A thyroid nodule detection method based on dynamic fuzzy adaptive fusion

Abstract

The application discloses a thyroid nodule detection method based on dynamic blur adaptive fusion, relates to the technical field of medical images, and takes into account the problems of real-time performance and image quality blur; the method does not suppress noise by regarding blurred images as noise interference, but learns and understands the blur structure, and the process is realized by outputting clear features and blur features through a double-branch perception network; a clear-blur switching unit is designed according to the change of the scanning speed and the average Laplacian variance of the image, the strategy can be adaptively adjusted according to the blur degree of the input image, and the clear features and the blur features are dynamically mixed. Moreover, the application is a plug-and-play component, and can be organically embedded in mainstream target detection frameworks, such as the YoLO series and the DETR series.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion. Background Technology

[0002] Benefiting from the development of artificial intelligence technology, the use of AI models to achieve autonomous detection of thyroid nodules has become one of the mainstream research trends in medical ultrasound imaging. Accurate detection by AI models can help doctors further determine the nature of the nodules and formulate subsequent treatment measures such as follow-up observation or fine-needle aspiration scanning.

[0003] Most existing R&D models are based on stored image data. In actual examinations, lesion areas are judged through single-frame inspection and post-processing algorithms. This similar processing paradigm (single-frame prediction plus multi-frame decision-making) is widely used in smooth video scenarios such as traffic flow detection and pedestrian detection. However, during ultrasound scanning, due to the doctor's scanning speed and the real-time changes in the scanning area, ultrasound images often produce model artifacts or discontinuities. These low-quality images have a large inconsistency with the distribution of stored image data, which will seriously affect the model's lesion detection performance.

[0004] Some researchers have attempted to address these issues in different ways, such as: 1. using image restoration / reconstruction techniques to generate original high-resolution images; 2. using optical flow techniques and video consistency to compensate for motion blur; although related research has made progress in specific medical scenarios.

[0005] However, the following problems still exist: Ultrasound scans of thyroid nodules emphasize real-time performance, requiring an extremely high detection frame rate per second. Image restoration or reconstruction techniques are often developed based on generative models, which require additional time to restore the images. The higher the resolution of the generated image, the longer the time required, which may not meet the real-time requirement. Secondly, during the scan, the probe's cross-section and the scanned area are constantly and rapidly switching, such as switching the probe direction from transverse to longitudinal, or from the isthmus to the left lobe. Therefore, the video is not always consistent, and fusing the video information from consecutive frames has no physical meaning. Summary of the Invention

[0006] To address the above technical problems, this invention provides a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, comprising the following steps: S1. Collect different thyroid ultrasound images and corresponding ultrasound reports from multiple hospitals and public datasets, and perform data annotation and dataset partitioning; S2. Collect relevant videos of thyroid nodule areas during actual scans at multiple hospitals, and then annotate the video data frame by frame to form a video test set; S3. For each image, generate a motion trajectory pool to store multiple blurred images during motion; during model training, randomly select blurred images from the motion trajectory pool and combine them with the original images to form image pairs for input into the model. S4. Construct a dynamic fuzzy adaptive fusion component and insert it between the backbone network and the detection head of the existing target detection network framework to form a nodule detection model. The dynamic fuzzy adaptive fusion component includes a dual-branch perception network, which includes a clear image feature extraction branch and a dynamic fuzzy perception branch, respectively used to extract feature information of the original clear image and the fuzzy image. In the dynamic fuzzy perception branch, the original clear image is simulated by different degrees of probe jitter through a predefined optical flow data enhancement method to generate a fuzzy image. The fuzzy image is further processed by Gaussian filtering. Before being input into the detection head, the clear features and fuzzy features are dynamically weighted and fused layer by layer. S5. Train the nodule detection model and update the model parameters through backpropagation using the Adam optimizer; after each round of training, save the best model according to the map50 optimal principle of the validation set. S6. Verify the inference performance of the model using a video test set. During inference, the input for each subsequent frame, except for the first frame, is the image of the current frame and the image of the previous frame. Measure the inference performance of the model on real videos by calculating the recall and precision of nodule detection.

