Thyroid cancer lung metastatic focus iodine intake state prediction method and system based on deep learning

Abstract

The invention discloses a thyroid cancer lung metastasis focus iodine taking state prediction method and system based on deep learning, is used for identifying the iodine taking state of a lung metastasis focus of a thyroid cancer patient in a medical scene, and relates to the field of image processing, medical treatment and deep learning, and the method comprises the following steps: 1) obtaining lung CT image data of the patient, converting the data into a three-dimensional file, and labeling the focus of part of the patient; 2) preprocessing lung CT image data, and dividing the lung CT image data into a segmentation data set and a classification data set; (3) constructing a depth image segmentation model; (4) constructing a depth image classification model; (5) using the segmentation model to obtain the position and size of a focus of an unlabeled patient; and (6) classifying the iodine intake state of the focus obtained in the step (5) by using a classification model to obtain a classification result. According to the method, the focus iodine intake state of the thyroid cancer lung metastasis patient is identified based on deep learning, so that unnecessary radioactive iodine treatment is prevented, an effective treatment scheme is formulated, and finally the personalized treatment scheme and patient prognosis are improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing and relates to a method and system for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning. Background Art

[0002] Medical imaging provides doctors with information about diseased tissues and anatomical views of special organs, and is a key foundation in the process of medical diagnosis and treatment. With the rapid development and popularization of medical imaging equipment, various medical imaging methods have emerged in an endless stream. Compared with other imaging technologies, MRI can provide high-resolution images more accurately without causing any harm to the human body, especially in the detection of human soft tissues and tumor sites. CT images are the best medical imaging method for diagnosing lung metastases of thyroid cancer.

[0003] With the development of deep learning algorithms and corresponding hardware devices, in 2014, Long et al. replaced all fully connected layers in the CNN structure with fully convolutional layers, and proposed a fully convolutional neural network FCN that can achieve the purpose of image semantic segmentation, proving that deep neural networks can perform semantic segmentation in an end-to-end manner on images of non-fixed size. In 2015, Ronneberger proposed the Vnet model, which used an encoding-decoding architecture. The encoding process extracted image features and the decoding process restored the image, achieving the best performance in the segmentation competition at the time. In 2015, He proposed the ResNet model, which used residual connections to solve the problem of poor performance of deep neural networks and made deep neural networks easy to train. In 2016, FaVsto Milletari proposed the Vnet model, which improved the Vnet so that it could train 3D images, proposed Diceloss as the loss function, and added residual connections.

[0004] The present invention provides a VNet segmentation network and a Resnet classification network with a branched convolution module and a compression excitation module, which accelerates the convergence speed of the original network and obtains higher performance. In the inference stage, the branched convolution module can obtain a larger field of view, greatly improving the inference speed and efficiency. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and propose a method and system for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning. Through a deep convolutional network that associates global multi-scale upper and lower feature information, the feature dependency of the channel level at each scale is judged, the image feature information is effectively integrated and the influence of redundant features is reduced, and finally the network's ability to segment and classify targets is improved.

[0006] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided, comprising the following steps:

[0007] Step 1: Obtain the patient's lung CT image data, convert it into a three-dimensional file, and label the lesions of some patients;

[0008] Step 2: preprocess the lung CT image data into a segmentation data set and a classification data set;

[0009] Step 3: Build a deep image segmentation model, use the training set in the segmentation data set to train the segmentation model, and use the test set to detect the DICE coefficient of the segmentation model;

[0010] Step 4: Build a deep image classification model, use the training set classification model in the classification data set for training, and use the test set to detect the accuracy of the classification model;

[0011] Step 5: Use the lesion segmentation model in step 3 above to obtain the lesion location and size of the unlabeled patient.

[0012] Step 6: Use the classification model in the above step 4 to classify the iodine uptake status of the lesion obtained in the above step 5 to obtain a classification result.

