A small sample ultrasonic image thyroid function intelligent auxiliary diagnosis system
By constructing a small-sample thyroid ultrasound image dataset and improving the EfficientNet-B0 model, combined with the CBAM attention module and dynamic learning rate strategy, the shortcomings of existing technologies in intelligent auxiliary diagnosis of hypothyroidism and Hashimoto's thyroiditis are addressed, enabling comprehensive screening and accurate identification of thyroid diseases.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG LIGONG UNIV
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-10
AI Technical Summary
Existing ultrasound image-assisted diagnostic systems mainly focus on the identification and analysis of thyroid nodules, neglecting common, prevalent, and important disease types such as hypothyroidism or Hashimoto's thyroiditis. This results in an inability to fully cover the diverse manifestations of thyroid diseases in clinical diagnosis, especially lacking effective intelligent support in inflammation-related lesions.
A small dataset of thyroid ultrasound images was constructed. The EfficientNet-B0 model was used, with the addition of a CBAM attention module, layer normalization layer, embedding layer, and Dropout layer. The model was trained using a dynamic learning rate strategy to form a thyroid function classification model, enabling intelligent auxiliary diagnosis of thyroid function.
It enables comprehensive screening for thyroid diseases, especially accurate identification of inflammatory lesions, improves the accuracy and stability of diagnosis, fills the gap in existing technology, and supports more comprehensive clinical decision-making.
Smart Images
Figure CN122367940A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of auxiliary diagnostic technology, and in particular to an intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images. Background Technology
[0002] The classification and diagnosis of thyroid diseases, that is, distinguishing between normal thyroid function, hypothyroidism, or Hashimoto's thyroiditis based on ultrasound images, faces a core challenge in extracting effective features from ultrasound images and differentiating the categories using deep learning models. Traditional thyroid disease assessment relies on physicians' interpretation of ultrasound image features and analysis of thyroid function blood test indicators. Current research primarily focuses on intelligent auxiliary diagnostic methods for thyroid nodules based on ultrasound images, while the urgent clinical need is to develop intelligent auxiliary diagnostic methods for thyroid function based on ultrasound images.
[0003] In 2024, the team led by Li Xianyan from the Department of Ultrasound Medicine at the Affiliated Hospital of Binzhou Medical University published a paper titled "Optimization of C-TIRADS Class 4a and 4b Nodules Based on Deep Learning Features" in the *Chinese Journal of Ultrasound in Medicine*. This study, targeting class 4a and 4b nodules with high diagnostic uncertainty in the C-TIRADS system, proposed extracting and reducing the dimensionality of deep learning features from ultrasound images. A multilayer perceptron model was then constructed to generate a deep learning score, which was used to correct the original C-TIRADS classification. This improved the AUC of the training and validation sets from 0.693 and 0.685 to 0.814 and 0.810, respectively.
[0004] In her 2024 paper, "Research on Ultrasound Image Classification of Thyroid Nodules Based on APS-Net," Huang Simei proposed an active-passive sonar network (APS-Net) that integrates global and local features. This study addresses the subjective differences inherent in doctors' manual ROI annotation by introducing a self-attention mechanism to enhance nodule recognition and employing six semantic segmentation models of varying sizes to simulate annotation by physicians with different experience levels. Experiments used a local dataset from the Second Affiliated Hospital of Guangxi University of Science and Technology and an open-source dataset from Columbia University, evaluating model performance through 10-fold cross-validation. The results show that APS-Net maintains stable performance across different semantic segmentation methods, achieving an accuracy of 93.97% on the local dataset and 92.59% on the open-source dataset.
[0005] In the diagnosis of thyroid inflammation, Zhao Wanjun's team at West China Hospital of Sichuan University published "Computer-aided diagnosis of Hashimoto's thyroiditis based on convolutional neural networks" in The Journal of Clinical Endocrinology & Metabolism in 2022. They integrated nine types of convolutional neural networks to build an HT-CAD system, which achieved an accuracy of 89.2% based on 39,280 ultrasound images, which is nearly 9 percentage points better than that of senior physicians, and maintained stable performance in a multi-center environment.
