Thyroid ultrasound image diagnosis method based on deep learning

Abstract

The invention discloses a thyroid ultrasound image diagnosis method based on deep learning, and the method comprises the following steps: collecting a thyroid ultrasound original image set, and carrying out the preprocessing; performing focus segmentation on the standardized thyroid ultrasound image set; performing morphological constraint and boundary refinement; calculating the blood flow direction, blood flow velocity and blood flow power of each thyroid focus area and neighborhood; generating a preliminary fusion feature map based on a feature adaptive deep learning network, and fusing the preliminary fusion feature map with the thyroid focus blood flow feature vector set; obtaining a thyroid focus detection list through thyroid focus benign and malignant discrimination branches; and generating thyroid focus structured diagnosis data records based on the thyroid focus detection list, and writing the thyroid focus structured diagnosis data records into a computer-aided diagnosis system. According to the method, deep learning and multi-modal blood flow features are fused, intelligent diagnosis of the thyroid focus is achieved, and the method has the advantages of being high in precision, high in interference resistance and structured in result.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and artificial intelligence diagnosis, and in particular to a deep learning-based method for thyroid ultrasound image diagnosis. Background Technology

[0002] Thyroid ultrasound, a commonly used clinical imaging method, offers advantages such as being non-invasive, convenient, and real-time, enabling the detection and benign / malignant assessment of thyroid nodules and lesions. However, current ultrasound diagnostic methods largely rely on manual interpretation by physicians, making them susceptible to factors such as operator skill, image quality, and equipment differences, resulting in high subjectivity and poor repeatability. Furthermore, speckle noise, background artifacts, and differences in resolution and intensity between devices can interfere with lesion boundary identification and structural analysis, leading to insufficient accuracy in lesion detection and quantitative measurement.

[0003] In recent years, deep learning methods have been gradually applied to ultrasound image analysis, using convolutional neural networks to extract and classify image features, demonstrating certain advantages in thyroid lesion identification. However, existing methods mostly focus on two-dimensional image features or single-modal information, lacking comprehensive utilization of lesion region segmentation accuracy, blood flow feature fusion, and multimodal information. This results in insufficiently refined lesion boundaries and inadequate modeling of hemodynamic features, ultimately affecting the ability to distinguish between benign and malignant thyroid lesions and the reliability of clinical applications.

[0004] Therefore, how to provide a deep learning-based diagnostic method for thyroid ultrasound images is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based thyroid ultrasound image diagnostic method. This invention integrates deep learning with multimodal blood flow features to achieve intelligent diagnosis of thyroid lesions, and has the advantages of high accuracy, strong anti-interference, and structured results.

[0006] A deep learning-based thyroid ultrasound image diagnosis method according to an embodiment of the present invention includes the following steps:

[0007] A set of raw thyroid ultrasound images was acquired and preprocessed to obtain a standardized set of thyroid ultrasound images.

[0008] Lesion segmentation is performed on a standardized set of thyroid ultrasound images to generate an initial segmentation mask;

[0009] Morphological constraints and boundary refinement are applied to the initial segmentation mask to obtain the set of thyroid lesion regions and their boundary contours.

[0010] Calculate the blood flow direction, blood flow velocity, and blood flow power of each thyroid lesion region and its surrounding area to generate a set of blood flow feature vectors for thyroid lesions;

[0011] A standardized set of thyroid ultrasound images is input into a feature-adaptive deep learning network trained on a thyroid ultrasound dataset to generate a preliminary fusion feature map, which is then fused with a set of thyroid lesion blood flow feature vectors to generate a thyroid lesion fusion feature tensor.

[0012] The thyroid fusion feature tensor is input into the benign and malignant thyroid lesion discrimination branch, and the benign and malignant probability value and classification label of each thyroid lesion are output. After confidence filtering and non-maximum suppression, the list of thyroid lesion detection is obtained.

[0013] Structured diagnostic data records for thyroid lesions are generated based on the thyroid lesion detection list and written into the computer-aided diagnostic system.

[0014] Optionally, the preprocessing includes, in sequence, speckle noise suppression, background artifact suppression, contrast-controlled enhancement, spatial resolution unification, and cross-device intensity normalization.

[0015] Optionally, the generation of the initial segmentation mask specifically includes:

[0016] A standardized set of thyroid ultrasound images is input into a deep convolutional neural network to extract multi-layer feature representations for each pixel in the image and output a lesion probability map corresponding to each pixel.

[0017] The probability value of each pixel in the lesion probability map is compared with a preset threshold to form an initial segmentation mask.

