NL2029876B1 - Deep residual network-based classification system for thyroid cancer computed tomography (ct) images - Google Patents
Deep residual network-based classification system for thyroid cancer computed tomography (ct) images Download PDFInfo
- Publication number
- NL2029876B1 NL2029876B1 NL2029876A NL2029876A NL2029876B1 NL 2029876 B1 NL2029876 B1 NL 2029876B1 NL 2029876 A NL2029876 A NL 2029876A NL 2029876 A NL2029876 A NL 2029876A NL 2029876 B1 NL2029876 B1 NL 2029876B1
- Authority
- NL
- Netherlands
- Prior art keywords
- image
- tumor
- deep
- images
- lymph node
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration by the use of more than one image, e.g. averaging, subtraction
-
- 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
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
-
- 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
Abstract
The present disclosure provides a deep residual network-based classification system for thyroid cancer computed tomography (CT) images, including: a thyroid cancer CT image acquisition module, configured to acquire labeled CT images of plurality of thyroid cancer patients; a multi-scale segmentation module, configured to segment CT images of each of the plurality of thyroid cancer patients according to different scales, and sequentially intercept a cubic tumor area, a cubic tumor area expanded by 5 mm and a cubic tumor area expanded by 10 mm, to obtain a tumor image, a tumor image expanded by 5 mm and a tumor image expanded by 10 mm; a preprocessing module, configured to preprocess the images to obtain a training set; a deep residual network training module, configured to train and optimize the deep residual network by using the training set; and a thyroid cancer CT image classification module, configured to input thyroid cancer CT images to be classified into the optimized deep residual network for classification, to obtain a classification result of the thyroid cancer CT images. The present disclosure can accurately classify the thyroid cancer CT images.
Description
DEEP RESIDUAL NETWORK-BASED CLASSIFICATION SYSTEM FOR THYROID
CANCER COMPUTED TOMOGRAPHY (CT) IMAGES
[01] The present disclosure relates to the technical field of medical imaging and artificial intelligence, and in particular to a deep residual network-based classification system for thyroid cancer computed tomography (CT) images.
[02] In recent years, computer technology has been widely used in the medical field. In particular, computer-aided diagnosis technology can assist radiologist in diagnosis using medical imaging and medical image processing technologies together with computer-related algorithms, to improve the accuracy and efficiency of diagnosis.
[03] Thyroid cancer has a relatively high incidence. It is reported that up to 60-70% of thyroid cancer patients have lymph node metastasis. Therefore, it is necessary to accurately determine the area required for lymph node dissection before the initial surgery to determine the risk of lymph node metastasis. Clinically, the area is generally determined by CT examination, and CT images need to be differentiated to help radiologists make judgements.
[04] At present, artificial intelligence -assisted diagnosis technology mainly includes radiomics-based methods and deep learning-based methods. Radiomics method extracts manually-designed features from medical images, and constructs models through feature selection and traditional machine learning methods. However, the manually-designed features are difficult to accurately characterize the inherent features of the image.
[05] Deep learning method can automatically extract high-dimensional features of the images, has great advantages over traditional machine learning methods, and can avoid problems caused by manually extracting image features. Although there are many frameworks for image classification with the development of deep learning, there is still no deep learning model for classifying CT images of thyroid cancer patients. Due to lesion images contained, the CT images of thyroid cancer patients are more complicated and have more features than ordinary images. The current frameworks for classification of ordinary images cannot accurately classify the thyroid cancer CT images, and thus cannot assist radiologists in determining whether there is lymph node metastasis in thyroid cancer CT images. Therefore, there is an urgent need in the art for a deep learning model for classifying the CT images of thyroid cancer patients to solve the above problems.
[06] The purpose of the present disclosure is to provide a deep residual network-based classification system for thyroid cancer CT images. The system can accurately classify the thyroid cancer CT images, and thus can assist radiologists in determining whether there is lymph node metastasis through the thyroid cancer CT images.