[0007] The technical solution further defined in this invention is: Furthermore, in step S1, the extent of the thyroid nodules is manually delineated based on the description in the ultrasound report; the data is divided into a training set, a validation set, and a test set in a ratio of 7:2:1.

[0008] As described above, in a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, step S4 includes a dynamic fuzzy perception branch comprising a clear-fuzzy switching unit and a feature pyramid network. The clear-fuzzy switching unit is used to select Gaussian kernels of different scales to filter the blurred image according to different probe jitter amplitudes, and to adaptively adjust the fusion weights of clear features and blurred features.

[0009] As described above, in a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, in step S4, the clear image feature extraction branch is set as a feature pyramid network, and multi-scale clear features are extracted using the following formula: ; In this context, FPN stands for Feature Pyramid Network, which processes the original sharp image I and outputs sharp multi-scale features. .

[0010] As described above, in a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, step S4 involves generating a blurred image based on optical flow motion simulation, as shown in the following formula: ; Here, MoveBlur() represents a predefined optical flow data enhancement method that simulates motion in a single image; for a single image, the blurred image generated by randomly simulating the jitter of the probe at different speeds is... .

[0011] As described above, in the dynamic fuzzy adaptive fusion-based thyroid nodule detection method, in the dynamic fuzzy perception branch of step S4, the fuzzy features of the image are modeled using Gaussian filtering, as shown in the following formula: ; ; ; Wherein, G represents the sharp-to-blur switching unit, which processes the original sharp image and the blurred image, and outputs a series of scale values. The larger the value, the stronger the Gaussian filter smooths the image; x and y represent the spatial coordinates of the blurred image; This represents the blurred image after Gaussian filtering, which is then processed using a feature pyramid network. Output fuzzy multi-scale features .

[0012] As described above, in a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, in step S4, clear features and fuzzy features are dynamically weighted and fused layer by layer, as shown in the following formula: ; in, This represents the weighting factor generated by the sharp-to-fuzzy switching unit. This represents the dot product.

[0013] As described above, a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, with a scale value... Generated by a sharp-blur switching unit, as shown in the following formula: ; ; ; in, Let H(x) represent the normalized histogram of the i-th gray level, L represent the number of gray levels, H(x) represent the estimated information entropy of a single image, and H(I) represent the information entropy of the original sharp image. The information entropy of a blurred image, The information gain represents the original sharp image and the blurred image; G represents the sharp-blur switching module, with different subscripts indicating different internal components; CNN represents a single-layer convolutional neural network; CA represents the interactive attention calculation module; MLP represents a multilayer perceptron, i.e., linear projection; Laplacian(I) represents the Laplacian operator mean of the original sharp image. This represents a multi-scale Gaussian kernel used for Gaussian filtering.

[0014] As described above, a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion adjusts the fusion strategy by real-time monitoring the mean value of the Laplacian operator in the original clear image and the information entropy of the corresponding fuzzy image, as shown in the following equation: ; in, Indicates convolution projection. This represents the weighting factor when fusing clear and fuzzy features.

[0015] As described above, in a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, step S5 involves a loss function during model training that includes classification loss, localization loss, and distribution loss, as shown in the following equation: ; in, Represents classification loss. Indicates location loss. This represents the distributed loss.