[0018] Optionally, the generation of the thyroid lesion region set and boundary contour specifically includes:

[0019] Morphological opening and closing operations are performed on the initial segmentation mask. The opening operation is used to remove low-probability isolated pixel clusters, and the closing operation is used to fill the tiny cavities inside the lesion. The structuring element adopts a circular template that matches the image resolution, and the template radius is set according to the resolution of the thyroid ultrasound image.

[0020] After morphological processing, the edge gradient operator is used to detect the boundary position of the initial segmentation mask, and the boundary is refined by applying a sub-pixel level interpolation method.

[0021] Perform connected component analysis on the initial segmentation mask after boundary refinement, count the area and aspect ratio of each connected component, remove regions with an area smaller than the minimum threshold or an aspect ratio exceeding the set range, and retain only connected components with reasonable shape as the lesion region set.

[0022] Extract the corresponding boundary contours from the set of lesion regions, record the boundary contours of each lesion region as a point set, and generate a set of thyroid lesion regions and boundary contours.

[0023] Optionally, the generation of the set of blood flow feature vectors for thyroid lesions specifically includes:

[0024] Based on the set of thyroid lesion areas, the lesion neighborhood area is formed by expanding outward by a preset pixel distance according to its boundary contour. The thyroid lesion area and the lesion neighborhood area are merged as the calculation range of blood flow parameters.

[0025] Within the range of blood flow parameter calculation, color Doppler codes are analyzed, hue is mapped to blood flow direction, linear velocity is calculated using Doppler frequency shift and ultrasound emission center frequency, and blood flow power intensity is converted based on Doppler signal amplitude and system gain. Blood flow direction, blood flow velocity and blood flow power are obtained for each thyroid lesion area and its neighborhood, forming blood flow characteristics.

[0026] The blood flow features are spliced ​​together in a set order to obtain a standardized blood flow feature vector. The standardized blood flow feature vectors of all lesions are then combined to form a set of blood flow feature vectors for thyroid lesions.

[0027] Optionally, the generation of the thyroid lesion fusion feature tensor specifically includes:

[0028] A standardized set of thyroid ultrasound images is used as a single-channel input tensor to be input into a feature-adaptive deep learning network. The height of the single-channel input tensor is the number of rows in the image, the width is the number of columns in the image, and the pixel value ranges from zero to one. The feature-adaptive deep learning network includes a multi-scale convolutional encoder, a speckle noise perception attention module, a pyramid feature fusion structure, and a benign or malignant thyroid lesion discrimination branch.

[0029] Multiple parallel convolutional paths are set in a multi-scale convolutional encoder. Each path uses a different convolutional kernel size and dilation rate to extract fine-grained texture features suitable for small lesion identification and structural boundary features suitable for structural localization from thyroid ultrasound images. The features output by all paths are concatenated by channels and compressed by one-to-one convolution to obtain fused multi-scale encoded features. The multi-scale convolutional encoder includes a parallel convolutional path module, a channel concatenation module, and a feature compression module. The parallel convolutional path module contains three convolutional paths with different convolutional kernel sizes and dilation rates. The channel concatenation module is used to concatenate the output features of the parallel convolutional path module according to the channel dimension. The feature compression module uses one-to-one convolution to compress the concatenated multi-scale features.

[0030] The multi-scale encoded features are input into the speckle noise perception attention module to calculate the speckle noise intensity score of each region. Differential weights are assigned to different regions based on the speckle noise intensity scores, and speckle noise suppression features are output. The speckle noise perception attention module includes a noise intensity estimation unit, an attention weight calculation unit, and a feature reweighting unit. The noise intensity estimation unit scores the speckle noise intensity of local regions with the multi-scale encoded features. The attention weight calculation unit generates a region-level weight distribution based on the speckle noise intensity scores. The feature reweighting unit multiplies the region-level weight distribution with the multi-scale encoded features channel by channel to output speckle noise suppression features.

[0031] The speckle noise suppression features are fused with different levels of encoded features through upsampling and skip connections, and convolution, batch normalization and nonlinear activation operations are performed until the spatial resolution is restored to the same as the input tensor to generate segmentation decoding features;

[0032] The segmented and decoded features are input into the pyramid feature fusion structure. After performing a one-to-one convolution on the features at each scale, they are aligned to a uniform resolution using bilinear interpolation. The summation is then performed at the corresponding pixel positions to generate a preliminary fused feature map.

[0033] The preliminary fusion feature map is fused with the set of blood flow feature vectors of thyroid lesions to generate the fusion feature tensor of thyroid lesions.

[0034] Optionally, the generation of the thyroid lesion fusion feature tensor specifically includes:

[0035] The preliminary fusion feature map is normalized to generate a normalized lesion feature set. The preliminary fusion feature map contains grayscale texture information and local detail texture information under different receptive field conditions.

