
| CN108776969A (en) * | 2018-05-24 | 2018-11-09 | 复旦大学 | Breast ultrasound image lesion segmentation approach based on full convolutional network |
| CN109087327A (en) * | 2018-07-13 | 2018-12-25 | 天津大学 | A kind of thyroid nodule ultrasonic image division method cascading full convolutional neural networks |
| CN109344883A (en) * | 2018-09-13 | 2019-02-15 | 西京学院 | Fruit tree diseases and pests recognition methods under a kind of complex background based on empty convolution |
Non-Patent Citations (1)
| Title |
|---|
| MARIOS ANTHIMOPOULOS 等: "Semantic Segmentation of Pathological Lung Tissue with Dilated Fully Convolutional Networks", ARXIV * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112381846A (en) * | 2020-12-11 | 2021-02-19 | 江南大学 | Ultrasonic thyroid nodule segmentation method based on asymmetric network |
| CN112381846B (en) * | 2020-12-11 | 2024-03-08 | 江南大学 | Ultrasonic thyroid nodule segmentation method based on asymmetric network |
| CN113610859A (en) * | 2021-06-07 | 2021-11-05 | 东北大学 | Automatic thyroid nodule segmentation method based on ultrasonic image |
| CN113610859B (en) * | 2021-06-07 | 2023-10-31 | 东北大学 | Automatic thyroid nodule segmentation method based on ultrasonic image |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111798455B (en) | 2023-07-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wu et al. | Predict pneumonia with chest X-ray images based on convolutional deep neural learning networks | |
| Ali et al. | Machine learning based automated segmentation and hybrid feature analysis for diabetic retinopathy classification using fundus image | |
| US11200982B2 (en) | Method for analysing medical treatment data based on deep learning and intelligence analyser thereof | |
| Liu et al. | Fine-tuning pre-trained convolutional neural networks for gastric precancerous disease classification on magnification narrow-band imaging images | |
| El-Shafai et al. | Efficient Deep-Learning-Based Autoencoder Denoising Approach for Medical Image Diagnosis. | |
| CN109447183A (en) | Model training method, device, equipment and medium | |
| CN109389171B (en) | Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology | |
| Lindsey et al. | Automated pneumothorax diagnosis using deep neural networks | |
| CN111798455B (en) | Thyroid nodule real-time segmentation method based on full convolution dense cavity network | |
| Shamrat et al. | Analysing most efficient deep learning model to detect COVID-19 from computer tomography images | |
| Nie et al. | Recent advances in diagnosis of skin lesions using dermoscopic images based on deep learning | |
| CN111784713A (en) | Attention mechanism-introduced U-shaped heart segmentation method | |
| Jin et al. | Machine learning techniques for pulmonary nodule computer-aided diagnosis using CT images: A systematic review | |
| Farhangi et al. | Automatic lung nodule detection in thoracic CT scans using dilated slice‐wise convolutions | |
| Kaya | Feature fusion-based ensemble CNN learning optimization for automated detection of pediatric pneumonia | |
| Zheng et al. | Pneumoconiosis identification in chest X-ray films with CNN-based transfer learning | |
| Junyue et al. | Breast cancer diagnosis using hybrid AlexNet-ELM and chimp optimization algorithm evolved by Nelder-mead simplex approach | |
| Jerbi et al. | Automatic classification of ultrasound thyroids images using vision transformers and generative adversarial networks | |
| Nainwal et al. | Convolution neural network based covid-19 screening model | |
| WO2022127500A1 (en) | Multiple neural networks-based mri image segmentation method and apparatus, and device | |
| Thomas et al. | Diabetic retinopathy detection using EADBSC and improved dilated ensemble CNN-based classification | |
| Pradeep et al. | A Certain Investigation on Intravascular Optical Coherence Tomography for Clustering Based Image Segmentation | |
| Chutia et al. | Classification of lung diseases using an attention-based modified DenseNet model | |
| Thomas et al. | Diabetic retinopathy detection using ensembled transfer learning based thrice CNN with SVM classifier | |
| Zang et al. | AmtNet: Attentional multi-scale temporal network for phonocardiogram signal classification |
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 |