[0016] The beneficial effects of this invention are: (1) This invention significantly improves the model's anti-motion blurring ability and lesion detection accuracy in real ultrasound scanning scenarios. Existing technologies rely on static image training, which is prone to missed detection when faced with motion artifacts and deformations caused by rapid probe sliding. This method breaks the traditional thinking of "treating blur as noise that needs to be suppressed" and innovatively constructs a dual-branch perception network. By performing optical flow motion simulation and Gaussian filtering on the original image, the model actively learns and extracts significant topological features under extreme blur conditions. This mechanism fundamentally bridges the domain distribution difference between the static training set and the dynamic inference environment. Even when high-quality images are missing for a moment, the model can still accurately locate the nodule region by relying on the blur branch, which greatly improves the robustness of clinical detection. (2) While ensuring high-precision lesion detection, this invention perfectly meets the stringent requirements of extremely high real-time performance for clinical diagnosis of thyroid ultrasound. Traditional image restoration techniques rely on complex generative models to reconstruct high-definition images, resulting in extremely large computational overhead. The resulting inference delay makes it impossible to adapt to the high frame rate detection standard of ultrasound equipment. This invention abandons the time-consuming pixel-level reconstruction paradigm and designs dynamic blur adaptive technology as a lightweight, plug-and-play component. This component is directly embedded between the feature extraction layer and the detection head of mainstream target detection frameworks such as YOLO, and performs efficient weighted operations only in the feature space. This architecture design avoids the heavy image generation time and controls hardware memory usage, truly achieving smooth and lag-free real-time clinical detection. (3) This invention realizes an adaptive mechanism that integrates physical priors and data-driven approaches, giving the model excellent generalization ability for complex and varied scanning techniques. Due to the huge differences in the scanning speed and the frequency of section switching between doctors in clinical practice, traditional fixed-weight networks are prone to failure. The "clear-fuzzy switching unit" of this invention cleverly introduces the image Laplacian variance and the information entropy gain of the previous and next frames to quantify the spatial fuzziness and temporal abrupt change in real time. The network adaptively outputs Gaussian kernel parameters and feature fusion weights in this way: focusing on clear details during smooth scanning; automatically increasing the proportion of fuzzy features to prevent semantic pollution during violent shaking. This mechanism enables the model to get rid of its dependence on specific data and can be highly intelligently and flexibly adapted to any clinical operating environment. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the dual-branch sensing network in an embodiment of the present invention; Figure 2 This is a schematic diagram of the clear-to-fuzzy switching unit in an embodiment of the present invention. Detailed Implementation

[0018] This embodiment provides a method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, which includes the following steps: S1. Collect diverse thyroid ultrasound images and corresponding ultrasound reports from multiple hospitals and public datasets; accurately delineate the extent of thyroid nodules using the imaging knowledge of professional doctors and the descriptions in the ultrasound reports, ensuring no omissions or mislabeling; divide the data into training, validation, and test sets in a 7:2:1 ratio.

[0019] S2. Collect relevant videos of thyroid nodule areas during actual scans at multiple hospitals, and then annotate the video data frame by frame. All video data are used as out-of-domain test data to measure the model's generalization performance and performance on real videos.

[0020] S3. During training, a single image is constructed into a clear-blurred image pair. To accelerate model training, this process can be implemented in the preprocessing stage. For each image, this embodiment generates a motion trajectory pool to store multiple blurred images during motion. In actual training, it is only necessary to randomly select and input the image pair with the original image into the model. The method in this embodiment preprocesses the images in advance to simulate a series of different motion trajectories and form a binary input, which can avoid the time consumption caused by data processing during training.

[0021] S4. Construct a dynamic blur adaptive fusion component and insert it between the backbone and the detection head of the YOLO11 training framework to form a nodule detection model. The dynamic blur adaptive fusion component includes a dual-branch perception network, which includes a clear image feature extraction branch and a dynamic blur perception branch. The dynamic blur perception branch includes a clear-blur switching unit.

[0022] like Figure 1 As shown, the dual-branch sensing network is constructed based on FPN (Feature Pyramid Network). One branch is the traditional FPN, used to extract feature information from the original image; the other branch is a dynamic blur sensing branch, used to simulate the rapid shaking of the probe, thus adapting to the real environment.

[0023] The original clear image is then processed by a predefined optical flow data enhancement method to simulate different degrees of probe jitter, generating a blurred image. The blurred image is then further processed by Gaussian filtering. Although some details are lost, the significant structural features within the image are highlighted, which is beneficial for the model's understanding of the lesion area. Before being input into the detection head, the clear and blurred features are dynamically weighted and fused layer by layer to further increase the robustness of the features while avoiding semantic pollution.

