Method for constructing thyroid nodule detection model based on intermediate frequency feature interaction and application thereof

Abstract

This application proposes a method and application for constructing a thyroid nodule detection model based on mid-frequency feature interaction. The method includes: acquiring multiple training images, wherein the training images are thyroid ultrasound images annotated with thyroid nodules; constructing a thyroid nodule detection architecture consisting of a backbone network, a feature fusion network, and a classification network connected sequentially; and training the thyroid nodule detection architecture using multiple training images to obtain the constructed thyroid nodule detection model. This scheme uses dilated convolutional units with depthwise separable convolutions at different dilation rates for feature processing, which can aggregate mid-range context while preserving details of small nodules. Furthermore, the mid-frequency feature interaction unit achieves efficient coupling between local details and mid-range semantics, thereby enabling the thyroid nodule detection model to achieve better classification performance.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and in particular to a method for constructing and applying a thyroid nodule detection model based on mid-frequency feature interaction. Background Technology

[0002] Thyroid nodules are one of the most common thyroid diseases worldwide, with a prevalence rate of 10% to 20%. Among them, 5% to 10% of nodules have a risk of malignancy. Early and accurate diagnosis is crucial for improving cure rates and patient prognosis. Ultrasound imaging, with its advantages of being non-invasive, convenient, and low-cost, has become the preferred imaging tool for diagnosing thyroid nodules and is widely used in clinical practice. However, the quality of thyroid ultrasound images is easily affected by various factors such as operator experience, patient size, and equipment differences. Common problems include noise interference, insufficient contrast, and blurred nodule boundaries. Especially when processing small nodules (<10mm accounting for more than 80%), complex-shaped nodules, and low-resolution images, traditional image analysis methods rely on manual annotation, which is not only inefficient but also prone to misdiagnosis and missed diagnosis, making it difficult to meet the needs of accurate clinical diagnosis.

[0003] With the rapid development of deep learning technology, AI-based medical image analysis has provided a new solution for the diagnosis of thyroid nodules. It can reduce human intervention, improve automation, and significantly enhance detection accuracy and efficiency, providing powerful assistance to clinicians. However, existing deep learning models still face key bottlenecks: on the one hand, in environments with low computing resources, models struggle to balance detection accuracy and computational efficiency, making them difficult to adapt to low-computing-power clinical equipment; on the other hand, existing feature extraction mechanisms have limitations in addressing the specific problems of low signal-to-noise ratios (typically <15dB), large differences in nodule scale, weak edge signals, and susceptibility to speckle noise in thyroid ultrasound images. Global self-attention mechanisms are computationally redundant and resource-intensive, fixed-window local attention mechanisms cannot flexibly adapt to lesion morphology, and single-path convolution cannot simultaneously capture fine-grained textures and multi-scale context. This results in insufficient detection accuracy and high false negative and false positive rates in low-quality images, small nodules, and blurred boundary scenarios. Therefore, developing an intelligent detection solution for thyroid nodule ultrasound images that balances lightweight design with high-precision detection capabilities and adapts to complex clinical scenarios has become an urgent need in the field of medical image processing. Summary of the Invention

[0004] This application provides a method and application for constructing a thyroid nodule detection model based on mid-frequency feature interaction. It uses dilated convolution units with depth separable convolutions at different dilation rates for feature processing, which can aggregate mid-range context while preserving the details of small nodules. Then, the mid-frequency feature interaction unit achieves efficient coupling between local details and mid-range semantics, thereby enabling the thyroid nodule detection model to have better classification performance.

[0005] In a first aspect, embodiments of this application provide a method for constructing a thyroid nodule detection model based on mid-frequency feature interaction, the method comprising:

[0006] Multiple training images are acquired, wherein the training images are thyroid ultrasound images with thyroid nodules labeled.

[0007] A thyroid nodule detection architecture is constructed, consisting of a backbone network, a feature fusion network, and a classification network connected sequentially. The backbone network extracts features from training images using a feature pyramid structure, obtaining multiple first feature maps with different channel numbers. The feature fusion network includes an input branch, a first fusion branch, and a second fusion branch. The input branch processes each first feature map using dilated convolutional units to obtain corresponding first input features. The first fusion branch includes multiple cascaded first fusion modules, each consisting of a dilated convolutional unit and an upsampling layer connected sequentially. The first input features with different channel numbers are fused based on these multiple first fusion modules to obtain a first fusion output. The second fusion branch includes multiple cascaded second fusion modules and a second output module. The two-way fusion module consists of a mid-frequency feature interaction unit and a dilated convolution unit connected sequentially. The second output module is the mid-frequency feature interaction unit. Based on multiple second fusion modules, the first fusion output is fused with the output of each dilated convolution unit in the first fusion branch to obtain an intermediate fusion result. The intermediate fusion result is input into the second output module to obtain the second fusion output. The second fusion output is then fused with the output of the mid-frequency feature interaction unit in each second fusion module to obtain the feature fusion result. The dilated convolution unit processes the input features with dilation convolutions of different dilation rates. The mid-frequency feature interaction unit preserves the boundary and texture details of the nodules by capturing the local contextual dependencies of the input features. The classification network obtains the thyroid nodule diagnosis result based on the feature fusion result.

[0008] The thyroid nodule detection architecture is trained using multiple training images to obtain a well-constructed thyroid nodule detection model.

[0009] Secondly, embodiments of this application provide an application method for a thyroid nodule detection model based on mid-frequency feature interaction, including:

[0010] A thyroid ultrasound image is obtained as the image to be detected. The image to be detected is then input into the thyroid nodule detection model to obtain the diagnostic result of the thyroid nodule corresponding to the image to be detected.

