A medical image analysis method, electronic device, and computer program product
By using a dual-frequency uncertainty network model, we have addressed the shortcomings of existing models in defining the weight differences of morphological ambiguity and blurred edges in the diagnosis of breast nodules. This model enables the accurate capture and uncertainty assessment of high-frequency and low-frequency pathological features, thereby improving the accuracy of diagnosis and clinical suitability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 脉得智能科技(无锡)有限公司
- Filing Date
- 2026-02-13
- Publication Date
- 2026-06-02
AI Technical Summary
Existing deep learning models lack an effective mechanism to define the weight differences between morphological ambiguity and blurred edges in the diagnosis of breast nodules. Their confidence assessment is out of touch with the actual difficulty of clinical diagnosis and lacks clinical guidance significance.
A dual-frequency uncertainty network model is adopted. Through the collaborative design of encoder, dual-frequency masking processing module and decision module, high-frequency and low-frequency pathological features are accurately captured, uncertainty scores are generated, and the final prediction results are output.
It achieves precise capture and uncertainty quantification of high-frequency and low-frequency pathological features, and the assessment results are highly consistent with the actual clinical diagnostic difficulty, thus improving the accuracy and reliability of diagnosis.
Smart Images
Figure CN122134668A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a medical image analysis method, electronic device, and computer program product. Background Technology
[0002] As an important exocrine gland in the human body, the breast is finely regulated by endocrine hormones, and its physiological structure dynamically changes with physiological processes such as the menstrual cycle, pregnancy, and lactation. Currently, breast cancer has become the most common malignant tumor among women worldwide, seriously threatening women's lives and health. Breast nodules are the most common clinical manifestation of breast diseases, and their pathological nature is complex and diverse. In clinicopathology, visual features at different frequencies carry different diagnostic information: macroscopic morphological structures (such as aspect ratio and overall growth direction) usually correspond to the low-frequency components of the image, while microscopic edge features (such as spiculations, microcalcifications, and angularity) correspond to the high-frequency components of the image. Studies have confirmed that these two types of features have equally important but distinctly different clinical value in differentiating benign and malignant breast nodules. Accurately analyzing and comprehensively evaluating both is the core key to achieving precise diagnosis.
[0003] In the clinical screening and diagnosis of breast diseases, ultrasound imaging (US) has become the preferred examination method due to its advantages such as no radiation, real-time imaging, and high sensitivity. However, ultrasound images are easily affected by speckle noise, artifacts, and sectional angle deviations, often resulting in blurred lesion edges (caused by high-frequency noise interference) or incomplete morphological display (caused by missing low-frequency information). In clinical practice, the qualitative diagnosis of breast nodules highly depends on the physician's subjective clinical experience, and the imaging characteristics of some atypical benign nodules and early malignant nodules overlap significantly, further blurring the boundary between benign and malignant nodules. For such difficult cases, different physicians have significant differences in their risk weighting of "morphological ambiguity" and "blurred edges," resulting in low diagnostic consistency and directly affecting diagnostic efficiency.
[0004] In recent years, deep learning technologies, represented by convolutional neural networks (CNNs), have made groundbreaking progress in the field of medical image analysis, demonstrating performance comparable to or even surpassing that of human experts in the automatic detection and classification of breast nodules. Existing deep learning models mostly employ deterministic point estimation, outputting only a single classification probability value (e.g., 0.85). However, this prediction mode has significant limitations: the model is prone to "overconfidence," tending to give high-confidence predictions even when faced with out-of-distribution samples or complex cases with ambiguous features. Furthermore, existing models struggle to distinguish between "data uncertainty" (caused by poor image quality) and "model uncertainty" (caused by complex lesion morphology and ambiguous features). More importantly, there is a lack of effective mechanisms to define the weight difference between "morphological ambiguity" and "marginal ambiguity" in diagnostic decisions, resulting in a final confidence assessment that is often disconnected from the actual difficulty of clinical diagnosis (e.g., BI-RADS classification logic), lacking practical clinical guidance value. Summary of the Invention
[0005] The purpose of this application is to provide a medical image analysis method, electronic device, and computer program product to solve the problems that existing models lack an effective mechanism to define the weight difference between "morphological ambiguity" and "marginal blur" in diagnostic decision-making, and that confidence assessment is often out of touch with the actual clinical diagnostic difficulty and lacks clinical guidance significance.
