Image segmentation method, gestational week prediction method and diagnostic report determination method and device

Abstract

The invention provides an image segmentation method, a gestational week prediction method and a diagnosis report determination method and device, and belongs to the field of medical artificial intelligence. The image segmentation method comprises the steps that a U-shaped image segmentation network is constructed, and the U-shaped image segmentation network comprises an encoder and a decoder; wherein the encoder and the decoder are connected through a jump connection layer; a self-adaptive feature gating module of each jump connection layer embedded layer of the U-shaped image segmentation network is used for strengthening the expression ability of a key anatomical structure and filtering irrelevant information; a bottleneck layer of the U-shaped image segmentation network is embedded into a bottleneck domain adaptive feature calibration module which is used for guiding deep semantic positioning by utilizing shallow space details and improving multi-section organ segmentation robustness; training the U-shaped image segmentation network by adopting the training set to obtain an image segmentation model; and inputting a to-be-segmented image into the image segmentation model to obtain an image segmentation result. According to the method, standardized, multi-section and multi-structure collaborative segmentation of the image can be realized.

Description

Technical Field

[0001] This application relates to the field of medical artificial intelligence, and in particular to image segmentation methods, gestational age prediction methods and diagnostic report determination methods, image segmentation devices, gestational age prediction devices, diagnostic report determination devices, electronic devices, computer-readable storage media, and computer program products. Background Technology

[0002] As a crucial aspect of perinatal medicine, fetal growth and development monitoring directly impacts maternal and infant health outcomes. Ultrasound technology, due to its non-invasive and repeatable advantages, has become the preferred method for assessing fetal anatomy and screening for congenital defects. Its measurement data plays a key role in gestational age estimation and developmental assessment.

[0003] Currently, the main methods for segmenting ultrasound images include model prediction. However, model prediction-based methods are limited by simple network structures or static tasks, making it difficult to meet the needs of multi-slice and multi-structure collaborative segmentation. Summary of the Invention

[0004] This application aims to at least partially solve one of the aforementioned technical problems or at least provide a useful commercial solution. In view of this, this application provides an image segmentation method, a gestational age prediction method, and a diagnostic report determination method and apparatus, capable of achieving standardized, multi-section, and multi-structure collaborative segmentation of images.

[0005] In a first aspect, this application proposes an image segmentation method, comprising: constructing a U-shaped image segmentation network, the U-shaped image segmentation network including an encoder and a decoder; wherein the encoder and the decoder are connected through skip connection layers; each skip connection layer of the U-shaped image segmentation network embeds a layer adaptive feature gating module, the layer adaptive feature gating module including: a channel attention extraction submodule and a spatial attention extraction submodule, used to enhance the expressive power of key anatomical structures and filter irrelevant information; the bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module, the bottleneck domain adaptive feature calibration module including: a feature fusion unit, a multi-layer spatial attention extraction submodule, a multi-layer channel attention extraction submodule, and a feature residual connection unit, used to utilize shallow spatial details to guide deep semantic localization and improve the robustness of multi-faceted organ segmentation; training the U-shaped image segmentation network using a training set to obtain an image segmentation model; and inputting the image to be segmented into the image segmentation model to obtain an image segmentation result.

[0006] In some implementations, the layer-adaptive feature gating module integrates channel attention and spatial attention mechanisms, which are used to enhance the expressive power of key anatomical structures and filter irrelevant information, respectively, including: for input features Channel attention extraction is performed, and the input feature map is subjected to max pooling and average pooling operations to obtain the feature map. and Will and The input features are fed into a shared multilayer perceptron, and their output features are fused through a concatenation operation. Then, they are mapped using a sigmoid activation function to generate channel attention feature weights. These channel attention feature weights are then compared with the input features. Multiply to obtain a weighted output feature map; apply spatial attention to extract spatial attention weights. This includes applying average pooling and max pooling operations along the channel dimension of the feature map to obtain the features. and Connecting them together yields a feature descriptor. Then, a convolution operation is applied to this feature to generate a spatial attention feature; the channel attention feature and the spatial attention feature are multiplied to obtain the output feature.

[0007] In some possible implementations, this further includes: introducing a hierarchical weight parameter α. l ∈[0,1], used to quantify the contribution of the target layer to the decoder; weighted features of channel attention and spatial attention. After hierarchical weight α l In conjunction with the gating unit, feature fusion is performed to obtain the final features. Figure X l .

[0008] In some implementations, the bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module to guide deep semantic localization using shallow spatial details, thereby improving the robustness of multi-faceted organ segmentation. This includes: constructing shallow composite features by upsampling the shallow features processed by the layer adaptive feature gating module after the predetermined coding layer is processed by transposed convolution to the bottleneck layer size, performing 1×1 convolution dimensionality reduction, and then concatenating them to obtain the composite feature F. G Composite feature F G and the deep features F of the bottleneck layer A Spatial attention weights are obtained by passing through the spatial attention module, multiplying the obtained weights, processing them with the Softmax activation function, and then multiplying them with F. A Multiplication yields feature weights Will and F A Perform residual connection to obtain spatial feature F S ; F S F GThe channel attention weights are obtained by inputting them into the channel attention module. These weights are then multiplied and processed using the Softmax activation function, and finally multiplied by F. S Multiplication yields feature weights Will and F S Perform residual connections to obtain the output feature F. C .

[0009] In a second aspect of this application, a method for predicting gestational age is proposed. The method includes: segmenting an ultrasound image of the gestational age to be predicted using the method described in the first aspect to obtain an image segmentation result; extracting measurement values ​​of fetal anatomical structures based on the image segmentation result; and inputting the image segmentation result and the measurement values ​​of the fetal anatomical structures into a prediction model to obtain a fetal gestational age prediction result.

[0010] In some implementations, the prediction model is trained as follows: A training dataset is acquired, comprising a first training set and a second training set, wherein the first training set includes measurements of fetal anatomical structures and the second training set includes image segmentation results; the gestational age regression model is trained using the training dataset to obtain the prediction model; wherein constructing the gestational age regression model includes: inputting the first training set into a multilayer perceptron to obtain parameter feature vectors, wherein the multilayer perceptron contains three fully connected layers and a ReLU activation function; inputting the second training set into a pre-trained ResNet-18 network to obtain feature vectors of multiple predetermined dimensions; concatenating the multiple feature vectors of predetermined dimensions and inputting them into an attention module to obtain an attention feature map; concatenating the parameter feature vectors with the attention feature map and inputting them into the multilayer perceptron, activating the feature map using the Sigmoid function to obtain the predicted gestational age value.