[0036] Extract the blood flow feature vector of the thyroid lesion corresponding to the current lesion, and splice and fuse it with the normalized lesion feature set in the feature channel dimension to generate lesion-level fusion features;

[0037] After channel-weighted processing of the lesion-level fusion features, they are input into a fully connected mapping layer, which outputs a thyroid lesion fusion feature tensor that integrates texture, morphology, multi-scale and blood flow information. The channel-weighted weights are obtained by a weight generation network from the statistical response generated by global average pooling.

[0038] Optionally, the generation of the thyroid lesion detection list specifically includes:

[0039] The thyroid fusion feature tensor is input into the thyroid lesion benign or malignant discrimination branch and the output is generated with the probability value of benign or malignant thyroid lesion and the classification label. The thyroid lesion benign or malignant discrimination branch includes several convolutional layers, global average pooling layers and fully connected layers. The probability value of benign or malignant thyroid lesion is a real number ranging from zero to one, and the classification label is benign or malignant.

[0040] The benign or malignant probability value and classification label of each thyroid lesion are evaluated for confidence. Lesions with a confidence level below the preset confidence threshold are removed, and only lesions that meet the confidence requirements are retained. The preset confidence threshold is set based on a balance between clinical sensitivity and specificity.

[0041] After screening, non-maximum suppression is performed on the lesions according to the classification labels. The overlap ratio between candidate lesions is calculated and combined with the confidence priority. Only the candidate lesions with the highest confidence priority are retained, and the remaining candidate lesions are removed. The final list of thyroid lesion detection is then output.

[0042] Optionally, the writing to the computer-aided diagnostic system specifically includes:

[0043] Extract the location information and boundary contour of each lesion in the thyroid lesion detection list. The location information includes the horizontal and vertical coordinates and depth information of the lesion in the standardized thyroid ultrasound image coordinate system. The boundary contour is a set of continuous closed contour points obtained by fitting the lesion segmentation results.

[0044] Blood flow direction, blood flow velocity, and blood flow power are extracted from the corresponding lesion area and its neighborhood to generate a blood flow parameter set;

[0045] The location information, boundary contour, benign and malignant probability values ​​of thyroid lesions, classification labels and blood flow parameter sets are combined to form a structured diagnostic data record of thyroid lesions;

[0046] The generated structured diagnostic data of thyroid lesions is recorded and written into the computer-aided diagnostic system.

[0047] The beneficial effects of this invention are:

[0048] This invention proposes a deep learning-based diagnostic method for thyroid ultrasound images, overcoming the limitations of traditional methods that rely on manual interpretation. It constructs a fully automated diagnostic system encompassing image preprocessing, lesion segmentation, boundary extraction, blood flow feature analysis, deep fusion, benign / malignant differentiation, and structured output. By introducing a multi-scale convolutional coding structure and a speckle noise-aware attention mechanism, the model's ability to identify lesions of different sizes and its noise robustness are significantly enhanced, solving the technical challenge of speckle noise affecting the accuracy of lesion detection in ultrasound images. Simultaneously, by utilizing pyramid feature fusion and multi-channel information integration, a unified expression of texture, morphology, and structural features is achieved, improving the accuracy of image semantic information expression and providing stable input for subsequent discrimination modules.

[0049] Compared to existing methods that rely solely on grayscale information for discrimination, this invention introduces blood flow direction, velocity, and power information extracted using color Doppler encoding during the feature fusion stage. This constructs a standardized set of blood flow feature vectors for thyroid lesions, which are then deeply fused with image features. This effectively enhances the ability to distinguish lesions at the microscopic blood flow level, significantly improving the accuracy of benign / malignant classification. Furthermore, confidence assessment and non-maximum suppression strategies are employed to screen and optimize candidate lesions, ensuring the final lesion detection list possesses high reliability and clinical applicability. Finally, this invention also constructs a structured diagnostic data generation module, systematically integrating lesion location information, boundary contours, discrimination results, and blood flow features. This provides a readable, storable, and traceable standardized diagnostic data format for computer-aided diagnostic systems, further supporting subsequent case analysis and treatment decisions. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of a deep learning-based thyroid ultrasound image diagnosis method proposed in this invention;

[0052] Figure 2 This is a schematic diagram of the lesion-level fusion feature generation and channel weighting processing flow of a deep learning-based thyroid ultrasound image diagnosis method proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1-2A deep learning-based diagnostic method for thyroid ultrasound images includes the following steps:

[0055] A set of raw thyroid ultrasound images was acquired and preprocessed to obtain a standardized set of thyroid ultrasound images.

[0056] Lesion segmentation is performed on a standardized set of thyroid ultrasound images to generate an initial segmentation mask;

[0057] Morphological constraints and boundary refinement are applied to the initial segmentation mask to obtain the set of thyroid lesion regions and their boundary contours.