[0006] A research team led by Yasaman Sharifi at Mashhad University of Medicine published a study in *WFUMB UltrasoundOpen* in 2025 entitled "Using deep learning for thyroid nodule risk stratification from ultrasound images." They constructed a fully automated diagnostic system that integrates convolutional autoencoders for image artifact removal and employs Faster R-CNN, ResNet-101, and Xception networks for nodule detection and ACR-TIRADS multi-feature classification. In testing, the system achieved an accuracy of 0.98, an AUC of 0.99, and a Kappa value of 0.85, higher than the three radiologists tested.
[0007] It is evident that existing diagnostic aids primarily focus on the identification and analysis of thyroid nodules, neglecting common, prevalent, and important disease types such as hypothyroidism or Hashimoto's thyroiditis. This limitation prevents the comprehensive coverage of the diverse manifestations of thyroid diseases during clinical diagnosis, particularly for inflammatory lesions, where effective intelligent support is lacking. Currently, clinical identification of thyroid inflammatory lesions based on ultrasound relies on subjective experience, easily leading to misdiagnosis or missed diagnosis, requiring the addition of other relevant examinations such as blood tests. Therefore, an intelligent auxiliary system is urgently needed to overcome this deficiency. Summary of the Invention
[0008] The technical problem to be solved by the present invention is to provide a smart auxiliary diagnostic system for thyroid function based on small sample ultrasound images, which addresses the shortcomings of the prior art and enables intelligent auxiliary diagnosis of thyroid function.
[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images, comprising:
[0010] Obtain a thyroid dataset consisting of n cases and construct a small sample thyroid ultrasound image dataset;
[0011] A thyroid function classification model was constructed based on the EfficientNet-B0 model. The thyroid function classification model added a CBAM attention module, a layer normalization layer, an embedding layer, and a Dropout layer to the original EfficientNet-B0 model. The classifier used an independent linear layer.
[0012] A thyroid function classification model was trained based on a thyroid ultrasound image dataset.
[0013] Unlabeled thyroid ultrasound images were acquired, and thyroid function was classified using a trained thyroid function classification model.
[0014] Furthermore, in the thyroid dataset consisting of n cases, each case includes two standardized transverse ultrasound images of the thyroid gland; and the LabelImg annotation tool is used to annotate the bounding box of the thyroid region in the ultrasound images to obtain the region of interest, generate a series of standardized local images, and construct a small sample thyroid ultrasound image dataset.
[0015] Furthermore, the thyroid function classification model introduces a CBAM attention module between the last convolutional layer final_conv and the pooling layer of the original EfficientNet-B0 model, and adds a layer normalization layer after feature extraction of the original EfficientNet-B0 model; an embedding layer is added between the pooling layer and the classifier to compress the original feature dimension; in terms of regularization, the thyroid function classification model integrates a Dropout layer and adopts a configurable dropout rate.
[0016] Furthermore, the training process of the thyroid function classification model adopts a dynamically adjusted learning rate strategy, which includes three stages: linear warm-up, main training stage, and later fine-tuning. In the warm-up stage, a linear learning rate warm-up strategy is used to increase the learning rate. In the main training stage, a cosine annealing scheduler with restart is used to adjust the learning rate to maintain the model's learning vitality. In the later fine-tuning stage, the learning rate is finely tuned through exponential decay.
[0017] Furthermore, end-to-end training was performed on the thyroid function classification model, specifically using the following method:
[0018] At the beginning of training, all backbone feature extraction layers, CBAM attention modules, layer normalization layers, and embedding layers of the model are frozen, and only the topmost classifier is allowed to update its parameters.