[0011] In a third aspect of this application, a method for determining a diagnostic report is proposed. This method includes: determining a gestational age prediction result for the target subject based on gestational ultrasound images of the target subject using the method described in the second aspect. The gestational age prediction result includes: a predicted gestational age value and measured values ​​of fetal anatomical structures; constructing a fetal growth standard and medical diagnostic knowledge base; matching the gestational age prediction result with the fetal growth standard and medical diagnostic knowledge base, retrieving a reference range for the gestational age in the knowledge base, and determining each measurement item; constructing a three-layer evidence chain of "structure-measurement-clinical phenotype," and outputting a structured diagnostic report. Specifically, constructing the structured diagnostic evidence chain and outputting the structured diagnostic report includes: displaying a segmented mask image of key fetal anatomical regions; listing the actual measured values, reference ranges, and percentage deviations for each structure; and outputting possible diagnostic suggestions, disease categories, and their clinical interpretations according to knowledge base rules.

[0012] In a fourth aspect, this application proposes an image segmentation apparatus, comprising: a network construction module for constructing a U-shaped image segmentation network, the U-shaped image segmentation network including an encoder and a decoder; wherein the encoder and the decoder are connected through skip connection layers; each skip connection layer of the U-shaped image segmentation network embeds a layer adaptive feature gating module, the layer adaptive feature gating module including a channel attention extraction submodule and a spatial attention extraction submodule, for enhancing the expressive power of key anatomical structures and filtering irrelevant information; the bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module, the bottleneck domain adaptive feature calibration module including a feature fusion unit, a multi-layer spatial attention extraction submodule, a multi-layer channel attention extraction submodule, and a feature residual connection unit, for using shallow spatial details to guide deep semantic localization and improve the robustness of multi-faceted organ segmentation; a training module for training the U-shaped image segmentation network using a training set to obtain an image segmentation model; and a first prediction module for inputting the image to be segmented into the image segmentation model to obtain an image segmentation result.

[0013] In a fifth aspect of this application, a gestational age prediction device is proposed, comprising: an image segmentation module for segmenting an ultrasound image of the gestational age to be determined using the method described in the first aspect, to obtain an image segmentation result; a parameter determination module for extracting measurement values ​​of fetal anatomical structures based on the image segmentation result; and a second prediction module for inputting the image segmentation result and the measurement values ​​of the fetal anatomical structures into a prediction model to obtain a fetal gestational age prediction result.

[0014] In a sixth aspect of this application, a diagnostic report determination device is proposed, comprising: a gestational age result determination module, used to determine the gestational age prediction result of the target object based on the gestational age ultrasound image of the target object, using the method described in the second aspect, the gestational age prediction result including: gestational age prediction value and measured values ​​of fetal anatomical structures; a knowledge base construction module, used to construct a fetal growth standard and medical diagnostic knowledge base; a retrieval module, used to match the gestational age prediction result with the fetal growth standard and medical diagnostic knowledge base, retrieve the reference range for the gestational age in the knowledge base, and determine each measured value; and a report output module, used to construct a three-layer evidence chain of "structure-measurement-clinical phenotype" and output a structured diagnostic report; wherein, constructing the structured diagnostic evidence chain and outputting the structured diagnostic report includes: displaying a segmented mask image of key fetal anatomical regions; listing the actual measured values, reference ranges, and percentage deviations of each structure; and outputting possible diagnostic suggestions, disease categories, and their clinical interpretations according to knowledge base rules.

[0015] In an eighth aspect of this application, an electronic device is provided, comprising: a processor and a memory; the memory for storing a computer program; and the processor for executing the computer program to implement the image segmentation method as described in the first aspect, the gestational age prediction method as described in the second aspect, or the diagnostic report determination method as described in the third aspect.

[0016] In a ninth aspect of this application, a computer-readable storage medium is provided that stores computer instructions or programs that, when executed on a computer, cause the image segmentation method as described in the first aspect, the gestational age prediction method as described in the second aspect, or the diagnostic report determination method as described in the third aspect to be performed.

[0017] In a tenth aspect of this application, a computer program product is proposed, comprising computer instructions that, when some or all of the computer instructions are executed on a computer, cause the image segmentation method as described in the first aspect, the gestational age prediction method as described in the second aspect, or the diagnostic report determination method as described in the third aspect to be performed.

[0018] Compared with the prior art, the solution according to the embodiments of this application can achieve at least one of the following technical advantages:

[0019] To address the limitations of existing methods, which are constrained by single network structures or static tasks and cannot meet the demands of multi-faceted and multi-structure collaborative segmentation, this application employs a feature-enhanced structural segmentation network. This network introduces a layer-adaptive feature gating module at the encoder-decoder skip connection, dynamically selecting effective features through channel attention and spatial attention to suppress noise artifacts and improve segmentation accuracy. Furthermore, a bottleneck domain adaptive feature calibration module is introduced at the bottleneck layer, utilizing shallow spatial details to guide deep semantic localization and enhance the robustness of multi-faceted organ segmentation.

[0020] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic flowchart of an image segmentation method according to an embodiment of this application;

[0023] Figure 2 This is a schematic diagram of the structure of a U-shaped image segmentation network according to an embodiment of this application;

[0024] Figure 3 This is a schematic diagram of the network structure of a layer-adaptive feature-gated module according to an embodiment of this application;

[0025] Figure 4 This is a schematic diagram of the network structure of a bottleneck domain adaptive feature calibration module according to an embodiment of this application;

[0026] Figure 5 This is a schematic flowchart of a gestational age prediction method according to an embodiment of this application;

[0027] Figure 6 This is a schematic diagram of the predictive model network structure according to an embodiment of this application;

[0028] Figure 7 This is a schematic flowchart of a diagnostic report determination method according to an embodiment of this application;

[0029] Figure 8 This is a schematic diagram illustrating an example of a diagnostic report result according to an embodiment of this application;

[0030] Figure 9 This is a schematic diagram of an image segmentation apparatus according to an embodiment of the present application;

[0031] Figure 10 This is a schematic diagram of a gestational age prediction device according to an embodiment of this application;

[0032] Figure 11 This is a schematic diagram of a diagnostic report determination device according to an embodiment of this application;

[0033] Figure 12 This is a schematic block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0034] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0036] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0037] Before introducing the technical solution of this application, the relevant knowledge of this application will be introduced below:

[0038] In this embodiment, the "U-shaped image segmentation network" is a deep neural network based on an encoder-decoder structure, typically including a downsampling path (encoder) and an upsampling path (decoder), with a structure resembling a "U". The encoder extracts high-dimensional semantic features from the input image, while the decoder progressively restores the spatial resolution of the image and combines these features with shallow features from the encoder to achieve accurate image segmentation.