[07] To achieve the above objective, the present disclosure provides the following solutions:
[08] A deep residual network-based classification system for thyroid cancer CT image includes:
[09] athyroid cancer CT image acquisition module, configured to acquire labeled CT images of multiple thyroid cancer patients;
[19] a multi-scale segmentation module, connected to the thyroid cancer CT image acquisition module, and configured to segment CT images of each of the multiple thyroid cancer patients according to different scales, and sequentially intercept a cubic tumor area, a cubic tumor area expanded by 5 mm and a cubic tumor area expanded by 10 mm, to obtain a tumor image, a tumor image expanded by 5 mm and a tumor image expanded by 10 mm;
[11] a preprocessing module, connected to the multi-scale segmentation module, and configured to preprocess the tumor image, the tumor image expanded by 5 mm and the tumor image expanded by 10 mm, respectively, to obtain a training set;
[12] a deep residual network training module, connected to the preprocessing module, and configured to train and optimize a deep residual network by using the training set to obtain an optimized deep residual network; and
[13] a thyroid cancer CT image classification module, connected to the deep residual network training module, and configured to input thyroid cancer CT images to be classified into the optimized deep residual network for classification, to obtain a classification result of the thyroid cancer CT images; where the classification result includes lymph node metastasis and lymph node non-metastasis through the thyroid cancer CT images.
[14] In some embodiments, the CT image of each of the multiple thyroid cancer patients may be composed of multiple consecutive image slices corresponding to different phases; and the different phases may include a plain scan phase, an arterial phase and a venous phase.
[15] In some embodiments, the CT image of each of the multiple thyroid cancer patients may include a region of interest (ROI); the ROI may be delineated slice by slice along an edge of a thyroid primary lesion in the plain scan phase, the arterial phase and the venous phase; and the ROL in each phase may be superimposed slice by slice to form a three-dimensional volume of interest (VOI).
[16] In some embodiments, the multi-scale segmentation module may specifically include:
[17] a voxel spacing conversion unit, connected to the thyroid cancer CT image acquisition module, and configured to convert a voxel spacing of the CT image of each of the multiple thyroid cancer patients to obtain a converted CT image;
[18] a VOI determination unit, connected to the voxel spacing conversion unit, and configured to determine a length, a width, a height and a center point coordinate of the VOI according to a position of the VOL in the converted CT image; and
[19] a cropping unit, connected to the VOI determination unit, and configured to intercept the cubic tumor area, the cubic tumor area expanded by 5 mm and the cubic tumor area expanded by 10 mm from the converted CT image according to the length, the width, the height and the center point coordinate of the VOI, to obtain the tumor image, the tumor image expanded by 5 mm and the tumor image expanded by 10 mm.
[20] In some embodiments, the preprocessing module may specifically include:
[21] a normalization unit, connected to the multi-scale segmentation module, and configured to normalize each voxel in the tumor image, the tumor image expanded by 5 mm and the tumor image expanded by 10 mm, respectively, to obtain normalized tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm;
[22] a data scaling unit, connected to the normalization unit, and configured to unity the normalized tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm to a set image size, respectively, to obtain image size-set tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm; and
[23] a data augmentation unit, connected to the data scaling unit, and configured to conduct data augmentation on the image size-set tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm via flipping, rotation, translation and zooming, to obtain the training set, where the training set may include tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm after data augmentation.
[24] In some embodiments, the deep residual network training module may specifically include:
[25] a deep residual network construction unit, connected to the preprocessing module, and configured to construct the deep residual network; and
[26] a deep residual network training unit, connected to the deep residual network construction unit, and configured to receive the training set sent by the preprocessing module, and to train and optimize the deep residual network using the training set to obtain the optimized deep residual network.