[0058] Calculate the blood flow direction, blood flow velocity, and blood flow power of each thyroid lesion region and its surrounding area to generate a set of blood flow feature vectors for thyroid lesions;

[0059] A standardized set of thyroid ultrasound images is input into a feature-adaptive deep learning network trained on a thyroid ultrasound dataset to generate a preliminary fusion feature map, which is then fused with a set of thyroid lesion blood flow feature vectors to generate a thyroid lesion fusion feature tensor.

[0060] The thyroid fusion feature tensor is input into the benign and malignant thyroid lesion discrimination branch, and the benign and malignant probability value and classification label of each thyroid lesion are output. After confidence filtering and non-maximum suppression, the list of thyroid lesion detection is obtained.

[0061] Structured diagnostic data records for thyroid lesions are generated based on the thyroid lesion detection list and written into the computer-aided diagnostic system.

[0062] In this embodiment, the preprocessing includes, in sequence, speckle noise suppression, background artifact suppression, contrast-controlled enhancement, spatial resolution unification, and cross-device intensity normalization.

[0063] In this embodiment, the generation of the initial segmentation mask specifically includes:

[0064] A standardized set of thyroid ultrasound images is input into a deep convolutional neural network to extract multi-layer feature representations for each pixel in the image and output a lesion probability map corresponding to each pixel.

[0065] The probability value of each pixel in the lesion probability map is compared with a preset threshold to form an initial segmentation mask.

[0066] In this embodiment, the generation of the thyroid lesion region set and boundary contour specifically includes:

[0067] Morphological opening and closing operations are performed on the initial segmentation mask. The opening operation is used to remove low-probability isolated pixel clusters, and the closing operation is used to fill the tiny cavities inside the lesion. The structuring element adopts a circular template that matches the image resolution, and the template radius is set according to the resolution of the thyroid ultrasound image.

[0068] After morphological processing, the edge gradient operator is used to detect the boundary position of the initial segmentation mask, and the boundary is refined by applying a sub-pixel level interpolation method.

[0069] Connectivity analysis is performed on the initial segmentation mask after boundary refinement. The area and aspect ratio of each connected component are calculated. Regions with an area smaller than the minimum threshold or an aspect ratio exceeding the set range are removed. Only connected components with reasonable shapes are retained as the lesion region set. The minimum threshold is an area threshold set based on image resolution and the minimum size of clinically visible lesions. The aspect ratio set range is a reasonable interval determined based on the lesion morphology distribution in the training dataset and clinical experience.

[0070] Extract the corresponding boundary contours from the set of lesion regions, record the boundary contours of each lesion region as a point set, and generate a set of thyroid lesion regions and boundary contours.

[0071] In this embodiment, the generation of the thyroid lesion blood flow feature vector set specifically includes:

[0072] Based on the set of thyroid lesion areas, the lesion neighborhood area is formed by expanding outward by a preset pixel distance according to its boundary contour. The thyroid lesion area and the lesion neighborhood area are merged as the calculation range of blood flow parameters.

[0073] Within the range of blood flow parameter calculation, color Doppler codes are analyzed, hue is mapped to blood flow direction, linear velocity is calculated using Doppler frequency shift and ultrasound emission center frequency, and blood flow power intensity is converted based on Doppler signal amplitude and system gain. Blood flow direction, blood flow velocity and blood flow power are obtained for each thyroid lesion area and its neighborhood, forming blood flow characteristics.

[0074] The blood flow features are spliced ​​together in a set order to obtain a standardized blood flow feature vector. The standardized blood flow feature vectors of all lesions are then combined to form a set of blood flow feature vectors for thyroid lesions.

[0075] In this embodiment, the generation of the thyroid lesion fusion feature tensor specifically includes:

[0076] A standardized set of thyroid ultrasound images is used as a single-channel input tensor to be input into a feature-adaptive deep learning network. The height of the single-channel input tensor is the number of rows in the image, the width is the number of columns in the image, and the pixel value ranges from zero to one. The feature-adaptive deep learning network includes a multi-scale convolutional encoder, a speckle noise perception attention module, a pyramid feature fusion structure, and a benign or malignant thyroid lesion discrimination branch.

[0077] Multiple parallel convolutional paths are set in a multi-scale convolutional encoder. Each path uses a different convolutional kernel size and dilation rate to extract fine-grained texture features suitable for small lesion identification and structural boundary features suitable for structural localization from thyroid ultrasound images. The features output by all paths are concatenated by channels and compressed by one-to-one convolution to obtain fused multi-scale encoded features. The multi-scale convolutional encoder includes a parallel convolutional path module, a channel concatenation module, and a feature compression module. The parallel convolutional path module contains three convolutional paths with different convolutional kernel sizes and dilation rates. The channel concatenation module is used to concatenate the output features of the parallel convolutional path module according to the channel dimension. The feature compression module uses one-to-one convolution to compress the concatenated multi-scale features.