[0039] In the embodiments of this application, the "encoder" is a neural network module that converts the original input data into a high-dimensional feature representation with semantic information, used to extract key information from the input and compress redundant content.

[0040] In the embodiments of this application, the "decoder" is a neural network module that restores or converts the high-dimensional feature representation generated by the encoder into the target form output (such as an image, sequence, or label map) to generate the final result.

[0041] In this embodiment, the "training set" is the data used to train the model. It contains a series of input samples (independent variables) and corresponding known outputs (dependent variables). During the training phase, the deep learning model uses samples in the training set to learn and adjust the model's parameters (weights) to minimize the error between the predicted result and the true label. By continuously trying different parameters and algorithms, the model gradually learns the patterns and features between samples, thereby determining a more accurate predictive ability.

[0042] In this embodiment, the "Test Set" is a dataset used to evaluate the performance of a deep learning model. It also contains input samples (independent variables) and corresponding true outputs (dependent variables), but the samples in the test set are not used during the model's training phase, ensuring that the data in the test set is unknown to the model. After training, the test set is used to evaluate the model's performance, that is, the model predicts the samples in the test set and compares the predictions with the true labels. By comparing the prediction results with the true labels, the model's performance on unknown data is evaluated, and its generalization ability and prediction accuracy are determined.

[0043] The technical problems to be solved, the inventive concept and the system architecture of the embodiments of this application will be described below:

[0044] In recent years, deep learning technology has made breakthroughs in the field of medical imaging. U-Net and its derivative networks achieve end-to-end automatic segmentation through encoder-decoder structures and skip connections, significantly improving the detection accuracy of targets such as tumors and organs. Attention mechanisms (such as CBAM and SE modules) further enhance the model's ability to focus on key features. However, existing networks are mostly designed for single-structure or static tasks, making it difficult to meet the needs of multi-faceted and multi-structure collaborative segmentation. At the same time, the lack of automatic integration with quantitative measurements limits their application in specific scenarios (such as gestational age prediction).

[0045] To address the technical problems of the aforementioned prior art, this application proposes an image segmentation method. This method embeds a layer adaptive feature gating module and a bottleneck domain adaptive feature calibration module into a U-shaped image segmentation network. These modules are used to enhance the expressive power of key anatomical structures and filter irrelevant information, respectively, and to utilize shallow spatial details to guide deep semantic localization, thereby improving the robustness of multi-section organ segmentation.

[0046] The technical solution of this application will be described in detail below:

[0047] Figure 1 This is a flowchart of an image segmentation method according to an embodiment of this application. The method may include:

[0048] S110, Construct a U-shaped image segmentation network, which includes an encoder and a decoder; wherein the encoder and the decoder are connected through skip connection layers; each skip connection layer of the U-shaped image segmentation network embeds a layer adaptive feature gating module, which includes a channel attention extraction submodule and a spatial attention extraction submodule, used to enhance the expressive power of key anatomical structures and filter irrelevant information; the bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module, which includes a feature fusion unit, a multi-layer spatial attention extraction submodule, a multi-layer channel attention extraction submodule, and a feature residual connection unit, used to utilize shallow spatial details to guide deep semantic localization and improve the robustness of multi-section organ segmentation;

[0049] In some implementations, the U-shaped image segmentation network is selected from the U-Net basic network framework.

[0050] For example, combined Figure 2 As shown, the U-shaped image segmentation network uses U-Net as its basic framework, with the backbone consisting of an encoder, a decoder, and skip connections. The encoder extracts features at different scales, the decoder progressively restores spatial resolution, and the skip connections fuse features from different levels to compensate for detail loss caused by high-level abstract semantic information. Preferably, the encoder contains 8 layers, each consisting of two 3×3 convolutional modules, a ReLU activation function, and a 2×2 max-pooling layer. With each downsampling level, the number of channels doubles, and the spatial size is halved, progressively constructing a pyramid-shaped feature representation. Let the output features of the l-th layer encoder be... l∈{1, 2, ..., 8}. In the decoder part, each layer uses transposed convolution for upsampling, skips connections with the encoder features of the same scale, and then fuses them. Subsequently, it is refined through two 3×3 convolutional layers to gradually restore the spatial resolution of the segmentation map. Let the decoder features of the l-th layer be... To enhance the expressive power of key anatomical structures and filter irrelevant information, while simultaneously improving information flow and structural semantic preservation between different feature levels, a preferred approach is to use a layer-adaptive feature gating module to mine spatial and channel features of the image at each layer's skip connections. Furthermore, a bottleneck domain adaptive feature calibration module is designed to aggregate features at different levels, utilizing shallow spatial details to guide deep semantic feature localization, thereby achieving high-precision segmentation and localization of various anatomical structures (such as the femur, thalamus, and abdominal girth). For clarity, this structure will be described in detail below:

[0051] 1) Regarding the layer-adaptive feature gating module

[0052] Specifically, in combination Figure 3 As shown, firstly, the input features will be analyzed. Channel attention extraction is performed, and the input feature map is subjected to max pooling and average pooling operations respectively to obtain two features. and Max pooling divides the input feature map into non-overlapping pooling windows and selects the maximum value from each window as the output, while average pooling calculates the average value within each pooling window as the output. Pooling operations can compress the features of each channel into a single scalar, which can be used to measure the importance of the channel.