[0013] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided. The three-dimensional conversion and preprocessing process of the lung CT image data in the above step 1 and the above step 2 is as follows:

[0014] 1) Convert the slice DICOM file of lung CT images into a three-dimensional NIFTI file;

[0015] 4) Resample the image to voxel [1,1,1] and normalize the value to the interval (0,1);

[0016] 5) Read the original image and remove the redundant background pixels;

[0017] 4) Perform various operations on the training data, such as random noise, multi-directional flipping, random cropping, and other data enhancement operations that do not destroy spatial data;

[0018] 5) The final image size fed into the segmentation neural network model is 32*128*128;

[0019] 6) The data size of the classification neural network is determined according to the location and size of the patient's lesions, and is finally normalized to 32*48*48.

[0020] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided, and the obtained data set is divided into training set and test set in a ratio of 7:3.

[0021] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided. Step 3 constructs a deep image segmentation model.

[0022] The constructed deep image segmentation model is an improved Vnet model, which includes a downsampling encoder and a corresponding upsampling decoder. Each layer of the decoder obtains the features of the corresponding layer of the encoder through jump connections. The workflow of the neural network is as follows:

[0023] 1) The preprocessed 3D lung CT image dataset is first input into a convolution module containing 4 groups of downsampling, and 2 3*3*3 convolution operations are performed;

[0024] 2) The feature map after the convolution operation is input into the compression excitation module to perform compression excitation operation and encode and filter the feature map channel level;

[0025] 3) The feature map filtered by the output code is accumulated with the original input feature map of this layer to complete the residual operation, and down-sampled using a 3*3*3 convolution with a step size of 2;

[0026] 4) The feature maps after 4 downsampling encoding are input into the upsampling decoder composed of 4 global context aggregation modules for decoding;

[0027] 5) Use the deconvolution module to upsample the image and concatenate it with the corresponding downsampled encoding feature map at the channel layer;

[0028] 6) Perform 2 3*3*3 convolution operations and input the encoded feature map channel level in the compression excitation module;

[0029] 7) The feature map filtered by the output code is accumulated with the original input feature map of this layer to complete the residual operation;

[0030] 8) Finally output the image segmentation result.

[0031] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided, a deep image segmentation model, and a deconvolution module is specifically:

[0032] MixLayer(X)=Concat(DowmSampling(X'),Deconv(X))

[0033] Where X' is the input feature map of the corresponding downsampling encoder, X is the input feature map, DowmSampling is the downsampling operation, Deconv is the deconvolution operation, and Concat is the channel merging operation;

[0034] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided. Step 4 constructs a deep image classification model:

[0035] The constructed deep image segmentation model is an improved Resnet model, which includes multiple downsampling encoders and a fully connected output layer. The workflow of the neural network is as follows:

[0036] 1) The preprocessed 3D lung CT image dataset is first input into a convolution module containing 4 groups of downsampling, and 2 3*3*3 convolution operations are performed;

[0037] 2) The feature map after the convolution operation is input into the compression excitation module to perform compression excitation operation and encode and filter the feature map channel level;

[0038] 3) The feature map filtered by the output code is accumulated with the original input feature map of this layer to complete the residual operation, and pooling is performed using a 3*3*3 convolution with a step size of 2;

[0039] 4) Complete the feature map of 4 downsampling encodings, perform global pooling and send it to the fully connected layer for prediction to obtain the final classification result.

[0040] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided. The fully connected layer in the deep image classification model is specifically:

[0041] FC(X)=Sigmoid(MLP(AvgPooling(X)))

[0042] Among them, MLP is a multi-layer perceptron module, AvgPooling is a global pooling, X is an input feature map, and Sigmoid is an activation function.

[0043] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided. The deep image classification model of the sixth aspect and the deep image segmentation model of the fourth aspect, the convolutional layer is specifically:

[0044] Conv(X) = Norm(ReLU(Conv 3×3×3 (X)))

[0045] Among them, Norm is the normalization layer, Conv 3×3×3It is a 3*3*3 convolution operation, X is the input feature map, and ReLV is the activation function.