[0006] This application provides a medical image analysis method, including: The target medical image is input into a dual-frequency uncertainty network model to obtain the final prediction result and uncertainty score; The dual-frequency uncertainty network model includes an encoder, a dual-frequency masking processing module, and a decision module; the dual-frequency masking processing module is used to capture high-frequency and low-frequency pathological features.
[0007] The aforementioned technical solution utilizes a dual-frequency uncertainty network model to accurately capture and quantify the uncertainty of high-frequency and low-frequency pathological features at the architectural level. This not only outputs the final prediction result but also simultaneously generates an uncertainty score, ensuring that model predictions are not limited to a single probability value but consider both diagnostic results and risk confidence levels. This approach allows the assessment results to closely align with the actual difficulty of clinical diagnosis, effectively addressing the shortcomings of existing models in terms of clinical adaptability. The core dual-frequency uncertainty network model, through the collaborative design of the encoder, dual-frequency masking processing module, and decision module, constructs an analytical framework that conforms to the patterns of clinical pathological features. The dual-frequency masking processing module is specifically designed for the high-frequency and low-frequency features of breast imaging, accurately capturing microscopic edge features (high-frequency components such as spiculations, microcalcifications, and angularity) and macroscopic morphological structures (low-frequency components such as aspect ratio and growth direction), achieving full-dimensional coverage of pathological features.
[0008] In some alternative implementations, the encoder is used to: extract deep, high-dimensional features F through hierarchical convolution and downsampling. ; Where X represents the target medical image. The feature map output by the encoder. , , These represent the number of channels, height, and width of the feature, respectively.
[0009] In the above technical solution, the encoder can adopt the ResNet-50 backbone network, and the deep high-dimensional feature F implies the macroscopic morphology and microscopic edge information of the breast nodules.
[0010] In some alternative implementations, the dual-frequency masking processing module includes: The feature decoupling unit is used to decompose deep high-dimensional features into low-frequency features and high-frequency features; The adaptive unit is used to obtain adaptive morphological weights based on low-frequency and high-frequency features. and edge weights ; The mask processing unit is used to: randomly generate N masks, and perform the following processing on each mask: configure the mask as a first scale parameter, process the low-frequency features to obtain the low-frequency features after masking corresponding to the mask; configure the mask as a second scale parameter, process the high-frequency features to obtain the high-frequency features after masking corresponding to the mask; wherein, the first scale parameter is greater than the second scale parameter; Feature fusion unit, used for: based on morphological weights and edge weights We perform weighted fusion of the low-frequency and high-frequency features after masking for each mask to obtain the final feature for each mask.
[0011] In the above technical solution, the feature decoupling unit first explicitly decomposes the hybrid feature F output by the encoder into low-frequency morphological features (corresponding to aspect ratio and growth direction) and high-frequency edge features (corresponding to burrs and microcalcifications) to avoid mutual interference between features of different dimensions; the adaptive unit dynamically calculates the morphological weight α and edge weight β according to the feature distribution of the current case (such as a case with very obvious burr features but atypical morphological features), rather than fixed values, so that the fused features can better reflect the real clinical lesion features.
[0012] This scheme also employs a structured dual-scale masking mechanism. For low-frequency morphological features, the mask is configured with the first scale parameter (larger scale) because low-frequency features are macroscopic and global, requiring a larger receptive field coverage. For high-frequency edge features, the mask is configured with the second scale parameter (smaller scale) because high-frequency features are microscopic and local, and a smaller scale mask can more accurately capture detailed features. This differentiated scale design ensures that masking processing does not destroy the inherent properties of high-frequency / low-frequency features, thereby improving the accuracy of feature processing.
[0013] In some alternative implementations, deep high-dimensional features are decomposed into low-frequency features and high-frequency features, including: The Laplace pyramid principle is used to orthogonally decompose deep high-dimensional features to obtain the decomposed features. After decomposition, low-frequency features are extracted using adaptive average pooling and upsampling operations. ; Subtracting low-frequency features from the decomposed features yields high-frequency features. .
[0014] In the above technical solution, the Laplace pyramid principle is used to analyze the features. Perform orthogonal decomposition to eliminate frequency domain entanglement.
[0015] For low-frequency feature flows (morphological flows), adaptive average pooling is used. and upsampling operation Extracting low-frequency features Used to characterize the overall outline and aspect ratio of nodules: ; Where k represents the kernel size of the average pooling (AvgPool) operation.
[0016] For high-frequency feature flows (edge flows), the concept of residuals is used to subtract low-frequency features from the original features to obtain high-frequency features. Used to characterize the edge burrs and microcalcifications of nodules: .