[0053] Next, and The input is fed into a Shared Multi-Layer Perceptron (Shared MLP), which consists of two convolutional layers and an activation function to learn the feature representations and importance weights of each channel dimension. During training, the Shared MLP shares the same weight parameters, allowing the model to fully utilize information from all channels and thus improve the accuracy of channel importance estimation. The output features of the Shared MLP are concatenated and then mapped using a sigmoid activation function to finally generate channel attention feature weights, which can be expressed as:

[0054]

[0055] Then, the channel weights are combined with the original input features. Calibration is performed by multiplying the weights by the original image to obtain a weighted output feature map, and then spatial attention is applied to extract the spatial attention weights. To compute spatial attention, average pooling and max pooling operations are applied along the channel dimension of the feature map, resulting in two features. and Concatenating these two feature maps yields a feature descriptor. Then, a convolution operation is applied to this feature to generate spatial attention features. This process can be represented as follows:

[0056]

[0057] Finally, the channel attention features and spatial attention features are multiplied together to obtain the output features. This process can be represented by the following formula:

[0058]

[0059] To focus on the contribution of features from a particular encoder layer to the decoder, a layer weight parameter α is introduced. l ∈[0,1], to quantify the contribution of this layer to the decoder, can be expressed as:

[0060]

[0061] Where GAP(·) is the global average pooling operation, W l and b l The weights and biases of the gating units are defined. Through the GAP operation, the spatial information of the feature map is compressed into a scalar, representing the overall importance of the feature map at that layer. Finally, weighted features based on channel attention and spatial attention are applied. After hierarchical weight α l In conjunction with the gating unit, feature fusion is performed to obtain the final features. Figure X l , can be represented as:

[0062]

[0063] in β represents element-wise multiplication. l Generated by the feature complexity estimator:

[0064]

[0065] Here, Var(·) calculates the channel variance, and γ is the temperature hyperparameter. By using variance operations, the complexity of the feature map can be measured, thereby dynamically adjusting the weights of feature fusion to ensure that it can degenerate into an identity mapping when the features are simple, reducing computational redundancy.

[0066] 2) Regarding the bottleneck domain adaptive feature calibration module:

[0067] To further alleviate the spatial information loss problem of deep features, as a preferred approach, a bottleneck domain adaptive feature calibration module is introduced between the deepest layer of the encoder (layer 8) and the decoder. Combined with... Figure 4 As shown, this module first constructs shallow composite features. The shallow features from layers 1 to 7, after being processed by the layer adaptive feature gating module, are upsampled to the bottleneck layer size through transposed convolution, and then concatenated after 1×1 convolution dimensionality reduction to obtain the composite feature F. G It can be expressed by the following formula:

[0068] F G =Conv(cat(Tranconv(X)) 1 ),Tranconυ(X 2 ),...)), Equation 7);

[0069] F G and the deep features F of the final bottleneck layer ASpatial attention weights are obtained by passing the spatial attention module separately. These two weights are then multiplied, and the resulting weights are processed using the Softmax activation function before being multiplied by F. A Multiplication yields feature weights Finally and F A Perform residual connections to obtain the final feature F. S The entire process can be represented by the following formula:

[0070]

[0071] Then, F S F G The channel attention modules are input separately to obtain channel attention weights. These two weights are then multiplied, and the resulting weights are processed using the Softmax activation function before being multiplied by F. S Multiplication yields feature weights Finally and F S Perform residual connections to obtain the final feature F. C The entire process can be represented by the following formula, F C That is, the output F of the final module bottleneck domain adaptive feature calibration module. M :

[0072]

[0073] S120, The U-shaped image segmentation network is trained using the training set to obtain an image segmentation model;

[0074] In some implementations, the training set is selected from publicly available gestational ultrasound images.

[0075] It is understood that the training set of this application is a dataset after preprocessing the original data. The preprocessing includes: normalizing the acquired fetal ultrasound images, such as normalizing the pixels to the [0,1] interval, and using data augmentation techniques (such as random rotation, horizontal flipping, random scaling and cropping) to expand the sample diversity in order to enhance the robustness of the model under different fetal positions and imaging conditions.

[0076] The training method used in this application adopts conventional training methods in the field, which will not be described in detail here.

[0077] S130, the image to be segmented is input into the image segmentation model to obtain the image segmentation result.

[0078] Figure 5 This is a flowchart of a gestational age prediction method according to an embodiment of this application. The method may include:

[0079] S510, The ultrasound image of the gestational age to be tested is segmented using the aforementioned image segmentation method to obtain the image segmentation result;

[0080] In this embodiment, a detailed description of obtaining the image segmentation result can be found above, and will not be repeated here.

[0081] S520, Based on the image segmentation results, extract the measurement values ​​of the fetal anatomical structure;

[0082] In some feasible implementations, the fetal anatomical structures include, but are not limited to, biparietal diameter (BPD), head circumference (HC), abdominal circumference (AC), and femur length (FL). The extraction of fetal anatomical measurements based on image segmentation results can be automatically performed using region marker extraction, edge detection, and geometric analysis.

[0083] For example, the biparietal diameter (BPD): In a transverse section mask at the thalamic level, the midline of the brain is extracted using a skeletalization algorithm, and the maximum transverse distance is determined using an iterative rotational projection method. The BPD measurement axis must be defined as perpendicular to the midline of the brain and passing through the anterior 1 / 3 of the cavum septum pellucidum, with the following geometric relationships:

[0084]

[0085] Where M is the set of points on the skull mask boundary, and θ is the rotation angle of the measurement line.

[0086] Head circumference (HC) and abdominal circumference (AC): An improved ellipse fitting algorithm was used to measure the contour point set. The elliptic parameter estimation is transformed into a weighted least squares optimization problem:

[0087]

[0088] Wherein, weight w i The final perimeter is calculated using the Ramanujan approximation formula, determined by the local gradient magnitude:

[0089] Femur Length (FL): Extract the bone endpoints in the long axis section and calculate the straight-line distance.

[0090] The measurement process can be completed automatically based on region marker extraction, edge detection and geometric analysis without human intervention, ensuring objectivity and consistency.

[0091] S530, the image segmentation results and the measured values ​​of the fetal anatomy are input into the prediction model to obtain the fetal gestational age prediction results.