[0046] According to one aspect of the present invention, a method for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided. The deep image classification model of the sixth aspect and the deep image segmentation model of the fourth aspect, the compression excitation module are specifically:

[0047] SE(X)=X+MLP(AvgPooling(X))

[0048] Among them, MLP is a multi-layer perceptron module, AvgPooling is a global pooling, and X is an input feature map.

[0049] According to one aspect of the present invention, a method system for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning is provided, the system comprising a computer-readable storage medium, a computer program is stored on the computer-readable storage medium, and when the computer program is executed in a computer, the computer executes any one of the above methods.

[0050] In addition to the above solutions, the present invention also has some other technical solutions:

[0051] Step S1: Preprocessing of patient lung CT images

[0052] The DICOM file of the patient's lung CT is obtained and converted into a three-dimensional image in NIFTI format. The NIFTI format image is resampled and the region of interest (ROI) image is intercepted from the resampled image according to the lesion area.

[0053] Step S2: Build a data set

[0054] The ROI image obtained by preprocessing in the above step S1 is subjected to data enhancement processing to generate a data set for model training.

[0055] Data enhancement processing includes random noise, multi-directional flipping, random cropping and other data enhancement operations that do not destroy spatial data.

[0056] Step S3: Segmentation network construction and training

[0057] The segmentation network consists of 4 groups of downsampling convolutional modules and 4 groups of upsampling modules to form its encoding and decoding parts, and the encoding and decoding parts use a skip connection method;

[0058] Each group of down-sampling encoding modules includes two convolution modules with a convolution kernel size of 3 and a compression excitation module; the shallow encoding module has a small receptive field to obtain local information of the image, and the deep encoding module captures global information;

[0059] Each group of decoding modules obtains the corresponding encoding feature map through skip connection, merges the feature map with the encoding feature map through deconvolution operation, and then obtains the information through two convolution modules with a convolution kernel size of 3 and a compression excitation module;

[0060] The downsampling operation uses a 3*3*3 convolution with a stride of 2 for pooling, and a 2*2*2 deconvolution module with a stride of 2 for upsampling;

[0061] After completing the convolution operation, each group of encoding or decoding modules needs to perform a compression excitation operation. First, complete the global pooling to obtain a feature vector with a length equal to the number of channels. After the first MLP layer, it is compressed to 1 / 4 of the number of channels to complete the compression operation. After the second MLP layer, the number of channels is restored to the original size to complete the excitation operation, and a weight vector with a value range of 0 to 1 is obtained. Finally, the vector is multiplied with the original feature map. The channels with low weights reduce the feature expression, and the channels with high weights increase the feature expression, thereby completing the screening of feature channels.

[0062] The feature map filtered by the output encoding is accumulated with the original input feature map of this layer to complete the residual operation. Compared with the original VNet network architecture, the actual speed and performance are improved without adding a large number of parameters.

[0063] Step S4: Obtaining classification data set

[0064] After the labeled image is processed in step S1, the region of interest (ROI) image is directly intercepted. Based on the segmentation network of step S3, the unlabeled image obtains the mask image of each nodule, obtains the ROI image through the edge detection method, detects and deletes the image with too few pixels, and then obtains the nodule ROI image.

[0065] Step S6: Classification network construction and training

[0066] The classification network contains 4 groups of down-sampling convolution modules. Each group of down-sampling encoding modules includes two convolution modules with a convolution kernel size of 3 and a compression excitation module. The shallow encoding module has a small receptive field to obtain local information of the image, while the deep encoding module captures global information.

[0067] After completing the 4 groups of downsampling, global pooling is performed and then sent to the fully connected layer for prediction to obtain the final classification result; compared with the original architecture of the Resnet network, the actual speed and performance have been improved without adding a large number of parameters.