[0011] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to execute a method for constructing a thyroid nodule detection model based on mid-frequency feature interaction or an application method for a thyroid nodule detection model based on mid-frequency feature interaction.

[0012] The main contributions and innovations of this invention are as follows:

[0013] This scheme employs a multi-dimensional interactive enhancement unit as the backbone network. By adding the input features through deep convolutions and standard convolutions within the multi-dimensional interactive enhancement unit, combined with a multi-scale contextual attention mechanism and a residual network, it efficiently extracts first feature maps with different channel numbers, enhancing feature representation capabilities and improving model training stability and learning efficiency, laying a high-quality foundation for subsequent feature fusion. The dilated convolution unit uses parallel dilated convolutional layers with multiple dilation rates to respond to nodules of different scales. While preserving details of small nodules, it aggregates mid-range context, mitigating grid effects. Combined with a group-level channel-aware attention mechanism and channel shuffling operations, it suppresses noise, enhances nodule signals, and achieves efficient coupling of multi-scale features. Furthermore, this scheme places mid-frequency feature interaction units between the first fusion modules. Mid-frequency deep convolution branches preserve nodule boundaries and texture details, while mid-frequency context branches aggregate mid-range semantics, achieving efficient coupling of local details and mid-range semantics, and strengthening key mid-scale visual features.

[0014] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1 This is an overall structural diagram of a thyroid nodule detection model based on mid-frequency feature interaction according to an embodiment of this application;

[0017] Figure 2 This is a structural diagram of a multi-dimensional interaction enhancement unit according to an embodiment of this application;

[0018] Figure 3 This is a structural diagram of a multi-scale contextual attention mechanism layer according to an embodiment of this application;

[0019] Figure 4 This is a structural diagram of a dilated convolutional unit according to an embodiment of this application;

[0020] Figure 5This is a structural diagram of a position-driven adaptive attention mechanism unit according to an embodiment of this application;

[0021] Figure 6 This is a structural diagram of a mid-frequency feature interaction unit according to an embodiment of this application;

[0022] Figure 7 This is a visual comparison of the detection results of the thyroid nodule detection model according to the embodiments of this application and other models on the TN3K dataset;

[0023] Figure 8 This is a visual comparison of the detection results of the thyroid nodule detection model according to the embodiments of this application and other models on the DDTI dataset;

[0024] Figure 9 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0026] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0027] Example 1

[0028] This application provides a method for constructing a thyroid nodule detection model based on mid-frequency feature interaction. It uses dilated convolutional units with depthwise separable convolutions at different dilation rates for feature processing, which can aggregate mid-range context while preserving details of small nodules. Then, through mid-frequency feature interaction units, it achieves efficient coupling between local details and mid-range semantics, thereby enabling the thyroid nodule detection model to have better classification performance. Specifically, refer to... Figure 1 The method includes:

[0029] Multiple training images are acquired, wherein the training images are thyroid ultrasound images with thyroid nodules labeled.

[0030] A thyroid nodule detection architecture is constructed, consisting of a backbone network, a feature fusion network, and a classification network connected sequentially. The backbone network extracts features from training images using a feature pyramid structure, obtaining multiple first feature maps with different channel numbers. The feature fusion network includes an input branch, a first fusion branch, and a second fusion branch. The input branch processes each first feature map using dilated convolutional units to obtain corresponding first input features. The first fusion branch includes multiple cascaded first fusion modules, each consisting of a dilated convolutional unit and an upsampling layer connected sequentially. The first input features with different channel numbers are fused based on these multiple first fusion modules to obtain a first fusion output. The second fusion branch includes multiple cascaded second fusion modules and a second output module. The two-way fusion module consists of a mid-frequency feature interaction unit and a dilated convolution unit connected sequentially. The second output module is the mid-frequency feature interaction unit. Based on multiple second fusion modules, the first fusion output is fused with the output of each dilated convolution unit in the first fusion branch to obtain an intermediate fusion result. The intermediate fusion result is input into the second output module to obtain the second fusion output. The second fusion output is then fused with the output of the mid-frequency feature interaction unit in each second fusion module to obtain the feature fusion result. The dilated convolution unit processes the input features with dilation convolutions of different dilation rates. The mid-frequency feature interaction unit preserves the boundary and texture details of the nodules by capturing the local contextual dependencies of the input features. The classification network obtains the thyroid nodule diagnosis result based on the feature fusion result.

[0031] The thyroid nodule detection architecture is trained using multiple training images to obtain a well-constructed thyroid nodule detection model.

[0032] In the current embodiment, the training images are first converted into feature embeddings using an embedding layer in the backbone network, and then the feature pyramid structure is used to extract features from the feature embeddings of the training images.

[0033] Specifically, the embedding layer consists of two 3×3 convolutional layers and a Gelu activation function layer between the two 3×3 convolutional layers, with a stride of 4.

[0034] In the current embodiment, the feature pyramid structure of the backbone network consists of multiple serially connected feature extraction modules. Each feature extraction module includes multiple serially connected multidimensional interaction enhancement units. The step size of the multidimensional interaction enhancement unit in the first feature extraction module is 1, and the step size of the multidimensional interaction enhancement unit in the remaining feature extraction modules is 2. The output of each non-first feature extraction module is used as the first feature map.

[0035] For example, this solution consists of four feature extraction modules connected in series to form a feature pyramid structure. The first feature extraction module contains three multi-dimensional interaction enhancement units, the second feature extraction module contains four multi-dimensional interaction enhancement units, the third feature extraction module contains fifteen multi-dimensional interaction enhancement units, and the fourth feature extraction module contains two multi-dimensional interaction enhancement units.

