Thyroid cancer early diagnosis model construction method based on convolutional neural network

Abstract

The invention provides a thyroid cancer early diagnosis model construction method based on a convolutional neural network. The thyroid cancer early diagnosis model construction method based on the convolutional neural network comprises the following steps: S1, data collection and preprocessing; S2, constructing a convolutional neural network model according to the preprocessed data in S1; and S3, carrying out model training on the obtained convolutional neural network model. The thyroid cancer early diagnosis model construction method based on the convolutional neural network provided by the invention has the advantages of getting rid of dependence on high-level color Doppler ultrasound doctors to the greatest extent, realizing high-level diagnosis of primary hospitals and reducing waste of medical data.

Description

Technical Field

[0001] The present invention relates to the technical field of model construction, and particularly to a method for constructing an early diagnosis model of thyroid cancer based on a convolutional neural network. Background Art

[0002] The existing diagnosis mode mainly relies on subjective diagnosis by ultrasound physicians. The diagnosis level is highly related to personal experience, resulting in relatively large limitations.

[0003] Convolutional neural network (CNN) is an important part of deep learning in the field of artificial intelligence, and has played a key role in tasks such as image processing, pattern recognition, and computer vision. It enables a computer to automatically learn meaningful features from a large amount of data, achieving accurate understanding and classification of complex data such as images and videos. CNN can automatically extract subtle and discriminative features from various modalities of data such as thyroid ultrasound images, CT scans, and pathological sections.

[0004] Therefore, it is necessary to provide a new method for constructing an early diagnosis model of thyroid cancer based on a convolutional neural network to solve the above technical problems. Summary of the Invention

[0005] The technical problem solved by the present invention is to provide a method for constructing an early diagnosis model of thyroid cancer based on a convolutional neural network, which can maximize the independence from highly skilled color Doppler ultrasound physicians, achieve high-level diagnosis in primary hospitals, and reduce the waste of medical resources.

[0006] To solve the above technical problem, the method for constructing an early diagnosis model of thyroid cancer based on a convolutional neural network provided by the present invention includes the following steps: S1: Data collection and preprocessing: S2: Construct a convolutional neural network model according to the preprocessed data in S1; S3: Train the obtained convolutional neural network model.

[0007] Preferably, in S1, the data collection and preprocessing include the following steps: S11: Data collection: Image data: Collect a large number of thyroid ultrasound images and thyroid fine needle aspiration biopsy images from the imaging departments of multiple hospitals. At the same time, record information such as the acquisition device, parameters, and acquisition time of the images.

[0008] Pathological diagnosis data: Corresponding to the image data, collect detailed pathological diagnosis reports to clarify the type, grade, and stage information of the lesions, as accurate labels for the image data.

[0009] Clinical information: Collect the basic information of patients, including age, gender, family medical history, past medical history, and symptom manifestations; S12: Data preprocessing: Image normalization: Perform size normalization on the collected ultrasound images and FNA images, adjust all images to the same size. At the same time, unify the resolution and color mode of the images, convert color images to grayscale images or perform standardized color space conversion to eliminate image feature differences caused by device differences and different acquisition conditions; Geometric transformation: Rotate, flip, and translate the images to increase image diversity, expand the dataset size, and improve the generalization ability of the model; Gray-level transformation: Use methods such as histogram equalization and contrast stretching to adjust the gray-level distribution of the images, enhance the contrast of the images, and make the differences between the lesion areas and normal tissues more obvious; Filtering process: Adopt Gaussian filtering and median filtering methods to remove the noise in the images and improve the quality of the images; Data annotation: Professional pathologists and radiologists annotate the images in detail. For complex or difficult-to-judge cases, multidisciplinary experts are organized for consultation to ensure the accuracy of the annotation.

[0010] Preferably, it is characterized in that in S11, the ultrasound images should cover different scanning angles and modes to obtain comprehensive information of the thyroid; the thyroid fine needle aspiration biopsy images need to contain cell sample images with different lesion degrees and types.