[27] In some embodiments, the deep residual network may specifically include:
[28] a shallow feature extraction layer, connected to the preprocessing module, and configured to use a 64-channel 3x3x3 convolution kernel and a rectified linear unit (ReLU) connected to the 3x3x3 convolution kernel to extract shallow features of the images in the training set to obtain a shallow feature map with 64-channel;
[29] a deep feature extraction layer, connected to the shallow feature extraction layer, and configured to extract deep features in the shallow feature map to obtain a deep feature map:
[30] a skip connection layer separately connected with the shallow feature extraction layer and the deep feature extraction layer, and configured to connect the shallow feature map and the deep feature map;
[31] a convolutional layer, connected to the skip connection layer, and configured to further extract features from the connected shallow feature map and deep feature map with a 7x7x7 convolution kernel and an ReLU connected to the 7x7x7 convolution kernel, to generate a 128-channel feature map; and
[32] a classification layer, connected to the convolutional layer, and configured to conduct a 3D global average pooling operation on the 128-channel feature map, calculate probability of the lymph node metastasis and the lymph node non-metastasis through the thyroid cancer CT images, and take a category with the highest probability as a classification result. [331 In some embodiments, the deep feature extraction layer may specifically include:
[34] a plurality of residual dense blocks (RDBs) connected to the shallow feature extraction layer, where each RDB may be connected sequentially, and may be configured to extract the deep features in the shallow feature map using nine 3x3x3 convolution kernels and an ReLU separately connected to each 3x3x3 convolution kernel; and
[35] a IxIx1 convolutional layer connected with a plurality of the RDBs, and configured to fuse the deep features extracted by each RDB to obtain the deep feature map.
[36] In some embodiments, the classification layer may specifically include a fully connected (FC) and a Softmax layer that are mutually connected; the FC layer may be connected to the convolutional layer and may be used for conducting the 3D global average pooling operation on the 128-channel feature map; and the Softmax layer may be used for calculating the probability of the lymph node metastasis and the lymph node non-metastasis in the thyroid cancer CT images, and taking the category with the highest probability as the classification result of the thyroid cancer CT images.
[37] Based on specific examples provided in the present disclosure, the present disclosure discloses the following technical effects:
[38] The present invention discloses a deep residual network-based classification system for thyroid cancer CT images. In the system, the multi-scale segmentation module is provided to segment the CT images of each of the multiple thyroid cancer patients according to different scales, and sequentially intercept the cubic tumor area, the cubic tumor area expanded by 5 mm and the cubic tumor area expanded by 10 mm, to obtain the tumor image, the tumor image expanded by 5 mm and the tumor image expanded by 10 mm. The deep residual network training module is provided to train and optimize the deep residual network using the images of different scales, and to classity the thyroid cancer CT images using the optimized deep residual network. The present disclosure extracts multi-scale information of the thyroid cancer tumor in the thyroid cancer CT images through a combination of multi-scale segmentation and deep residual network, and fuses 5 features of the tumor and peritumoral to improve the accuracy of model classification. Compared with traditional deep learning frameworks such as ResNet and DenseNet, the present disclosure further strengthens the fusion and propagation of features, and fully learns the high-frequency and detailed features of the images, such that the thyroid cancer CT images can be accurately classified to assist radiologists in determining whether there is lymph node metastasis through the thyroid cancer
CT images.
[39] To describe the technical solutions in the embodiments of the present disclosure or in the prior art more clearly, the accompanying drawings required for the embodiments are briefly described below. Apparently, the accompanying drawings in the following description show merely some embodiments of the present disclosure, and a person of ordinary skill in the art may still derive other accompanying drawings from these accompanying drawings without creative efforts.
[40] FIG. | is a structural diagram of an example of a deep residual network-based classification system for thyroid cancer CT image of the present disclosure.
[41] FIG. 2 is a structural diagram of classification based on a deep residual network of the present disclosure.
[42] FIG. 3 is a structural diagram of a residual dense network of the present disclosure.
[43] The technical solutions of the examples of the present disclosure are clearly and completely described below with reference to the accompanying drawings. Apparently, the described examples are merely a part rather than all of the examples of the present disclosure. All other examples obtained by a person of ordinary skill in the art based on the examples of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.
[44] The purpose of the present disclosure is to provide a deep residual network-based classification system for thyroid cancer CT images. The system can accurately classify the thyroid cancer CT images, and thus can assist radiologists in determining whether there is lymph node metastasis in thyroid cancer CT images.
[45] To make the above objectives, features, and advantages of the present disclosure clearer and more comprehensible, the present disclosure will be further described in detail below with reference to the accompanying drawings and the specific implementation.
