Thyroid nodule recognition system based on self-distillation convolutional neural network

Abstract

The invention discloses a thyroid nodule recognition system based on a self-distillation convolutional neural network, belongs to the technical field of intelligent medical detection, and particularly relates to a thyroid nodule recognition system. The objective of the invention is to solve the problem of low thyroid nodule recognition accuracy based on deep learning in the prior art. The system comprises an image processing main module used for collecting a thyroid nodule ultrasonic image and preprocessing the collected thyroid nodule ultrasonic image to obtain a preprocessed thyroid nodule ultrasonic image; the neural network main module is used for building a self-distillation convolutional neural network model; the training main module uses the preprocessed thyroid nodule ultrasound image to train a built self-distillation convolutional neural network model to obtain a trained self-distillation convolutional neural network model; and the identification main module is used for loading the trained self-distillation convolutional neural network model, detecting a to-be-identified thyroid nodule ultrasonic image and identifying the category of the thyroid nodule.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent medical detection, and in particular relates to a thyroid nodule recognition system. Background Art

[0002] As one of the most common nodular lesions and endocrine cancers, thyroid nodules are frequently diagnosed in people of different ages. Although most thyroid nodules are benign (non-cancerous), some of them contain malignant thyroid nodules that require early diagnosis and treatment. The incidence of thyroid cancer has increased dramatically in the past 20 years, the largest increase among all cancer types. Early diagnosis of thyroid nodules is crucial to reducing the morbidity and mortality of thyroid cancer. Ultrasound technology, as a rapid, low-cost, and non-invasive imaging method, has become a common method for detecting and diagnosing thyroid nodules. In clinical examinations, radiologists usually assess the malignant risk of thyroid nodules based on ultrasound features such as echogenicity, shape, and margins, and recommend patients for aspiration biopsy when the estimated risk of malignancy is high. However, clinical evaluation of nodules based on ultrasound images of thyroid nodules is challenging for radiologists, with a positive predictive value of only 36%. On the one hand, ultrasound images are usually very noisy due to the influence of speckle noise and echo disturbance, which may cause non-related areas to reflect abnormal echoes, causing confusion in diagnosis; on the other hand, the characteristic differences of nodules in ultrasound images are subtle, and the interpretation of these differences by radiologists is subjective and heavily depends on the experience and cognitive ability of the diagnosticians. Developing an automatic thyroid nodule recognition model for ultrasound images plays an important role.

[0003] Computer-aided identification of thyroid nodules based on deep learning helps make diagnosis more reliable and objective. Although there are computer-aided diagnosis methods for thyroid nodule identification, these methods still have limitations: 1) The two-stage method of first locating nodules and then classifying them is directly limited by the segmentation or detection results in the first stage, and the imperfect first-stage results have a greater impact on the classification task in the second stage; 2) The end-to-end method does not fully utilize feature information during the forward propagation of features, resulting in a weak ability to learn the subtle and diverse features of thyroid nodules in ultrasound images. Therefore, in order to improve the accuracy of thyroid nodule identification, designing an effective thyroid nodule recognition system has important clinical application value. Summary of the invention

[0004] The purpose of the present invention is to solve the problem that the existing thyroid nodule recognition methods based on deep learning do not fully utilize the features, making it difficult to fully mine the discriminative feature information, resulting in low accuracy in thyroid nodule recognition. A thyroid nodule recognition system based on self-distillation convolutional neural network is proposed.

[0005] The thyroid nodule recognition system based on self-distilled convolutional neural network includes:

[0006] Image processing main module, neural network main module, training main module and recognition main module;

[0007] The image processing main module is used to collect thyroid nodule ultrasonic images, preprocess the collected thyroid nodule ultrasonic images, and obtain the preprocessed thyroid nodule ultrasonic images;

[0008] The neural network main module is used to build a self-distillation convolutional neural network model; the specific process is:

[0009] Step B1, constructing a self-distillation convolutional neural network model;

[0010] Step B2, designing a self-distillation convolutional neural network model loss function;

[0011] The main training module uses the preprocessed thyroid nodule ultrasound image to train the built self-distillation convolutional neural network model to obtain a trained self-distillation convolutional neural network model;

[0012] The main recognition module is used to load the trained self-distillation convolutional neural network model, detect the ultrasound image of the thyroid nodule to be identified, and identify the category of the thyroid nodule.

