An ovarian cyst recognition method and system based on ultrasound images

Abstract

The application discloses an ovary cyst recognition method and system based on an ultrasonic image, relates to the field of ultrasonic image recognition, and comprises the following steps: a collecting module is used for collecting a continuous dynamic ultrasonic scanning sequence of a user's ovary part, and simultaneously performing sequence image normalization processing, invalid frame elimination and frame time sequence number calibration; a modeling module is used for receiving the continuous ultrasonic sequence after number calibration, constructing an interframe time sequence correlation matrix, extracting motion and gray change features of adjacent frames and cross frames in the sequence, and recording the features as interframe time sequence features; by fusing single-frame spatial texture information and interframe time sequence correlation features, the application effectively makes up for the limitations of single static image recognition, can accurately complete ovary anatomical structure segmentation and cyst candidate region positioning, and effectively avoids misjudgment problems that are prone to occur in static recognition based on time sequence consistency verification.

Description

Technical Field

[0001] This invention relates to the field of ultrasound image recognition technology, specifically to a method and system for identifying ovarian cysts based on ultrasound images. Background Technology

[0002] Ovarian cyst identification based on ultrasound images is a gynecological auxiliary diagnostic technology that integrates medical imaging and artificial intelligence. Using pelvic ultrasound images as the core data source, it automatically performs cyst region segmentation, feature quantification, and benign / malignant classification through image processing and deep learning algorithms. This effectively overcomes the clinical shortcomings of manual image interpretation, which heavily relies on physician experience, is prone to missed or misdiagnosed cases, and suffers from insufficient screening efficiency. It can assist clinicians in achieving rapid and accurate screening for ovarian cysts, improve the accuracy of early diagnosis, and provide objective quantitative evidence for disease triage and prognostic assessment.

[0003] The invention patent application with application number 202211009842.4 discloses a method and program product for identifying ovarian cysts based on ultrasound medical images. The application aims to achieve the technical goal of "identifying and predicting the type of ovarian cyst based on the patient's abdominal ultrasound".

[0004] However, existing technologies are mostly based on single-frame static ultrasound images for identification, without utilizing the temporal correlation features of dynamic ultrasound scanning. As a result, they are not sensitive enough to identify small cysts and atypical cysts, and are prone to missed diagnosis and misdiagnosis.

[0005] Therefore, we propose a method and system for identifying ovarian cysts based on ultrasound images. Summary of the Invention In view of the above-mentioned shortcomings of the existing technology, the present invention provides a method and system for identifying ovarian cysts based on ultrasound images, which can effectively solve the problems of the existing technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions; This invention discloses an ovarian cyst identification system based on ultrasound images, comprising: The system comprises the following modules: an acquisition module for acquiring continuous dynamic ultrasound scan sequences of the user's ovarian region, simultaneously performing sequence image normalization, invalid frame removal, and frame temporal numbering; a modeling module for receiving the numbered and calibrated continuous ultrasound sequences, constructing an inter-frame temporal correlation matrix, and extracting motion and grayscale change features between adjacent frames and across frames, denoted as inter-frame temporal features; a fusion module for fusing the spatial texture features of single-frame ultrasound images with inter-frame temporal features to generate a multi-dimensional fusion feature map corresponding to the sequence; a localization module for performing ovarian anatomical structure region segmentation based on the multi-dimensional fusion feature map, locating and extracting candidate regions within the sequence that conform to cyst morphological characteristics; a judgment module for performing temporal consistency verification and category discrimination of each candidate region based on inter-frame temporal features; and a storage module for receiving the judgment results from the judgment module, generating a corresponding diagnostic report, and storing the diagnostic report based on the temporal association and user personal information. The acquisition module is interconnected with the modeling module via a wireless network. The modeling module is interconnected with the fusion module via a wireless network. The fusion module is interconnected with the positioning module via a wireless network. The positioning module is interconnected with the decision module via a wireless network. The decision module is interconnected with the storage module via a wireless network.

[0007] Furthermore, during the operation phase of the acquisition module, a continuous dynamic ultrasound scan sequence of the user's ovarian region is acquired at a preset frame sampling rate, and a unique incremental time sequence number is assigned to each frame image in the sequence simultaneously. During image normalization processing, the acquisition module maps the grayscale value range of each single frame image to a preset unified range. Then, based on the grayscale distribution dispersion and the proportion of effective anatomical structures of the single frame image, it performs an invalid frame removal operation, removing invalid frames whose grayscale distribution dispersion exceeds the preset range or whose proportion of effective anatomical structures is lower than the preset threshold. The remaining valid frames maintain the original relative order of their time sequence numbers.

[0008] Furthermore, the modeling module receives a series of valid ultrasound sequences after numbering and calibration, assuming the valid frame sequence is... Where n is the total number of valid frames. For a valid frame image with time sequence number t, construct an inter-frame temporal correlation matrix M of dimension n×n, where any element in the matrix... The calculation formula is: ; In the formula: for and Normalized mutual information; for and The gradient field cosine similarity; The time interval between two frames of images; The modeling module extracts the inter-frame grayscale change gradient and structural motion displacement features between adjacent frames and within a preset cross-frame interval based on the inter-frame temporal correlation matrix M, and integrates them to generate inter-frame temporal features.

[0009] Furthermore, during the fusion module's operation phase, spatial texture features of each valid single-frame image within the sequence are extracted to generate a corresponding single-frame spatial texture feature map. Then, feature dimension alignment processing is performed between the spatial texture feature map corresponding to the single-frame image and the inter-frame temporal feature map to generate a dimension-matched spatial feature map. With time series feature map Finally, based on the spatial-temporal feature intermodulation mechanism, a multi-dimensional fused feature map is generated. ; ; In the formula: ⊙ represents the Hadamard product operation; is the Sigmoid activation function used to generate feature modulation weights; GAP is the global average pooling operation; LN is the layer normalization operation.