[0078] The multi-scale encoded features are input into the speckle noise perception attention module to calculate the speckle noise intensity score of each region. Differential weights are assigned to different regions based on the speckle noise intensity scores, and speckle noise suppression features are output. The speckle noise perception attention module includes a noise intensity estimation unit, an attention weight calculation unit, and a feature reweighting unit. The noise intensity estimation unit scores the speckle noise intensity of local regions with the multi-scale encoded features. The attention weight calculation unit generates a region-level weight distribution based on the speckle noise intensity scores. The feature reweighting unit multiplies the region-level weight distribution with the multi-scale encoded features channel by channel to output speckle noise suppression features.

[0079] The speckle noise suppression features are fused with different levels of encoded features through upsampling and skip connections, and convolution, batch normalization and nonlinear activation operations are performed until the spatial resolution is restored to the same as the input tensor to generate segmentation decoding features;

[0080] The segmented and decoded features are input into the pyramid feature fusion structure. After performing a one-to-one convolution on the features at each scale, they are aligned to a uniform resolution using bilinear interpolation. The summation is then performed at the corresponding pixel positions to generate a preliminary fused feature map.

[0081] The preliminary fusion feature map is fused with the set of blood flow feature vectors of thyroid lesions to generate the fusion feature tensor of thyroid lesions.

[0082] In this embodiment, the generation of the thyroid lesion fusion feature tensor specifically includes:

[0083] The preliminary fusion feature map is normalized to generate a normalized lesion feature set. The preliminary fusion feature map contains grayscale texture information and local detail texture information under different receptive field conditions.

[0084] Extract the blood flow feature vector of the thyroid lesion corresponding to the current lesion, and splice and fuse it with the normalized lesion feature set in the feature channel dimension to generate lesion-level fusion features;

[0085] After channel-weighted processing of the lesion-level fusion features, they are input into a fully connected mapping layer, which outputs a thyroid lesion fusion feature tensor that integrates texture, morphology, multi-scale and blood flow information. The channel-weighted weights are obtained by a weight generation network from the statistical response generated by global average pooling.

[0086] In this embodiment, the generation of the thyroid lesion detection list specifically includes:

[0087] The thyroid fusion feature tensor is input into the thyroid lesion benign or malignant discrimination branch and the output is generated with the probability value of benign or malignant thyroid lesion and the classification label. The thyroid lesion benign or malignant discrimination branch includes several convolutional layers, global average pooling layers and fully connected layers. The probability value of benign or malignant thyroid lesion is a real number ranging from zero to one, and the classification label is benign or malignant.

[0088] The benign or malignant probability value and classification label of each thyroid lesion are evaluated for confidence. Lesions with a confidence level below the preset confidence threshold are removed, and only lesions that meet the confidence requirements are retained. The preset confidence threshold is set based on a balance between clinical sensitivity and specificity.

[0089] After screening, non-maximum suppression is performed on the lesions according to the classification labels. The overlap ratio between candidate lesions is calculated and combined with the confidence priority. Only the candidate lesions with the highest confidence priority are retained, and the remaining candidate lesions are removed. The final list of thyroid lesion detection is then output.

[0090] In this embodiment, the writing to the computer-aided diagnostic system specifically includes:

[0091] Extract the location information and boundary contour of each lesion in the thyroid lesion detection list. The location information includes the horizontal and vertical coordinates and depth information of the lesion in the standardized thyroid ultrasound image coordinate system. The boundary contour is a set of continuous closed contour points obtained by fitting the lesion segmentation results.

[0092] Blood flow direction, blood flow velocity, and blood flow power are extracted from the corresponding lesion area and its neighborhood to generate a blood flow parameter set;

[0093] The location information, boundary contour, benign and malignant probability values ​​of thyroid lesions, classification labels and blood flow parameter sets are combined to form a structured diagnostic data record of thyroid lesions;

[0094] The generated structured diagnostic data of thyroid lesions is recorded and written into the computer-aided diagnostic system.

[0095] Example 1:

[0096] To verify the feasibility of this invention in practice, it was applied to the radiology department of a top-tier hospital to assist in optimizing the screening and diagnosis process for thyroid diseases. This hospital undertakes large-scale thyroid health checkups in the region, with doctors reviewing nearly a hundred images daily, facing high workload and diagnostic pressure. Especially when processing ultrasound images with blurred boundaries, complex echoes, or low contrast, traditional experience-based manual interpretation methods are easily subject to subjective interference, posing a risk of missed diagnoses or misdiagnoses.