[0017] In some alternative implementations, N masks are randomly generated, including: Assume the total number of channels in the feature map is For each mask The number of channels with a value of 1 represents the feature channels that need to be activated, denoted as . The feature channels with a value of 0 are those that need to be discarded.
[0018] Scale parameters The relationship between the number of activations and activations is inverse:
[0019] in, The larger, The smaller the value, the more sparse the mask becomes, as a large number of bits in the mask are 0.
[0020] When generated When using a random mask, two masks are needed. and The probability of activating the same channel simultaneously is approximately .therefore, The larger the value, the lower the overlap probability, the fewer features shared between sub-models, and the stronger their independence. In the initial training phase, features are randomly generated. A fixed set of binary mask matrices is formed from several masks. Each mask dimension is (Right now ).
[0021] In some optional implementations, the mask is configured as a first scale parameter to process low-frequency features, obtaining the masked low-frequency features corresponding to the mask; the mask is configured as a second scale parameter to process high-frequency features, obtaining the masked high-frequency features corresponding to the mask, specifically including: Low-frequency information processing: Targeted Characterized by strong semantic ambiguity, it employs highly independent configurations (larger scale parameters). ,like This results in a low mask overlap rate, forcing the sub-models to learn diverse morphological representations.
[0022] ; High-frequency information processing: Targeting Due to its susceptibility to noise interference, a highly stable configuration (smaller scale parameters) is adopted. ,like This results in a high mask overlap rate, enhancing the consensus of the sub-model on weak edge signals.
[0023] .
[0024] In some alternative implementations, adaptive shape weights are obtained based on low-frequency and high-frequency features. and edge weights ,include: Low-frequency features and high frequency characteristics Global average pooling is performed in the spatial dimension, and the result is concatenated in the channel dimension to obtain the context vector; The context vector is input into a lightweight multilayer perceptron to obtain the shape weights. and edge weights .
[0025] In the above technical solution, a task-driven implicit supervision mechanism is introduced to dynamically integrate information based on the feature dependencies of the current sample. Simultaneously, to ensure the weights... and It can objectively reflect the feature dependencies of the input image itself, without being affected by the randomness of the mask, and uses complete low-frequency and high-frequency features for weight calculation. Specifically, it assumes the original low-frequency features after Laplacian decoupling. and original high-frequency features All dimensions are ,in For batch size, For the number of channels, This represents the spatial resolution of the feature map.
[0026] First, respectively and Global average pooling (GAP) is performed along the spatial dimension to compress two-dimensional spatial information into channel descriptors, with a dimension of... Then, concatenation is performed along the channel dimension, resulting in a concatenated dimension. ; ; Then, the context vector Input a lightweight multilayer perceptron (MLP) with a bottleneck structure to generate attention weights. This MLP consists of two fully connected layers: the first layer maps features to a low-dimensional latent space with a compression ratio of [value missing]. The first layer activates the latent features using the ReLU activation function; the second layer maps the latent features to two output nodes and normalizes them using the Softmax function to obtain the morphological weights. and edge weights ;
[0027] in Dimensions The dimension of the first layer output (hidden layer) of an MLP is: Second layer output (weights): .
[0028] Using the generated weights to apply to the masked random features and (All dimensions are) The broadcast-weighted fusion is performed to obtain the final feature representation used for classification. : ; This approach ensures that the more a model relies on features of a certain frequency band in a classification task, the greater the weight of that feature in subsequent uncertainty assessment.
[0029] In some alternative implementations, the decision module is used for: Based on the final features corresponding to each mask, the corresponding prediction results are obtained; The average of the prediction results corresponding to the N masks is taken as the final prediction result.
[0030] In the above technical solution, the decision module adopts a combination of Global Average Pooling (GAP) and Fully Connected (FC) layers to adapt to the needs of medical image classification scenarios. Global Average Pooling compresses the final feature map (dimension C×H×W) corresponding to each mask into a one-dimensional vector of length C. Compared to the direct flattening by FC layers, GAP preserves the spatial global information of the features, avoids overfitting, and its output dimension is only related to the number of channels C, not changing with the feature map size, thus adapting to medical image inputs of different sizes. The FC layer maps the one-dimensional vector output by GAP to predicted probabilities (e.g., in a binary classification scenario, outputting [0.1, 0.9], corresponding to benign and malignant probabilities, respectively).