[46] FIG. 1is a structural diagram of an example of the deep residual network-based classification system for thyroid cancer CT image of the present disclosure. As shown in FIG. 1, the deep residual network-based classification system for thyroid cancer CT image includes a thyroid cancer CT image acquisition module 101, a multi-scale segmentation module 102 connected to the thyroid cancer CT image acquisition module 101, a preprocessing module 103 connected to the multi-scale segmentation module 102, a deep residual network training module 104 connected to the preprocessing module 103, and a thyroid cancer CT image classification module 105 connected to the deep residual network training module 104.
[47] The thyroid cancer CT image acquisition module (thyroid cancer CT image collection module) 101 is used for acquiring labeled CT images of thyroid cancer patients. The CT image of each of patient is composed of multiple consecutive image slices corresponding to different phases.
The different phases include a plain scan phase, an arterial phase and a venous phase. The CT image of each of patient includes an ROL, the ROI is delineated slice by slice along an edge of a thyroid primary lesion in the plain scan phase, the arterial phase and the venous phase; and the ROI in each phase are superimposed slice by slice to form a three-dimensional VOL. The ROI in CT images, are obtained by a radiologist with more than 10 years of diagnostic experience. Each patient's image is labeled with label information (lymph node metastasis/lymph node non-metastasis).
[48] The thyroid cancer CT images were collected from 913 thyroid cancer patients subjected to thyroid CT examination in Yantai Yuhuangding Hospital from 2017 to 2020, and label of lymph node metastasis/lymph node non-metastasis in thyroid cancer CT images is obtained through pathological sample detection, that is, the label whether there is lymph node metastasis is determined by pathology result. Since the three-dimensional (3D) structure of each phase of each patient is represented by multi-layer continuous slices, a 3D original CT image matrix is obtained. To reduce the interference of peritumoral on the network model, the present disclosure segments the original
CT image in different sizes according to a position of the delineated VOL
[49] The multi-scale segmentation module (multi-scale clipping module) 102 is used for segmenting CT images (three-dimensional images) of each of the multiple thyroid cancer patients according to different scales, and sequentially intercepting a cubic tumor area, a cubic tumor area expanded by 5 mm and a cubic tumor area expanded by 10 mm, to obtain a tumor image, a tumor image expanded by 5 mm and a tumor image expanded by 10 mm. The multi-scale segmentation module 102 segments the tumor area, the tumor area expanded by 5 mm and the tumor area expanded by 10 mm according to coordinates and a center point of the tumor (coordinates and center position of the tumor in the original image), to obtain a multi-scale three-dimensional image input to the deep residual network, that is, three-dimensional images with three different scales in the thyroid tumor area.
[50] The multi-scale segmentation module 102 specifically includes a voxel spacing conversion unit connected to the thyroid cancer CT image acquisition module 101, a VOI determination unit connected to the voxel spacing conversion unit, and a cropping unit connected to the VOI determination unit.
[51] The voxel spacing conversion unit is used for converting the voxel spacing of the CT images (original CT images) of each of patient to (1 mm, 1 mm, 5 mm) to obtain the converted CT images.
[52] The VOI determination unit is used for determining the length, width, height and center point coordinates L (x, y, z) of the VOI according to the position of the VOI in the converted CT image.
[53] The cropping unit is used for intercepting the cubic tumor area, the cubic tumor area expanded by 5 mm and the cubic tumor area expanded by 10 mm from the converted CT image according to the length, the width, the height and the center point coordinates L (x, y, z) of the VOI, to obtain the tumor image, the tumor image expanded by 5 mm and the tumor image expanded by 10 mm. The cropping unit intercepts the tumor area from the 3D image according to the center position of the tumor. Studies have shown that the peritumoral is also meaningful, and the cubic tumor area expanded by 5 mm and the cubic tumor area expanded by 10 mm are also successively intercepted to obtain the multi-scale three-dimensional image.
[54] The CT images are segmented according to different scales, and the cubic tumor area, the cubic tumor area expanded by 5 mm (including the cubic tumor area) and the cubic tumor area expanded by 10 mm (including the cubic tumor area) are successively intercepted, to obtain the three-dimensional image of three different scales.
[55] The preprocessing module 103 is used for preprocessing the tumor image, the tumor image expanded by 5 mm and the tumor image expanded by 10 mm, respectively, to obtain a training set.