[0097] The deployment process of this method consists of multiple modules working together. After receiving the raw set of thyroid ultrasound images, the system first performs speckle noise suppression, background artifact suppression, contrast-controlled enhancement, spatial resolution unification, and cross-device intensity standardization through the image preprocessing module to generate a standardized set of thyroid ultrasound images, providing consistent input for subsequent feature extraction.

[0098] Standardized images are input into a trained deep convolutional neural network to generate a lesion probability map and form an initial segmentation mask based on a preset threshold. Morphological opening and closing operations are then performed on the mask, edge gradient operators are applied to detect boundaries, and edges are refined through subpixel-level interpolation. Combined with connected component analysis, morphologically abnormal regions are eliminated, and a set of structurally reasonable thyroid lesion regions and their boundary contours are extracted.

[0099] Based on the lesion area, indicators such as blood flow direction, blood flow velocity, and blood flow power are extracted using color Doppler encoding and encoded into a set of thyroid lesion blood flow feature vectors according to a set of rules. Compared with traditional methods that rely solely on grayscale images, this invention effectively enhances the ability to distinguish the physiological features of lesions by introducing the dimension of blood flow features.

[0100] The system's feature-adaptive deep learning network includes a multi-scale convolutional encoder, a speckle noise-aware attention module, and a pyramid feature fusion structure. These are used to extract multi-scale texture structure information, perform noise-aware weighting, and integrate semantic features at different levels, respectively. After fusing image features and blood flow vectors, a thyroid lesion fusion feature tensor is generated, which is then processed through channel weighting and a fully connected mapping layer to output the final representation.

[0101] In the benign / malignant differentiation stage, the system outputs benign / malignant probability values ​​and classification labels based on the fused feature tensor. Through confidence filtering and non-maximum suppression mechanisms, it removes candidate lesions with low confidence or high overlap, generating a final list of detected thyroid lesions. The system then extracts lesion location information, boundary contours, classification results, and key parameters to generate structured diagnostic data records for thyroid lesions, which are then written into the computer-aided diagnostic system for clinical reference.

[0102] During actual deployment, the physician feedback system demonstrated excellent performance in assisting the interpretation of cases with complex echoes or atypical blood flow patterns, quickly identifying suspicious areas and reducing false positives and false negatives. Simultaneously, structured data recording improved the efficiency of case follow-up and retrospection, resulting in a more efficient and stable overall diagnostic process.

[0103] Table 1. Performance Comparison of the Invention and Traditional Thyroid Ultrasound Diagnostic Methods

[0104] Indicator Categories Traditional methods Method of the present invention Lesion segmentation accuracy (IoU) 72.8％ 78.6％ Average error of lesion boundary (mm) 1.85 1.42 Accuracy of benign / malignant classification 81.4％ 85.1％ Sensitivity 79.6％ 83.4％ Specificity 82.3％ 86.7％ Confidence stability (variance) 0.148 0.123 Average diagnosis time (seconds / case) 15.2 12.4

[0105] Table 1 shows that the present invention improves the lesion segmentation accuracy (IoU) from 72.8% to 78.6% compared to the traditional method, while reducing the average error of lesion boundaries from 1.85mm to 1.42mm. This performance improvement is mainly due to the systematic optimization of the segmentation process. The image preprocessing module effectively improves the consistency and clarity of the input image through speckle noise suppression, background artifact suppression, and cross-device intensity normalization. The multi-scale convolutional encoder and the speckle noise-aware attention module work together to enable the network to extract fine-grained textures and structural boundaries under different receptive fields, while reducing the impact of noise. In addition, morphological opening and closing operations, sub-pixel interpolation methods to refine boundaries, and connected component analysis further improve the geometric integrity and boundary accuracy of the segmentation mask.

[0106] In benign / malignant differentiation, the accuracy of this invention increased from 81.4% to 85.1%, while sensitivity and specificity also improved to 83.4% and 86.7%, respectively. This performance advantage is attributed to the innovative design of the feature fusion strategy: the fusion not only includes grayscale texture and morphological features, but also integrates physiological feature vectors such as blood flow direction, blood flow velocity, and blood flow power from color Doppler images. Through the pyramid feature fusion structure and channel weighting mechanism, the expression efficiency of multimodal features and the discriminative ability of the classifier are effectively improved.

[0107] Furthermore, the confidence stability (variance) of this invention is reduced from 0.148 to 0.123, indicating that the system is more stable and reliable in terms of output probability distribution. This change benefits from the synergistic control of the fusion feature representation and confidence screening strategy, effectively avoiding interference from low-confidence candidates in the final output.