[0036] Furthermore, the structure of the multidimensional interaction enhancement unit is as follows: Figure 2 As shown, in the multi-dimensional interaction enhancement unit, the input features are subjected to depthwise convolution and standard convolution respectively. The input features, the depthwise convolution result and the standard convolution result are summed to obtain the first multi-dimensional interaction output. The first multi-dimensional interaction output is calculated by a multi-scale contextual attention mechanism to obtain the second multi-dimensional interaction output. The second multi-dimensional interaction output is passed through a residual network to obtain the third multi-dimensional interaction output. The third multi-dimensional interaction output is the output feature of the multi-dimensional interaction enhancement unit. The output feature of the last multi-dimensional interaction enhancement unit in the feature extraction module is the output of the feature extraction module.

[0037] Specifically, the input features of the multidimensional interaction enhancement unit are: Input features For a three-dimensional tensor, the formula is expressed as:

[0038]

[0039] Where H is the height of the image, and W is the width of the image. This represents the number of channels.

[0040] The deep convolutions used in the multidimensional interaction enhancement unit are 3×3 in size. Deep convolutions process each channel of the input features using an independent convolution kernel, thereby reducing computational complexity and preserving the features of each channel. The deep convolution result remains unchanged in spatial size and number of channels compared to the input features. The formula for deep convolution is expressed as:

[0041]

[0042] in, This is the result of depthwise convolution. This represents depthwise convolution.

[0043] The standard convolution used in the multidimensional interaction unit is 1×1 in size. Standard convolution is used for feature fusion. Although 1×1 convolution does not change the spatial size, it enhances the expressive power of the input features through convolution operations along the channel dimension. The formula for standard convolution is expressed as:

[0044]

[0045] in, The result is a standard convolution; Conv represents the standard convolution.

[0046] Then, by adding the input features, the standard convolution result, and the depthwise convolution result, the multi-level feature information is fused, ensuring the transmission and flow of feature information. The formula is expressed as:

[0047]

[0048] in, This is the first multi-dimensional interactive output.

[0049] Specifically, this scheme calculates the multi-scale contextual attention mechanism for the first multi-dimensional interactive output after batch normalization. The batch normalization formula is expressed as:

[0050]

[0051] in, Indicates batch normalization, This is the result after batch normalization.

[0052] Specifically, this scheme uses a multi-scale contextual attention mechanism layer to perform the computation of the multi-scale contextual attention mechanism. The structure of the multi-scale contextual attention mechanism layer (MLCA) is as follows: Figure 3 As shown, the multi-scale contextual attention mechanism layer performs tensor reshaping on the local pooling results and global pooling results of the input features respectively. The tensor reshaping results of the local pooling results and the tensor reshaping results of the global pooling results are added together to obtain the reshaping intermediate features. The reshaping intermediate features are then multiplied element-wise with the input features to obtain the second multidimensional interactive output.

[0053] The formula for calculating multi-scale contextual attention mechanism is expressed as follows:

[0054]

[0055] in, For the second multi-dimensional interactive output, This represents the computation of a multi-scale contextual attention mechanism.

[0056] In the multi-scale contextual attention mechanism layer, local pooling is performed on the input features to obtain the local pooling result, and then global pooling is performed on the local pooling result to obtain the global pooling result. Through the multi-scale contextual attention mechanism, the expression of local region features is enhanced, and the overall perception capability of the image is improved by modeling global information.

[0057] Specifically, in the residual network, the second multidimensional interactive output is passed through a 1×1 convolution, a Gelu activation function, and another 1×1 convolution to obtain the residual intermediate features. The residual intermediate features are then residually concatenated with the second multidimensional interactive output to obtain the third multidimensional interactive output.

[0058] The nonlinear expressive power of the second multidimensional interactive output can be further enhanced by using 1×1 convolution and the Gelu activation function, thereby capturing more complex patterns and relationships. The formula is expressed as follows:

[0059]

[0060] in, The output of the second multidimensional interactive output is obtained through a 1×1 convolution and the Gelu activation function. For 1×1 convolution, This is the Gelu activation function.

[0061] Then use another 1×1 convolution to The number of channels from Adjust back To maintain a consistent number of channels in residual connections, the formula is expressed as:

[0062]

[0063] in, As an intermediate feature of the residual, It is a 1×1 convolution.

[0064] Finally, the intermediate residual features are concatenated with the second multidimensional interactive output to improve the training stability and learning efficiency of the model. The formula is as follows:

[0065]

[0066] in, This is the third multi-dimensional interactive output.

[0067] In the current embodiment, the number of dilated convolution units within the input branch is the same as the number of first feature maps; that is, one dilated convolution unit is allocated to each first feature map. The structure of the dilated convolution unit is as follows: Figure 4As shown, the dilated convolution unit uses parallel first, second, and third dilated convolution layers to process the input features to obtain the first, second, and third dilated convolution results, respectively. The first, second, and third dilated convolution results are added element-wise to obtain the comprehensive dilated convolution result. The comprehensive dilated convolution result is then concatenated with the input features and sequentially subjected to a group-level channel-aware attention mechanism and channel shuffling operation to obtain the output. The dilation rate of the first dilated convolution layer is 1, the dilation rate of the second dilated convolution layer is 2, and the dilation rate of the third dilated convolution layer is 3.