[0013] The beneficial effects of the present invention are:

[0014] 1. This invention proposes for the first time a thyroid nodule recognition system based on a self-distillation convolutional neural network. The system mines the distinctive feature information and rich contextual information of thyroid nodules in features at different levels to improve the reliability of the recognition system.

[0015] 2. An adjacent feature interactive learning module and self-distillation loss are designed for the proposed thyroid nodule recognition system based on self-distillation convolutional neural network. The adjacent feature interactive learning module flexibly performs multi-level feature aggregation through interactive learning to enhance the discriminative ability of the extracted features. The self-distillation loss guides the network to fully learn high-level features and aggregated adjacent features.

[0016] 3. Experiments show that the present invention has good performance. The self-distillation convolutional neural network can achieve an accuracy of 82.42%, a precision of 81.46%, and a recall of 86.90% for identifying thyroid nodules in ultrasound images, which is at an advanced level. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the system of the present invention;

[0018] Figure 2This is a schematic diagram of the self-distillation convolutional neural network architecture;

[0019] Figure 3 Schematic diagram of the adjacent feature interaction module. DETAILED DESCRIPTION

[0020] Specific implementation method 1: The thyroid nodule recognition system based on the self-distillation convolutional neural network in this implementation method includes:

[0021] Image processing main module, neural network main module, training main module and recognition main module;

[0022] The image processing main module is used to collect thyroid nodule ultrasonic images, preprocess the collected thyroid nodule ultrasonic images, and obtain the preprocessed thyroid nodule ultrasonic images;

[0023] The neural network main module is used to build a self-distillation convolutional neural network model; the specific process is:

[0024] Step B1, constructing a self-distillation convolutional neural network model;

[0025] Step B2, designing a self-distillation convolutional neural network model loss function;

[0026] The main training module uses the preprocessed thyroid nodule ultrasound image to train the built self-distillation convolutional neural network model to obtain a trained self-distillation convolutional neural network model;

[0027] The main recognition module is used to load the trained self-distillation convolutional neural network model, detect the ultrasound image of the thyroid nodule to be identified, and identify the category of the thyroid nodule (whether it is a benign thyroid nodule or a malignant nodule).

[0028] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that: the image processing main module is used to collect thyroid nodule ultrasound images, pre-process the collected thyroid nodule ultrasound images, and obtain the pre-processed thyroid nodule ultrasound images; the specific process is:

[0029] Step A1, obtaining ultrasound images of thyroid nodules, filling each image into a square, and adjusting each square-filled image to a size of 512×512;

[0030] Step A2: Normalize the image adjusted to 512×512 to obtain the normalized image I norm ;

[0031] For the normalized image I norm Data enhancement operations were performed to obtain preprocessed thyroid nodule ultrasound images as training sets.

[0032] The preprocessed images are used as input to the self-distillation convolutional neural network model.

[0033] The other steps and parameters are the same as those in the first embodiment.

[0034] Specific implementation method three: This implementation method is different from specific implementation method one or two in that: in step A2, the image adjusted to a size of 512×512 is normalized to obtain a normalized image I norm ; The expression is:

[0035]

[0036] Among them, I represents the image resized to 512×512, I min and I max Represent the minimum and maximum pixel values ​​in the image resized to 512×512.

[0037] The other steps and parameters are the same as those in the first or second embodiment.

[0038] Specific implementation method 4: This implementation method is different from any one of specific implementation methods 1 to 3 in that the data enhancement operation includes random horizontal flipping, random scaling, and random rotation.