[0031] In some alternative implementations, the decision module is also used for: Based on the low-frequency features behind each mask, the corresponding low-frequency prediction result is obtained; Based on the high-frequency features behind each mask, the corresponding high-frequency prediction result is obtained.
[0032] In some alternative implementations, the dual-frequency uncertainty network model further includes: The uncertainty assessment module is used for: Calculate the first standard deviation of the low-frequency prediction results corresponding to N masks; Calculate the second standard deviation of the high-frequency prediction results corresponding to N masks; The uncertainty score is obtained based on the first and second standard deviations.
[0033] In the above technical solution, during the inference phase, the model executes... Each forward propagation (or parallel processing) uses a different mask from the mask set each time.
[0034] Low-frequency uncertainty ( ): calculate In the subsequent sampling, different masks were selected. Obtaining low-frequency features Its corresponding low-frequency prediction results The standard deviation is the largest possible deviation. A larger standard deviation indicates a greater discrepancy in morphological predictions under different masks, meaning that low-frequency features (morphological features) are more unstable.
[0035] ; High-frequency uncertainty ( ): calculate In the subsequent sampling, different masks were selected. High frequency characteristics Its corresponding high-frequency prediction results The standard deviation of the features. The larger the standard deviation, the lower the model's recognition of edge details (e.g., it cannot distinguish between noise and burrs).
[0036] .
[0037] In some alternative implementations, uncertainty scoring is performed:
[0038] in, The first standard deviation, This is the second standard deviation.
[0039] In the above technical solution, the morphological weights learned by the adaptive unit are reused. and edge weights Based on the orthogonal composition rule, , Two independent uncertainty indicators are merged into a single comprehensive uncertainty score. This indicator not only reflects the dispersion of the model's predictions, but also incorporates the image's own feature attributes, thus achieving a weighted assessment of risk.
[0040] In some alternative implementations, the dual-frequency uncertainty network model employs a cross-entropy loss optimization model during training.
[0041] The cross-entropy loss is as follows: ; Among them, y k Let K be the true label of the k-th sample, and K be the number of samples. Let be the predicted probability of the k-th sample. Attached Figure Description
[0042] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0043] Figure 1 This is a schematic diagram of a dual-frequency uncertainty network model provided in an embodiment of this application; Figure 2 This is a schematic diagram of the dual-frequency masking processing module provided in an embodiment of this application; Figure 3 This is a schematic diagram of mask generation provided for an embodiment of this application; Figure 4 This is a schematic diagram of a possible structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0044] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.
[0045] Please refer to Figure 1 , Figure 1 A schematic diagram of a dual-frequency uncertainty network model provided in this application embodiment. Based on this model, this application embodiment provides a medical image analysis method, including: The target medical image is input into a dual-frequency uncertainty network model to obtain the final prediction result and uncertainty score; The dual-frequency uncertainty network model includes an encoder, a dual-frequency masking processing module, and a decision module; the dual-frequency masking processing module is used to capture high-frequency and low-frequency pathological features.
[0046] The aforementioned technical solution utilizes a dual-frequency uncertainty network model to accurately capture and quantify the uncertainty of high-frequency and low-frequency pathological features at the architectural level. This not only outputs the final prediction result but also simultaneously generates an uncertainty score, ensuring that model predictions are not limited to a single probability value but consider both diagnostic results and risk confidence levels. This approach allows the assessment results to closely align with the actual difficulty of clinical diagnosis, effectively addressing the shortcomings of existing models in terms of clinical adaptability. The core dual-frequency uncertainty network model, through the collaborative design of the encoder, dual-frequency masking processing module, and decision module, constructs an analytical framework that conforms to the patterns of clinical pathological features. The dual-frequency masking processing module is specifically designed for the high-frequency and low-frequency features of breast imaging, accurately capturing microscopic edge features (high-frequency components such as spiculations, microcalcifications, and angularity) and macroscopic morphological structures (low-frequency components such as aspect ratio and growth direction), achieving full-dimensional coverage of pathological features.
[0047] In some alternative implementations, the encoder is used to: extract deep, high-dimensional features F through hierarchical convolution and downsampling. ; Where X represents the target medical image. The feature map output by the encoder. , , These represent the number of channels, height, and width of the feature, respectively.
[0048] In the above technical solution, the encoder can adopt the ResNet-50 backbone network, and the deep high-dimensional feature F implies the macroscopic morphology and microscopic edge information of the breast nodules.