[0010] Furthermore, the generation process of the spatial texture feature map is as follows: After obtaining a single frame of valid image that has undergone normalization, speckle noise suppression based on neighborhood grayscale distribution consistency is first performed to preserve the anatomical structure edges and texture details in the image. The image with noise suppression is subjected to a preset number of non-downsampling multi-scale decompositions to generate multi-channel sub-image groups with different spatial resolutions. Each channel sub-image corresponds to a structural texture component in a different direction. For each channel sub-image, extract the gray-level co-occurrence statistical features, gradient magnitude distribution features, and local structure regularity features within a preset neighborhood range pixel by pixel to generate a single-channel texture feature map for the corresponding channel. Perform channel-dimensional feature aggregation and spatial-dimensional size restoration on all single-channel texture feature maps to finally generate a spatial texture feature map that matches the spatial size of the original single-frame effective image and has a preset fixed number of channels. Each pixel position in the spatial texture feature map corresponds one-to-one with the pixel position of the corresponding single frame of valid image.

[0011] Furthermore, during the operation phase of the localization module, pixel-level semantic segmentation of the ovarian parenchyma and ovarian capsule regions is performed based on the multi-dimensional fused feature map. Each pixel in the image is assigned three mutually exclusive classification labels: ovarian parenchyma, ovarian capsule, and background. Based on the classification labels, an ovarian anatomical structure segmentation mask that perfectly matches the size of the original ultrasound frame image is generated. In the segmentation mask, the ovarian parenchyma and ovarian capsule regions are assigned unique fixed pixel values, and the background region is assigned a uniform background pixel value. The localization module locks the pixel range of the ovarian parenchyma region completely wrapped by the ovarian capsule based on the closed boundary contour of the ovarian capsule region within the segmentation mask. It performs candidate region extraction only within this pixel range to exclude background interference outside the ovarian capsule. Within the locked ovarian parenchyma region, it extracts four cyst morphological features: boundary closure, morphological regularity, internal gray-level uniformity, and gray-level difference with surrounding tissues. When all four features of a certain region meet the preset cyst feature constraints, the region is marked as a cyst candidate region. Among them, boundary closure is the proportion of continuous closed pixels of the region boundary contour to the total contour pixels, morphological regularity is the fit degree of the region contour to the contour of a standard circle or standard ellipse, internal gray uniformity is the coefficient of variation of the gray values ​​of all pixels inside the region, and gray difference with surrounding tissue is the absolute difference between the average gray value of pixels inside the region and the average gray value of pixels in the ovarian parenchyma region with a preset width. For all marked cyst candidate regions, a pixel-level region mask is generated for each region. At the same time, a minimum bounding rectangle that fits the contour of each candidate region is generated. The pixel coordinates and size parameters of the minimum bounding rectangle in the corresponding ultrasound frame image are recorded, and the candidate region is finally located.

[0012] Furthermore, during the operation phase of the determination module, for each candidate region, based on the inter-frame temporal features, the position and morphological changes of each candidate region in the continuous valid frame sequence are tracked, and temporal consistency verification is performed; The temporal consistency check includes checking whether the existence persistence, position offset, and morphological change amplitude of the candidate region in consecutive frames all meet the preset constraints. For candidate regions that pass the temporal consistency check, based on the location coordinates and region mask corresponding to the candidate region, the fusion features of the region of interest corresponding to the candidate region are extracted from the multi-dimensional fusion feature map corresponding to each consecutive valid frame tracked by the candidate region. Then, the temporal dimension feature aggregation is performed on the fusion features corresponding to each frame in combination with the inter-frame temporal correlation matrix to generate the global discriminant feature vector corresponding to the candidate region. The global discriminant feature vector is input into the preset cyst category discrimination branch, and the matching degree between the feature vector and the preset physiological cyst feature benchmark and pathological cyst feature benchmark are calculated respectively. The binary classification of cyst categories is performed to distinguish between physiological cysts and pathological cysts. Based on the normalization of the matching degree calculation results, the posterior probability values ​​of whether the candidate region belongs to physiological cysts or pathological cysts are generated. The cyst category corresponding to the maximum posterior probability is taken as the category determination result of the candidate region, and the maximum posterior probability value is taken as the confidence level of the corresponding category determination result.

[0013] Furthermore, the preset cyst category discrimination branch includes a feature standardization layer, a category matching degree calculation layer, and a posterior probability output layer; After receiving the input global discriminative feature vector, the discriminative branch first performs zero-mean, unit-variance standardization on the vector through a feature standardization layer to eliminate the interference of feature amplitude differences on the matching result; let the standardized global discriminative feature vector be... Where D is the fixed dimension of the feature vector, consistent with the channel dimension of the multi-dimensional fused feature map output by the fusion module; preset and Same-dimensional physiological cyst feature baseline vector Pathological cyst characteristic baseline vector Both types of baseline vectors are standardized fixed baseline vectors generated based on the statistical analysis of ultrasound features of large-sample cysts; let the temporal compliance coefficient output by the candidate region after temporal consistency verification be α, and let the value of α be (0,1]. The better the temporal consistency verification result, the closer the value of α is to 1. The matching degree between the candidate region and the feature benchmarks of the two types of cysts is calculated using the category matching degree calculation layer: ; Through the posterior probability output layer, probability normalization mapping is performed based on the two types of matching degrees to generate the posterior probability of the candidate region belonging to a physiological cyst and the posterior probability of it belonging to a pathological cyst, respectively: ; In the formula: These represent the matching degree between the candidate region and the physiological cyst feature benchmark and the pathological cyst feature benchmark, respectively. , To standardize the cosine similarity between the feature vector and the corresponding category baseline vector, p and m are the category identifiers; This is the KL divergence of the standardized feature vectors relative to the corresponding class baseline vectors. The smaller the KL divergence, the higher the overlap of the feature distributions. denoted as posterior probabilities of the candidate region belonging to physiological cysts and posterior probabilities of it belonging to pathological cysts, respectively; exp is the natural exponential function.