[0039] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0040] Specific implementation mode 5: This implementation mode is different from any one of the specific implementation modes 1 to 4 in that a self-distillation convolutional neural network model is constructed in step B1; the specific process is as follows:

[0041] The self-distilled convolutional neural network model includes a basic multi-stage network and a multi-scale feature interaction network;

[0042] The basic multi-stage network includes, in sequence: a first convolution unit, a second convolution unit, a third convolution unit, a fourth convolution unit, a fifth convolution unit and a first output unit;

[0043] The first convolution unit includes a first convolution layer, a batch normalization layer, a ReLU activation layer, and a maximum pooling layer in sequence;

[0044] The second convolution unit includes, in sequence, a second convolution layer, a batch normalization layer, a ReLU activation layer, a third convolution layer, a batch normalization layer, a ReLU activation layer, a fourth convolution layer, a batch normalization layer, and a ReLU activation layer;

[0045] The third convolution unit sequentially includes a fifth convolution layer, a batch normalization layer, a ReLU activation layer, a sixth convolution layer, a batch normalization layer, a ReLU activation layer, a seventh convolution layer, a batch normalization layer, and a ReLU activation layer;

[0046] The fourth convolution unit sequentially includes an eighth convolution layer, a batch normalization layer, a ReLU activation layer, a ninth convolution layer, a batch normalization layer, a ReLU activation layer, a tenth convolution layer, a batch normalization layer, and a ReLU activation layer;

[0047] The fifth convolution unit sequentially includes an eleventh convolution layer, a batch normalization layer, a ReLU activation layer, a twelfth convolution layer, a batch normalization layer, a ReLU activation layer, a thirteenth convolution layer, a batch normalization layer, and a ReLU activation layer;

[0048] The first output unit is a fully connected layer;

[0049] The multi-scale feature interaction network includes, in sequence: a first adjacent feature interaction learning module, a second adjacent feature interaction learning module, a third adjacent feature interaction learning module, a fourth adjacent feature interaction learning module, a fifth adjacent feature interaction learning module, and a second output unit;

[0050] The second output unit is a fully connected layer.

[0051] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0052] Specific implementation method 6: This implementation method is different from any one of the specific implementation methods 1 to 5 in that the working process of the self-distillation convolutional neural network model is:

[0053] 1) The preprocessed thyroid nodule ultrasound image is sequentially input into the first convolution unit, the second convolution unit, the third convolution unit, the fourth convolution unit, the fifth convolution unit, and the first output unit in the basic multi-stage network, and the first output unit outputs the prediction result of the basic multi-stage network;

[0054] 2) The output features of the first convolution unit and the second convolution unit in the basic multi-stage network are input into the first adjacent feature interaction learning module in the multi-scale feature interaction network, and the first adjacent feature interaction learning module outputs features;

[0055] The output features of the first convolution unit, the second convolution unit, and the third convolution unit in the basic multi-stage network are input into the second adjacent feature interactive learning module in the multi-scale feature interactive network, and the second adjacent feature interactive learning module outputs features;

[0056] The output features of the second convolution unit, the third convolution unit, and the fourth convolution unit in the basic multi-stage network are input into the third adjacent feature interaction learning module in the multi-scale feature interaction network, and the third adjacent feature interaction learning module outputs features;

[0057] The output features of the third convolution unit, the fourth convolution unit, and the fifth convolution unit in the basic multi-stage network are input into the fourth adjacent feature interaction learning module in the multi-scale feature interaction network, and the fourth adjacent feature interaction learning module outputs features;

[0058] The output features of the fourth convolutional unit and the fifth convolutional unit in the basic multi-stage network are input into the fifth adjacent feature interactive learning module in the multi-scale feature interactive network, and the fifth adjacent feature interactive learning module outputs features;

[0059] 3) Add the output features of the first adjacent feature interactive learning module, the output features of the second adjacent feature interactive learning module, the output features of the third adjacent feature interactive learning module, the output features of the fourth adjacent feature interactive learning module, and the output features of the fifth adjacent feature interactive learning module, and input the added features into the second output unit, which outputs the multi-scale feature interaction network prediction result;

[0060] 4) Add the basic multi-stage network prediction result output by the first output unit and the multi-scale feature interaction network prediction result output by the second output unit and take the average, and use the average value as the output result of the self-distillation convolutional neural network model.