[0019] A linear learning rate warm-up strategy was used for the first 80 training cycles, gradually increasing the learning rate from a base value. Throughout the training process, the Dropout layer randomly masked neurons before the embedding layer with a probability of 0.5.
[0020] After the 81st training cycle, a cosine annealing learning rate scheduler with hot restart is used to periodically decrease and restart the learning rate; and an exponential decay fine-tuning strategy is adopted to multiply the current learning rate by a set decay factor in each training cycle to make the learning rate decrease smoothly.
[0021] Unfreeze the embedding layer on the 120th epoch of training;
[0022] During the 200th-300th training cycle, the normalization layer and the CBAM attention module are unfrozen sequentially.
[0023] When training reaches the 400th cycle, the intermediate feature blocks of the deep backbone network are unfrozen, and the learning rate is simultaneously reduced to 0.6 times the original value.
[0024] On the other hand, this application proposes an electronic device comprising: one or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to execute the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images.
[0025] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images.
[0026] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images.
[0027] The beneficial effects of adopting the above technical solution are as follows: This invention provides an intelligent auxiliary diagnostic system for thyroid function based on small-sample ultrasound images. It constructs an intelligent thyroid function classification model based on small-sample thyroid ultrasound images, enabling comprehensive screening for thyroid diseases. This model aims to utilize limited image data, based on deep learning algorithms, to extract key features and learn lesion patterns, thereby helping doctors quickly and accurately determine whether a patient has a thyroid disease. This invention emphasizes the identification of inflammatory lesions to fill the gap in existing technology in this field, ensuring that the auxiliary diagnostic system can more comprehensively support clinical decision-making. Attached Figure Description
[0028] Figure 1 This is a schematic diagram illustrating the implementation process of a small-sample ultrasound image-based intelligent auxiliary diagnostic system for thyroid function provided in Embodiment 1 of the present invention.
[0029] Figure 2The original ultrasound image and the standardized local image provided in Embodiment 1 of the present invention are shown, wherein (a) is the original ultrasound image and (b) is the standardized local image;
[0030] Figure 3 This is a schematic diagram of the thyroid function classification model provided in Embodiment 1 of the present invention;
[0031] Figure 4 The loss curve of the EfficientNetB0+ model provided in Embodiment 1 of the present invention;
[0032] Figure 5 The following are ROC curves of different models provided in Embodiment 1 of the present invention, wherein (a) is the ROC curve of the EfficientNet-B0 model and (b) is the ROC curve of the EfficientNetB0+ model. Detailed Implementation
[0033] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0034] Example 1:
[0035] This embodiment obtained a dataset comprising 204 cases, including 102 cases with normal thyroid function and 102 cases with abnormal thyroid function (51 cases of hypothyroidism and 51 cases of Hashimoto's thyroiditis). The dataset was divided proportionally, with 144 cases in the training set, 38 cases in the validation set, and 22 cases in the test set. Two transverse ultrasound images of the thyroid gland were selected for each case. The thyroid region was outlined using LabelImg, and the region of interest was cropped and output based on the LabelImg XML file.
[0036] The model utilizes the EfficientNet-B0 network architecture, with a training strategy centered on progressive learning. A layered unfreezing mechanism gradually releases model capacity. Initially, all backbone networks are frozen, and only the classifier is trained. Subsequently, network parameters are unfrozen in stages, following the order of embedding layers, normalization layers, CBAM attention modules, and deep features. This allows the model to gradually adapt to data features without overfitting. A dynamically adjusted learning rate strategy is employed during training, comprising three stages: linear warm-up, cosine annealing restart, and late-stage fine-tuning. The warm-up stage steadily increases the learning rate; the main training stage uses cosine annealing with restart to maintain learning vitality; and the later stage uses exponential decay for fine-tuning. A learning rate decay mechanism is also introduced when unfreezing deep features to ensure training stability. This results in an intelligent thyroid disease auxiliary diagnostic system.