[0014] On the other hand, a method for identifying ovarian cysts based on ultrasound images includes: The system acquires continuous dynamic ultrasound scan sequences of the ovarian region at a preset frame sampling rate, assigns a unique incremental temporal sequence number to each frame, performs image grayscale normalization, removes invalid frames, and retains valid frames while maintaining the relative order of their original temporal sequence numbers. It receives the numbered and labeled continuous valid ultrasound sequences, constructs an inter-frame temporal correlation matrix, extracts grayscale gradient changes and structural motion displacement features within adjacent frames and preset cross-frame intervals, and integrates these to generate inter-frame temporal features. It extracts spatial texture features from single-frame valid images, generates corresponding spatial texture feature maps, aligns the dimensions of spatial and temporal feature maps, and generates a multi-dimensional fusion feature map through intermodulation of spatial and temporal features. Based on the multi-dimensional fusion feature map, it performs pixel-level semantic segmentation of the ovarian anatomical structure, generates a corresponding segmentation mask, locks the ovarian parenchymal region, extracts candidate regions that conform to cyst morphological characteristics, and completes localization. It tracks the position and morphological changes of candidate regions in continuous frames to complete temporal consistency verification. It extracts fusion features from regions that pass verification and performs temporal aggregation, inputs them into a discriminant branch to complete classification, and generates a judgment result and confidence level. Finally, it generates a standardized diagnostic report that binds the judgment result, temporal information, and user personal information.

[0015] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: This invention effectively overcomes the limitations of single static image recognition by fusing single-frame spatial texture information with inter-frame temporal correlation features. It can accurately segment ovarian anatomical structures and locate candidate cyst regions. Based on temporal consistency verification, it effectively avoids the misjudgment problems that are prone to occur in static recognition. It can clearly distinguish between physiological and pathological ovarian cysts, thereby improving the accuracy and stability of cyst recognition. While suppressing ultrasound speckle noise and filtering out interference from invalid tissues around the ovary, it can accurately capture the feature information of small cysts with weak boundaries and low contrast, reducing the missed detection of small cysts. The overall processing flow fits the actual clinical ultrasound scanning scenario, and the recognition results are objective and reliable. It can also automatically generate standardized diagnostic reports and complete the temporal and personal information association storage to simplify the clinical diagnostic operation process. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0017] Figure 1 This is a schematic diagram of an ovarian cyst identification system based on ultrasound images. Figure 2 This is a flowchart illustrating a method for identifying ovarian cysts based on ultrasound images. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] The present invention will be further described below with reference to embodiments.