[0061] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0062] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that the specific working process of each adjacent feature interactive learning module in the first adjacent feature interactive learning module, the second adjacent feature interactive learning module, the third adjacent feature interactive learning module, the fourth adjacent feature interactive learning module, and the fifth adjacent feature interactive learning module is as follows:

[0063] 1) Through the convolution operation, the feature f i-1 and f i+1 The number of channels and features f i The number of channels is consistent, and the feature f after adjusting the number of channels is obtained i-1 and f i+1 ;

[0064] The feature f i-1 , f i , f i+1 It is the features of adjacent scales output by adjacent convolution units (three adjacent convolution units) in the basic multi-stage network in the self-distillation convolutional neural network model; 2)

[0066] The feature f is adjusted downward through pooling, interpolation and convolution operations in sequence i-1 The resolution and f i The same, get the feature f after adjusting the resolution i-1;

[0067] The feature f is adjusted upward through pooling, interpolation and convolution operations in sequence i+1 The resolution and f i The same, get the feature f after adjusting the resolution i+1 ;

[0068] The feature f after adjusting the resolution i-1 , f i+1 and f i Add together to get the new feature f i ;

[0069] The original feature f is adjusted downward through pooling, interpolation and convolution operations in sequence i The resolution of the feature f obtained after adjusting the number of channels in 1) i+1 The same, get the feature f after adjusting the resolution i , the feature f after adjusting the resolution i And 1) the feature f after adjusting the number of channels i+1 Add together to get the new feature f i+1 ;

[0070] The original feature f is adjusted upward through pooling, interpolation and convolution operations in sequence i The resolution of the feature f obtained after adjusting the number of channels in 1) i-1 The same, get the feature f after adjusting the resolution i , the feature f after adjusting the resolution i And 1) the feature f after adjusting the number of channels i-1 Add together to get the new feature f i-1 ;

[0071] 3) Obtain the new feature f i-1 , f i , f i+1 The sum is performed and then input into a convolutional layer. The convolutional layer outputs the fused features and reduces the number of channels.

[0072] The original feature f i The features are added after fusion with the output of the convolutional layer, and the summed features are used as the output features of the adjacent feature interaction learning module.

[0073] The complete process is as follows:

[0074] f IA,i =M i (f T,i )+I i (f i )

[0075]

[0076] Among them, i is the convolution unit index of the basic multi-stage network, and j is the weight index of the feature combination;

[0077] f T,i represents the feature fusion result after passing β(·);

[0078] f IA,i Output features of the adjacent feature interactive learning module;

[0079] M(·) indicates feature merging;

[0080] β(·) represents the sequential pooling, interpolation and convolution operations on the features;

[0081] I(·) represents the identity mapping, that is, for any element x, x=I(x).

[0082] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0083] Specific implementation eight: This implementation is different from any one of specific implementations one to seven in that a self-distillation convolutional neural network model loss function is designed in step B3; the specific process is:

[0084] The loss function of the self-distillation convolutional neural network model consists of two parts: category loss and self-distillation loss;

[0085] The first output unit in the basic multi-stage network and the second output unit in the multi-scale feature interaction network jointly use the category loss;

[0086] The first output unit in the base multi-stage network and the second output unit in the multi-scale feature interaction network both use a self-distillation loss;

[0087] Class loss L cls Defined as:

[0088]

[0089] Where N is the number of samples, y i represents the true category of the i-th sample, represents the prediction result of the first output unit in the basic multi-stage network for the i-th sample, Represents the prediction result of the second output unit in the multi-scale feature interaction network for the i-th sample;

[0090] Self-distillation loss L sd Defined as:

[0091]

[0092] Where D KLis the Kullback-Leibler divergence, T is the distillation temperature, · is the dot product operation, and || is the separator;

[0093] The loss function of the self-distillation convolutional neural network model is the category loss L cls and self-distillation loss L sd The sum of L = L cls +L sd .