[0037] In this embodiment, a small-sample ultrasound image-based intelligent auxiliary diagnostic system for thyroid function is described, such as... Figure 1 As shown, it includes the following steps:
[0038] Step 1: Obtain a thyroid dataset consisting of n cases, each case including two standardized transverse ultrasound images of the thyroid gland; and use the professional annotation tool LabelImg, with experienced physicians annotating the thyroid region in the ultrasound images with bounding boxes to obtain the region of interest, generate a series of standardized local images, and construct a small sample thyroid ultrasound image dataset.
[0039] A cross-sectional image of the thyroid gland can stably display the overall morphology and internal structure of the thyroid gland, making it the most commonly used and informative perspective in clinical diagnosis.
[0040] In this embodiment, the coordinate information of the region of interest (ROI) is read from the XML annotation file output by LabelImg, and the thyroid region in the original ultrasound image is automatically and accurately cropped out based on this information, generating a series of standardized local images (i.e., labeled images). Thus, a high-quality thyroid ultrasound image dataset focused on the ROI is constructed, laying a solid data foundation for subsequent training of a high-performance intelligent auxiliary diagnostic model. Specific effects are as follows... Figure 1 and Figure 2 As shown;
[0041] Step 2: Construct a thyroid function classification model based on the EfficientNet-B0 model. This thyroid function classification model incorporates several important architectural improvements based on the original EfficientNet-B0 model to better adapt to small-sample classification tasks of thyroid ultrasound images. First, a CBAM attention module is introduced between the last convolutional layer (final_conv) and the pooling layer of the EfficientNet-B0 model. This module includes two branches: channel attention and spatial attention. Channel attention learns the importance weights of each channel through global average pooling and fully connected layers. Spatial attention focuses on key spatial regions by combining max pooling and average pooling with convolutional operations, enabling the model to adaptively emphasize diagnostically valuable feature regions.
[0042] Meanwhile, the thyroid function classification model adds a layer normalization layer after feature extraction from the original EfficientNet-B0 model. This design significantly improves training stability, especially performance on small batches of data. By adding an embedding layer between the pooling layer and the classifier, the original feature dimension is compressed from 1280 dimensions to 128 dimensions. This not only reduces model parameters and prevents overfitting but also facilitates further metric learning or feature analysis. The classifier part uses an independent linear layer, separating the classifier from the embedding layer. This design allows the thyroid function classification model to be used for classification tasks or to extract feature vectors independently for other downstream applications.
[0043] In terms of regularization, the model integrates a Dropout layer and employs a configurable dropout rate, effectively enhancing generalization ability. The newly added embedding layers and classifier are initialized using a Kaiming normal distribution during initialization to ensure numerical stability in the early stages of training. Features are enhanced through an attention mechanism, and normalization and regularization techniques are combined to improve training robustness, providing a more accurate and stable model foundation for the intelligent assisted diagnosis of thyroid diseases. The model network architecture diagram is shown below. Figure 3 As shown.
[0044] Step 3: Train a thyroid function classification model based on a thyroid ultrasound image dataset;
[0045] In this embodiment, end-to-end training is performed on the EfficientNet-B0 model, which integrates the CBAM attention module, embedding layer, and Dropout regularization. Initially, all backbone feature extraction layers, the CBAM attention module, layer normalization layers, and embedding layers are frozen; only the topmost classifier is allowed to update its parameters. This initialization strategy ensures that the model can first learn the basic classification decision boundary based on a very stable set of features, avoiding getting trapped in local optima or overfitting due to inappropriate updates to deeper features.
[0046] A linear learning rate warm-up strategy was employed for the first 80 training epochs, allowing the learning rate to gradually and steadily increase from a very small base value. This effectively ensured numerical stability in the early stages of training. Throughout the training process, the Dropout layer randomly masked neurons before the embedding layer with a probability of 0.5. This ensured that the model did not overly rely on any single feature, but instead constructed a distributed feature representation, thereby greatly enhancing the model's generalization ability and solving the overfitting problem in small-sample medical image tasks.