[0020] Example 1: This embodiment provides an ovarian cyst identification system based on ultrasound images, such as... Figure 1 As shown, it includes: The acquisition module is used to acquire continuous dynamic ultrasound scan sequences of the user's ovarian region, and simultaneously perform sequence image normalization processing, invalid frame removal, and frame time sequence numbering. During the acquisition module operation phase, a continuous dynamic ultrasound scan sequence of the user's ovarian region is acquired at a preset frame sampling rate, and a unique incremental time sequence number is assigned to each frame image in the sequence simultaneously. When the acquisition module performs image normalization processing, it maps the grayscale value range of each single frame image to a preset unified range. Then, based on the grayscale distribution dispersion and the proportion of effective anatomical structures of the single frame image, it performs invalid frame removal operation, removing invalid frames whose grayscale distribution dispersion exceeds the preset range and whose proportion of effective anatomical structures is lower than the preset threshold. The remaining valid frames maintain the original relative order of their time sequence numbers. The modeling module is used to receive the numbered and calibrated continuous ultrasound sequences, construct the inter-frame temporal correlation matrix, and extract the motion and grayscale change features of adjacent frames and cross-frames of the sequence, which are recorded as inter-frame temporal features. The modeling module receives a series of valid ultrasound sequences after numbering and calibration. Let the valid frame sequence be... Where n is the total number of valid frames, For a valid frame image with time sequence number t, construct an inter-frame temporal correlation matrix M of dimension n×n, where any element in the matrix... The calculation formula is: ; In the formula: for and The normalized mutual information is used to characterize the consistency of gray-level distribution between two frames of images; for and The gradient field cosine similarity is used to characterize the consistency between the anatomical structure edges and motion trends of two frames of images; The time interval between two frames of images; This formula multiplies the normalized mutual information with the gradient field cosine similarity to comprehensively reflect the consistency of gray-level distribution and the consistency of anatomical structure edges and motion trends between two ultrasound images. Then, it uses the time sequence number interval as a weight constraint to accurately construct the temporal correlation between ultrasound sequence frames. This not only fits the temporal logic of dynamic scanning, but also captures tissue motion and gray-level changes more comprehensively than the single feature correlation method, making the temporal feature extraction more suitable for the actual scenario of ovarian ultrasound scanning. The modeling module extracts the inter-frame temporal correlation matrix M, extracts the inter-frame grayscale change gradient and structural motion displacement features within adjacent frames and preset cross-frame intervals, and integrates them to generate inter-frame temporal features. The fusion module is used to fuse the spatial texture features and inter-frame temporal features of a single frame image of an ultrasound sequence to generate a multi-dimensional fused feature map corresponding to the sequence. During the fusion module's operation phase, spatial texture features of each valid single-frame image within the sequence are extracted to generate a corresponding single-frame spatial texture feature map. Then, feature dimension alignment processing is performed between the single-frame spatial texture feature map and the inter-frame temporal feature map to generate a dimension-matched spatial feature map. With time series feature map Finally, based on the spatial-temporal feature intermodulation mechanism, a multi-dimensional fused feature map is generated. ; ; In the formula: ⊙ represents the Hadamard product operation; Sigmoid activation function is used to generate feature modulation weights; GAP is global average pooling operation, used to extract global context information of feature maps; LN is layer normalization operation, used to unify the numerical distribution of fused features. This formula first extracts global contextual information of spatial and temporal features through global average pooling, generates feature modulation weights through Sigmoid activation, then achieves mutual modulation of the two types of features through Hadamard product, and finally uses layer normalization to regularize the feature value distribution, so that spatial texture features and inter-frame temporal features can mutually reinforce and complement each other, effectively distinguishing cyst features from normal tissue features, while suppressing ultrasound speckle noise, thereby solving the problems of incomplete single-dimensional feature representation and susceptibility to noise interference. The Sigmoid activation function is: ; In the formula: x is the input value of the activation function; k is a preset gain coefficient used to adjust the gradient change rate of the activation function; This is a preset bias coefficient used to adjust the response threshold of the activation function; Where k∈[1,3], the value is larger when it is necessary to improve the weight differentiation between cyst features and normal ovarian tissue features and enhance the feature modulation effect of high-confidence cyst regions, and smaller when it is necessary to suppress feature fluctuations caused by ultrasound speckle noise and smooth the modulation weight difference between consecutive frames. Since the Sigmoid curve is too flat when the gain coefficient k is less than 1, the feature modulation differentiation is insufficient and effective feature weighting cannot be achieved; when it is greater than 3, the curve is too steep and close to the step function, which will amplify the small fluctuations of ultrasound speckle noise and cause mismodulation, so it is set in this range. The value ∈ [-1,1] is larger when it is necessary to increase the feature activation threshold and filter out feature interference from invalid background tissues such as ovarian periphery fat and intestines, and smaller when it is necessary to decrease the feature activation threshold and amplify the weak feature modulation effect of weakly bordered small cysts and low-contrast cyst regions within the ovary. When the value exceeds the range of [-1,1], it will either cause the background noise to be fully activated and lose its feature filtering effect, or it will filter out the effective features of small cysts with weak boundaries. Therefore, it is set within this range. The process of generating spatial texture feature maps is as follows: After obtaining a single frame of valid image that has undergone normalization, speckle noise suppression based on neighborhood grayscale distribution consistency is first performed to preserve the anatomical structure edges and texture details in the image. The image with noise suppression is subjected to a preset number of non-downsampling multi-scale decompositions to generate multi-channel sub-image groups with different spatial resolutions. Each channel sub-image corresponds to a structural texture component in a different direction. For each channel sub-image, extract the gray-level co-occurrence statistical features, gradient magnitude distribution features, and local structure regularity features within a preset neighborhood range pixel by pixel to generate a single-channel texture feature map for the corresponding channel. Perform feature aggregation in the channel dimension and size restoration in the spatial dimension on all single-channel texture feature maps, and finally generate a spatial texture feature map that matches the spatial size of the original single-frame effective image and has a preset fixed number of channels. Each pixel position in the spatial texture feature map corresponds one-to-one with the pixel position of the corresponding single frame of valid image. The localization module is used to perform regional segmentation of ovarian anatomy based on multi-dimensional fused feature maps, and to locate and extract candidate regions within the sequence that conform to the morphological features of cysts. During the localization module operation phase, pixel-level semantic segmentation of the ovarian parenchyma and ovarian capsule regions is performed based on multi-dimensional fused feature maps. Each pixel in the image is assigned a mutually exclusive classification label of ovarian parenchyma, ovarian capsule, and background. Based on the classification labels, an ovarian anatomical structure segmentation mask that perfectly matches the size of the original ultrasound frame image is generated. Within the segmentation mask, the ovarian parenchyma and ovarian capsule regions are assigned unique fixed pixel values, while the background region is assigned a uniform background pixel value. The localization module locks the pixel range of the ovarian parenchyma region completely wrapped by the ovarian capsule based on the closed boundary contour of the ovarian capsule region within the segmentation mask. It performs candidate region extraction only within this pixel range to exclude background interference outside the ovarian capsule. Within the locked ovarian parenchyma region, it extracts four cyst morphological features: boundary closure, morphological regularity, internal gray-level uniformity, and gray-level difference with surrounding tissues. When all four features of a certain region meet the preset cyst feature constraints, the region is marked as a cyst candidate region. Among them, boundary closure is the proportion of continuous closed pixels of the region boundary contour to the total contour pixels, morphological regularity is the fit degree of the region contour to the contour of a standard circle or standard ellipse, internal gray uniformity is the coefficient of variation of the gray values ​​of all pixels inside the region, and gray difference with surrounding tissue is the absolute difference between the average gray value of pixels inside the region and the average gray value of pixels in the ovarian parenchyma region with a preset width. For all marked cyst candidate regions, a pixel-level region mask is generated for each region. At the same time, a minimum bounding rectangle that fits the contour of each candidate region is generated. The pixel coordinates and size parameters of the minimum bounding rectangle in the corresponding ultrasound frame image are recorded to finally complete the localization of the candidate region. The determination module is used to perform temporal consistency verification and category discrimination of each candidate region in consecutive frames based on inter-frame temporal characteristics. During the operation of the judgment module, for each candidate region, the position and morphological changes of each candidate region in the continuous valid frame sequence are tracked based on the inter-frame temporal features, and temporal consistency verification is performed. Temporal consistency verification includes verifying whether the existence persistence, position offset, and morphological change amplitude of the candidate region in consecutive frames meet the preset constraints. For candidate regions that pass the temporal consistency check, based on the location coordinates and region mask corresponding to the candidate region, the fusion features of the region of interest corresponding to the candidate region are cropped and extracted from the multi-dimensional fusion feature map corresponding to each consecutive valid frame tracked by the candidate region. Then, the temporal dimension feature aggregation is performed on the fusion features corresponding to each frame in combination with the inter-frame temporal correlation matrix to generate the global discriminant feature vector corresponding to the candidate region. The global discriminant feature vector is input into the preset cyst category discrimination branch, and the matching degree between the feature vector and the preset physiological cyst feature benchmark and pathological cyst feature benchmark are calculated respectively. The binary classification of cyst categories is performed to distinguish between physiological cysts and pathological cysts. Based on the normalization of the matching degree calculation results, the posterior probability value of the candidate region belonging to physiological cysts or pathological cysts is generated. The cyst category corresponding to the maximum posterior probability is taken as the category determination result of the candidate region, and the maximum posterior probability value is taken as the confidence level of the corresponding category determination result. The pre-defined cyst category discrimination branch includes a feature standardization layer, a category matching degree calculation layer, and a posterior probability output layer; After receiving the input global discriminative feature vector, the discriminative branch first performs zero-mean, unit-variance standardization on the vector through a feature standardization layer to eliminate the interference of feature amplitude differences on the matching results; let the standardized global discriminative feature vector be... Where D is the fixed dimension of the feature vector, and R is the internationally accepted standard mathematical symbol for the set of real numbers, consistent with the channel dimension of the multi-dimensional fused feature map output by the fusion module; preset and Same-dimensional physiological cyst feature baseline vector Pathological cyst characteristic baseline vector Both types of baseline vectors are standardized fixed baseline vectors generated based on the statistical analysis of ultrasound features of large-sample cysts; let the temporal compliance coefficient output by the candidate region after temporal consistency verification be α, and let the value of α be (0,1]. The better the temporal consistency verification result, the closer the value of α is to 1. The matching degree between the candidate region and the feature benchmarks of the two types of cysts is calculated using the category matching degree calculation layer: ; Through the posterior probability output layer, probability normalization mapping is performed based on the two types of matching degrees to generate the posterior probability of the candidate region belonging to a physiological cyst and the posterior probability of it belonging to a pathological cyst, respectively: ; In the formula: These represent the matching degree between the candidate region and the physiological cyst feature benchmark and the pathological cyst feature benchmark, respectively. , To standardize the cosine similarity between the feature vector and the corresponding category baseline vector, p and m are the category identifiers; This is the KL divergence of the standardized feature vectors relative to the corresponding class baseline vectors. The smaller the KL divergence, the higher the overlap of the feature distributions. , , respectively, represent the posterior probabilities of the candidate region belonging to physiological cysts and the posterior probabilities of it belonging to pathological cysts; exp is the natural exponential function; The above formula combines cosine similarity and KL divergence to calculate the matching degree between the candidate region and the baseline features of the two types of cysts. It incorporates a temporal compliance coefficient to enhance the temporal verification effect, and then completes probability normalization through a natural exponential function. This formula can accurately measure the similarity and overlap of feature distributions and incorporate temporal consistency into the discrimination logic, which greatly improves the reliability of distinguishing between physiological and pathological cysts and makes the confidence of the judgment results more in line with actual diagnostic needs. The storage module is used to receive the judgment results from the judgment module, generate corresponding diagnostic result reports, and store the diagnostic result reports based on the association between time series and user personal information. The data acquisition module is interconnected with the modeling module via a wireless network. The modeling module is interconnected with the fusion module via a wireless network. The fusion module is interconnected with the positioning module via a wireless network. The positioning module is interconnected with the decision module via a wireless network. The decision module is interconnected with the storage module via a wireless network.