[0068] The specific convolution module in the classification segmentation network is:

[0069] Conv(X) = Norm(ReLU(Conv 3×3×3 (X)))

[0070] Among them, Norm is the normalization layer, Conv 3×3×3 It is a 3*3*3 convolution operation, X is the input feature map, and ReLV is the activation function;

[0071] The specific structure of the compression excitation module in the classification and segmentation network is:

[0072] SE(X)=X+MLP(AvgPooling(X))

[0073] Among them, MLP is a multi-layer perceptron module, AvgPooling is a global pooling, and X is an input feature map.

[0074] The specific upsampling module in the segmentation network is:

[0075] MixLayer(X)=Concat(DowmSampling(X'),Deconv(X))

[0076] Where X' is the input feature map of the corresponding downsampling encoder, X is the input feature map, DowmSampling is the downsampling operation, Deconv is the deconvolution operation, and Concat is the channel merging operation;

[0077] Specifically, the fully connected module in the classification network is as follows:

[0078] FC(X)=Sigmoid(MLP(AvgPooling(X)))

[0079] Among them, MLP is a multi-layer perceptron module, AvgPooling is a global pooling, X is an input feature map, and Sigmoid is an activation function.

[0080] The present invention provides a system for predicting the iodine uptake status of thyroid cancer lung metastases, comprising:

[0081] Data acquisition and preprocessing module, used for preprocessing CT images of thyroid cancer lung metastases and obtaining images of regions of interest;

[0082] The segmentation module uses the improved VNet network to predict the thyroid cancer lung metastasis area and obtain the image segmentation result. It is used for the patient's lung CT images without labels.

[0083] The classification module uses the improved Resnet network to predict the lung metastases of thyroid cancer and obtain the iodine uptake classification results of the image.

[0084] The data preprocessing module of the present invention processes the DICOM format data of CT, obtains NIFTI format images for resampling and data enhancement operations, which are used for network training to improve the effects of image segmentation and classification. The compression excitation module is introduced into the VNet and Resnet networks to improve the speed and accuracy of the algorithm application, solve the problem of information not being shared between channels, aggregate the features between channels, and further improve the accuracy of the network.

[0085] In the same batch of test sets, the comparison between the improved V-Net and the original V-Net is as follows:

[0086] Model DICE coefficient V-Net 0.793 Compression-Inspired V-Net 0.821

[0087] In the same batch of test sets, the comparison between improved ResNet and traditional imaging omics is as follows:

[0088] Model ACC AUC SENS SPEC PPV NPV KNN 69.76％ 0.733 0.59779 0.77616 0.67782 0.71011 Logistic Regression 72.03％ 0.759 0.65683 0.77035 0.69261 0.74022 Random Forest 69.59％ 0.756 0.67528 0.71221 0.64894 0.73574 SVM 70.24％ 0.749 0.64207 0.75000 0.66923 0.72676 Compressed Excitation ResNet 78.86％ 0.879 0.76753 0.80523 0.75636 0.81471

[0089] Under the external validation set, the improved ResNet results are as follows:

[0090] Model ACC AUC SENS SPEC PPV NPV Compressed Excitation ResNet 69.11％ 0.713 0.72289 0.625 0.8 0.52083

[0091] The present invention solves the shortcomings of the original network such as slow convergence speed and low precision without adding a large number of parameters.

[0092] The method and system for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning are used to identify the iodine uptake status of thyroid cancer lung metastases in medical scenarios, and involve the fields of image processing, medical treatment, and deep learning. The method and system of the present invention prevent unnecessary radioactive iodine treatment and formulate effective treatment plans, ultimately improving personalized treatment plans and patient prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 is a flow chart of the method of the present invention;

[0094] Figure 2 The main structural diagram of the improved VNet network proposed by the present invention;

[0095] Figure 3 This is the main structure diagram of the improved Resnet network point proposed by the present invention;

[0096] Figure 4 This is the main structural diagram of the compression excitation module in the segmentation and classification model proposed in the present invention. DETAILED DESCRIPTION

[0097] The following embodiments are intended to further illustrate some preferred embodiments of the present invention, but are not all embodiments. Other embodiments based on the present invention made by professionals in the field without creative work all belong to the scope of protection of the present invention. The present invention will be further described below in conjunction with the accompanying drawings.