[56] The preprocessing module 103 specifically includes a normalization unit connected to the multi-scale segmentation module 102, a data scaling unit connected to the normalization unit. and a data augmentation unit connected to the data scaling unit.
[57] The normalization unit is used for normalizing each voxel in the tumor image, the tumor image expanded by 5 mm and the tumor image expanded by 10 mm, respectively, to obtain normalized tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm.
The normalization unit (standardization unit) normalizes each voxel to [0,1] according to a formula
Ny = €; — BVO quch that all images can be scaled to a uniform size for network learning. In the formula, ¥; represents an unstandardized CT value of the i-th voxel, g and 4 represent a mean and a standard deviation of the CT values of each voxel in an unstandardized first image block, respectively; and A; represents a standardized (normalized) CT value of the i-th voxel.
[58] The data scaling unit is used for unifying the normalized tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm to a set image size, respectively, to obtain image size-set tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm.
Specifically, the data scaling unit unifies images of different scales to an average size of all images of the scale, such that all images are scaled to a uniform size, which is convenient for network learning.
After that, the data set is separated into a training set and a testing set randomly with a ratio of 8:2, where the training set is used for model training, and the testing set is used for testing performance of the model.
[59] The data augmentation unit is used for conducting data augmentation on the image size-set tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm via flipping, rotation, translation and zooming, to obtain the training set. The training set includes data-enhanced tumor image, tumor image expanded by 5 mm and tumor image expanded by 10 mm. To improve generalization ability of the model, the data augmentation unit enhances the data samples of the training set via tlipping, rotation, translation and zooming to prevent over-fitting. The data augmentation is conducted on the training set, but not on the testing set.
[60] The preprocessing module 103 obtains a training set used for training and optimizing the deep residual network by conducting preprocessing operations such as standardization, data scaling, data augmentation and data set separation on images of different scales.
[61] The deep residual network training module (deep learning network training module) 104 is used for training and optimizing the deep residual network by using the training set to obtain an optimized deep residual network. The deep residual network training module 104 inputs the CT images of three scales into a constructed deep residual classification model to obtain the probability whether there is lymph node metastasis in each scale of lesion or not.
[62] The deep residual network training module 104 specifically includes a deep residual network construction unit connected to the preprocessing module 103, and a deep residual network training unit connected to the deep residual network construction unit.
[63] The deep residual network construction unit is used for constructing a deep residual network. FIG. 2 is a structural diagram of classification based on the deep residual network of the present disclosure. As shown in FIG. 2, the deep residual network is constructed to classify the thyroid CT images.
[64] The deep residual network specifically includes a shallow feature extraction layer connected to the preprocessing module 103, a deep feature extraction layer connected to the shallow feature extraction layer, a skip connection layer separately connected to the shallow feature extraction layer and the deep feature extraction layer, a convolutional layer connected to the skip connection layer, and a classification layer connected to the convolutional layer.
[65] The shallow feature extraction layer is used for using a 64-channel 3x3x3 convolution kernel and an ReLU connected to the 3x3x3 convolution kernel to extract shallow features of the images in the training set to obtain a shallow feature map. The 64-channel 3x3x3 convolution kernel conducts convolution operation, and the ReLU added later conducts nonlinear mapping. Since the shallow feature extraction layer uses only one convolution kernel, only the shallow features of the image are extracted.
[66] The deep feature extraction layer (deep characteristics extraction layer) is used for extracting deep features in the shallow feature map to obtain a deep feature map.
[67] The deep feature extraction layer specifically includes a plurality of RDBs connected with the shallow feature extraction layer, and a 1x1x1 convolutional layer connected with a plurality of the RDBs.