[0049] Please refer to Figure 2 , Figure 2 This is a schematic diagram of the dual-frequency masking processing module provided in an embodiment of this application.
[0050] In some alternative implementations, the dual-frequency masking processing module includes: The feature decoupling unit is used to decompose deep high-dimensional features into low-frequency features and high-frequency features; The adaptive unit is used to obtain adaptive morphological weights based on low-frequency and high-frequency features. and edge weights ; The mask processing unit is used to: randomly generate N masks, and perform the following processing on each mask: configure the mask as a first scale parameter, process the low-frequency features to obtain the low-frequency features after masking corresponding to the mask; configure the mask as a second scale parameter, process the high-frequency features to obtain the high-frequency features after masking corresponding to the mask; wherein, the first scale parameter is greater than the second scale parameter; Feature fusion unit, used for: based on morphological weights and edge weights We perform weighted fusion of the low-frequency and high-frequency features after masking for each mask to obtain the final feature for each mask.
[0051] In the above technical solution, the feature decoupling unit first explicitly decomposes the hybrid feature F output by the encoder into low-frequency morphological features (corresponding to aspect ratio and growth direction) and high-frequency edge features (corresponding to burrs and microcalcifications) to avoid mutual interference between features of different dimensions; the adaptive unit dynamically calculates the morphological weight α and edge weight β according to the feature distribution of the current case (such as a case with very obvious burr features but atypical morphological features), rather than fixed values, so that the fused features can better reflect the real clinical lesion features.
[0052] This scheme also employs a structured dual-scale masking mechanism. For low-frequency morphological features, the mask is configured with the first scale parameter (larger scale) because low-frequency features are macroscopic and global, requiring a larger receptive field coverage. For high-frequency edge features, the mask is configured with the second scale parameter (smaller scale) because high-frequency features are microscopic and local, and a smaller scale mask can more accurately capture detailed features. This differentiated scale design ensures that masking processing does not destroy the inherent properties of high-frequency / low-frequency features, thereby improving the accuracy of feature processing.
[0053] In some alternative implementations, deep high-dimensional features are decomposed into low-frequency features and high-frequency features, including: The Laplace pyramid principle is used to orthogonally decompose deep high-dimensional features to obtain the decomposed features. After decomposition, low-frequency features are extracted using adaptive average pooling and upsampling operations. ; Subtracting low-frequency features from the decomposed features yields high-frequency features. .
[0054] In the above technical solution, the Laplace pyramid principle is used to analyze the features. Perform orthogonal decomposition to eliminate frequency domain entanglement.
[0055] For low-frequency feature flows (morphological flows), adaptive average pooling is used. and upsampling operation Extracting low-frequency features Used to characterize the overall outline and aspect ratio of nodules: ; Where k represents the kernel size of the average pooling (AvgPool) operation.
[0056] For high-frequency feature flows (edge flows), the concept of residuals is used to subtract low-frequency features from the original features to obtain high-frequency features. Used to characterize the edge burrs and microcalcifications of nodules: .
[0057] In some alternative implementations, N masks are randomly generated, including: Assume the total number of channels in the feature map is For each mask The number of channels with a value of 1 represents the feature channels that need to be activated, denoted as . The feature channels with a value of 0 are those that need to be discarded.
[0058] Scale parameters The relationship between the number of activations and activations is inverse:
[0059] in, The larger, The smaller the value, the more sparse the mask becomes, as a large number of bits in the mask are 0.
[0060] When generated When using a random mask, two masks are needed. and The probability of activating the same channel simultaneously is approximately .therefore, The larger the value, the lower the overlap probability, the fewer features shared between sub-models, and the stronger the independence.
[0061] The generation of the mask set is as follows Figure 3 As shown, assuming the number of feature channels C is 8, the scale parameters S of the two generated mask sets are 4 and 1.5 respectively, and the number of generated masks N is 4. Each mask M^((i)) corresponds to a column in the matrix. Figure 3 In the mask set on the left, each mask will have 2 feature channels activated, and in the mask set on the right, each mask will have 5 feature channels activated.
[0062] In the initial training phase, random generation A fixed set of binary mask matrices is formed from several masks. Each mask dimension is (Right now ).