[0021] In this embodiment, the acquisition module runs a continuous dynamic ultrasound scan sequence of the user's ovarian region, simultaneously performing sequence image normalization, invalid frame removal, and frame temporal numbering. The modeling module then runs the received and numbered continuous ultrasound sequence, constructs an inter-frame temporal correlation matrix, and extracts motion and grayscale change features of adjacent frames and cross-frames, which are recorded as inter-frame temporal features. The fusion module simultaneously fuses the spatial texture features of single-frame images of the ultrasound sequence with the inter-frame temporal features to generate a multi-dimensional fusion feature map corresponding to the sequence. Then, the localization module performs regional segmentation of the ovarian anatomical structure based on the multi-dimensional fusion feature map, locates and extracts candidate regions within the sequence that conform to the morphological characteristics of cysts, and the judgment module performs temporal consistency verification and category discrimination of each candidate region in continuous frames based on the inter-frame temporal features. Finally, the storage module receives the judgment results from the judgment module, generates a corresponding diagnostic result report, and stores the diagnostic result report based on the temporal association with the user's personal information.

[0022] When the system described in the above embodiments is applied to the ultrasound screening and diagnosis of ovarian cysts, it can perform standardized processing on continuous dynamic ultrasound images, effectively remove invalid images and retain temporal information, suppress ultrasound speckle noise by fusing spatial texture and inter-frame temporal features, accurately segment ovarian anatomical structures and lock cyst candidate regions, reduce misjudgments based on temporal consistency verification, accurately distinguish between physiological and pathological cysts, and automatically generate standardized diagnostic reports and complete information association storage, thereby improving the efficiency and accuracy of clinical diagnosis and reducing problems such as missed detection of small cysts and interference from background tissue.