[0068] Specifically, setting the dilation rate of the first dilated convolutional layer to 1 can respond to small nodules with a diameter of less than 3 mm in thyroid ultrasound images, preserving high-resolution local details. Setting the dilation rate of the second dilated convolutional layer to 2 can respond to small nodules of 3-5 mm in thyroid ultrasound images, capturing the main structure of the nodule while taking into account the surrounding tissue context. Setting the dilation rate of the third dilated convolutional layer to 3 can respond to larger nodules greater than 5 mm in thyroid ultrasound images and their surrounding infiltrative areas, effectively aggregating long-distance semantics.

[0069] Specifically, a group-level channel-aware attention mechanism is used to process the combined dilated convolution result and the input feature concatenation result. The weight generation introduces a priori local energy distribution, explicitly models the Rayleigh distribution characteristics of speckle noise in ultrasound images, dynamically suppresses false responses in high-energy background regions, and enhances the activation intensity of low-intensity nodule signals. Finally, cross-path information interaction is achieved through channel shuffling operations to ensure that multi-scale features are fully coupled under lightweight constraints.

[0070] Specifically, the input features are processed by 1×1 convolution in the dilated convolution unit, thereby performing channel compression and linear mapping on the input features, reducing computational cost and unifying feature dimensions. The formula is expressed as:

[0071]

[0072] in, This is the result after a 1×1 convolution. The input features are those of the dilated convolution unit. It is the number of channels after convolution adjustment.

[0073] Processing the input features using parallel first, second, and third dilated convolutional layers can capture contextual information at different scales with extremely low computational overhead. The formula is as follows:

[0074]

[0075] in, This is the result of the first dilated convolution. This is the result of the second dilated convolution. This is the result of the third dilated convolution. This is the first dilated convolutional layer with a dilation rate of 1. This is the second dilated convolutional layer with a dilation rate of 2. It is the third dilated convolutional layer with an inflation rate of 3.

[0076] Adding the results of the first, second, and third dilated convolutions element-wise can effectively alleviate the grid effect of a single dilated convolution and enhance the model's boundary awareness. The formula is as follows:

[0077]

[0078] in, This is to synthesize the results of dilated convolution.

[0079] Concatenating the combined dilated convolution result with the input features can fuse shallow detail features with deep contextual information, forming a more discriminative multi-scale feature representation, expressed by the formula:

[0080]

[0081] in, It is a concatenation result of the combined dilated convolution and the input features. This represents the number of channels after splicing.

[0082] The group-level channel-aware attention mechanism adaptively emphasizes key feature regions and suppresses background noise by learning the weights of channels and spatial dimensions. The formula is expressed as:

[0083]

[0084] in, The output of the group-level channel-aware attention mechanism is CAA.

[0085] Finally, a channel shuffling operation is used to break the fixed grouping relationship between channels, thereby improving the effectiveness of feature fusion and the generalization ability of the model. The formula is expressed as:

[0086]

[0087] in, This is the output of the channel shuffling operation, i.e., the output of the dilated convolution unit. Shuffling operation for the channel.

[0088] Specifically, the dilated convolutional units in different branches of this scheme have the same structure, so they will not be described in detail here.

[0089] In the current embodiment, the first fusion module in the first fusion branch performs feature processing on the first input feature with the largest number of channels and then concatenates it with another first input feature of the corresponding size to obtain the first fusion intermediate feature. The first fusion intermediate feature is then input into the next first fusion module. After all the first input features are concatenated, the first fusion output is obtained.

[0090] Specifically, since the backbone network extracts features from the training images using a feature pyramid structure, the image size of the multiple first input features it acquires becomes smaller and smaller, while the number of channels becomes larger and larger. In the first fusion module, an upsampling layer is used to process the first input feature with the largest number of channels, increase its image size, and then concatenate it with the first input feature of the corresponding image size to complete the fusion of first input features with different numbers of channels.

[0091] Furthermore, before the first fusion module processes the first input feature with the largest number of channels, a position-driven adaptive attention mechanism unit is used to process the first input feature with the largest number of channels. The structure of the position-driven adaptive attention mechanism unit is as follows: Figure 5 As shown, in the position-driven adaptive attention mechanism unit, a local sliding candidate window is defined for the input features to obtain a window feature map. The window feature map is then expanded into a window feature sequence by row. The position offset and position attention weight of each window in the window feature sequence are predicted by a dynamic position offset attention mechanism. Based on the position offset of each window, the input features are adjusted to obtain offset features. The offset features are then weighted by the position attention weights to obtain weighted offset features. After reshaping the weighted offset features, they are concatenated with the window feature map as the output of the position-driven adaptive attention mechanism unit. The output of the position-driven adaptive attention mechanism unit is the input of the first fusion module in the first fusion branch.

[0092] Specifically, the position-driven adaptive attention mechanism unit implements dynamic sampling and weighting within the feature layer, combining a variable sampling domain with single convolutional-level computational overhead. It provides continuous and sharpened attention features for blurred edges and tiny nodules, avoiding the redundancy of global traversal and breaking through the rigid limitations of fixed windows.

[0093] Specifically, in the adaptive attention mechanism unit, a 1×1 convolution is first used to compress the input features through channels and unify their dimensions, reducing the computational cost of subsequent offset prediction. The formula is as follows:

[0094]

[0095] in, This is the result after a 1×1 convolution. The input features are for the position-driven adaptive attention mechanism unit.

[0096] This scheme defines a 20×20 local sliding candidate window to obtain a window feature map. The window feature map is then expanded into a window feature sequence, where each token corresponds to a spatial location. Two learnable parameters, position offset and position attention weights, are obtained by predicting the window feature sequence using a dynamic position offset attention mechanism. In this mechanism, the relative distance between two windows is received as input, and their positional deviation and position attention weights are output. Then, bilinear interpolation is used to adjust the position of the input features based on the positional deviation, as expressed by the formula:

[0097]

[0098] in, This represents the position offset of window i. Let be the coordinates of the feature within window i in the window feature map. This represents the position offset of window i. For bilinear interpolation, This is an offset feature.