[0063] In some optional implementations, the mask is configured as a first scale parameter to process low-frequency features, obtaining the masked low-frequency features corresponding to the mask; the mask is configured as a second scale parameter to process high-frequency features, obtaining the masked high-frequency features corresponding to the mask, specifically including: Low-frequency information processing: Targeted Characterized by strong semantic ambiguity, it employs highly independent configurations (larger scale parameters). ,like This results in a low mask overlap rate, forcing the sub-model to learn diverse morphological representations. The low-frequency features after masking are: ; High-frequency information processing: Targeting Due to its susceptibility to noise interference, a highly stable configuration (smaller scale parameters) is adopted. ,like This results in a high mask overlap rate, enhancing the consensus of the sub-model regarding weak edge signals. The high-frequency features after masking are: .
[0064] In some alternative implementations, adaptive shape weights are obtained based on low-frequency and high-frequency features. and edge weights ,include: Low-frequency features and high frequency characteristics Global average pooling is performed in the spatial dimension, and the result is concatenated in the channel dimension to obtain the context vector; The context vector is input into a lightweight multilayer perceptron to obtain the shape weights. and edge weights .
[0065] In the above technical solution, a task-driven implicit supervision mechanism is introduced to dynamically integrate information based on the feature dependencies of the current sample. Simultaneously, to ensure the weights... and It can objectively reflect the feature dependencies of the input image itself, without being affected by the randomness of the mask, and uses complete low-frequency and high-frequency features for weight calculation. Specifically, it assumes the original low-frequency features after Laplacian decoupling. and original high-frequency features All dimensions are ,in For batch size, For the number of channels, This represents the spatial resolution of the feature map.
[0066] First, respectively and Global average pooling (GAP) is performed along the spatial dimension to compress two-dimensional spatial information into channel descriptors, with a dimension of... Then, concatenation is performed along the channel dimension, resulting in a concatenated dimension. ; ; Then, the context vector Input a lightweight multilayer perceptron (MLP) with a bottleneck structure to generate attention weights. This MLP consists of two fully connected layers: the first layer maps features to a low-dimensional latent space with a compression ratio of [value missing]. The first layer activates the latent features using the ReLU activation function; the second layer maps the latent features to two output nodes and normalizes them using the Softmax function to obtain the morphological weights. and edge weights ;
[0067] in Dimensions The dimension of the first layer output (hidden layer) of an MLP is: Second layer output (weights): .
[0068] Using the generated weights to apply to the masked random features and (All dimensions are) The broadcast-weighted fusion is performed to obtain the final feature representation used for classification. : ; This approach ensures that the more a model relies on features of a certain frequency band in a classification task, the greater the weight of that feature in subsequent uncertainty assessment.
[0069] In some alternative implementations, the decision module is used for: Based on the final features corresponding to each mask, the corresponding prediction results are obtained; The average of the prediction results corresponding to the N masks is taken as the final prediction result.
[0070] In the above technical solution, the decision module adopts a combination of Global Average Pooling (GAP) and Fully Connected (FC) layers to adapt to the needs of medical image classification scenarios. Global Average Pooling compresses the final feature map (dimension C×H×W) corresponding to each mask into a one-dimensional vector of length C. Compared to the direct flattening by FC layers, GAP preserves the spatial global information of the features, avoids overfitting, and its output dimension is only related to the number of channels C, not changing with the feature map size, thus adapting to medical image inputs of different sizes. The FC layer maps the one-dimensional vector output by GAP to predicted probabilities (e.g., in a binary classification scenario, outputting [0.1, 0.9], corresponding to benign and malignant probabilities, respectively).
[0071] In some alternative implementations, the decision module is also used for: Based on the low-frequency features behind each mask, the corresponding low-frequency prediction result is obtained; Based on the high-frequency features behind each mask, the corresponding high-frequency prediction result is obtained.
[0072] In some alternative implementations, the dual-frequency uncertainty network model further includes: The uncertainty assessment module is used for: Calculate the first standard deviation of the low-frequency prediction results corresponding to N masks; Calculate the second standard deviation of the high-frequency prediction results corresponding to N masks; The uncertainty score is obtained based on the first and second standard deviations.
[0073] In the above technical solution, during the inference phase, the model executes... Each forward propagation (or parallel processing) uses a different mask from the mask set each time.
[0074] Low-frequency uncertainty ( ): calculate In the subsequent sampling, different masks were selected. Obtaining low-frequency features Its corresponding low-frequency prediction results The standard deviation is the largest possible deviation. A larger standard deviation indicates a greater discrepancy in morphological predictions under different masks, meaning that low-frequency features (morphological features) are more unstable.