[0108] In terms of diagnostic efficiency, the average time has been reduced from the traditional 15.2 seconds to 12.4 seconds, demonstrating the comprehensive advantages of this invention in process automation, network structure simplification and inference path optimization, ensuring that rapid response capability is maintained while improving accuracy.

[0109] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based diagnostic method for thyroid ultrasound images, characterized in that, Includes the following steps: A set of raw thyroid ultrasound images was acquired and preprocessed to obtain a standardized set of thyroid ultrasound images. Lesion segmentation is performed on a standardized set of thyroid ultrasound images to generate an initial segmentation mask; Morphological constraints and boundary refinement are applied to the initial segmentation mask to obtain the set of thyroid lesion regions and their boundary contours. Calculate the blood flow direction, blood flow velocity, and blood flow power of each thyroid lesion region and its surrounding area to generate a set of blood flow feature vectors for thyroid lesions; A standardized set of thyroid ultrasound images is input into a feature-adaptive deep learning network trained on a thyroid ultrasound dataset to generate a preliminary fusion feature map, which is then fused with a set of thyroid lesion blood flow feature vectors to generate a thyroid lesion fusion feature tensor. The thyroid fusion feature tensor is input into the benign and malignant thyroid lesion discrimination branch, and the benign and malignant probability value and classification label of each thyroid lesion are output. After confidence filtering and non-maximum suppression, the list of thyroid lesion detection is obtained. Structured diagnostic data records for thyroid lesions are generated based on the thyroid lesion detection list and written into the computer-aided diagnostic system.

2. The deep learning-based thyroid ultrasound image diagnostic method according to claim 1, characterized in that, The preprocessing includes speckle noise suppression, background artifact suppression, contrast-controlled enhancement, spatial resolution unification, and cross-device intensity normalization.

3. The deep learning-based thyroid ultrasound image diagnostic method according to claim 1, characterized in that, The generation of the initial segmentation mask specifically includes: A standardized set of thyroid ultrasound images is input into a deep convolutional neural network to extract multi-layer feature representations for each pixel in the image and output a lesion probability map corresponding to each pixel. The probability value of each pixel in the lesion probability map is compared with a preset threshold to form an initial segmentation mask.

4. The deep learning-based thyroid ultrasound image diagnostic method according to claim 1, characterized in that, The generation of the thyroid lesion region set and boundary contour specifically includes: Morphological opening and closing operations are performed on the initial segmentation mask. The opening operation is used to remove low-probability isolated pixel clusters, and the closing operation is used to fill the tiny cavities inside the lesion. The structuring element adopts a circular template that matches the image resolution, and the template radius is set according to the resolution of the thyroid ultrasound image. After morphological processing, the edge gradient operator is used to detect the boundary position of the initial segmentation mask, and the boundary is refined by applying a sub-pixel level interpolation method. Perform connected component analysis on the initial segmentation mask after boundary refinement, count the area and aspect ratio of each connected component, remove regions with an area smaller than the minimum threshold or an aspect ratio exceeding the set range, and retain only connected components with reasonable shape as the lesion region set. Extract the corresponding boundary contours from the set of lesion regions, record the boundary contours of each lesion region as a point set, and generate a set of thyroid lesion regions and boundary contours.

5. The deep learning-based thyroid ultrasound image diagnostic method according to claim 1, characterized in that, The generation of the set of blood flow feature vectors for thyroid lesions specifically includes: Based on the set of thyroid lesion areas, the lesion neighborhood area is formed by expanding outward by a preset pixel distance according to its boundary contour. The thyroid lesion area and the lesion neighborhood area are merged as the calculation range of blood flow parameters. Within the range of blood flow parameter calculation, color Doppler codes are analyzed, hue is mapped to blood flow direction, linear velocity is calculated using Doppler frequency shift and ultrasound emission center frequency, and blood flow power intensity is converted based on Doppler signal amplitude and system gain. Blood flow direction, blood flow velocity and blood flow power are obtained for each thyroid lesion area and its neighborhood, forming blood flow characteristics. The blood flow features are spliced ​​together in a set order to obtain a standardized blood flow feature vector. The standardized blood flow feature vectors of all lesions are then combined to form a set of blood flow feature vectors for thyroid lesions.