[0099] The offset features are then weighted using positional attention weights, as expressed by the formula:

[0100]

[0101] in, The weighted offset feature is represented by i, where i is the window index. This represents the total number of sampling points within the window. Let i be the position attention weight for window i.

[0102] The formula for reshaping the weighted offset features is expressed as follows:

[0103]

[0104] in, The result is the reshaping of the weighted offset features. For reshaping operations.

[0105] Concatenating the reshaped weighted offset features with the window feature map can fuse shallow details and dynamic contextual information, as expressed by the formula:

[0106] ，

[0107] in, This is the result of concatenating the reshaped weighted offset features with the window feature map. This is a window feature map.

[0108] The reconstructed weighted offset features are concatenated with the window feature map using 1×1 convolution and channel shuffling operations, thereby breaking grouping artifacts and improving cross-channel information flow. The formula is expressed as:

[0109] ，

[0110] in, The output of the position-driven adaptive attention mechanism unit. Shuffling operations for the channel. It is a 1×1 convolution.

[0111] Furthermore, an intermediate frequency (IF) feature interaction unit is arranged between every two first fusion modules. The IF feature interaction unit takes the first fusion intermediate feature output from the previous first fusion module as input and inputs its output into the next first fusion module. The structure of the IF feature interaction unit is as follows: Figure 6 As shown, the mid-frequency feature interaction unit includes a mid-frequency deep convolution branch and a mid-frequency context branch. The mid-frequency deep convolution branch processes the input features with a deep convolution layer to obtain mid-frequency deep convolution features. The mid-frequency context branch processes the input features with multiple cascaded dilated convolution units to obtain mid-frequency context features. The mid-frequency deep convolution features and the mid-frequency context features are concatenated as the output of the mid-frequency feature interaction unit.

[0112] Specifically, to further explore the diagnostically valuable intermediate-scale semantic information in thyroid nodule ultrasound images—such as spiculated edges and microcalcification distribution—a mid-frequency feature interaction unit was used for feature processing. The mid-frequency feature interaction unit within the mid-frequency feature interaction unit employed a 3×3 low-dilation-rate deep convolutional layer to preserve the boundary and texture details of small nodules <5mm. By stacking multiple dilated convolutional units on the mid-frequency context branch, the receptive field was gradually expanded to a range of 5×5-15×15 pixels, thereby accurately matching the spatial span of clinically common nodules and capturing local contextual dependencies.

[0113] Specifically, in the mid-frequency feature interaction unit, a 1×1 convolution is first used to compress the input features and perform preliminary feature fusion, as expressed by the formula:

[0114]

[0115] in, The output features are those of a 1×1 convolution. The input features are for the mid-frequency feature interaction unit. This represents the number of channels after convolution adjustment.

[0116] Deep convolutional layers are used to extract local spatial structure information from input features, ensuring improved edge and texture responsiveness while maintaining lightweight computation. The formula is expressed as:

[0117]

[0118] in, The mid-frequency deep convolutional features are the output of the deep convolutional layer, and DWC is the deep convolution.

[0119] The mid-frequency context branch, by concatenating multiple dilated convolutional units, can significantly expand the receptive field and model multi-scale context, while adaptively emphasizing key information and suppressing redundant noise. The formula is as follows:

[0120]

[0121] in, For mid-frequency context features, is a dilated convolution unit, and n is the number of dilated convolution units.

[0122] Concatenating mid-frequency contextual features with mid-frequency deep convolutional features can preserve both local detail features and global contextual information, achieving complementary fusion of features from different receptive fields. The formula is as follows:

[0123]

[0124] in, This is the result of concatenating mid-frequency contextual features and mid-frequency deep convolutional features. This represents the number of channels after splicing.

[0125] In the output stage of the mid-frequency feature interaction unit, the concatenation result of the mid-frequency context features and the mid-frequency deep convolution features is compressed and fused through 1×1 convolution, outputting a structural feature map with stronger discriminative power and structural stability, as expressed by the formula:

[0126]

[0127] This is the output of the mid-frequency feature interaction unit. This is a convolution operation.

[0128] Specifically, all the mid-frequency feature interaction units in this scheme have the same structure, so they will not be described again.

[0129] In the current embodiment, the first second fusion module in the second fusion branch performs feature processing on the first fusion output and concatenates it with the dilated convolution feature of the corresponding size to obtain the second fusion intermediate feature. The second fusion intermediate feature is then input into the next second fusion module, and the output of the last second fusion module is the fusion intermediate result. The dilated convolution feature is the output of the dilated convolution unit in the first fusion branch, and the number of second fusion modules is the same as that of the first fusion module.

[0130] In the current embodiment, the second fusion output is fused with the output of the intermediate frequency feature interaction unit in each second fusion module using a feature splicing method.

[0131] In the current embodiment, the classification network can use any classification network structure. In this scheme, a query selection layer and a decoder are used to classify the feature fusion results to obtain the diagnostic results of thyroid nodules.

[0132] In the current embodiment, during the training of the thyroid nodule detection architecture, the classification error between the classification result of the training image and the corresponding annotation information is calculated. The parameters of the thyroid nodule detection architecture are adjusted based on the classification error. When the number of iterations or the loss function meets the preset conditions, the training is completed and a trained thyroid nodule detection model is obtained. The thyroid nodule diagnosis result in this scheme is the location and confidence level of each thyroid structure in the thyroid ultrasound image.