[0047] From the 81st cycle onwards, a cosine annealing learning rate scheduler with hot restart is employed, causing the learning rate to periodically decrease and restart (within each cycle, the learning rate smoothly decreases from its initial value to its minimum value according to a cosine function; at the end of the cycle, the learning rate suddenly restarts to its initial value; at the start of a new cycle, it again smoothly decreases from high to low according to a cosine function), helping the model escape potential local optima. The exponential decay fine-tuning mechanism gradually approaches a minimum value, ensuring the model converges to the loss region, meaning better generalization performance. Therefore, in this cycle, the model also enables an exponential decay fine-tuning strategy, multiplying the current learning rate by a carefully set decay factor of 0.995 in each training cycle, causing the learning rate to glide smoothly downwards. This decay process continues until the learning rate reaches a preset minimum threshold of 1e-7.
[0048] The embedding layer is unfrozen at the 120th epoch. As a linear transformation from 1280 dimensions to 128 dimensions, the embedding layer's role is to extract high-dimensional, sparse features from the EfficientNet backbone and CBAM attention module, compressing them into a lower-dimensional, denser semantic space. This compression process not only reduces the parameter burden on subsequent classifiers but is also an important feature refinement process itself. With the embedding layer unfrozen, the model begins to learn how to retain the most discriminative information in this compact representation.
[0049] In subsequent cycles 200-300, the normalization layer and the CBAM attention module are unfrozen sequentially. The unfreezing of the CBAM attention module is particularly crucial, as it allows the model to dynamically adjust its focus. This module balances the importance of different feature maps through channel attention and focuses on key regions of interest in a two-dimensional plane through spatial attention, thereby improving the quality and discriminative power of feature representation.
[0050] At the 400th training epoch, the intermediate feature blocks of the deep backbone network were unfrozen, and the learning rate was simultaneously reduced to 0.6 times its original value. The deep backbone network implements feature knowledge from general to specific, and large-scale updates could destroy the prior knowledge embedded in the pre-trained weights. Therefore, slightly reducing the learning rate aims to allow these powerful feature extractors to fine-tune their parameters in a gentle way, making them better suited to the texture and structural characteristics of thyroid images.
[0051] Based on the EfficientNet-B0 model, this invention integrates the attention focusing ability of CBAM, the structural stability brought by Dropout, and the dense semantic representation provided by the embedding layer to construct a thyroid function classification model for small sample thyroid ultrasound images.
[0052] In this embodiment, the loss curve of the EfficientNet-B0 model (represented by EfficientNetB0+) that integrates the CBAM attention mechanism, embedding layer transformation, and Dropout regularization is as follows: Figure 4 As shown in the figure, a comparative analysis is performed between EfficientNetB0+ and the native EfficientNet-B0 model without incorporating the above three improvements. The ROC curves are shown in the figure. Figure 5 As shown in (b), the ROC curve generated after training with the native EfficientNet-B0 is as follows: Figure 5 As shown in (a), the evaluation metrics for the two models are shown in Table 1. The model accuracy improved from 0.7955 to 0.9773, and the AUC value improved from 0.8409 to 0.9793. Except for the recall, all other metrics improved.
[0053] Table 1 Evaluation metrics for the two models
[0054] Example 2:
[0055] This embodiment proposes an electronic device, including: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to execute the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images.
[0056] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.
[0057] The processor is used to execute all or part of the functions of the intelligent assisted diagnostic system for thyroid function based on small sample ultrasound images as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.
[0058] The processor can be implemented using an Application Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a controller, a microcontroller, a microprocessor, or other electronic components, and is used to execute the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images described in the above embodiments.
[0059] Example 3:
[0060] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.
[0061] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the functions of the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images described in the various embodiments of this application.
[0062] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various functions of the aforementioned intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images.
[0063] Example 4:
[0064] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images.