[0092] In some implementations, the prediction model is trained as follows: A training dataset is acquired, comprising a first training set and a second training set, wherein the first training set includes measurements of fetal anatomical structures and the second training set includes image segmentation results; the gestational age regression model is trained using the training dataset to obtain the prediction model; wherein constructing the gestational age regression model includes: inputting the first training set into a multilayer perceptron to obtain parameter feature vectors, wherein the multilayer perceptron contains three fully connected layers and a ReLU activation function; inputting the second training set into a pre-trained ResNet-18 network to obtain feature vectors of multiple predetermined dimensions; concatenating the multiple feature vectors of predetermined dimensions and inputting them into an attention module to obtain an attention feature map; concatenating the parameter feature vectors with the attention feature map and inputting them into the multilayer perceptron, activating the feature map using the Sigmoid function to obtain the predicted gestational age value.

[0093] Specifically, refer to Figure 6 As shown, the specific structure of the gestational age regression model is as follows:

[0094] The first branch: Input the measured value vector M into a multi-layer perceptron (MLP), which contains three fully connected layers and a ReLU activation function, and outputs a feature vector F. mea ;

[0095] The second branch uses a pre-trained ResNet-18 as the base backbone network. As a preferred approach, the original three-channel input layer is replaced with a single-channel input layer, resulting in an output feature map size of 512×1×1. Each of the three input backbones has its own parameters to adapt to features from different cross-sections, ultimately yielding three 512-dimensional feature vectors representing the feature maps F of the abdomen, femur, and thalamus, respectively. Abdominal F Femur and F Cranium The concatenated features are then input as Q, K, and V into the attention module to obtain the attention feature map F. mask :

[0096] F concat =[F Abdominal ,F Femur, ,F Cranium ],

[0097] F mask =Attention(F concat F concat F concat Equation 12);

[0098] Cross-modal feature fusion: First, the attention-weighted mask features and parameter flow features are concatenated and input into a multilayer perceptron. Finally, the sigmoid function is used to activate the feature map to obtain a standardized gestational age prediction value.

[0099] GA = MLP([F mask ,F mea ]), Equation 13);

[0100] Loss function design: The model is optimized using a weighted sum of the Smooth L1 loss function and the mean squared error loss function, specifically expressed as the following formula:

[0101] L total =αL smooth +(1-α)L mse

[0102]

[0103] The Smooth L1 loss function is quadratic when the error is small, but becomes linear when the error is large. It is less sensitive to outliers and converges to the global optimum better in the later stages of training. The mean squared error loss function averages the squared difference between the predicted and true values ​​for each sample. By weighting and combining these two loss functions, the model's fitting ability and robustness to outliers can be balanced during training. Preferably, the value of α is 0.3. Through the above end-to-end modeling strategy, this invention achieves a seamless connection from image segmentation to gestational age prediction, significantly reducing the impact of human intervention on prediction accuracy. It is particularly suitable for predicting fetuses with high growth heterogeneity in the mid-to-late stages of pregnancy.

[0104] Current AI-assisted diagnostic systems mostly employ "black box" models, only outputting risk scores and failing to provide traceable, structured diagnostic evidence, thus reducing clinicians' trust and operability. Therefore, how to achieve standardized anatomical measurements based on automatic ultrasound image segmentation and combine this with an interpretable diagnostic process has become a key technical challenge urgently needing to be addressed in the field of prenatal screening. To this end, this application proposes an interpretable diagnostic process. Specifically,

[0105] Figure 7 This is a flowchart of a diagnostic report determination method according to one embodiment of this application. The method may include:

[0106] S710, Based on the gestational age ultrasound image of the target object, the gestational age prediction result of the target object is determined by the gestational age prediction method, the gestational age prediction result includes: gestational age prediction value and measurement value of fetal anatomy;

[0107] Understandably, the methods used to determine the gestational age prediction results for the target population can be found above, and will not be repeated here.

[0108] S720, constructing a knowledge base for fetal growth standards and medical diagnostics;

[0109] In this embodiment, constructing a knowledge base for fetal growth standards and medical diagnostics includes:

[0110] First, a multi-level fetal development knowledge base is constructed, including: key biometric reference ranges corresponding to normal gestational age (such as BPD, HC, AC, FL, etc.); structural measurement characteristics corresponding to common developmental abnormalities (such as hydrocephalus, microcephaly, short limbs, etc.); and medical literature rules such as clinical manifestations, risk ranges, and auxiliary imaging characteristics corresponding to typical diseases.

[0111] Preferably, the knowledge base sources include WHO international fetal growth curves, screening standards of large domestic hospitals, authoritative guidelines and textbooks (such as ISUOG and ACOG recommendations), and the structured knowledge obtained is stored in the neo4j graph database to support fast retrieval.

[0112] S730, Match the gestational age prediction result with the fetal growth standard and medical diagnostic knowledge base, retrieve the reference range for the gestational age in the knowledge base, and make a judgment on each measurement item;

[0113] In this embodiment, the gestational age prediction result is matched and searched against the fetal growth standards and medical diagnostic knowledge base. The reference range for the gestational age is retrieved from the knowledge base, and each measurement item is evaluated. This includes: for each fetal ultrasound image, the predicted gestational age GA_pred and various structural measurements M = [BPD, HC, AC, FL, ...] have been obtained through the aforementioned steps. The system determines the gestational age based on the gestational age prediction model and retrieves the reference range R for that gestational age from the knowledge base. i =[min i max i The system will determine the outcome of each measurement. For example, if the GA is predicted to be 26 weeks and the FL is far below the normal range and the AC is small, it can be matched with abnormal patterns such as "small for gestational age" or "lethal achondroplasia".

[0114] S740 constructs a three-layer evidence chain of "structure-measurement-clinical phenotype" and outputs a structured diagnostic report. The construction of the structured diagnostic evidence chain and the output of the structured diagnostic report include: displaying a masked image of the segmented key anatomical regions of the fetus; listing the actual measurement values, reference intervals, and percentage deviations of each structure; and outputting possible diagnostic suggestions, disease categories, and their clinical interpretations based on knowledge base rules.

[0115] To enhance the interpretability and traceability of the system output, preferably, a structured diagnostic evidence chain is constructed, wherein, referring to Figure 8 The content includes: displaying a masked image of the segmented key anatomical regions of the fetus; listing the actual measurements, reference ranges, and percentage deviations for each structure; and outputting possible diagnostic suggestions, disease categories, and their clinical interpretations based on knowledge base rules.