[0098] like Figure 1 As shown, the present invention provides a method and system for predicting the iodine uptake status of thyroid cancer lung metastases based on deep learning, and the flowchart includes the following steps:

[0099] Step 1: Obtain the patient's lung CT image data, convert it into a three-dimensional file, and label the lesions of some patients;

[0100] Step 2: preprocessing the sample lung CT image data into a segmentation data set and a classification data set;

[0101] Step 3: Build a deep image segmentation model, use the training set in the segmentation data set to train the segmentation model, and use the test set to detect the DICE coefficient of the segmentation model;

[0102] Step 4: Build a deep image classification model, use the training set classification model in the classification data set for training, and use the test set to detect the accuracy of the classification model;

[0103] Step 5: Use the lesion segmentation model in step 3 to obtain the lesion location and size of the unlabeled patient.

[0104] Step 6: Use the classification model in step 4 to classify the iodine uptake status of the lesion obtained in step 5 to obtain a classification result.

[0105] Specifically, the lung CT images in step 1 and step 2 are converted from slice DICOM files to three-dimensional NIFTI files; the images are resampled to voxels [1,1,1], and the values ​​are normalized to the interval (0,1), and redundant background pixels are removed;

[0106] Specifically, the segmentation data set and classification data set are separated in step 2, and the ratio of the training set to the test set is 7:3. The image size fed into the segmentation neural network model is 32*128*128; the data size of the classification neural network is determined according to the location and size of the patient's lesion, and is finally normalized to 32*48*48.

[0107] Specifically, step 2 performs diversified operations on the training data, such as random noise, multi-directional flipping, random cropping and other data enhancement operations that do not destroy the spatial data;

[0108] Specifically, the segmentation network diagram of step 3 is as follows Figure 2As shown in the figure, it belongs to the improved VNet model, and the network structure includes an encoding part, a decoding part and a jump connection.

[0109] Specifically, the image in step 3 is encoded four times after entering the network. Each encoding process includes two branch convolution modules and a compression excitation force module in sequence. The feature map filtered by the output channel is accumulated with the original encoded input feature map of this layer to complete the residual operation. Downsampling is performed using a 3*3*3 convolution with a step size of 2. Each downsampling reduces the image space to 1 / 2 and increases the number of channels to 2 times, so that the receptive field of the next encoding is increased.

[0110] Specifically, each encoded point image in step 3 is sent to the decoding module through a skip connection before downsampling;

[0111] Specifically, the image encoded in step 3 will be decoded four times. Each decoding process includes an upsampling module, two branch convolution modules, and a compression excitation module. The upsampling module performs a channel layer splicing operation on the previous layer decoding output and the corresponding encoding output, and then performs convolution decoding and compression excitation screening channels. The feature map after channel selection is accumulated with the decoded feature map of the original input of this layer to complete the residual operation;

[0112] Specifically, the upsampling module of each decoding process in step 3 is specifically,

[0113] MixLayer(X)=Concat(DowmSampling(X'),Deconv(X))

[0114] Where X' is the input feature map of the corresponding downsampling encoder, X is the input feature map, DowmSampling is the downsampling operation, Deconv is the deconvolution operation, and Concat is the channel merging operation;

[0115] Specifically, the classification network diagram of step 4 is as follows Figure 3 As shown in the figure, it belongs to the improved Resnet model, and the network structure includes 4 downsampling convolutional encoding and fully connected layer output;