[68] The plurality of the RDBs are connected sequentially, and each RDB is used for extracting the deep features in the shallow feature map using nine 3x3x3 convolution kernels and an ReLU separately connected to each 3x3x3 convolution kernel. The deep features from the original image are fully extracted by using a plurality of the RDBs. A network structure of each RDB is shown in
FIG. 3, the RDB includes nine 3x3x3 convolution and a ReLU operation, each layer is closely connected to increase a receptive field inside each network layer, such that the network can fully learn features of each layer. S RDBs are provided, and an output of the s-th RDB can be obtained by
FE =L AF _}=L0L tit {F;}1)} Inthe formula, £, is the s-th RDB operation, which is equivalent to the convolution operation and Rel. U operation in a convolutional neural network, and
F. is the s-th RDB completely-generated by each convolutional layer inside the RDB. The previous layers can access to the following layer. The output of the i-th convolutional layer of the s-th RDB can be expressed by Fy; = max{d,w, ; » 12 Fem Fais) + B;} . In the formula, W, ; represents a weight of the i-th convolutional layer in the RDB, and IF arg, Fais) represents a feature map in the s-th RDB generated by the convolutional layer [1,2 {i — 13] of the s-1-th
RDB.
[69] The Ix1x1 convolutional layer is used for fusing the deep features extracted by each RDB to obtain the deep feature map. All RDB outputs are cascaded and input to one 1x1x1 convolutional layer, and the outputs are fused with features of each RDB to reduce the number and parameters of the feature map, where the feature maps are reduced to 64. Finally, an identity mapping of the residual network is introduced to improve the convergence speed of the network and improve the gradient of the information flow. A deeper network leads to an easier extraction of richer and deeper features. With the increase in the number of RDBs and convolutional layers, a better performance is easily to be achieved, and a high growth rate also improves the performance of the model, such that 16 RDB blocks can be provided.
[70] The skip connection layer is used for connecting the shallow feature map and deep feature map. To fuse the shallow features with the deep features, the shallow feature map is added to the output of all RDB-cascaded features by using the skip connection. In this way, all feature maps are connected to extract rich discriminative image features. The present disclosure extracts highly-discriminative deep features of CT images, classities the thyroid cancer CT images, and has great value for use.
[71] The convolutional layer is used for further extracting features from the connected shallow feature map and deep feature map with a 7x7x7 convolution kernel and an ReLU connected to the 7x7x7 convolution kernel, to generate a 128-channel feature map. The image features are further extracted through the 7x7x7 convolutional layer to generate the 128-channel feature map, and nonlinear mapping is conducted by using the ReLU.
[72] The classification layer is used for conducting a 3D global average pooling operation on the 128-channel feature map, calculating probability of the lymph node metastasis and the lymph node non-metastasis through the thyroid cancer CT images, and taking a category with the highest probability as a classification result. The classification layer specifically includes an FC layer and a
Softmax that are mutually connected. The FC layer is connected to the convolutional layer and is used for conducting the 3D global average pooling operation on the 128-channel feature map; and the Softmax is used for calculating the probability of the lymph node metastasis and the lymph node non-metastasis through the thyroid cancer CT images, and taking the category with the highest probability as the classification result. The extracted feature map (feature map) is subjected to a 3D global average pooling operation, the classification probability of the lymph node metastasis and the lymph node non-metastasis is obtained using the FC layer and the Softmax, and the category with the largest classification probability is used as a final classification result of metastasis to determine whether there is lymph node metastasis in thyroid cancer CT images.
[73] The residual network alleviates the problem of gradient vanishing and solves feature redundancy, but has poor connection between the features. The dense network can solve the above problem by receiving feature maps of all layers to enhance the propagation between the features.
Moreover, the RDB network perfectly integrates the advantages of the above two networks, thereby maximizing the mining of highly-discriminative deep features.
[74] The deep residual network training unit is used for receiving the training set sent by the preprocessing module 103, and training and optimizing the deep residual network by using the training set to obtain the optimized deep residual network. The deep residual network training unit inputs training sets of different scales into the deep residual network, and conducts model training through an Softmax activation function and network parameters in the deep residual network. The model training uses cross-entropy loss as a loss function, and Adam as an optimization algorithm for iterative solution. The network parameters are initialized by He initialization, with an iteration epoch set to 200 and a network batch size of 32. An initial learning rate is set to le-5, and the learning rate is reduced by 10% at 1/2 of the epoch and by 1% at 3/4 of the epoch. If the data category ratio is quite different, the data set can be subjected to category unbalance processing. For three-dimensional images, resampling can be conducted, that is, the number of randomly-extracted images in each batch is controlled during the training, such that the extracted two types are the same. Error fitting between the training result and a true value can be conducted by using the loss function to minimize the loss function. When the loss function gradually converges, a model corresponding to the lowest point is the best classification model; the optimal network parameters are selected to obtain the best classification model (three optimal network models). The training set of the corresponding scale is input into the corresponding network model, the optimal model is trained to obtain the prediction probability of each multi-scale network, that is, a prediction probability of images of three scales. After obtaining the prediction probability of images of three scales, multi-scale network weighted fusion is conducted, that is, the predicted probabilities of the obtained images of three scales are subjected to weighted fusion to obtain a final prediction probability whether there is lymph node metastasis through the thyroid cancer CT images.