[0116] It is understood that the platform for implementing the above method can be based on a Python environment, combined with a deep learning framework (such as PyTorch or TensorFlow). The basic operations involved, such as data reading and writing, neural network training and inference, are well-known technologies in the field and will not be elaborated here. In specific implementation, the method of this application can achieve automated operation through software technology. It should be noted that the embodiments described in this specification are merely exemplary and not limiting, and the parts not described in detail belong to the prior art.

[0117] The three methods described in this application are combined into an end-to-end multi-task medical image analysis system guided by structural segmentation, aiming to improve the prediction accuracy of fetal gestational age and the intelligent screening capability for developmental abnormalities in ultrasound images. The overall system architecture includes three core components: a structure-enhanced segmentation network, an automated structural measurement and joint regression model, and a knowledge-prior-guided abnormality diagnosis module. The segmentation network adopts the U-Net network framework, integrating channel attention and spatial attention mechanisms to enhance the expressive power of key anatomical structures and filter irrelevant information, respectively. To enhance information flow and structural semantic preservation between different levels of features, a bottleneck domain feature calibration mechanism is further introduced to construct a deep and shallow feature fusion structure, thereby achieving high-precision segmentation and localization of various anatomical structures (such as the femur, thalamus, and abdominal circumference) in complex ultrasound backgrounds. Based on segmentation, this invention proposes a structure-guided automatic measurement method that directly extracts biometric values ​​of key fetal structures from the segmentation mask and inputs them into the regression model in conjunction with depth image features, realizing an end-to-end gestational age prediction process and avoiding information loss caused by the separation of segmentation and regression tasks in traditional processes. In addition, to improve the clinical interpretability of the system output, a priori retrieval engine based on a medical knowledge base was constructed, establishing a three-layer evidence chain of "image structure → anatomical measurement → clinical phenotype". The model's diagnostic results were derived in a structured manner, enhancing the cognitive consistency between the system and doctors.

[0118] The methods provided in the embodiments of this application have been described above. The apparatus provided in the embodiments of this application will be described below.

[0119] Figure 9 This is a schematic diagram of an image segmentation apparatus 900 according to an embodiment of this application, as shown below. Figure 9As shown, the device 900 includes: a network construction module 910, a training module 920, and a first prediction module 930. Among them,

[0120] The network construction module 910 is used to construct a U-shaped image segmentation network, which includes an encoder and a decoder. The encoder and decoder are connected via skip connection layers. Each skip connection layer of the U-shaped image segmentation network embeds a layer adaptive feature gating module, which includes a channel attention extraction submodule and a spatial attention extraction submodule, used to enhance the expressive power of key anatomical structures and filter irrelevant information. The bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module, which includes a feature fusion unit, a multi-layer spatial attention extraction submodule, a multi-layer channel attention extraction submodule, and a feature residual connection unit, used to utilize shallow spatial details to guide deep semantic localization and improve the robustness of multi-faceted organ segmentation. The training module 920 is used to train the U-shaped image segmentation network using a training set to obtain an image segmentation model. The first prediction module 930 is used to input the image to be segmented into the image segmentation model to obtain the image segmentation result.

[0121] In some implementations, the layer-adaptive feature gating module integrates channel attention and spatial attention mechanisms, which are used to enhance the expressive power of key anatomical structures and filter irrelevant information, respectively, including: for input features Channel attention extraction is performed, and the input feature map is subjected to max pooling and average pooling operations to obtain the feature map. and Will and The input features are fed into a shared multilayer perceptron, and their output features are fused through a concatenation operation. Then, they are mapped using a sigmoid activation function to generate channel attention feature weights. These channel attention feature weights are then compared with the input features. Multiply to obtain a weighted output feature map; apply spatial attention to extract spatial attention weights. This includes applying average pooling and max pooling operations along the channel dimension of the feature map to obtain the features. and Connecting them together yields a feature descriptor. Then, a convolution operation is applied to this feature to generate a spatial attention feature; the channel attention feature and the spatial attention feature are multiplied to obtain the output feature.

[0122] In some possible implementations, this further includes: introducing a hierarchical weight parameter α. l∈[0,1], used to quantify the contribution of the target layer to the decoder; weighted features of channel attention and spatial attention. After hierarchical weight α l In conjunction with the gating unit, feature fusion is performed to obtain the final features. Figure X l .

[0123] In some implementations, the bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module to guide deep semantic localization using shallow spatial details, thereby improving the robustness of multi-faceted organ segmentation. This includes: constructing shallow composite features by upsampling the shallow features processed by the layer adaptive feature gating module after the predetermined coding layer is processed by transposed convolution to the bottleneck layer size, performing 1×1 convolution dimensionality reduction, and then concatenating them to obtain the composite feature F. G Composite feature F G and the deep features F of the bottleneck layer A Spatial attention weights are obtained by passing through the spatial attention module, multiplying the obtained weights, processing them with the Softmax activation function, and then multiplying them with F. A Multiplication yields feature weights Will and F A Perform residual connection to obtain spatial feature F S ; F S F G The channel attention weights are obtained by inputting them into the channel attention module. These weights are then multiplied and processed using the Softmax activation function, and finally multiplied by F. S Multiplication yields feature weights Will and F S Perform residual connections to obtain the output feature F. C .

[0124] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, further details will not be provided here. Specifically, Figure 9 The device shown can perform Figure 1 The corresponding method embodiments, and the foregoing and other operations and / or functions of each module in the device, are respectively for implementing Figure 1 For the sake of brevity, the corresponding processes in each method are not described in detail here.

[0125] The apparatus of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0126] Figure 10 This is a schematic diagram of a gestational age prediction device 1000 according to one embodiment of this application, as shown below. Figure 10 As shown, the device 1000 includes: an image segmentation module 1010, a parameter determination module 1020, and a second prediction module 1030, wherein,

[0127] The image segmentation module 1010 is used to segment the ultrasound image of the gestational age to be tested using an image segmentation method to obtain the image segmentation result; the parameter determination module 1020 is used to extract the measurement values ​​of the fetal anatomical structure based on the image segmentation result; the second prediction module 1030 is used to input the image segmentation result and the measurement values ​​of the fetal anatomical structure into the prediction model to obtain the fetal gestational age prediction result.