[0023] It should be noted that: When implementing the above system, the preset frame sampling rate can be selected as 15-30 frames / second, which is the standard for dynamic scanning of ovarian ultrasound in clinical practice. The image normalization process maps the gray value range to a standard 8-bit gray value range of 0-255. The preset range for gray value distribution dispersion is 50-300 for image gray value variance. The preset threshold for the proportion of effective anatomical structures is that the proportion of pixels in the ovarian anatomical region in a single frame image is not less than 30%. This completes the invalid frame removal and time sequence numbering, ensuring that the effective frame sequence is complete and accurate in time.

[0024] In the modeling module, the preset cross-frame interval is 1 to 5 frames, covering the range of routine cyst movement and displacement during ovarian ultrasound scanning. When integrating inter-frame temporal features, the gray-level change gradient and structural motion displacement features within adjacent frames and the preset cross-frame interval are concatenated into a fixed-dimensional feature vector according to temporal weights. The gray-level change gradient is represented by the mean and variance of the gray-level difference between frames, and the structural motion displacement features are obtained by calculating the inter-frame pixel offset using the optical flow method. The normalized mutual information is calculated based on the joint entropy of the gray-level histogram, and the gradient field cosine similarity is achieved by the dot product operation after the gradient vector is normalized. This can stably construct the inter-frame temporal correlation matrix and extract inter-frame temporal features.

[0025] In the fusion module, speckle noise suppression is achieved by combining 3×3 neighborhood median filtering with adaptive mean filtering. The non-downsampling multi-scale decomposition has a preset number of 3 layers, and the single-channel texture feature extraction has a preset neighborhood of 5×5 pixels. The spatial texture feature map has a fixed number of channels set to 16. Feature dimension alignment is achieved by unifying the spatial size through bilinear interpolation and unifying the channel dimension through 1×1 convolution. The sigmoid activation function has a default gain coefficient k of 2 and a default bias coefficient μ of 0, which balances the discriminative power of cyst feature weights with the effect of ultrasound speckle noise suppression, and can stably generate multi-dimensional fused feature maps.

[0026] The localization module uses a lightweight U-Net model to perform pixel-level semantic segmentation of the ovarian parenchyma, ovarian capsule, and background. The model training samples are clinically labeled ovarian ultrasound image datasets. The morphological features of the cysts are subject to the following pre-set constraints: boundary closure degree not less than 90%, morphological regularity not less than 0.85, internal gray-level uniformity coefficient of variation not higher than 0.15, and gray-level difference with surrounding tissues not less than 20. When locating candidate regions, the pixel coordinates, width and height dimensions, and corresponding time frame number of the smallest bounding rectangle are recorded to accurately locate the candidate cyst regions within the ovarian parenchyma.

[0027] The temporal consistency verification of the judgment module has the following preset constraints: the candidate region must exist continuously for no less than 5 consecutive frames, the position offset must not exceed 10% of the image pixel size, and the morphological change must not exceed 15%. In the cyst category discrimination branch, the feature vector has a fixed dimension D of 16. The feature baseline vectors of physiological cysts and pathological cysts are generated by statistically averaging the features of 1000 clinically diagnosed samples. The temporal compliance coefficient α is linearly mapped to the (0,1] interval according to the temporal consistency verification score, which can accurately complete the binary classification of cyst categories and output the confidence level.

[0028] The storage module generates a standard PDF format diagnostic report, which includes user information, ultrasound scan sequence number, cyst location information, category determination result and confidence level. The diagnostic data is stored using a hash index with the user's unique identifier as the primary key and the time sequence number as the secondary key. The diagnostic results are securely stored and quickly retrieved using a relational database.

[0029] Example 2: At the implementation level, based on Example 1, this example refers to... Figure 2 A further detailed description of the ovarian cyst identification system based on ultrasound images in Example 1 is provided below: A method for identifying ovarian cysts based on ultrasound images, comprising: A continuous dynamic ultrasound scan sequence of the ovarian region was acquired at a preset frame sampling rate. Each frame was assigned a unique incremental time sequence number. Image grayscale normalization was performed, invalid frames were removed, and the remaining valid frames maintained the relative order of their original time sequence numbers. Receive continuous valid ultrasound sequences after numbering and calibration, construct inter-frame temporal correlation matrix, extract gray-level change gradient and structural motion displacement features in adjacent frames and preset cross-frame intervals, and integrate them to generate inter-frame temporal features. Extract spatial texture features from a single effective image, generate a corresponding spatial texture feature map, complete the dimensional alignment of spatial and temporal feature maps, and generate a multi-dimensional fusion feature map through intermodulation of spatial and temporal features. Based on multi-dimensional fusion feature maps, pixel-level semantic segmentation of ovarian anatomical structure is completed, corresponding segmentation masks are generated, the range of ovarian parenchymal region is locked, candidate regions that conform to cyst morphological features are extracted and localization is completed. Track the position and shape changes of candidate regions in consecutive frames to complete temporal consistency verification; extract fusion features from regions that pass the verification and perform temporal aggregation, input them into the discriminant branch to complete classification, and generate judgment results and confidence scores; Generate standardized diagnostic reports that link judgment results, time-series information, and user personal information.

[0030] In summary, the system and method in the above embodiments effectively compensate for the limitations of single static image recognition by fusing single-frame spatial texture information with inter-frame temporal correlation features. It can accurately segment ovarian anatomical structures and locate candidate cyst regions. Based on temporal consistency verification, it effectively avoids the misjudgment problem that is prone to occur in static recognition. It can clearly distinguish between physiological and pathological ovarian cysts, thereby improving the accuracy and stability of cyst recognition. While suppressing ultrasound speckle noise and filtering out interference from invalid tissues around the ovary, it can accurately capture the feature information of small cysts with weak boundaries and low contrast, reducing the missed detection of small cysts. The overall processing flow fits the actual clinical ultrasound scanning scenario, and the recognition results are objective and reliable. It can also automatically generate standardized diagnostic reports and complete the temporal and personal information association storage to simplify the clinical diagnostic operation process.