[0094] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0095] Specific embodiment nine: This embodiment is different from any one of specific embodiments one to eight in that the main training module uses the preprocessed thyroid nodule ultrasound image to train the built self-distillation convolutional neural network model to obtain the trained self-distillation convolutional neural network model; the specific process is:

[0096] Step C1, initialize the self-distillation convolutional neural network model parameters (weights and biases);

[0097] Set the initial learning rate and the learning rate decay rule. The decay rule is that the learning rate gradually decreases during the training process.

[0098] Set the number of training rounds;

[0099] After all training-related hyperparameters (learning rate, decay law, number of training rounds) are set, start training;

[0100] Step C2: Use the preprocessed thyroid nodule ultrasound image data as a training sample set to train the self-distillation convolutional neural network model until the maximum number of training rounds is reached or the model loss function converges, obtain the trained self-distillation convolutional neural network model, and save the trained self-distillation convolutional neural network model parameters.

[0101] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0102] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that the main recognition module is used to load the trained self-distillation convolutional neural network model, detect the ultrasound image of the thyroid nodule to be identified, and identify the category of the thyroid nodule; the specific process is:

[0103] Step D1, obtaining ultrasound images of thyroid nodules to be detected, filling each image into a square, adjusting each image filled into a square to a size of 512×512, and normalizing the images adjusted to a size of 512×512 to obtain normalized images;

[0104] Step D2: input the normalized image obtained in step D1 into the trained self-distillation convolutional neural network model, and the trained self-distillation convolutional neural network model outputs the category of thyroid nodules.

[0105] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0106] The performance index of the self-distillation convolutional neural network is calculated for the thyroid nodule recognition result obtained from the network output in step D2; the performance index calculation method is:

[0107]

[0108] TP refers to the true positive in the recognition result, FP refers to the false positive in the recognition result, TN refers to the true negative in the recognition result, and FN refers to the false negative in the recognition result.

[0109] The following examples are used to verify the beneficial effects of the present invention:

[0110] Embodiment 1:

[0111] The specific implementation of the present invention is described below in conjunction with specific embodiments and drawings. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0112] The thyroid nodule ultrasound dataset comes from actual hospital cases, including 3598 ultrasound images from 3598 patients, including 1621 benign cases and 1977 malignant cases. The ground truth classification labels of all ultrasound images are obtained based on the biopsy reports of the corresponding patients. They are randomly divided into training and test sets in a ratio of 6:4. Among them, 2159 cases are used as training sets to update model parameters; 1439 cases are used as test sets to test the final model.

[0113] Perform step 1:

[0114] Execute step A1, read 2159 grayscale breast ultrasound videos of the training set, fill each image into a square and resize it to 512×512;

[0115] Execute step A2, perform normalization operation on the image of uniform size, randomly flip the image horizontally with a probability of 0.5, randomly scale the image to 0.5 to 1.5 times the original image size with a probability of 0.5, and randomly rotate the image 90 degrees clockwise or counterclockwise with a probability of 0.5;

[0116] Perform step 2:

[0117] Execute step B1 and construct a self-distillation convolutional neural network model. The self-distillation convolutional neural network model includes a basic multi-stage network and a multi-scale feature interaction network. The basic multi-stage network is composed of a first convolution unit, a second convolution unit, a third convolution unit, a fourth convolution unit, a fifth convolution unit and a first output unit connected in sequence. The multi-scale feature interaction network consists of five adjacent feature interaction learning modules and a second output unit. The five adjacent feature interaction learning modules are connected in parallel and by addition, and then connected to the second output unit. The prediction result of the basic multi-stage network output by the first output unit and the prediction result of the multi-scale feature interaction network output by the second output unit are added and averaged, and the average value is used as the output result of the self-distillation convolutional neural network model.

[0118] Execute step B2 and design the loss function of the self-distillation convolutional neural network model. The loss function calculates the loss between the recognition result and the true result, and optimizes the overall network based on this. The overall loss function is the sum of the category loss and the self-distillation loss. The prediction results of the basic multi-stage network and the multi-scale feature interaction network are used separately to calculate the category loss, and the results of the two are used together to calculate the self-distillation loss.