Multi-scale network weighted fusion finds the weights by parameter search, and gives a weight to the output probability of each network. Finally, a weighted sum of the output probabilities of each network is a final fusion probability, as shown in a formula
Score=a*Modell+b*Model2+c*Model3. In the formula, a+b+c=1, and 1>a>0, 1>b>0 and 1>¢>0;
Modell, Model2 and Model3 are the predicted probabilities for prediction of tumor inside, tumor and peritumoral of 5 mm, and tumor and peritumoral of 10 mm in predicting lymph node metastasis, respectively. Score is a probability of a finally-fused lymph node being malignant.
Preferably, it is possible to search for the largest value of area under the curve (AUC) by traversing all parameter spaces from 0 to 1 at an interval of 0.01. The final prediction probability of lymph node metastasis is obtained through weighted fusion of the prediction results of the multi-scale network.
[75] The thyroid cancer CT image classification module 105 is used for inputting thyroid cancer CT images to be classified into the optimized deep residual network for classification, to obtain a classification result of the thyroid cancer CT images. The classification result includes the lymph node metastasis and the lymph node non-metastasis through the thyroid cancer CT images.
[76] The present disclosure provides a deep residual network-based classification system for thyroid cancer CT image that overcomes the shortcomings in the existing diagnosis by the thyroid cancer CT images. Prediction of the thyroid cancer CT images by computer improves the accuracy of prediction and assists radiologists” diagnosis. The present disclosure proposes a new deep learning network classification framework (a model based on deep learning) using the deep learning to extract highly-discriminative deep features of the thyroid cancer CT images.
Therefore, the thyroid cancer CT images can be accurately classified to assist radiologists in lymph node metastasis prediction through the thyroid cancer CT images, thereby assisting radiologists to conduct automated analysis and diagnosis on the thyroid cancer CT images.
[77] Compared with the prior studies, the present disclosure has the following advantages:
[78] 1) The present disclosure proposes a new deep learning-based image classification technology; the present disclosure extracts the deep layered features of the thyroid cancer CT images using residual dense network, where each layer is closely connected such that the features and relationships between the layers can be fully learned; the introduction of skip connections solves the problem of gradient vanishing/gradient explosion, and global residual makes the shallow features and the deep features fully fused.
[79] 2) Compared with common deep learning networks such as ResNet and DenseNet, the present disclosure further strengthens the fusion and propagation of features, and fully learns the high-frequency and detailed features of the images.
[89] 3) The present disclosure extracts the multiple-scale information of thyroid cancer tumors in the thyroid cancer CT images, fuses the features of the tumor and peritumoral and improves the accuracy of model classification.
[81] Each example of the present specification is described in a progressive manner, each example focuses on the difference from other examples, and the same and similar parts between the examples may refer to each other.
[82] In this specification, several examples are used for illustration of the principles and implementations of the present disclosure. The description of the foregoing examples is used to help illustrate the method and the core principles of the present disclosure. In addition, those of ordinary skill in the art can make various modifications in terms of specific implementations and scope of application in accordance with the teachings of the present disclosure. In conclusion, the content of the present specification shall not be construed as a limitation to the present disclosure.