[0075] ; High-frequency uncertainty ( ): calculate In the subsequent sampling, different masks were selected. High frequency characteristics Its corresponding high-frequency prediction results The standard deviation of the features. The larger the standard deviation, the lower the model's recognition of edge details (e.g., it cannot distinguish between noise and burrs).
[0076] .
[0077] In some alternative implementations, uncertainty scoring is performed:
[0078] in, The first standard deviation, This is the second standard deviation.
[0079] In the above technical solution, the morphological weights learned by the adaptive unit are reused. and edge weights Based on the orthogonal composition rule, , Two independent uncertainty indicators are merged into a single comprehensive uncertainty score. This indicator not only reflects the dispersion of the model's predictions, but also incorporates the image's own feature attributes, thus achieving a weighted assessment of risk.
[0080] In some alternative implementations, the dual-frequency uncertainty network model employs a cross-entropy loss optimization model during training.
[0081] The cross-entropy loss is as follows: ; Among them, y k Let K be the true label of the k-th sample, and K be the number of samples. Let be the predicted probability of the k-th sample.
[0082] The training and inference process of the dual-frequency uncertainty network model is as follows: Obtain a dataset of breast ultrasound images, including benign and malignant nodule samples and their corresponding pathological labels (gold standard).
[0083] Data preprocessing is performed, including image normalization, size standardization, and data augmentation (rotation, flipping, brightness adjustment), etc.
[0084] The dataset is divided into three parts: a training set (70%) for parameter optimization; a validation set (15%) for monitoring overfitting and adjusting hyperparameters; and a test set (15%) for final performance evaluation.
[0085] To train the model, for each input sample, a random sampling mask is applied and the classification loss is calculated to update the network parameters.
[0086] During the testing phase, a complete [process] was used for each sample. Inference is performed using a set of masks. First, the statistical standard deviation of the characteristics of each frequency band is calculated. and Then reuse the final model weights. It outputs the final benign or malignant prediction probability (mean) and the comprehensive uncertainty score (weighted standard deviation) to assist doctors in making more cautious clinical decisions.
[0087] In summary, this proposal suggests a deep learning network architecture based on dual-frequency decoupling and mask ensemble. The core technology lies in using a Laplacian frequency domain decoupling mechanism to forcibly separate the low-frequency morphological flow from the high-frequency edge flow, enabling independent capture of the macroscopic morphology and microscopic edge features of breast nodules. For the physical properties of different frequency band features (such as semantic ambiguity of morphology and noise sensitivity of edges), each branch is designed with parameterized mask ensemble units with differentiated parameters, significantly improving the robustness and stability of the model in extracting specific pathological features from complex ultrasound images.
[0088] This solution employs a task-driven adaptive uncertainty fusion mechanism, which scientifically integrates morphological uncertainty and edge uncertainty by introducing an implicitly supervised attention weight generation strategy. This mechanism utilizes unmasked raw global features to generate fixed feature importance weights, and then reuses them for statistical dispersion analysis of multi-masked random sampling (i.e., uncertainty calculation based on standard deviation). This effectively solves the problem that existing technologies cannot distinguish between "morphological ambiguity" and "edge blurring" risk weights, leading to a lack of clinical interpretability in confidence assessments.
[0089] This solution is not only applicable to the classification and risk assessment of benign and malignant breast nodules, but can also be extended to other medical image analysis tasks with significant differences in frequency domain characteristics (such as thyroid nodule classification and lung nodule diagnosis). It is particularly suitable for difficult diagnostic scenarios with complex lesion morphology, blurred edges, and the phenomenon of "different images for the same disease". Its plug-and-play modular design (such as the Laplacian decoupling unit and statistical uncertainty assessment module) allows for seamless integration with other mainstream classification networks (such as ResNet, DenseNet, etc.), and has broad applicability, flexibility, and clinical translational value.
[0090] Figure 4 This illustration shows a possible structure of an electronic device provided in an embodiment of this application. (Refer to...) Figure 4 The electronic device includes a processor, memory, and a communication interface, which are interconnected and communicate with each other via a communication bus and / or other forms of connection mechanism (not shown).
[0091] The memory includes one or more (only one is shown in the figure), which can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The processor and other possible components can access the memory to read and / or write data to it.
[0092] The processor comprises one or more (only one is shown in the figure), which can be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Microcontroller Unit (MCU), a Network Processor (NP), or other conventional processors; it can also be a special-purpose processor, including a Neural-network Processing Unit (NPU), a Graphics Processing Unit (GPU), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Furthermore, when there are multiple processors, some can be general-purpose processors, and others can be special-purpose processors.