[0011] Preferably, it is characterized in that in S12, when annotating the data, according to the pathological diagnosis results, the images are annotated as normal, benign lesions, and different types of thyroid cancer.

[0012] Preferably, in S2, a convolutional neural network model is constructed according to the data preprocessed in S1, including the following steps: S21: Select the base model: Use VGG16, ResNet, and Inception as the base models; S22: Model structure adjustment: Convolution layer adjustment: In the convolution layer part of the base model, appropriately increase or decrease the number of convolution layers and the size of the convolution kernels according to the characteristics of the thyroid images; Pooling layer optimization: Adjust the type and parameters of the pooling layer, and select the appropriate pooling method and pooling window size according to the characteristics of the dataset and the performance of the model to reduce the dimension of the data while retaining important features.

[0013] Introduction of attention mechanism: The attention mechanism is introduced into the model. The attention mechanism is used to enable the model to automatically learn the importance of different regions in the image, focus on the key features related to thyroid cancer, suppress the interference of irrelevant information, and thus improve the diagnostic accuracy of the model; S23: Design of fully connected layer and output layer: Fully connected layer: After the convolutional layer and pooling layer, multiple fully connected layers are connected. The number of neurons in the fully connected layer is adjusted according to the size and complexity of the dataset; Output layer: The output layer uses the softmax activation function to convert the feature vector output by the fully connected layer into the probability distribution of each category.

[0014] Preferably, in S3, the obtained convolutional neural network model is trained, including the following steps: S31: Divide the dataset: Division ratio: The preprocessed dataset is divided into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%; Data balancing: Since thyroid cancer may belong to the minority class in the actual data, in order to avoid overfitting of the model to the majority class during training, it is necessary to balance the dataset; S32: Training process: Forward propagation: Input the training set data into the model. The data passes through the convolutional layer, pooling layer, and fully connected layer network layers in sequence, and after convolution, pooling, and activation operations, the prediction result of the model is finally obtained at the output layer; Loss calculation: According to the prediction result and the true label, use the cross-entropy loss function to calculate the loss value; Backward propagation: Calculate the gradient of the loss function with respect to the model parameters through the backward propagation algorithm; Parameter update: According to the calculated gradient, use the Adam optimizer to update the model parameters; Training epochs and early stopping strategy: During training, the training set data is repeatedly input into the model for training. Each complete traversal of the training set is called an epoch.

[0015] Preferably, in S35, after each epoch ends, use the validation set data to evaluate the performance of the model. If the performance of the model on the validation set does not improve in consecutive multiple epochs, it indicates that the model may have overfitted. At this time, the early stopping strategy can be adopted to stop the training of the model to avoid overtraining.

[0016] Compared with the related technologies, the method for constructing an early diagnosis model of thyroid cancer based on convolutional neural network provided by the present invention has the following beneficial effects: The present invention provides a method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network, which integrates artificial intelligence with the diagnosis technology of thyroid nodules to form a new diagnosis model; through continuous optimization, an artificial intelligence technology that can continuously self-learn and improve is formed; it can get rid of the dependence on high-level color ultrasound physicians to the greatest extent, achieve high-level diagnosis in grassroots hospitals, and reduce the waste of medical information. DETAILED DESCRIPTION

[0017] The present invention will be further described below in conjunction with the embodiments.

[0018] The method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network includes the following steps: S1: Data collection and preprocessing S11: Data collection: Image data: A large number of thyroid ultrasound images and thyroid fine needle aspiration (FNA) images were collected from the imaging departments of multiple hospitals. Ultrasound images should cover different scanning angles and modes to obtain comprehensive information about the thyroid gland; FNA images should include cell sample images of different degrees and types of lesions. At the same time, information such as the image acquisition equipment, parameters, and acquisition time should be recorded.