[0024] When performing optical flow data enhancement, firstly, a simulated two-dimensional motion trajectory of the probe is generated using a Markov random walk algorithm. The intensity of the simulated jitter is controlled by adjusting the maximum pixel step size of the random walk. Next, a dense two-dimensional optical flow field is generated using Burmester noise to simulate the non-rigid deformation of soft tissue caused by the ultrasound probe pressing. Secondly, the input original high-resolution image is continuously translated N times along the motion trajectory, or a grid resampling is performed based on the dense optical flow field to generate N virtual intermediate transition images. Finally, a weighted average is applied to the pixels corresponding to the N intermediate transition images to generate a simulated blurred image with a realistic physical motion tail effect. In practical applications, the Albumentations library and the OpenCV2 library can be used to simulate the motion trajectory of ultrasound images.

[0025] The specific workflow of the dual-branch perceptron is shown in the following formula: (1); (2); (3); (4); (5); (6); Here, MoveBlur() represents a predefined method for simulating motion in a single image. For a single image, it randomly simulates the blurred image generated by the probe shaking at different speeds. FPN, or Feature Pyramid Network, processes the original image I to obtain multi-scale features. G is the sharp-blur switching unit, which processes sharp-blurred image pairs and outputs a series of scale values. The smaller the value, the less smooth the Gaussian filter is on the image, preserving more boundary details; the larger the value, the more intense the Gaussian filter is on the image, preserving only key structural information.

[0026] Formula (4) represents the Gaussian filtering process for image processing, where x and y represent the spatial coordinates of the image, ultimately yielding the input image for the dynamic blur perception branch. Similar to the original image, FPN is used to extract fuzzy multi-scale features. Finally, the clear features and fuzzy features are weighted and fused using formula (6), with the weighting factors being... Generated by a sharp-blur switching unit. This represents the dot product.

[0027] During video detection, if the probe scans smoothly, the video will remain clear and continuous. Conversely, if the probe experiences significant shaking, the video content will be blurry and discontinuous. To accommodate both scenarios, this embodiment incorporates a Clear-Blur Switching Unit for the following judgments:

[0028] 1. Whether to use a large-scale or small-scale Gaussian kernel to filter the image. If the probe jitter is large (the image is relatively blurry), it is desirable to use a series of large scale factors to extract large-scale structural information. If the probe jitter is small (the image is relatively clear), it is desirable to use a series of small scale factors to highlight boundary details. It is worth mentioning that during the training process, this embodiment will predict a next frame image for the captured image in each training process, forming a clear-blurred image pair input into the network. That is, through optical flow data enhancement, a motion image is simulated for the current image data. In the actual inference process, the input to the network is the image of the current frame and the previous frame.

[0029] 2. When performing feature fusion, should more clear features or blurred features be retained? This embodiment designs an adaptive decision-making mechanism through a network. By monitoring the mean of the Laplacian operator of the feature map and the information entropy of the blurred image (the information entropy of the two frames before and after during inference) in real time, the fusion strategy is adjusted. If a highly blurred region is detected, the fusion weight of the blurred features will be automatically increased, that is... (As mentioned in Formula 6) will automatically decrease; conversely, if a low-blur region is detected, then It will automatically upgrade.

[0030] like Figure 2 As shown, the sharp-to-fuzzy switching unit is the core adaptive control hub of the method in this embodiment. Its main function is to dynamically output a Gaussian filter scaling factor sequence and sharp-to-fuzzy feature fusion weights by sensing the spatiotemporal continuity and spatial fuzziness of the input video frames in real time. The specific processing flow of the clear-to-fuzzy switching unit can be divided into three stages: state quantity extraction, feature interaction, and adaptive parameter decoding.

[0031] Phase 1, State Variable Extraction: The Laplacian operator variance of the original image is calculated. The Laplacian variance is highly sensitive to high-frequency edge information in the image. When the probe shakes rapidly, causing image blurring, high-frequency details are lost, and the Laplacian variance decreases significantly. This variance value is normalized and used as a spatial blur prior feature input into the subsequent network. Next, the gain between the information entropy of the original image and the motion-blurred image (current frame and previous frame for prediction) is calculated. The information gain reflects the degree of drastic change in content between adjacent frames. When there are abrupt changes in the cross-section or rapid probe movement, the information gain fluctuates drastically. The mathematical expression for this process is as follows: (7); (8); in, Let L represent the normalized histogram of the i-th gray level, L represent the number of gray levels, which is usually 256, and Equation (7) is the expression for estimating the information entropy of a single image. This represents the information gain of the two images.