[0128] In some implementations, the prediction model is trained as follows: A training dataset is acquired, comprising a first training set and a second training set, wherein the first training set includes measurements of fetal anatomical structures and the second training set includes image segmentation results; the gestational age regression model is trained using the training dataset to obtain the prediction model; wherein constructing the gestational age regression model includes: inputting the first training set into a multilayer perceptron to obtain parameter feature vectors, wherein the multilayer perceptron contains three fully connected layers and a ReLU activation function; inputting the second training set into a pre-trained ResNet-18 network to obtain feature vectors of multiple predetermined dimensions, concatenating the multiple feature vectors of predetermined dimensions, and inputting them into an attention module to obtain an attention feature map; concatenating the parameter feature vectors with the attention feature map, inputting them into the multilayer perceptron, and activating the feature map using the Sigmoid function to obtain the predicted gestational age value.

[0129] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, further details will not be provided here. Specifically, Figure 10 The device 1000 shown can perform Figure 5 The corresponding method embodiments, and the foregoing and other operations and / or functions of each module in the device 1000 are respectively for implementing Figure 5 For the sake of brevity, the corresponding processes in each method are not described in detail here.

[0130] The apparatus of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0131] Figure 11 This is a schematic diagram of a diagnostic report determining device 1100 according to an embodiment of this application, as shown below. Figure 11 As shown, the device 1100 includes: a gestational age determination module 1110, a knowledge base construction module 1120, a retrieval module 1130, and a report output module 1140, wherein,

[0132] The gestational age determination module 1110 is used to determine the gestational age prediction result of the target object based on the gestational age ultrasound image of the target object through a gestational age prediction method. The gestational age prediction result includes: the predicted gestational age value and the measured values ​​of fetal anatomical structures. The knowledge base construction module 1120 is used to construct a fetal growth standard and medical diagnosis knowledge base. The retrieval module 1130 is used to match and retrieve the gestational age prediction result with the fetal growth standard and medical diagnosis knowledge base, retrieve the reference range for the gestational age in the knowledge base, and make a judgment on each measured value. The report output module 1140 is used to construct a three-layer evidence chain of "structure-measurement-clinical phenotype" and output a structured diagnostic report. Among them, constructing a structured diagnostic evidence chain and outputting a structured diagnostic report includes: displaying a masked image of the segmented key anatomical regions of the fetus; listing the actual measured values, reference ranges, and percentage deviations of each structure; and outputting possible diagnostic suggestions, disease categories, and their clinical interpretations according to the knowledge base rules.

[0133] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, further details will not be provided here. Specifically, Figure 11 The device 1100 shown can perform Figure 7 The corresponding method embodiments, and the foregoing and other operations and / or functions of each module in the device 1100 are respectively for implementing Figure 7 For the sake of brevity, the corresponding processes in each method are not described in detail here.

[0134] The apparatus of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0135] Figure 12 This is a schematic block diagram of the electronic device 1200 provided in an embodiment of this application. The electronic device 1200 may be the training device or execution device described above, but is not limited thereto. Figure 12 As shown, the electronic device 1200 may include:

[0136] The system includes a memory 1210 and a processor 1220. The memory 1210 stores a computer program 1230 and transfers the computer program 1230 to the processor 1220. In other words, the processor 1220 can retrieve and run the computer program 1230 from the memory 1210 to implement the methods described in the embodiments of this application.

[0137] For example, the processor 1220 can be used to execute the steps in the above method according to the instructions in the computer program 1230.

[0138] In some embodiments of this application, the processor 1220 may include, but is not limited to:

[0139] General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0140] In some embodiments of this application, the memory 1210 includes, but is not limited to:

[0141] Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0142] In some embodiments of this application, the computer program 1230 may be divided into one or more modules, which are stored in the memory 1210 and executed by the processor 1220 to complete the method provided in this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 1230 in the electronic device.

[0143] like Figure 12 As shown, the electronic device 1200 may further include:

[0144] Transceiver 1240, which can be connected to processor 1220 or memory 1210.

[0145] The processor 1220 can control the transceiver 1240 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 1240 may include a transmitter and a receiver. The transceiver 1240 may further include antennas, and the number of antennas may be one or more.

[0146] It should be understood that the various components in the electronic device 1200 are connected through a bus system, which includes a data bus, a power bus, a control bus, and a status signal bus.

[0147] According to one aspect of this application, a computer-readable storage medium is provided that stores computer instructions or programs thereon, which, when executed by a computer, enable the computer to perform the methods of the above-described method embodiments. Alternatively, embodiments of this application also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods of the above-described method embodiments.

[0148] According to another aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method described in the above-described method embodiments.

[0149] In other words, when implemented using software, it can be implemented wholly or partially in the form of a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0150] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0151] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0152] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs. For example, the functional modules in the various embodiments of this application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0153] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of image segmentation, characterized in that, include: A U-shaped image segmentation network is constructed, comprising an encoder and a decoder. The encoder and decoder are connected via skip connection layers. Each skip connection layer of the U-shaped image segmentation network embeds a layer adaptive feature gating module, which includes a channel attention extraction submodule and a spatial attention extraction submodule, used to enhance the expressive power of key anatomical structures and filter irrelevant information. The bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module, which includes a feature fusion unit, a multi-layer spatial attention extraction submodule, a multi-layer channel attention extraction submodule, and a feature residual connection unit, used to utilize shallow spatial details to guide deep semantic localization and improve the robustness of multi-section organ segmentation. The U-shaped image segmentation network is trained using the training set to obtain an image segmentation model; The image to be segmented is input into the image segmentation model to obtain the image segmentation result.

2. The method of claim 1, wherein, The layer-adaptive feature gating module integrates channel attention and spatial attention mechanisms, which are used to enhance the expressive power of key anatomical structures and filter irrelevant information, respectively, including: input features channel attention extraction is performed, and the input feature maps are respectively subjected to maximum pooling and average pooling operations to obtain features and will be described below with reference to the accompanying drawings. and The input is input to a shared multi-layer perception, and the output features are fused through a concatenation operation, mapped through a Sigmoid activation function, and channel attention feature weights are generated. multiplying the channel attention feature weight with the input feature to obtain a weighted output feature map; Applying spatial attention to extract spatial attention weights comprises: applying average pooling and maximum pooling operations on a channel dimension of the feature map to obtain a feature and connecting them to obtain a feature descriptor Then apply a convolution operation on the feature to generate a spatial attention feature; The output feature is obtained by multiplying the channel attention feature and the spatial attention feature.