[0133] In the current embodiment, the feasibility of the thyroid nodule detection model is verified:

[0134] To validate the effectiveness of our method, we conducted comprehensive experiments on two publicly available thyroid nodule ultrasound image datasets: TN3K and DDTI. The TN3K dataset contains 3493 images, including 1210 benign nodules and 2283 malignant nodules, while the DDTI dataset contains 637 thyroid ultrasound images from a single device. For dataset partitioning, we strictly followed a 7:2:1 ratio, dividing each dataset into training, validation, and test sets. This partitioning strategy ensured sufficient training data and made the validation and test sets statistically representative, allowing for objective evaluation of model performance. All experiments were conducted based on the same data partitioning to guarantee the comparability and consistency of the experimental results.

[0135] To comprehensively evaluate the performance of the proposed model, we used several standard evaluation metrics, including F1, mAP@50, mAP@0.5:0.95, computational cost (GFLOPs), and parameter count (Params), and compared it with existing mainstream object detection models. All models were tested on the same workstation equipped with an NVIDIA RTX 4080 GPU. All models were trained using the same training settings. Specifically, we used stochastic gradient descent (SGD) with momentum as the optimizer, with an initial learning rate of 0.0001 and a momentum coefficient of 0.8. During training, the batch size was set to 4, the maximum training epochs were 200, and an early stopping mechanism (Patience=50) was introduced. This mechanism automatically terminated training when the validation set performance did not improve within 50 consecutive epochs, thus avoiding overfitting and effectively conserving computational resources.

[0136] The test results of the thyroid nodule detection model and the SOTA detector in this scheme on TN3K are shown in Table 1:

[0137] Table 1. Test results of the thyroid nodule detection model and the SOTA detector on TN3K.

[0138]

[0139] As shown in Table 1, in terms of core detection accuracy, our method achieves an AP50 of 0.868, an AP50:95 of 0.534, and an F1 score of 0.852, comprehensively surpassing mainstream detection models such as Faster R-CNN, the RT-DETR series, YOLO-x, and TOOD. Compared to the currently superior RT-DETR-R18 (AP50=0.853), the thyroid nodule detection model achieves a 1.75 percentage point improvement in AP50 while maintaining a lower parameter count (11.994M) and less computational overhead (GFLOPs=31.512), fully demonstrating the effectiveness of its architectural design. This performance breakthrough stems from the synergistic effect of the position-driven adaptive attention mechanism unit, the dilated convolution unit, and the mid-frequency feature interaction unit. The position-driven adaptive attention mechanism unit focuses on high-probability regions based on anatomical priors, significantly suppressing background interference. The dilated convolution unit dynamically responds to different receptive field requirements according to the nodule scale distribution, enhancing multi-scale modeling capabilities. The mid-frequency feature interaction unit specifically strengthens edge continuity and microstructural details. Together, these three elements improve the robustness of discrimination against low-contrast, small-sized lesions.

[0140] Table 2 shows the performance comparison results of our proposed thyroid nodule detection model with several other state-of-the-art (SOTA) detection models on the DDTI dataset:

[0141] Table 2. Test results of the thyroid nodule detection model and the SOTA detector on TN3K.

[0142]

[0143] To verify the effectiveness and independent contribution of each module in TNSpot-Net, this study conducted systematic ablation experiments. Using the backbone network in our proposed scheme as the baseline model, we tested the impact of different modules on model performance on the TN3K dataset. In the experiments, we replaced the modules (dilated convolutional unit, mid-frequency feature interaction unit, and position-driven adaptive attention mechanism unit) in the thyroid nodule detection model with their corresponding modules (SConv, RepC3, and AIFI) in RT-DETR, respectively, to examine the contribution of each module. The experimental results are shown in Table 3.

[0144] Table 3. Ablation experimental results of the thyroid nodule detection model on TN3K.

[0145]

[0146] The results of the ablation experiment showed that:

[0147] Validation of the effectiveness of the dilated convolutional unit: After replacing the standard Conv with adaptively expanded perceptual convolution, the AP50 was improved by 0.59% and the number of parameters was reduced by 5.2%. Three parallel depthwise separable convolutions (d=1,2,3) correspond to the clinical nodule scale layering. After suppressing Rayleigh noise with CAA, the mid-range context can be aggregated while preserving the details of small nodules, thus solving the "scale-receptive field" mismatch problem of fixed convolution kernel.

[0148] Validation of the effectiveness of the mid-frequency feature interaction unit: After replacing RepC3 with the mid-frequency feature interaction unit, AP50 improved by 0.95% and GFLOPs decreased by 12.2%. The fine-grained branch preserves edge texture, and the context branch expands the receptive field to 15×15 pixels by stacking AEPConv. The two paths are stitched together at the same resolution and then compressed to achieve coupling of "local details-mid-range semantics", which significantly reduces multi-scale feature aliasing.

[0149] Validation of the effectiveness of the position-driven adaptive attention mechanism unit: Based on the dilated convolution unit and the mid-frequency feature interaction unit, the AIFI was replaced by the position-driven adaptive attention mechanism unit. The AP50 increased by 0.81%, and the Params and GFLOPs continued to decrease by 0.6% and 2.5%, respectively. The position-driven adaptive attention mechanism unit elastically fits the 20×20 candidate window to the nodule contour through a learnable offset field, and generates spatial-channel coupling weights in parallel. It first suppresses speckle noise and then enhances lesion contrast, completing the integration of "noise suppression-feature enhancement".