[0031] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An ovarian cyst identification system based on ultrasound images, characterized in that, include: The acquisition module is used to acquire continuous dynamic ultrasound scan sequences of the user's ovarian region, and simultaneously perform sequence image normalization processing, invalid frame removal, and frame time sequence numbering. The modeling module is used to receive the numbered and calibrated continuous ultrasound sequences, construct the inter-frame temporal correlation matrix, and extract the motion and grayscale change features of adjacent frames and cross-frames of the sequence, which are recorded as inter-frame temporal features. The fusion module is used to fuse the spatial texture features and inter-frame temporal features of a single frame image of an ultrasound sequence to generate a multi-dimensional fused feature map corresponding to the sequence. The localization module is used to perform regional segmentation of ovarian anatomy based on multi-dimensional fused feature maps, and to locate and extract candidate regions within the sequence that conform to the morphological features of cysts. The determination module is used to perform temporal consistency verification and category discrimination of each candidate region in consecutive frames based on inter-frame temporal characteristics. The storage module is used to receive the judgment results from the judgment module, generate corresponding diagnostic result reports, and store the diagnostic result reports based on the association between time sequence and user personal information.

2. The ovarian cyst identification system based on ultrasound images according to claim 1, characterized in that, During the operation phase of the acquisition module, a continuous dynamic ultrasound scan sequence of the user's ovarian region is acquired at a preset frame sampling rate, and a unique incremental time sequence number is assigned to each frame image in the sequence. During image normalization processing, the acquisition module maps the grayscale value range of each single frame image to a preset unified range. Then, based on the grayscale distribution dispersion and the proportion of effective anatomical structures of the single frame image, it performs an invalid frame removal operation, removing invalid frames whose grayscale distribution dispersion exceeds the preset range or whose proportion of effective anatomical structures is lower than the preset threshold. The remaining valid frames maintain the original relative order of their time sequence numbers.

3. The ovarian cyst identification system based on ultrasound images according to claim 1, characterized in that, The modeling module receives a series of valid ultrasound sequences after numbering and calibration, assuming the valid frame sequence is... Where n is the total number of valid frames, For a valid frame image with time sequence number t, construct an inter-frame temporal correlation matrix M of dimension n×n, where any element in the matrix... The calculation formula is: ； In the formula: for and Normalized mutual information; for and The gradient field cosine similarity; The time interval between two frames of images; The modeling module extracts the inter-frame grayscale change gradient and structural motion displacement features between adjacent frames and within a preset cross-frame interval based on the inter-frame temporal correlation matrix M, and integrates them to generate inter-frame temporal features.

4. The ovarian cyst identification system based on ultrasound images according to claim 1, characterized in that, During the fusion module's operation phase, spatial texture features of each valid single-frame image within the sequence are extracted to generate a corresponding single-frame spatial texture feature map. Then, feature dimension alignment processing is performed on the spatial texture feature map corresponding to the single-frame image and the inter-frame temporal feature map to generate a dimension-matched spatial feature map. With time series feature map Finally, based on the spatial-temporal feature intermodulation mechanism, a multi-dimensional fused feature map is generated. ; ； In the formula: ⊙ represents the Hadamard product operation; is the Sigmoid activation function used to generate feature modulation weights; GAP is the global average pooling operation; LN is the layer normalization operation.

5. The ovarian cyst identification system based on ultrasound images according to claim 4, characterized in that, The process of generating the spatial texture feature map is as follows: After obtaining a single frame of valid image that has undergone normalization, speckle noise suppression based on neighborhood grayscale distribution consistency is first performed to preserve the anatomical structure edges and texture details in the image. The image with noise suppression is subjected to a preset number of non-downsampling multi-scale decompositions to generate multi-channel sub-image groups with different spatial resolutions. Each channel sub-image corresponds to a structural texture component in a different direction. For each channel sub-image, extract the gray-level co-occurrence statistical features, gradient magnitude distribution features, and local structure regularity features within a preset neighborhood range pixel by pixel to generate a single-channel texture feature map for the corresponding channel. Perform channel-dimensional feature aggregation and spatial-dimensional size restoration on all single-channel texture feature maps to finally generate a spatial texture feature map that matches the spatial size of the original single-frame effective image and has a preset fixed number of channels. Each pixel position in the spatial texture feature map corresponds one-to-one with the pixel position of the corresponding single frame of valid image.

6. The ovarian cyst identification system based on ultrasound images according to claim 1, characterized in that, During the operation phase of the localization module, pixel-level semantic segmentation of the ovarian parenchyma and ovarian capsule regions is performed based on the multi-dimensional fused feature map. Each pixel in the image is assigned a mutually exclusive classification label of ovarian parenchyma, ovarian capsule, and background. Based on the classification label, an ovarian anatomical structure segmentation mask that perfectly matches the size of the original ultrasound frame image is generated. In the segmentation mask, the ovarian parenchyma and ovarian capsule regions are assigned unique fixed pixel values, and the background region is assigned a uniform background pixel value. The localization module locks the pixel range of the ovarian parenchyma region completely wrapped by the ovarian capsule based on the closed boundary contour of the ovarian capsule region within the segmentation mask. It performs candidate region extraction only within this pixel range to exclude background interference outside the ovarian capsule. Within the locked ovarian parenchyma region, it extracts four cyst morphological features: boundary closure, morphological regularity, internal gray-level uniformity, and gray-level difference with surrounding tissues. When all four features of a certain region meet the preset cyst feature constraints, the region is marked as a cyst candidate region. Among them, boundary closure is the proportion of continuous closed pixels of the region boundary contour to the total contour pixels, morphological regularity is the fit degree of the region contour to the contour of a standard circle or standard ellipse, internal gray uniformity is the coefficient of variation of the gray values ​​of all pixels inside the region, and gray difference with surrounding tissue is the absolute difference between the average gray value of pixels inside the region and the average gray value of pixels in the ovarian parenchyma region with a preset width. For all marked cyst candidate regions, a pixel-level region mask is generated for each region. At the same time, a minimum bounding rectangle that fits the contour of each candidate region is generated. The pixel coordinates and size parameters of the minimum bounding rectangle in the corresponding ultrasound frame image are recorded, and the candidate region is finally located.