3. The method according to claim 2, characterized in that, Further comprising: introducing a hierarchy weight parameter a l ∈ [0, 1], for quantifying the contribution of the target layer to the decoder; Weighted features of channel attention and spatial attention After hierarchical weight α l In conjunction with the gating unit, feature fusion is performed to obtain the final feature map X. l .

4. The method according to claim 1, characterized in that, The bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module, which is used to guide deep semantic localization using shallow spatial details, thereby improving the robustness of multi-faceted organ segmentation, including: The shallow composite feature is constructed, the shallow feature after the predetermined coding layer is processed through the layer adaptive feature gate module is up-sampled to the bottleneck layer size through the transpose convolution, and is spliced after 1x1 convolution dimension reduction to obtain the composite feature F G ; Composite feature F G and the deep features F of the bottleneck layer A Spatial attention weights are obtained by passing through the spatial attention module, multiplying the obtained weights, processing them with the Softmax activation function, and then multiplying them with F. A Multiplication yields feature weights Will and F A Perform residual connection to obtain spatial feature F S ; F S F G The channel attention weights are obtained by inputting them into the channel attention module. These weights are then multiplied and processed using the Softmax activation function, and finally multiplied by F. S Multiplication yields feature weights Will and F S Perform residual connections to obtain the output feature F. C .

5. A method for predicting gestational age, characterized in that, include: The method described in any one of claims 1-4 is used to segment the ultrasound image of the gestational age to be measured, and the image segmentation result is obtained. Based on the image segmentation results, the measurements of the fetal anatomical structure are extracted; The image segmentation results and the measured values ​​of the fetal anatomy are input into the prediction model to obtain the fetal gestational age prediction results.

6. The method according to claim 5, characterized in that, The prediction model was trained in the following manner: Obtain a training dataset, which includes: a first training set and a second training set, wherein the first training set includes measurements of fetal anatomical structures and the second training set includes image segmentation results; The gestational age regression model was trained using the training dataset to obtain a prediction model; The construction of the gestational age regression model includes: The first training set is input into a multilayer perceptron to obtain a parameter feature vector, wherein the multilayer perceptron contains three fully connected layers and a ReLU activation function; The second training set is input into a pre-trained ResNet-18 network to obtain feature vectors of multiple predetermined dimensions. The feature vectors of multiple predetermined dimensions are concatenated and input into the attention module to obtain an attention feature map. The parameter feature vector is concatenated with the attention feature map and input into a multilayer perceptron. The feature map is activated using the Sigmoid function to obtain the gestational age prediction value.

7. A method for determining a diagnostic report, characterized in that, include: Based on the gestational age ultrasound images of the target object, the gestational age prediction result of the target object is determined by the method of claim 5 or 6, wherein the gestational age prediction result includes: gestational age prediction value and measurement value of fetal anatomy. Construct a knowledge base for fetal growth standards and medical diagnostics; The gestational age prediction results are matched and retrieved with the fetal growth standards and medical diagnostic knowledge base. The reference range for the gestational age is retrieved in the knowledge base, and each measurement item is judged. A three-layered evidence chain of "structure-measurement-clinical phenotype" is constructed to output a structured diagnostic report. The structured diagnostic evidence chain and the output of the structured diagnostic report include: displaying a masked image of the segmented key anatomical regions of the fetus; listing the actual measurement values, reference intervals, and percentage deviations of each structure; and outputting possible diagnostic suggestions, disease categories, and their clinical interpretations based on knowledge base rules.

8. An image segmentation apparatus, characterized in that, include: A network construction module is used to construct a U-shaped image segmentation network, which includes an encoder and a decoder. The encoder and decoder are connected via skip connection layers. Each skip connection layer of the U-shaped image segmentation network embeds a layer adaptive feature gating module, which includes a channel attention extraction submodule and a spatial attention extraction submodule, used to enhance the expressive power of key anatomical structures and filter irrelevant information. The bottleneck layer of the U-shaped image segmentation network embeds a bottleneck domain adaptive feature calibration module, which includes a feature fusion unit, a multi-layer spatial attention extraction submodule, a multi-layer channel attention extraction submodule, and a feature residual connection unit, used to utilize shallow spatial details to guide deep semantic localization and improve the robustness of multi-section organ segmentation. The training module is used to train the U-shaped image segmentation network using the training set to obtain an image segmentation model; The first prediction module is used to input the image to be segmented into the image segmentation model to obtain the image segmentation result.

9. A gestational age prediction device, characterized in that, include: An image segmentation module is used to segment an ultrasound image of the gestational age to be measured by the method described in any one of claims 1-4, and obtain an image segmentation result; The parameter determination module is used to extract measurement values ​​of the fetal anatomical structure based on the image segmentation results; The second prediction module is used to input the image segmentation results and the measured values ​​of the fetal anatomy into the prediction model to obtain the fetal gestational age prediction results.

10. A diagnostic report determining device, characterized in that, include: A gestational age determination module is used to determine the gestational age prediction result of the target object based on the gestational age ultrasound image of the target object, by means of the method described in claim 5 or 6, wherein the gestational age prediction result includes: gestational age prediction value and measurement value of fetal anatomy. The knowledge base construction module is used to build a knowledge base for fetal growth standards and medical diagnosis. The retrieval module is used to match the gestational age prediction results with the fetal growth standards and medical diagnostic knowledge base, retrieve the reference range for the gestational age in the knowledge base, and make a judgment on each measurement value. The report output module is used to construct a three-layer evidence chain of "structure-measurement-clinical phenotype" and output a structured diagnostic report. The construction of the structured diagnostic evidence chain and the output of the structured diagnostic report include: displaying a masked image of the segmented key anatomical regions of the fetus; listing the actual measurement values, reference intervals and percentage deviations of each structure; and outputting possible diagnostic suggestions, disease categories and their clinical interpretations based on knowledge base rules.

11. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the method as described in any one of claims 1 to 7.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions or programs that, when executed on a computer, cause the method as described in any one of claims 1 to 7 to be performed.

13. A computer program product, characterized in that, The computer program product includes computer instructions that, when some or all of the computer instructions are run on a computer, cause the method as described in any one of claims 1 to 7 to be performed.

Images