[0116] Specifically, each encoding process of the classification network in step 4 includes two branch convolution modules and a compression excitation force module in sequence, and the feature map filtered by the output channel is accumulated with the original encoded input feature map of this layer to complete the residual operation;

[0117] Specifically, the convolutional feature map of the classification network in step 4 needs to be globally pooled before being input into the fully connected layer to form a feature vector for prediction in the fully connected layer to obtain the final classification result. The specific fully connected module is as follows:

[0118] FC(X)=Sigmoid(MLP(AvgPooling(X)))

[0119] Where MLP is a multi-layer perceptron module, AvgPooling is global pooling, X is the input feature map, and Sigmoid is the activation function;

[0120] Specifically, the compression excitation module of step 3 and step 4 is as follows Figure 4 As shown, specifically,

[0121] SE(X)=X+MLP(AvgPooling(X))

[0122] Among them, MLP is a multi-layer perceptron module, AvgPooling is a global pooling, and X is an input feature map.

[0123] Specifically, in step 3, the network sets the model training parameters: the training BatchSize is set to 16, the AdamW optimizer is used, and the learning rate is 0.001. The loss function is set to the DICE loss function.

[0124] Specifically, in step 4, the network sets the model training parameters: the training BatchSize is set to 48, the AdamW optimizer is used, and the learning rate is 0.001. The loss function is set to the cross entropy loss function.

[0125] Specifically, steps 5 and 6 are verified using external data sets. The segmentation network in step 3 is used to segment the unlabeled external data to obtain the mask images of each nodule. The ROI image is obtained by edge detection. The image with too few pixels is detected and deleted, and then the nodule ROI image is obtained. According to the nodule ROI image, the classification network in step 4 is used to classify the iodine uptake of the lesion, and the results are verified by professional doctors.

[0126] As can be seen from the above, the present invention relates to the field of medical image segmentation technology, and predicts the iodine uptake status of thyroid cancer lung metastases based on a deep learning method. In order to solve the shortcomings of slow model training convergence speed, low accuracy and reasoning process speed, a method and system for predicting the iodine uptake status of thyroid cancer lung metastases based on an improved UNet network and Resnet network with a compressed excitation module are proposed.

[0127] This method is based on deep learning to identify the iodine uptake status of lesions in patients with thyroid cancer lung metastasis, so as to prevent unnecessary radioactive iodine therapy and develop effective treatment plans, ultimately improving personalized treatment plans and patient prognosis.

[0128] The above embodiments are intended to further illustrate some preferred embodiments of the present invention, but not all embodiments. Other embodiments based on the present invention made by professionals in this field without creative work are all within the scope of protection of the present invention.

Claims

1. A method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning, characterized in that: The following steps are involved: Step 1: Obtain the patient's lung CT image data, convert it into a three-dimensional file, and label the lesions of some patients; Step 2: preprocess the lung CT image data into a segmentation data set and a classification data set; Step 3: Build a deep image segmentation model, use the training set in the segmentation data set to train the segmentation model, and use the test set to detect the DICE coefficient of the segmentation model; Step 4: Build a deep image classification model, use the training set classification model in the classification data set for training, and use the test set to detect the accuracy of the classification model; Step 5: Use the lesion segmentation model in step 3 to obtain the lesion location and size of the unlabeled patient; Step 6: Use the classification model in step 4 to classify the iodine uptake status of the lesion obtained in step 5 to obtain a classification result.

2. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 1, characterized in that: The three-dimensional conversion and preprocessing process of the lung CT image data in step 1 and step 2 is as follows: 1) Convert the slice DICOM file of lung CT images into a three-dimensional NIFTI file; 2) Resample the image to voxel [1,1,1] and normalize the value to the interval (0,1); 3) Read the original image and remove the redundant background pixels; 4) Perform various operations on the training data, such as random noise, multi-directional flipping, random cropping, and other data enhancement operations that do not destroy spatial data; 5) The final image size fed into the segmentation neural network model is 32*128*128; 6) The data size of the classification neural network is determined according to the location and size of the patient's lesions, and is finally normalized to 32*48*48.

3. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 1, characterized in that: The ratio of the obtained dataset divided into training set and test set jumped to 7:

3.

4. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 1, characterized in that: The step 3 constructs a deep image segmentation model. The constructed deep image segmentation model is an improved Vnet model, which includes a downsampling encoder and a corresponding upsampling decoder. Each layer of the decoder obtains the features of the corresponding layer of the encoder through jump connections. The workflow of the neural network is as follows: 1) The preprocessed 3D lung CT image dataset is first input into a convolution module containing 4 groups of downsampling, and 2 3*3*3 convolution operations are performed; 2) The feature map after the convolution operation is input into the compression excitation module to perform compression excitation operation and encode and filter the feature map channel level; 3) The feature map filtered by the output code is accumulated with the original input feature map of this layer to complete the residual operation, and down-sampled using a 3*3*3 convolution with a step size of 2; 4) The feature maps after 4 downsampling encoding are input into the upsampling decoder composed of 4 global context aggregation modules for decoding; 5) Use the deconvolution module to upsample the image and concatenate it with the corresponding downsampled encoding feature map at the channel layer; 6) Perform 2 3*3*3 convolution operations and input the encoded feature map channel level in the compression excitation module; 7) The feature map filtered by the output code is accumulated with the original input feature map of this layer to complete the residual operation; 8) Finally output the image segmentation result.

5. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 4, characterized in that: The deep image segmentation model, the deconvolution module is specifically: MixLayer(X)=Concat(DowmSampling(X'),Deconv(X)) Where X' is the input feature map of the corresponding downsampling encoder, X is the input feature map, DowmSampling is the downsampling operation, Deconv is the deconvolution operation, and Concat is the channel merging operation.

6. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 5, characterized in that: Step 4: Build a deep image classification model The constructed deep image segmentation model is an improved Resnet model, which includes multiple downsampling encoders and a fully connected output layer. The workflow of the neural network is as follows: 1) The preprocessed 3D lung CT image dataset is first input into a convolution module containing 4 groups of downsampling, and 2 3*3*3 convolution operations are performed; 2) The feature map after the convolution operation is input into the compression excitation module to perform compression excitation operation and encode and filter the feature map channel level; 3) The feature map filtered by the output code is accumulated with the original input feature map of this layer to complete the residual operation, and pooling is performed using a 3*3*3 convolution with a step size of 2; 4) Complete the feature map of 4 downsampling encodings, perform global pooling and send it to the fully connected layer for prediction to obtain the final classification result.

7. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 6, characterized in that: The fully connected layer in the deep image classification model is specifically: FC(X)=Sigmoid(MLP(AvgPooling(X))) Among them, MLP is a multi-layer perceptron module, AvgPooling is a global pooling, X is an input feature map, and Sigmoid is an activation function.

8. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 6, characterized in that: The deep image classification model described in claim 6 and the deep image segmentation model described in claim 4, wherein the convolutional layer is specifically: Conv(X)=Norm(ReLU(Conv) 3×3×3 (X))) Among them, Norm is the normalization layer, Conv 3×3×3 It is a 3*3*3 convolution operation, X is the input feature map, and ReLV is the activation function.

9. The method for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning according to claim 6, characterized in that: The depth image classification model described in claim 6 and the depth image segmentation model described in claim 4, wherein the compression excitation module is specifically: SE(X)=X+MLP(AvgPooling(X)) Among them, MLP is a multi-layer perceptron module, AvgPooling is a global pooling, and X is an input feature map.

10. A method system for predicting iodine uptake status of thyroid cancer lung metastases based on deep learning, characterized in that: The system comprises a computer-readable storage medium, on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method described in any one of claims 1 to 9.

Images