[0032] Phase Two, Feature Interaction: To organically integrate physical priors (information gain) with the image's intrinsic features (Laplacian operator), the sharp-to-fuzzy switching unit employs a lightweight decision network. First, both pass through a shared perceptual module constructed from a convolutional neural network. Extract shallow vector representations containing spatiotemporal motion patterns; these are respectively represented as spatial fuzzy features. and temporal mutation characteristics Next, the features extracted by the model are concatenated and processed through a self-attention mechanism. study and Nonlinear mapping relationship; (9); (10); (11); in, and This represents the spatial imitation features and temporal abrupt change features after processing by the self-attention mechanism.

[0033] Phase 3, Adaptive Parameter Decoding and Output: The implicit features after interactive fusion are finally fed into a multilayer perceptron or projection network for dimensionality reduction and parameter decoding, and are divided into two parallel output branches: the Gaussian scaling factor output branch and the feature fusion weight output branch.

[0034] Gaussian scaling factor output branch: If the motion is determined to be violent, the network automatically outputs a larger scaling factor. The value combination (applied to Equation 4) drives the dynamic blur perception branch to smooth the image using a large-scale Gaussian kernel, filtering out cluttered artifacts caused by motion and forcing the model to focus on the large-scale topological structure of the lesion; if it is determined to be a smooth scan, a smaller output is given. Values ​​that preserve the details of lesion boundaries.

[0035] Feature fusion weight output branch: When a highly blurred region is detected, The value automatically decreases, thereby increasing the dominance of fuzzy features in the final prediction; conversely, if a low-fuzzy region is detected, then... It will automatically upgrade.

[0036] The mathematical expression for the above process is as follows: (12); (13); in, This represents the multilayer perceptron network in the module. This represents the projection network in the module.

[0037] S5. Train the nodule detection model and update the model parameters through backpropagation using the Adam optimizer. After each round of training, save the best model according to the optimal principle of map50 (a commonly used evaluation metric for detection) on the validation set.

[0038] The dynamic blur adaptive fusion component is a plug-and-play component, inserted between the model backbone and the detection head. Its purpose is to incorporate the image's dynamic blur into learnable structural information. Within the YOLO11 detection framework, the loss function during training includes classification loss, localization loss, and distribution loss, as shown in the following equation: (14); in, Represents classification loss. Indicates location loss. This represents the distributed loss.

[0039] S6. Verify the inference performance of the model using a video test set. During inference, except for the first frame which performs normal inference, the input for each subsequent frame is the image of the current frame and the image of the previous frame. The recall and precision of the node detection are calculated to measure the inference performance of the model on real videos.

[0040] This embodiment takes into account both real-time performance and image quality blurring issues, and designs a thyroid nodule detection method based on dynamic blur adaptive fusion. This method does not treat blurry images as noise interference and suppress noise, but learns to understand the blurry structure. This process can be achieved by outputting clear features and blurry features through a two-branch perceptual network.

[0041] Furthermore, this embodiment's method designs a sharp-blur switching unit based on changes in scanning speed and the average Laplacian variance of the image. This unit can adaptively adjust the strategy according to the blur level of the input image, dynamically mixing sharp and blurry features. It is worth noting that this invention is a plug-and-play component that can be organically embedded into mainstream object detection frameworks, such as the YOLO and DETR series. Considering hardware memory and real-time requirements, this embodiment chooses to embed it based on the YOLO11 series.