[0019] Pathological diagnosis data: Corresponding to the image data, detailed pathological diagnosis reports are collected to clarify the type of lesions (such as papillary carcinoma, follicular carcinoma, medullary carcinoma, etc.), grading, and staging information as accurate labels for the image data.

[0020] Clinical information: Collect basic information about the patient, including age, gender, family history, past medical history, symptoms, etc. This clinical information helps to discover potential associated factors in subsequent analysis and improve the diagnostic accuracy of the model.

[0021] S12: Data preprocessing: Image standardization: Normalize the size of collected ultrasound images and FNA images, adjust all images to the same size, such as 224×224 pixels. At the same time, unify the resolution and color mode of the images, convert color images to grayscale images or perform standardized color space conversion to eliminate image feature differences caused by equipment differences and acquisition conditions.

[0022] Image enhancement.

[0023] Geometric transformation: Rotate, flip, translate, and other operations on images to increase image diversity, expand the size of the data set, and improve the generalization ability of the model.

[0024] Gray-scale transformation: By using methods such as histogram equalization and contrast stretching, the gray-scale distribution of the image is adjusted to enhance the contrast of the image, making the difference between the lesion area and normal tissues more obvious.

[0025] Filtering process: Methods such as Gaussian filtering and median filtering are adopted to remove the noise in the image and improve the quality of the image.

[0026] Data annotation: Professional pathologists and radiologists annotate the images in detail. According to the pathological diagnosis results, the images are annotated as normal, benign lesions (such as nodular goiter, thyroid adenoma, etc.), and different types of thyroid cancer (such as papillary carcinoma, follicular carcinoma, etc.). For complex or difficult-to-judge cases, multi-disciplinary experts are organized for consultation to ensure the accuracy of the annotation.

[0027] S2: Construction of convolutional neural network model S21: Selection of the base model VGG16: VGG16 has a simple network structure, consisting of multiple convolutional layers and pooling layers. By continuously stacking convolutional layers, it extracts deep features of the image. Its advantages are simple structure, easy to understand and train, and it shows good performance in image classification tasks.

[0028] ResNet: ResNet introduces residual connections, solves the problems of gradient vanishing and gradient explosion in the training process of deep neural networks, and enables the network to be trained deeper. Through the design of residual modules, ResNet can better learn the features of images, especially suitable for processing complex image data.

[0029] Inception: The Inception model adopts a parallel structure of multi-scale convolutional kernels, which can extract image features of different scales simultaneously, thus describing image information more comprehensively. While improving the performance of the model, the Inception model can also effectively reduce the number of model parameters and the computational complexity.

[0030] S22: Model structure adjustment Adjustment of convolutional layers: In the convolutional layer part of the base model, according to the characteristics of thyroid images, the number of convolutional layers and the size of convolutional kernels are appropriately increased or decreased. For example, for smaller lesions in thyroid ultrasound images, smaller convolutional kernels may be needed to capture detailed features; while for larger lesion areas, larger convolutional kernels can be used to extract overall features.

[0031] Optimization of pooling layers: Adjust the type and parameters of pooling layers, such as using different methods like average pooling, max pooling or adaptive pooling. According to the characteristics of the dataset and the performance of the model, select the appropriate pooling method and pooling window size to reduce the data dimension while retaining important features.

[0032] Introduction of Attention Mechanism: Introduce an attention mechanism into the model, such as the SE module (Squeeze-and-Excitation Module) or the CBAM module (Convolutional Block Attention Module). The attention mechanism enables the model to automatically learn the importance of different regions in the image, pay more attention to the key features related to thyroid cancer, suppress the interference of irrelevant information, and thus improve the diagnostic accuracy of the model.