6. The deep learning-based thyroid ultrasound image diagnostic method according to claim 1, characterized in that, The generation of the thyroid lesion fusion feature tensor specifically includes: A standardized set of thyroid ultrasound images is used as a single-channel input tensor to be input into a feature-adaptive deep learning network. The height of the single-channel input tensor is the number of rows in the image, the width is the number of columns in the image, and the pixel value ranges from zero to one. The feature-adaptive deep learning network includes a multi-scale convolutional encoder, a speckle noise perception attention module, a pyramid feature fusion structure, and a benign or malignant thyroid lesion discrimination branch. Multiple parallel convolutional paths are set in a multi-scale convolutional encoder. Each path uses a different convolutional kernel size and dilation rate to extract fine-grained texture features suitable for small lesion identification and structural boundary features suitable for structural localization from thyroid ultrasound images. The features output by all paths are concatenated by channels and compressed by one-to-one convolution to obtain fused multi-scale encoded features. The multi-scale convolutional encoder includes a parallel convolutional path module, a channel concatenation module, and a feature compression module. The parallel convolutional path module contains three convolutional paths with different convolutional kernel sizes and dilation rates. The channel concatenation module is used to concatenate the output features of the parallel convolutional path module according to the channel dimension. The feature compression module uses one-to-one convolution to compress the concatenated multi-scale features. The multi-scale encoded features are input into the speckle noise perception attention module to calculate the speckle noise intensity score of each region. Differential weights are assigned to different regions based on the speckle noise intensity scores, and speckle noise suppression features are output. The speckle noise perception attention module includes a noise intensity estimation unit, an attention weight calculation unit, and a feature reweighting unit. The noise intensity estimation unit scores the speckle noise intensity of local regions with the multi-scale encoded features. The attention weight calculation unit generates a region-level weight distribution based on the speckle noise intensity scores. The feature reweighting unit multiplies the region-level weight distribution with the multi-scale encoded features channel by channel to output speckle noise suppression features. The speckle noise suppression features are fused with different levels of encoded features through upsampling and skip connections, and convolution, batch normalization and nonlinear activation operations are performed until the spatial resolution is restored to the same as the input tensor to generate segmentation decoding features; The segmented and decoded features are input into the pyramid feature fusion structure. After performing a one-to-one convolution on the features at each scale, they are aligned to a uniform resolution using bilinear interpolation. The summation is then performed at the corresponding pixel positions to generate a preliminary fused feature map. The preliminary fusion feature map is fused with the set of blood flow feature vectors of thyroid lesions to generate the fusion feature tensor of thyroid lesions.

7. The deep learning-based thyroid ultrasound image diagnostic method according to claim 6, characterized in that, The generation of the thyroid lesion fusion feature tensor specifically includes: The preliminary fusion feature map is normalized to generate a normalized lesion feature set. The preliminary fusion feature map contains grayscale texture information and local detail texture information under different receptive field conditions. Extract the blood flow feature vector of the thyroid lesion corresponding to the current lesion, and splice and fuse it with the normalized lesion feature set in the feature channel dimension to generate lesion-level fusion features; After channel-weighted processing of the lesion-level fusion features, they are input into a fully connected mapping layer, which outputs a thyroid lesion fusion feature tensor that integrates texture, morphology, multi-scale and blood flow information. The channel-weighted weights are obtained by a weight generation network from the statistical response generated by global average pooling.

8. The deep learning-based thyroid ultrasound image diagnostic method according to claim 1, characterized in that, The generation of the thyroid lesion detection list specifically includes: The thyroid fusion feature tensor is input into the thyroid lesion benign or malignant discrimination branch and the output is generated with the probability value of benign or malignant thyroid lesion and the classification label. The thyroid lesion benign or malignant discrimination branch includes several convolutional layers, global average pooling layers and fully connected layers. The probability value of benign or malignant thyroid lesion is a real number ranging from zero to one, and the classification label is benign or malignant. The benign or malignant probability value and classification label of each thyroid lesion are evaluated for confidence. Lesions with a confidence level below the preset confidence threshold are removed, and only lesions that meet the confidence requirements are retained. The preset confidence threshold is set based on a balance between clinical sensitivity and specificity. After screening, non-maximum suppression is performed on the lesions according to the classification labels. The overlap ratio between candidate lesions is calculated and combined with the confidence priority. Only the candidate lesions with the highest confidence priority are retained, and the remaining candidate lesions are removed. The final list of thyroid lesion detection is then output.

9. The deep learning-based thyroid ultrasound image diagnostic method according to claim 1, characterized in that, The writing to the computer-aided diagnostic system specifically includes: Extract the location information and boundary contour of each lesion in the thyroid lesion detection list. The location information includes the horizontal and vertical coordinates and depth information of the lesion in the standardized thyroid ultrasound image coordinate system. The boundary contour is a set of continuous closed contour points obtained by fitting the lesion segmentation results. Blood flow direction, blood flow velocity, and blood flow power are extracted from the corresponding lesion area and its neighborhood to generate a blood flow parameter set; The location information, boundary contour, benign and malignant probability values ​​of thyroid lesions, classification labels and blood flow parameter sets are combined to form a structured diagnostic data record of thyroid lesions; The generated structured diagnostic data of thyroid lesions is recorded and written into the computer-aided diagnostic system.

Images