[0150] Module Collaborative Validation: From a pure backbone to a complete thyroid nodule detection model, AP50 cumulatively improved by 2.72%, Params cumulatively decreased by 16.3%, and GFLOPs cumulatively decreased by 17.9%, with the reduction significantly exceeding the linear superposition of single-module gains. This confirms that the three modules form a progressive optimization link: the dilated convolution unit completes scale-adaptive feature extraction, the mid-frequency feature interaction unit achieves mid-frequency semantic purification, and the position-driven adaptive attention mechanism unit achieves precise regional focusing. Layer by layer, it overcomes the three major bottlenecks of "blurred edges, scale jumps, and complex backgrounds" in thyroid ultrasound images, ultimately achieving a simultaneous leap in accuracy and efficiency.

[0151] In this current embodiment, to demonstrate the performance of the thyroid nodule detection model in medical image target detection, the detection results of the thyroid nodule detection model are visualized using the TN3K and DDTI datasets. Figure 7 The results of the thyroid nodule detection model in this scheme and other models on the TN3K dataset are shown. Figure 7 The first column contains labeled medical images from the TN3K dataset, the second column shows the thyroid nodule detection model used in this scheme, and the other columns represent other models. Figure 7 It can be seen that the model of this scheme is superior to other comparative models in terms of detection accuracy and localization, especially in the detection of small nodules and nodules with blurred boundaries.

[0152] Figure 8 The results of the thyroid nodule detection model in this scheme and other models on the DDTI dataset are shown. Figure 8 The first column contains labeled medical images from the DDTI dataset, the second column shows the thyroid nodule detection model used in this scheme, and the other columns represent other models. Figure 8 As can be seen, the proposed model performs well on the DDTI dataset, accurately identifying and locating targets, especially in complex image backgrounds where it still extracts target information effectively. Compared to other models, the thyroid nodule detection model demonstrates high target detection accuracy and robustness, and can operate stably in diverse images.

[0153] Example 2

[0154] Based on the same concept, this application also proposes an application method for a thyroid nodule detection model based on mid-frequency feature interaction, including:

[0155] A thyroid ultrasound image is acquired as the image to be detected. The image to be detected is then input into the thyroid nodule detection model trained in Example 1 to obtain the thyroid nodule diagnosis result corresponding to the image to be detected.

[0156] Example 3

[0157] This embodiment also provides an electronic device, see reference. Figure 9 It includes a memory 404 and a processor 402, the memory 404 storing a computer program and the processor 402 being configured to run the computer program to perform the steps in any of the above method embodiments.

[0158] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0159] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0160] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0161] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the thyroid nodule detection model construction methods based on mid-frequency feature interaction in the above embodiments.

[0162] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0163] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0164] The input / output device 408 is used to input or output information. In this embodiment, the input information may be a thyroid ultrasound image, etc., and the output information may be a thyroid nodule diagnosis result, etc.

[0165] Optionally, in this embodiment, the processor 402 can be configured to perform the following steps via a computer program:

[0166] Multiple training images are acquired, wherein the training images are thyroid ultrasound images with thyroid nodules labeled.

[0167] A thyroid nodule detection architecture is constructed, consisting of a backbone network, a feature fusion network, and a classification network connected sequentially. The backbone network extracts features from training images using a feature pyramid structure, obtaining multiple first feature maps with different channel numbers. The feature fusion network includes an input branch, a first fusion branch, and a second fusion branch. The input branch processes each first feature map using dilated convolutional units to obtain corresponding first input features. The first fusion branch includes multiple cascaded first fusion modules, each consisting of a dilated convolutional unit and an upsampling layer connected sequentially. The first input features with different channel numbers are fused based on these multiple first fusion modules to obtain a first fusion output. The second fusion branch includes multiple cascaded second fusion modules and a second output module. The two-way fusion module consists of a mid-frequency feature interaction unit and a dilated convolution unit connected sequentially. The second output module is the mid-frequency feature interaction unit. Based on multiple second fusion modules, the first fusion output is fused with the output of each dilated convolution unit in the first fusion branch to obtain an intermediate fusion result. The intermediate fusion result is input into the second output module to obtain the second fusion output. The second fusion output is then fused with the output of the mid-frequency feature interaction unit in each second fusion module to obtain the feature fusion result. The dilated convolution unit processes the input features with dilation convolutions of different dilation rates. The mid-frequency feature interaction unit preserves the boundary and texture details of the nodules by capturing the local contextual dependencies of the input features. The classification network obtains the thyroid nodule diagnosis result based on the feature fusion result.

[0168] The thyroid nodule detection architecture is trained using multiple training images to obtain a well-constructed thyroid nodule detection model.