7. The ovarian cyst identification system based on ultrasound images according to claim 1, characterized in that, During the operation phase of the determination module, for each candidate region, the position and morphological changes of each candidate region in the continuous effective frame sequence are tracked based on the inter-frame temporal features, and temporal consistency verification is performed. The temporal consistency verification includes verifying whether the existence persistence, position offset, and morphological change amplitude of the candidate region in consecutive frames all meet the preset constraints. For candidate regions that pass the temporal consistency check, based on the location coordinates and region mask corresponding to the candidate region, the fusion features of the region of interest corresponding to the candidate region are extracted from the multi-dimensional fusion feature map corresponding to each consecutive valid frame tracked by the candidate region. Then, the temporal dimension feature aggregation is performed on the fusion features corresponding to each frame in combination with the inter-frame temporal correlation matrix to generate the global discriminant feature vector corresponding to the candidate region. The global discriminant feature vector is input into the preset cyst category discrimination branch, and the matching degree between the feature vector and the preset physiological cyst feature benchmark and pathological cyst feature benchmark are calculated respectively. The binary classification of cyst categories is performed to distinguish between physiological cysts and pathological cysts. Based on the normalization of the matching degree calculation results, the posterior probability values ​​of whether the candidate region belongs to physiological cysts or pathological cysts are generated. The cyst category corresponding to the maximum posterior probability is taken as the category determination result of the candidate region, and the maximum posterior probability value is taken as the confidence level of the corresponding category determination result.

8. The ovarian cyst identification system based on ultrasound images according to claim 7, characterized in that, The preset cyst category discrimination branch includes a feature standardization layer, a category matching degree calculation layer, and a posterior probability output layer; After receiving the input global discriminative feature vector, the discriminative branch first performs zero-mean, unit-variance standardization on the vector through a feature standardization layer to eliminate the interference of feature amplitude differences on the matching result; let the standardized global discriminative feature vector be... Where D is the fixed dimension of the feature vector, consistent with the channel dimension of the multi-dimensional fused feature map output by the fusion module; preset and Same-dimensional physiological cyst feature baseline vector Pathological cyst characteristic baseline vector Both types of baseline vectors are standardized fixed baseline vectors generated based on the statistical analysis of ultrasound features of large-sample cysts; let the temporal compliance coefficient output by the candidate region after temporal consistency verification be α, and let the value of α be (0,1]. The better the temporal consistency verification result, the closer the value of α is to 1. The matching degree between the candidate region and the feature benchmarks of the two types of cysts is calculated using the category matching degree calculation layer: ； Through the posterior probability output layer, probability normalization mapping is performed based on the two types of matching degrees to generate the posterior probability of the candidate region belonging to a physiological cyst and the posterior probability of it belonging to a pathological cyst, respectively: ； In the formula: These represent the matching degree between the candidate region and the physiological cyst feature benchmark and the pathological cyst feature benchmark, respectively. , To standardize the cosine similarity between the feature vector and the corresponding category baseline vector, p and m are the category identifiers; This is the KL divergence of the standardized feature vectors relative to the corresponding class baseline vectors. The smaller the KL divergence, the higher the overlap of the feature distributions. denoted as posterior probabilities of the candidate region belonging to physiological cysts and posterior probabilities of it belonging to pathological cysts, respectively; exp is the natural exponential function.

9. The ovarian cyst identification system based on ultrasound images according to claim 1, characterized in that, The acquisition module is interconnected with the modeling module via a wireless network. The modeling module is interconnected with the fusion module via a wireless network. The fusion module is interconnected with the positioning module via a wireless network. The positioning module is interconnected with the decision module via a wireless network. The decision module is interconnected with the storage module via a wireless network.

10. A method for identifying ovarian cysts based on ultrasound images, wherein the method is an embodiment of the ovarian cyst identification system based on ultrasound images as described in any one of claims 1-9, characterized in that... include: A continuous dynamic ultrasound scan sequence of the ovarian region was acquired at a preset frame sampling rate. Each frame was assigned a unique incremental time sequence number. Image grayscale normalization was performed, invalid frames were removed, and the remaining valid frames maintained the relative order of their original time sequence numbers. Receive continuous valid ultrasound sequences after numbering and calibration, construct inter-frame temporal correlation matrix, extract gray-level change gradient and structural motion displacement features in adjacent frames and preset cross-frame intervals, and integrate to generate inter-frame temporal features. Extract spatial texture features from a single effective image, generate a corresponding spatial texture feature map, complete the dimensional alignment of spatial and temporal feature maps, and generate a multi-dimensional fusion feature map through intermodulation of spatial and temporal features. Based on multi-dimensional fusion feature maps, pixel-level semantic segmentation of ovarian anatomical structure is completed, corresponding segmentation masks are generated, the range of ovarian parenchymal region is locked, candidate regions that conform to cyst morphological features are extracted and localization is completed. Track the position and shape changes of candidate regions in consecutive frames to complete temporal consistency verification; extract fusion features from regions that pass the verification and perform temporal aggregation, input them into the discriminant branch to complete classification, and generate judgment results and confidence scores; Generate standardized diagnostic reports that link judgment results, time-series information, and user personal information.

Images