[0065] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.
[0066] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.
Claims
1. A smart auxiliary diagnostic system for thyroid function based on small sample ultrasound images, characterized in that, include: Obtain a thyroid dataset consisting of n cases and construct a small sample thyroid ultrasound image dataset; A thyroid function classification model was constructed based on the EfficientNet-B0 model. The thyroid function classification model added a CBAM attention module, a layer normalization layer, an embedding layer, and a Dropout layer to the original EfficientNet-B0 model. The classifier used an independent linear layer. A thyroid function classification model was trained based on a thyroid ultrasound image dataset. Unlabeled thyroid ultrasound images were acquired, and thyroid function was classified using a trained thyroid function classification model.
2. The intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images according to claim 1, characterized in that, The thyroid dataset consisting of n cases includes two standardized transverse ultrasound images of the thyroid gland for each case. The LabelImg annotation tool is used to annotate the bounding boxes of the thyroid region in the ultrasound images to obtain the region of interest, generate a series of standardized local images, and construct a small sample thyroid ultrasound image dataset.
3. The intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images according to claim 2, characterized in that, The thyroid function classification model introduces a CBAM attention module between the last convolutional layer (final_conv) and the pooling layer of the original EfficientNet-B0 model, and adds a layer normalization layer after feature extraction in the original EfficientNet-B0 model; an embedding layer is added between the pooling layer and the classifier to compress the original feature dimension; in terms of regularization, the thyroid function classification model integrates a Dropout layer and adopts a configurable dropout rate.
4. The intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images according to claim 3, characterized in that, The training process of the thyroid function classification model adopts a dynamically adjusted learning rate strategy, which includes three stages: linear warm-up, main training stage, and later fine-tuning. In the warm-up stage, a linear learning rate warm-up strategy is used to increase the learning rate. In the main training stage, a cosine annealing scheduler with restart is used to adjust the learning rate to maintain the model's learning vitality. In the later fine-tuning stage, the learning rate is finely tuned through exponential decay.
5. The intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images according to claim 4, characterized in that, End-to-end training of the thyroid function classification model was performed using the following method: At the beginning of training, all backbone feature extraction layers, CBAM attention modules, layer normalization layers, and embedding layers of the model are frozen, and only the topmost classifier is allowed to update its parameters. A linear learning rate warm-up strategy was used for the first 80 training cycles, gradually increasing the learning rate from a base value. Throughout the training process, the Dropout layer randomly masked neurons before the embedding layer with a probability of 0.
5. After the 81st training cycle, a cosine annealing learning rate scheduler with hot restart is used to periodically decrease and restart the learning rate; It also employs an exponential decay fine-tuning strategy, multiplying the current learning rate by a set decay factor in each training cycle to ensure a smooth decrease in the learning rate. Unfreeze the embedding layer on the 120th epoch of training; During the 200th-300th training cycle, the normalization layer and the CBAM attention module are unfrozen sequentially. When training reaches the 400th cycle, the intermediate feature blocks of the deep backbone network are unfrozen, and the learning rate is simultaneously reduced to 0.6 times the original value.
6. An electronic device for executing a small-sample ultrasound image intelligent auxiliary diagnostic system for thyroid function according to any one of claims 1-5, characterized in that, include: One or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the intelligent auxiliary diagnostic system for thyroid diseases based on small sample ultrasound images.
7. A computer-readable storage medium storing executable instructions for executing the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images according to any one of claims 1-5, characterized in that, When the instruction is executed, it causes the processor to execute the intelligent auxiliary diagnostic system for thyroid diseases based on small sample ultrasound images.
8. A computer program product for executing a small-sample ultrasound image intelligent auxiliary diagnostic system for thyroid diseases according to any one of claims 1-5, characterized in that, Includes a computer program or instructions that, when executed by a processor, implement the intelligent auxiliary diagnostic system for thyroid function based on small sample ultrasound images.