[0042] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion, characterized in that: Includes the following steps: S1. Collect different thyroid ultrasound images and corresponding ultrasound reports from multiple hospitals and public datasets, and perform data annotation and dataset division; S2. Collect relevant videos of thyroid nodule areas during actual scans at multiple hospitals, and then annotate the video data frame by frame to form a video test set; S3. For each image, generate a motion trajectory pool to store multiple blurred images during motion; during model training, randomly select blurred images from the motion trajectory pool and combine them with the original images to form image pairs for input into the model. S4. Construct a dynamic fuzzy adaptive fusion component and insert it between the backbone network and the detection head of the existing target detection network framework to form a nodule detection model. The dynamic fuzzy adaptive fusion component includes a dual-branch perception network, which includes a clear image feature extraction branch and a dynamic fuzzy perception branch, respectively used to extract feature information of the original clear image and the blurred image. In the dynamic blur perception branch, the original clear image is simulated by different degrees of probe jitter through a predefined optical flow data enhancement method to generate a blurred image. The blurred image is further processed by Gaussian filtering. Before being input into the detection head, the clear features and blurred features are dynamically weighted and fused layer by layer. S5. Train the nodule detection model and update the model parameters through backpropagation using the Adam optimizer; after each round of training, save the best model according to the map50 optimal principle of the validation set. S6. Verify the inference performance of the model using a video test set. During inference, the input for each subsequent frame, except for the first frame, is the image of the current frame and the image of the previous frame. Measure the inference performance of the model on real videos by calculating the recall and precision of nodule detection.

2. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 1, characterized in that: In step S1, the extent of the thyroid nodules is manually delineated based on the description in the ultrasound report; the data is divided into a training set, a validation set, and a test set in a ratio of 7:2:

1.

3. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 1, characterized in that: In step S4, the dynamic blur perception branch includes a sharp-blur switching unit and a feature pyramid network. The sharp-blur switching unit is used to select Gaussian kernels of different scales to filter the blurred image according to different probe jitter amplitudes, and to adaptively adjust the fusion weights of sharp features and blurred features.

4. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 3, characterized in that: In step S4, the sharp image feature extraction branch is set as a feature pyramid network, and multi-scale sharp features are extracted using the following formula: ； In this context, FPN stands for Feature Pyramid Network, which processes the original sharp image I to output sharp multi-scale features. .

5. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 4, characterized in that: In step S4, a blurred image is generated based on optical flow motion simulation, as shown in the following formula: ； Here, MoveBlur() represents a predefined optical flow data enhancement method that simulates motion in a single image; for a single image, the blurred image generated by randomly simulating the jitter of the probe at different speeds is... .

6. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 5, characterized in that: In the dynamic blur perception branch of step S4, the blur features of the image are modeled by Gaussian filtering, as shown in the following formula: ； ； ； Wherein, G represents the sharp-to-blur switching unit, which processes the original sharp image and the blurred image, and outputs a series of scale values. The larger the value, the stronger the Gaussian filter smooths the image; x and y represent the spatial coordinates of the blurred image; This represents the blurred image after Gaussian filtering, which is then processed using a feature pyramid network. Output fuzzy multi-scale features .

7. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 6, characterized in that: In step S4, the clear features and the fuzzy features are dynamically weighted and fused layer by layer, as shown in the following formula: ； in, This represents the weighting factor generated by the sharp-to-fuzzy switching unit. This represents the dot product.

8. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 6, characterized in that: The scale value Generated by a sharp-blur switching unit, as shown in the following formula: ； ； ； in, Let H(x) represent the normalized histogram of the i-th gray level, L represent the number of gray levels, H(x) represent the estimated information entropy of a single image, and H(I) represent the information entropy of the original sharp image. The information entropy of a blurred image, The information gain represents the original sharp image and the blurred image; G represents the sharp-blur switching module, with different subscripts indicating different internal components; CNN represents a single-layer convolutional neural network; CA represents the interactive attention calculation module; MLP represents a multilayer perceptron, i.e., linear projection; Laplacian(I) represents the Laplacian operator mean of the original sharp image. This represents a multi-scale Gaussian kernel used for Gaussian filtering.

9. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 8, characterized in that: The sharp-blur switching unit adjusts the fusion strategy by real-time monitoring the mean value of the Laplacian operator of the original sharp image and the information entropy of the corresponding blurred image, as shown in the following formula: ； in, Indicates convolution projection. This represents the weighting factor when fusing clear and fuzzy features.

10. The method for detecting thyroid nodules based on dynamic fuzzy adaptive fusion according to claim 1, characterized in that: In step S5, the loss function during model training includes classification loss, localization loss, and distribution loss, as shown in the following formula: ； in, Represents classification loss, Indicates location loss. This represents the distributed loss.

Images