Claims (9)
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110812511.3A CN113537357A (en) | 2021-07-19 | 2021-07-19 | Thyroid cancer CT image classification system based on depth residual error network |
Publications (2)
Publication Number | Publication Date |
---|---|
NL2029876A NL2029876A (en) | 2023-01-23 |
NL2029876B1 true NL2029876B1 (en) | 2023-03-14 |
Family
ID=78128656
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
NL2029876A NL2029876B1 (en) | 2021-07-19 | 2021-11-23 | Deep residual network-based classification system for thyroid cancer computed tomography (ct) images |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN113537357A (en) |
NL (1) | NL2029876B1 (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114549413B (en) * | 2022-01-19 | 2023-02-03 | 华东师范大学 | Multi-scale fusion full convolution network lymph node metastasis detection method based on CT image |
CN116416239B (en) * | 2023-04-13 | 2024-03-12 | 中国人民解放军海军军医大学第一附属医院 | Pancreatic CT image classification method, image classification model, electronic equipment and medium |
CN116797879A (en) * | 2023-06-28 | 2023-09-22 | 脉得智能科技(无锡)有限公司 | Thyroid cancer metastasis lymph node prediction model construction method, system, equipment and medium |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107680678B (en) * | 2017-10-18 | 2020-12-01 | 北京航空航天大学 | Thyroid ultrasound image nodule diagnosis system based on multi-scale convolution neural network |
US11937973B2 (en) * | 2018-05-31 | 2024-03-26 | Mayo Foundation For Medical Education And Research | Systems and media for automatically diagnosing thyroid nodules |
-
2021
- 2021-07-19 CN CN202110812511.3A patent/CN113537357A/en active Pending
- 2021-11-23 NL NL2029876A patent/NL2029876B1/en active
Also Published As
Publication number | Publication date |
---|---|
NL2029876A (en) | 2023-01-23 |
CN113537357A (en) | 2021-10-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
NL2029876B1 (en) | Deep residual network-based classification system for thyroid cancer computed tomography (ct) images | |
Usman et al. | Volumetric lung nodule segmentation using adaptive roi with multi-view residual learning | |
Wang et al. | Automatic whole heart segmentation using deep learning and shape context | |
CN109523521B (en) | Pulmonary nodule classification and lesion positioning method and system based on multi-slice CT image | |
CN110309860B (en) | Method for classifying malignancy degree of lung nodule based on convolutional neural network | |
WO2021203795A1 (en) | Pancreas ct automatic segmentation method based on saliency dense connection expansion convolutional network | |
Pan et al. | Cell detection in pathology and microscopy images with multi-scale fully convolutional neural networks | |
CN114565761B (en) | Deep learning-based method for segmenting tumor region of renal clear cell carcinoma pathological image | |
CN111798424B (en) | Medical image-based nodule detection method and device and electronic equipment | |
Cao et al. | An automatic breast cancer grading method in histopathological images based on pixel-, object-, and semantic-level features | |
CN113261012B (en) | Method, device and system for processing image | |
Tripathi et al. | HematoNet: Expert level classification of bone marrow cytology morphology in hematological malignancy with deep learning | |
CN112686899B (en) | Medical image analysis method and apparatus, computer device, and storage medium | |
Manikandan et al. | Segmentation and Detection of Pneumothorax using Deep Learning | |
Pham et al. | Chest x-rays abnormalities localization and classification using an ensemble framework of deep convolutional neural networks | |
Li et al. | Structure convolutional extreme learning machine and case-based shape template for HCC nucleus segmentation | |
Abdulaal et al. | A self-learning deep neural network for classification of breast histopathological images | |
Zhang et al. | Lesion detection of computed tomography and magnetic resonance imaging image based on fully convolutional networks | |
CN115809988A (en) | Survival analysis method and system for brain tumor patient | |
Mir et al. | Artificial intelligence-based techniques for analysis of body cavity fluids: a review | |
Tripathi et al. | An Object Aware Hybrid U-Net for Breast Tumour Annotation | |
Yang et al. | Leveraging auxiliary information from emr for weakly supervised pulmonary nodule detection | |
Qiu et al. | MMA-Net: Multiple Morphology-Aware Network for Automated Cobb Angle Measurement | |
CN112233106A (en) | Thyroid cancer ultrasonic image analysis method based on residual capsule network | |
Saturi et al. | Modelling of deep learning enabled lung disease detection and classification on chest X-ray images |