[0169] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0170] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0171] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 9 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0172] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing a thyroid nodule detection model based on mid-frequency feature interaction, characterized in that, Includes the following steps: Multiple training images are acquired, wherein the training images are thyroid ultrasound images with thyroid nodules labeled. A thyroid nodule detection architecture is constructed, consisting of a backbone network, a feature fusion network, and a classification network connected sequentially. The backbone network extracts features from training images using a feature pyramid structure, obtaining multiple first feature maps with different channel numbers. The feature fusion network includes an input branch, a first fusion branch, and a second fusion branch. The input branch processes each first feature map using dilated convolutional units to obtain corresponding first input features. The first fusion branch includes multiple cascaded first fusion modules, each consisting of a dilated convolutional unit and an upsampling layer connected sequentially. Based on these multiple first fusion modules, the first input features with different channel numbers are fused to obtain a first fusion output. An intermediate frequency feature interaction unit is placed between every two first fusion modules. This intermediate frequency feature interaction unit takes the first fusion intermediate feature output from the previous first fusion module as input and inputs its output to the next first fusion module. The intermediate frequency feature interaction unit includes an intermediate frequency deep convolutional branch and an intermediate frequency context branch. The intermediate frequency deep convolutional branch processes the input features using a deep convolutional layer to obtain intermediate frequencies. The system employs a deep convolutional feature generation method. The mid-frequency context branch processes the input features using multiple cascaded dilated convolutional units to obtain mid-frequency contextual features. The deep convolutional features and mid-frequency contextual features are concatenated as the output of the mid-frequency feature interaction unit. The second fusion branch includes multiple cascaded second fusion modules and a second output module. Each second fusion module consists of a mid-frequency feature interaction unit and a dilated convolutional unit connected sequentially. The second output module is the mid-frequency feature interaction unit. Based on these multiple second fusion modules, the first fusion output is fused with the output of each dilated convolutional unit in the first fusion branch to obtain an intermediate fusion result. This intermediate fusion result is input into the second output module to obtain a second fusion output. The second fusion output is then fused with the output of the mid-frequency feature interaction unit in each second fusion module to obtain the feature fusion result. The dilated convolutional units process the input features using dilated convolutions with different dilation rates. The mid-frequency feature interaction unit preserves the boundaries and texture details of the nodules by capturing the local contextual dependencies of the input features. The classification network obtains the diagnostic result for thyroid nodules based on the feature fusion result. The thyroid nodule detection architecture is trained using multiple training images to obtain a well-constructed thyroid nodule detection model.

2. The method for constructing a thyroid nodule detection model based on mid-frequency feature interaction according to claim 1, characterized in that, The feature pyramid structure of the backbone network consists of multiple serial feature extraction modules. Each feature extraction module includes multiple serial multidimensional interaction enhancement units. The step size of the multidimensional interaction enhancement unit in the first feature extraction module is 1, and the step size of the multidimensional interaction enhancement unit in the remaining feature extraction modules is 2. The output of each non-first feature extraction module is used as the first feature map.

3. The method for constructing a thyroid nodule detection model based on mid-frequency feature interaction according to claim 2, characterized in that, In the multidimensional interaction enhancement unit, the input features are subjected to depthwise convolution and standard convolution respectively. The input features, the depthwise convolution result and the standard convolution result are summed to obtain the first multidimensional interaction output. The first multidimensional interaction output is calculated by a multi-scale contextual attention mechanism to obtain the second multidimensional interaction output. The second multidimensional interaction output is passed through a residual network to obtain the third multidimensional interaction output. The third multidimensional interaction output is the output feature of the multidimensional interaction enhancement unit. The output feature of the last multidimensional interaction enhancement unit in the feature extraction module is the output of the feature extraction module.

4. The method for constructing a thyroid nodule detection model based on mid-frequency feature interaction according to claim 1, characterized in that, The dilated convolution unit uses parallel first, second, and third dilated convolution layers to process the input features to obtain the first, second, and third dilated convolution results, respectively. The first, second, and third dilated convolution results are added element-wise to obtain the comprehensive dilated convolution result. The comprehensive dilated convolution result is then concatenated with the input features and sequentially subjected to a group-level channel-aware attention mechanism and channel shuffling operation to obtain the output. The dilation rate of the first dilated convolution layer is 1, the dilation rate of the second dilated convolution layer is 2, and the dilation rate of the third dilated convolution layer is 3.

5. The method for constructing a thyroid nodule detection model based on mid-frequency feature interaction according to claim 1, characterized in that, The first fusion module in the first fusion branch processes the first input feature with the largest number of channels and concatenates it with another first input feature of the corresponding size to obtain the first fusion intermediate feature. The first fusion intermediate feature is then input into the next first fusion module. After all the first input features are concatenated, the first fusion output is obtained.

6. The method for constructing a thyroid nodule detection model based on mid-frequency feature interaction according to claim 5, characterized in that, Before the first fusion module processes the first input feature with the largest number of channels, a position-driven adaptive attention mechanism unit is used to process the first input feature with the largest number of channels. In the position-driven adaptive attention mechanism unit, a local sliding candidate window is defined for the input feature to obtain a window feature map. The window feature map is expanded into a window feature sequence by row. The position offset and position attention weight of each window in the window feature sequence are predicted by a dynamic position offset attention mechanism. Based on the position offset of each window, the position of the input feature is adjusted to obtain the offset feature. The offset feature is then weighted with the position attention weight to obtain the weighted offset feature. The weighted offset feature is reshaped and then concatenated with the window feature map as the output of the position-driven adaptive attention mechanism unit.

7. The method for constructing a thyroid nodule detection model based on mid-frequency feature interaction according to claim 1, characterized in that, The first second fusion module in the second fusion branch processes the first fusion output and concatenates it with the dilated convolution feature of the corresponding size to obtain the second fusion intermediate feature. The second fusion intermediate feature is then input into the next second fusion module. The output of the last second fusion module is the fusion intermediate result. The dilated convolution feature is the output of the dilated convolution unit in the first fusion branch. The number of second fusion modules is the same as that of the first fusion module.

8. An application method for a thyroid nodule detection model based on mid-frequency feature interaction, characterized in that, include: A thyroid ultrasound image is acquired as the image to be detected. The image to be detected is then input into the thyroid nodule detection model trained by any of the methods described in claims 1-7 to obtain the thyroid nodule diagnosis result corresponding to the image to be detected.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to execute the method for constructing a thyroid nodule detection model based on mid-frequency feature interaction as described in any one of claims 1-7 or the method for applying a thyroid nodule detection model based on mid-frequency feature interaction as described in claim 8.

Images