[0033] S23: Design of Fully Connected Layer and Output Layer Fully Connected Layer: After the convolutional layer and pooling layer, connect multiple fully connected layers. The number of neurons in the fully connected layer is adjusted according to the size and complexity of the dataset. Generally, the first fully connected layer has a larger number of neurons to fully integrate the features extracted previously; the subsequent fully connected layers gradually reduce the number of neurons to further compress and abstract the features.

[0034] Output Layer: The output layer uses the softmax activation function to convert the feature vector output by the fully connected layer into the probability distribution of each class. The number of neurons in the output layer is equal to the number of classification classes. For example, for the diagnosis of thyroid cancer, it can be set to different classes such as normal, benign lesion, papillary carcinoma, follicular carcinoma, etc., and each neuron corresponds to the probability prediction value of a class.

[0035] S3: Model Training S31: Dataset Division Division Ratio: Divide the preprocessed dataset into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%. The training set is used for model training, the validation set is used to adjust the hyperparameters of the model and evaluate the model's performance, and the test set is used to finally evaluate the generalization ability of the model.

[0036] Data Balancing: Since thyroid cancer may be a minority class in the actual data, in order to avoid the model overfitting to the majority classes (normal and benign lesions) during training, it is necessary to balance the dataset. Methods such as oversampling, undersampling, or generative adversarial networks (GANs) can be used to balance the number of samples in different classes.

[0037] S32: Training Process Forward Propagation: Input the training set data into the model. The data passes through network layers such as convolutional layers, pooling layers, and fully connected layers in sequence, and after a series of operations such as convolution, pooling, and activation, finally obtain the prediction result of the model at the output layer.

[0038] Loss calculation: Based on the prediction results and the true labels, the cross entropy loss function is used to calculate the loss value. The loss value reflects the gap between the current prediction results of the model and the actual situation. The smaller the loss value, the closer the model's prediction results are to the true labels.

[0039] Back propagation: The gradient of the loss function to the model parameters is calculated through the back propagation algorithm. The back propagation algorithm starts from the output layer, calculates the gradient layer by layer, and passes the gradient back to the previous network layer to update the model parameters.

[0040] Parameter update: Based on the calculated gradient, the Adam optimizer is used to update the model parameters. The optimizer adjusts the value of each parameter based on the learning rate and the size of the gradient so that the loss function can be further reduced in the next iteration.

[0041] Training rounds and early stopping strategy: During the training process, the training set data is repeatedly input into the model for training. Each completed traversal of the training set is called an epoch. After each epoch, the performance of the model is evaluated using the validation set data. If the performance of the model on the validation set does not improve in multiple consecutive epochs, it means that the model may have been overfitted. At this time, the early stopping strategy can be used to stop the training of the model to avoid overtraining.

[0042] Compared with the related art, the method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network provided by the present invention has the following beneficial effects: The present invention provides a method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network, which integrates artificial intelligence with the diagnosis technology of thyroid nodules to form a new diagnosis model; through continuous optimization, an artificial intelligence technology that can continuously self-learn and improve is formed; it can get rid of the dependence on high-level color ultrasound physicians to the greatest extent, achieve high-level diagnosis in grassroots hospitals, and reduce the waste of medical information.

[0043] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network, characterized in that: The following steps are involved: S1: Data collection and preprocessing: S2: Construct a convolutional neural network model based on the preprocessed data in S1; S3: Perform model training on the obtained convolutional neural network model.