[0119] Class loss L cls Defined as:

[0120]

[0121] Where N is the number of samples, y i represents the true category of the i-th sample, represents the prediction result of the first output unit in the basic multi-stage network for the i-th sample, Represents the prediction result of the second output unit in the multi-scale feature interaction network for the i-th sample;

[0122] Self-distillation loss L sd Defined as:

[0123]

[0124] Where D KL is the Kullback-Leibler divergence, T is the distillation temperature, which is set to 10. · is the dot product operation, and || is the separator.

[0125] Execute step three:

[0126] Execute step C1, initialize the self-distillation convolutional neural network model parameters (weights and biases), use the Adam optimizer, set the initial learning rate to 0.0002, and the learning rate decays by 35% after every 20 rounds of training. Set the number of training rounds to 100 rounds, and use the batch iteration method for training. The batch size is set to 8, and one round is defined as one iteration of the entire training data set.

[0127] Execute step C2, use the training sample data set to train the built self-distillation convolutional neural network model. If the network reaches the set number of training rounds or the average accuracy of the network does not improve within 10 rounds, the network is considered to have converged, and the trained network parameters are saved.

[0128] Execute step 4:

[0129] Execute step D1, obtain the ultrasound image of the thyroid nodule to be detected, fill each image into a square and adjust the size to 512×512, and perform a normalization operation.

[0130] Execute step D2, input the thyroid nodule ultrasound image processed in step D1 into the trained self-distillation convolutional neural network model, and obtain the thyroid nodule recognition result output by the network. This process does not perform the update of network parameters.

[0131] After executing step D3 and completing the test of all images, the performance indicators of the self-distillation convolutional neural network are calculated based on the test results.

[0132] Table 1 Average accuracy of breast lesion detection

[0133] method Accuracy Accuracy Recall ResNet101 75.91％ 80.37％ 71.96％ CH-UNet 80.31％ 78.79％ 81.18％ ThyNet 79.60％ 81.21％ 78.62％ Self-distilled convolutional neural network 82.42％ 81.46％ 86.90％

[0134] Table 1 shows the recognition results of the proposed self-distilled convolutional neural network for thyroid nodules in ultrasound images. ResNet101, CH-UNet and ThyNet are currently commonly used recognition algorithms. Compared with the existing commonly used algorithms, the system of the present invention has improved the three evaluation indicators of thyroid nodule recognition accuracy, precision and recall.

[0135] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A thyroid nodule recognition system based on self-distilled convolutional neural network, characterized by: The system comprises: Image processing main module, neural network main module, training main module and recognition main module; The image processing main module is used to collect thyroid nodule ultrasonic images, preprocess the collected thyroid nodule ultrasonic images, and obtain the preprocessed thyroid nodule ultrasonic images; The neural network main module is used to build a self-distillation convolutional neural network model; the specific process is: Step B1, constructing a self-distillation convolutional neural network model; Step B2, designing a self-distillation convolutional neural network model loss function; The main training module uses the preprocessed thyroid nodule ultrasound image to train the built self-distillation convolutional neural network model to obtain a trained self-distillation convolutional neural network model; The main recognition module is used to load the trained self-distillation convolutional neural network model, detect the ultrasound image of the thyroid nodule to be identified, and identify the category of the thyroid nodule.

2. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 1, characterized in that: The image processing main module is used to collect thyroid nodule ultrasonic images, preprocess the collected thyroid nodule ultrasonic images, and obtain the preprocessed thyroid nodule ultrasonic images; the specific process is: Step A1, obtaining ultrasound images of thyroid nodules, filling each image into a square, and adjusting each square-filled image to a size of 512×512; Step A2: Normalize the image adjusted to 512×512 to obtain the normalized image I norm ; For the normalized image I norm Data enhancement operations were performed to obtain preprocessed thyroid nodule ultrasound images as training sets.

3. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 2, characterized in that: In step A2, the image adjusted to a size of 512×512 is normalized to obtain a normalized image I norm ; The expression is: Among them, I represents the image resized to 512×512, I min and I max Represent the minimum and maximum pixel values ​​in the image resized to 512×512.

4. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 3, characterized in that: The data enhancement operation includes random horizontal flipping, random scaling, and random rotation.

5. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 4, characterized in that: In step B1, a self-distillation convolutional neural network model is constructed; the specific process is: The self-distilled convolutional neural network model includes a basic multi-stage network and a multi-scale feature interaction network; The basic multi-stage network includes, in sequence: a first convolution unit, a second convolution unit, a third convolution unit, a fourth convolution unit, a fifth convolution unit and a first output unit; The first convolution unit includes a first convolution layer, a batch normalization layer, a ReLU activation layer, and a maximum pooling layer in sequence; The second convolution unit includes, in sequence, a second convolution layer, a batch normalization layer, a ReLU activation layer, a third convolution layer, a batch normalization layer, a ReLU activation layer, a fourth convolution layer, a batch normalization layer, and a ReLU activation layer; The third convolution unit sequentially includes a fifth convolution layer, a batch normalization layer, a ReLU activation layer, a sixth convolution layer, a batch normalization layer, a ReLU activation layer, a seventh convolution layer, a batch normalization layer, and a ReLU activation layer; The fourth convolution unit sequentially includes an eighth convolution layer, a batch normalization layer, a ReLU activation layer, a ninth convolution layer, a batch normalization layer, a ReLU activation layer, a tenth convolution layer, a batch normalization layer, and a ReLU activation layer; The fifth convolution unit sequentially includes an eleventh convolution layer, a batch normalization layer, a ReLU activation layer, a twelfth convolution layer, a batch normalization layer, a ReLU activation layer, a thirteenth convolution layer, a batch normalization layer, and a ReLU activation layer; The first output unit is a fully connected layer; The multi-scale feature interaction network includes, in sequence: a first adjacent feature interaction learning module, a second adjacent feature interaction learning module, a third adjacent feature interaction learning module, a fourth adjacent feature interaction learning module, a fifth adjacent feature interaction learning module, and a second output unit; The second output unit is a fully connected layer.

6. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 5, characterized in that: The working process of the self-distillation convolutional neural network model is: 1) The preprocessed thyroid nodule ultrasound image is sequentially input into the first convolution unit, the second convolution unit, the third convolution unit, the fourth convolution unit, the fifth convolution unit, and the first output unit in the basic multi-stage network, and the first output unit outputs the prediction result of the basic multi-stage network; 2) The output features of the first convolution unit and the second convolution unit in the basic multi-stage network are input into the first adjacent feature interaction learning module in the multi-scale feature interaction network, and the first adjacent feature interaction learning module outputs features; The output features of the first convolution unit, the second convolution unit, and the third convolution unit in the basic multi-stage network are input into the second adjacent feature interactive learning module in the multi-scale feature interactive network, and the second adjacent feature interactive learning module outputs features; The output features of the second convolution unit, the third convolution unit, and the fourth convolution unit in the basic multi-stage network are input into the third adjacent feature interaction learning module in the multi-scale feature interaction network, and the third adjacent feature interaction learning module outputs features; The output features of the third convolution unit, the fourth convolution unit, and the fifth convolution unit in the basic multi-stage network are input into the fourth adjacent feature interaction learning module in the multi-scale feature interaction network, and the fourth adjacent feature interaction learning module outputs features; The output features of the fourth convolutional unit and the fifth convolutional unit in the basic multi-stage network are input into the fifth adjacent feature interactive learning module in the multi-scale feature interactive network, and the fifth adjacent feature interactive learning module outputs features; 3) Add the output features of the first adjacent feature interactive learning module, the output features of the second adjacent feature interactive learning module, the output features of the third adjacent feature interactive learning module, the output features of the fourth adjacent feature interactive learning module, and the output features of the fifth adjacent feature interactive learning module, and input the added features into the second output unit, which outputs the multi-scale feature interaction network prediction result; 4) Add the basic multi-stage network prediction result output by the first output unit and the multi-scale feature interaction network prediction result output by the second output unit and take the average, and use the average value as the output result of the self-distillation convolutional neural network model.

7. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 6, characterized in that: The specific working process of each of the first adjacent feature interactive learning module, the second adjacent feature interactive learning module, the third adjacent feature interactive learning module, the fourth adjacent feature interactive learning module, and the fifth adjacent feature interactive learning module is as follows: 1) Through the convolution operation, the feature f i-1 and f i+1 The number of channels and features f i The number of channels is consistent, and the feature f after adjusting the number of channels is obtained i-1 and f i+1 ; The feature f i-1 , f i , f i+1 It is the features of adjacent scales output by adjacent convolution units in the basic multi-stage network in the self-distillation convolutional neural network model; 2)、 The feature f is adjusted downward through pooling, interpolation and convolution operations in sequence i-1 The resolution and f i The same, get the feature f after adjusting the resolution i-1 ; The feature f is adjusted upward through pooling, interpolation and convolution operations in sequence i+1 The resolution and f i The same, get the feature f after adjusting the resolution i+1 ; The feature f after adjusting the resolution i-1 , f i+1 and f i Add together to get the new feature f i ; The original feature f is adjusted downward through pooling, interpolation and convolution operations in sequence i The resolution of the feature f obtained after adjusting the number of channels in 1) i+1 The same, get the feature f after adjusting the resolution i , the feature f after adjusting the resolution i And 1) the feature f after adjusting the number of channels i+1 Add together to get the new feature f i+1 ; The original feature f is adjusted upward through pooling, interpolation and convolution operations in sequence i The resolution of the feature f obtained after adjusting the number of channels in 1) i-1 The same, get the feature f after adjusting the resolution i , the feature f after adjusting the resolution i And 1) the feature f after adjusting the number of channels i-1 Add together to get the new feature f i-1 ; 3) Obtain the new feature f i-1 , f i , f i+1 The sum is performed and then input into a convolutional layer, which outputs the fused features. The original feature f i The features are added after fusion with the output of the convolutional layer, and the summed features are used as the output features of the adjacent feature interaction learning module.

8. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 7, characterized in that: In step B3, a self-distillation convolutional neural network model loss function is designed; the specific process is: The loss function of the self-distillation convolutional neural network model consists of two parts: category loss and self-distillation loss; Class loss L cls Defined as: Where N is the number of samples, y i represents the true category of the i-th sample, represents the prediction result of the first output unit in the basic multi-stage network for the i-th sample, Represents the prediction result of the second output unit in the multi-scale feature interaction network for the i-th sample; Self-distillation loss L sd Defined as: Where D KL is the Kullback-Leibler divergence, T is the distillation temperature, · is the dot product operation, and || is the separator; The loss function of the self-distillation convolutional neural network model is the category loss L cls and self-distillation loss L sd The sum of L = L cls +L sd .

9. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 8, characterized in that: The main training module uses the preprocessed thyroid nodule ultrasound image to train the built self-distillation convolutional neural network model to obtain the trained self-distillation convolutional neural network model; the specific process is: Step C1, initializing the self-distillation convolutional neural network model parameters; Set the initial learning rate and learning rate decay rule; Set the number of training rounds; Step C2: Use the preprocessed thyroid nodule ultrasound image data as a training sample set to train the self-distillation convolutional neural network model until the maximum number of training rounds is reached or the model loss function converges, obtain the trained self-distillation convolutional neural network model, and save the trained self-distillation convolutional neural network model parameters.

10. The thyroid nodule recognition system based on self-distillation convolutional neural network according to claim 9, characterized in that: The identification main module is used to load the trained self-distillation convolutional neural network model, detect the ultrasound image of the thyroid nodule to be identified, and identify the category of the thyroid nodule; the specific process is: Step D1, obtaining ultrasound images of thyroid nodules to be detected, filling each image into a square, adjusting each image filled into a square to a size of 512×512, and normalizing the images adjusted to a size of 512×512 to obtain normalized images; Step D2: input the normalized image obtained in step D1 into the trained self-distillation convolutional neural network model, and the trained self-distillation convolutional neural network model outputs the category of thyroid nodules.

Images