[0093] The communication interface includes one or more (only one is shown in the figure), which can be used to communicate directly or indirectly with other devices to exchange data. The communication interface may include interfaces for wired and / or wireless communication.
[0094] One or more computer program instructions may be stored in the memory, and the processor may read and execute these computer program instructions to implement the methods provided in the embodiments of this application.
[0095] Understandable. Figure 4 The structure shown is for illustrative purposes only; the electronic device may also include structures that are more complex than those shown. Figure 4 The more or fewer components shown, or having the same Figure 4 The different structures shown. Figure 4 The components shown can be implemented using hardware, software, or a combination thereof. Electronic devices may be physical devices, such as PCs, laptops, tablets, mobile phones, servers, embedded devices, etc., or they may be virtual devices, such as virtual machines, virtualized containers, etc. Furthermore, electronic devices are not limited to a single device; they can also be a combination of multiple devices or a cluster of a large number of devices.
[0096] This application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of any of the methods described above.
[0097] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.
[0098] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0099] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0100] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.
[0101] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A medical image analysis method, characterized in that, include: The target medical image is input into a dual-frequency uncertainty network model to obtain the final prediction result and uncertainty score; The dual-frequency uncertainty network model includes an encoder, a dual-frequency masking processing module, and a decision module; wherein the dual-frequency masking processing module is used to capture high-frequency and low-frequency pathological features. The dual-frequency masking processing module includes: The feature decoupling unit is used to decompose deep high-dimensional features into low-frequency features and high-frequency features; The adaptive unit is used to obtain adaptive morphological weights based on low-frequency and high-frequency features. and edge weights ; A mask processing unit is configured to: randomly generate N masks, and perform the following processing on each mask: configure the mask as a first scale parameter, process the low-frequency features to obtain the masked low-frequency features corresponding to the mask; configure the mask as a second scale parameter, process the high-frequency features to obtain the masked high-frequency features corresponding to the mask; wherein, the first scale parameter is greater than the second scale parameter; Feature fusion unit, used for: based on morphological weights and edge weights We perform weighted fusion of the low-frequency and high-frequency features after masking for each mask to obtain the final feature for each mask.
2. The method as described in claim 1, characterized in that, The encoder is used to: extract deep, high-dimensional features F through hierarchical convolution and downsampling. ； Where X represents the target medical image. The feature map output by the encoder. , , These represent the number of channels, height, and width of the feature, respectively.
3. The method as described in claim 1, characterized in that, The process of decomposing deep high-dimensional features into low-frequency features and high-frequency features includes: The Laplace pyramid principle is used to orthogonally decompose deep high-dimensional features to obtain the decomposed features. After decomposition, low-frequency features are extracted using adaptive average pooling and upsampling operations. ; Subtracting low-frequency features from the decomposed features yields high-frequency features. .
4. The method as described in claim 3, characterized in that, The adaptive morphological weights are obtained based on low-frequency and high-frequency features. and edge weights ,include: Low-frequency features and high frequency characteristics Global average pooling is performed in the spatial dimension, and the result is concatenated in the channel dimension to obtain the context vector; The context vector is input into a lightweight multilayer perceptron to obtain the shape weights. and edge weights .
5. The method as described in claim 4, characterized in that, The decision module is used for: Based on the final features corresponding to each mask, the corresponding prediction result is obtained; The average of the prediction results corresponding to the N masks is taken as the final prediction result.
6. The method as described in claim 5, characterized in that, The decision-making module is also used for: Based on the low-frequency features behind each mask, the corresponding low-frequency prediction result is obtained; Based on the high-frequency features behind each mask, the corresponding high-frequency prediction result is obtained.
7. The method as described in claim 6, characterized in that, The dual-frequency uncertainty network model also includes: The uncertainty assessment module is used for: Calculate the first standard deviation of the low-frequency prediction results corresponding to N masks; Calculate the second standard deviation of the high-frequency prediction results corresponding to N masks; The uncertainty score is obtained based on the first and second standard deviations.
8. The method as described in claim 7, characterized in that, The uncertainty score: in, The first standard deviation, This is the second standard deviation.
9. The method as described in claim 1, characterized in that, The dual-frequency uncertainty network model employs a cross-entropy loss optimization model during training.
10. An electronic device, characterized in that, include: A processor and a memory, the memory storing machine-readable instructions executable by the processor, which, when executed by the processor, perform the method as described in any one of claims 1-9.
11. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-9.