2. The method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network according to claim 1, characterized in that: In S1, data collection and preprocessing include the following steps: S11: Data collection: Image data: A large number of thyroid ultrasound images and thyroid fine needle aspiration biopsy images were collected from the imaging departments of multiple hospitals. At the same time, information such as the image acquisition equipment, parameters, and acquisition time were recorded: Pathological diagnosis data: Corresponding to the image data, collect detailed pathological diagnosis reports to clarify the type, grade, and stage of the lesion as accurate labels for the image data: Clinical information: Collect the patient's basic information, including age, gender, family history, past medical history, and symptoms; S12: Data preprocessing: Image standardization: Normalize the size of collected ultrasound images and FNA images, adjust all images to the same size, unify the image resolution and color mode, convert color images to grayscale images or perform standardized color space conversion to eliminate image feature differences caused by equipment differences and acquisition conditions; Geometric transformation: rotate, flip, and translate images to increase image diversity, expand the size of the data set, and improve the generalization ability of the model; Grayscale transformation: Use methods such as histogram equalization and contrast stretching to adjust the grayscale distribution of the image, enhance the contrast of the image, and make the difference between the lesion area and normal tissue more obvious; Filtering: Use Gaussian filtering and median filtering methods to remove noise from the image and improve the image quality; Data annotation: Professional pathologists and radiologists annotate the images in detail. For complex or difficult cases, multidisciplinary experts are organized to consult to ensure the accuracy of the annotation.

3. The method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network according to claim 2, characterized in that: In S11, the ultrasound image should cover different scanning angles and modes to obtain comprehensive information about the thyroid gland; the thyroid fine needle aspiration biopsy image should include cell sample images of different lesion degrees and types.

4. The method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network according to claim 2, characterized in that: In S12, when labeling the data, the images are labeled as normal, benign lesions, and different types of thyroid cancer according to the pathological diagnosis results.

5. The method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network according to claim 1, characterized in that: In S2, a convolutional neural network model is constructed according to the preprocessed data in S1, including the following steps: S21: Select the base model: VGG16, ResNet and Inception are used as basic models; S22: Model structure adjustment: Convolution layer adjustment: In the convolution layer part of the basic model, the number of convolution layers and the size of the convolution kernel are appropriately increased or decreased according to the characteristics of the thyroid image; Pooling layer optimization: Adjust the type and parameters of the pooling layer, select the appropriate pooling method and pooling window size according to the characteristics of the dataset and the performance of the model, so as to reduce the dimension of the data while retaining important features: Introduction of attention mechanism: The attention mechanism is introduced into the model. The attention mechanism is used to allow the model to automatically learn the importance of different areas in the image, focus on key features related to thyroid cancer, and suppress the interference of irrelevant information, thereby improving the diagnostic accuracy of the model; S23: Fully connected layer and output layer design: Fully connected layer: After the convolutional layer and the pooling layer, multiple fully connected layers are connected. The number of neurons in the fully connected layer is adjusted according to the size and complexity of the dataset. Output layer: The output layer uses the softmax activation function to convert the feature vector output by the fully connected layer into the probability distribution of each category.

6. The method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network according to claim 1, characterized in that: In S3, the obtained convolutional neural network model is trained, including the following steps: S31: Divide the data set: Division ratio: The preprocessed data set is divided into training set, validation set and test set in the ratio of 70%, 15% and 15%; Data balance: Since thyroid cancer may belong to the minority class in actual data, the dataset needs to be balanced to avoid overfitting of the model to the majority class during training. S32: Training process: Forward propagation: The training set data is input into the model. The data passes through the convolution layer, pooling layer, and fully connected layer network layer in sequence. After convolution, pooling, and activation operations, the model's prediction results are finally obtained at the output layer. Loss calculation: According to the prediction results and the true labels, the loss value is calculated using the cross entropy loss function; Back propagation: Calculate the gradient of the loss function to the model parameters through the back propagation algorithm; Parameter update: Based on the calculated gradient, the Adam optimizer is used to update the model parameters; Training rounds and early stopping strategy: During the training process, the training set data is repeatedly input into the model for training. Each completed traversal of the training set is called an epoch.

7. The method for constructing an early diagnosis model for thyroid cancer based on a convolutional neural network according to claim 6, characterized in that: In S35, after each epoch, the performance of the model is evaluated using the validation set data. If the performance of the model on the validation set does not improve in multiple consecutive epochs, it means that the model may have been overfitted. At this time, an early stopping strategy can be used to stop the training of the model to avoid overtraining.