Construction method and application of endometrial cancer CT image segmentation model

Abstract

The application provides a method for constructing an endometrial carcinoma CT image segmentation model and application thereof, including the following steps: acquiring multiple endometrial CT images labeled with cancerous regions as a training data set; constructing an endometrial carcinoma CT image segmentation architecture, wherein the endometrial carcinoma CT image segmentation architecture is a U-Net structure including an encoding unit, a bottleneck layer and a decoding unit; and training the endometrial carcinoma CT image segmentation architecture with the training data set to obtain an endometrial carcinoma CT image segmentation model. The scheme explicitly captures multi-scale structural change information by constructing a multi-directional coding structure response operator and generating a coding structure gradient field, thereby strengthening the representation of thin and weak boundaries and small lesions to improve the segmentation accuracy of the model.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and in particular to a method for constructing and applying a segmentation model for CT images of endometrial cancer. Background Technology

[0002] Endometrial cancer, a common malignant tumor of the female reproductive system, requires early detection and accurate assessment for treatment planning and prognosis. In clinical practice, computed tomography (CT) has become the core imaging tool for staging and determining the extent of lesions in endometrial cancer due to its wide applicability. However, CT imaging of endometrial cancer presents significant technical challenges: the lesion area occupies a very small portion of the overall image, has weak grayscale contrast with surrounding normal tissue, and often exhibits a wide and blurred transition zone at its boundary. The difference between the lesion and normal or thickened endometrium is not obvious, making it difficult to clearly define the lesion outline and greatly increasing the difficulty of automatic segmentation.

[0003] In current clinical practice, lesion segmentation and volume assessment still mainly rely on manual layer-by-layer delineation by physicians. Although experienced physicians can obtain relatively precise results, this method has inherent drawbacks: it is time-consuming, labor-intensive, and the results are highly dependent on the operator's experience. Furthermore, the consistency of delineation between different physicians and even within the same physician at different times is poor, easily leading to incomplete boundary delineation and the omission of small lesions, failing to meet the urgent clinical need for rapid, stable, and repeatable objective segmentation results. Therefore, developing automated fine segmentation technology for endometrial cancer CT images has clear clinical application value and practical significance.

[0004] In recent years, deep learning-based medical image segmentation methods have made significant progress. Among them, segmentation networks with encoder-decoder structures, which obtain high-level semantic information through downsampling and restore spatial resolution through upsampling and cross-layer fusion, have become the mainstream technical approach for CT image segmentation. However, in the scenario of endometrial cancer CT image segmentation, such methods still have prominent limitations: On the one hand, the network relies excessively on high-level semantic responses to distinguish between the target and the background, lacking explicit internal expression of key information such as the intensity of local structural changes and boundary transition features. This leads to the easy smoothing attenuation of shallow edge cues during convolution and downsampling, resulting in blurred contours and morphological deviations in the final segmentation results. On the other hand, cross-layer fusion and loss design generally adopt spatial uniform processing strategies, with skip connections uniformly fusing different positions and pixel-level loss assigning approximately the same weight to each position. This makes gradient updates during training easily dominated by the dominant background regions, resulting in insufficient effective optimization of difficult-to-segment regions such as boundary transition zones and small-volume lesions, further exacerbating the problems of missed detections and boundary instability.

[0005] In summary, existing automated CT segmentation methods for endometrial cancer generally lack a geometric quantity that can explicitly characterize the intensity of local structural changes within the network, be continuously propagated with the feature flow, and be uniformly utilized. Furthermore, they lack a mechanism for synergistically regulating boundary reconstruction and error allocation based on this internal structural information. This core deficiency leads to difficulties in precisely restoring tumor boundaries and easily overlooking small lesions under complex conditions such as low contrast, blurred boundaries, and a very low proportion of small lesions, becoming a key bottleneck restricting segmentation accuracy and robustness. Summary of the Invention

[0006] This application provides a method for constructing and applying a segmentation model for CT images of endometrial cancer. By constructing a multi-directional coded structural response operator and generating a coded structural gradient field, multi-scale structural change information is explicitly captured, and the characterization of weak boundaries and small lesions is enhanced to improve the segmentation accuracy of the model.

[0007] In a first aspect, embodiments of this application provide a method for constructing a segmentation model of CT images of endometrial cancer, the method comprising:

[0008] Multiple endometrial CT images with cancerous areas marked were obtained as a training dataset.

[0009] A segmentation architecture for CT images of endometrial cancer is constructed, wherein the segmentation architecture for CT images of endometrial cancer is a U-Net structure including an encoding unit, a bottleneck layer and a decoding unit;

[0010] Each encoder in the coding unit includes a coding layer and a coding internal geometry field. The coding internal geometry field constructs coding structure response operators corresponding to different directions to perform group convolution on each channel of the coding features to obtain multiple coding structure response features. The multiple coding structure response features are aggregated by the second norm to obtain a coding structure gradient field. The coding structure gradient field is fused with the initial coding features to obtain the encoder output. The coding features are the output features of the coding layer.

[0011] The bottleneck layer has the same structure as the encoder. The output of the bottleneck layer is fused with the output of the encoder at the corresponding scale and then input into the decoding unit.

[0012] Each decoder in the decoding unit includes a decoding layer and a decoding internal geometry field. The decoding internal geometry field constructs decoding structure response operators corresponding to different directions, which convolve the decoding features to obtain multiple decoding structure response features. The multiple decoding structure response features are aggregated using the second norm to obtain a decoding structure gradient field. The decoding structure gradient field is mapped to a gating factor map. The decoding features are gated and modulated using the gating factor map to obtain a gating result. The gating result is concatenated with the multiple decoding structure response features along the channel dimension to obtain the decoder output. The output of the last decoder in the decoding unit is the segmentation result, and the decoding features are the output features of the decoding layer.

[0013] The endometrial cancer CT image segmentation model was obtained by training the training dataset on the endometrial cancer CT image segmentation architecture.

[0014] Secondly, embodiments of this application provide a method for applying a CT image segmentation model for endometrial cancer, including:

[0015] Acquire the CT image of the endometrium to be segmented, and input the CT image of the endometrium to be segmented into the trained endometrial cancer CT image segmentation model to obtain the segmentation result of the cancerous region.

[0016] Thirdly, embodiments of this application provide an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a method for constructing a CT image segmentation model for endometrial cancer.

[0017] The main contributions and innovations of this invention are as follows:

[0018] This scheme constructs a multi-directional coding structure response operator and generates a coding structure gradient field in the coding unit to explicitly capture multi-scale structural change information, enhance the representation of weak boundaries and small lesions, and avoid the attenuation of shallow edge clues. In the decoding unit, this scheme constructs a decoding structure response operator and maps the gradient field to a gating factor map for gating modulation, thereby enhancing the salient structural regions pixel by pixel, suppressing background interference, and improving the continuity and accuracy of boundary reconstruction. This scheme fuses the coding structure gradient fields to generate a global saliency field and constructs a pixel-level weighted loss function, thereby adaptively allocating training errors, allowing difficult-to-distinguish regions such as boundary transition zones and small lesions to obtain more optimization, reducing the false negative rate and boundary bias.

[0019] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0021] Figure 1 This is a structural diagram of a CT image segmentation model for endometrial cancer according to an embodiment of this application;

[0022] Figure 2 This is a schematic diagram of the structure of an encoder according to an embodiment of this application;

[0023] Figure 3 This is a schematic diagram of obtaining the gradient field of the coding structure based on the coding structure response operator according to an embodiment of this application;

[0024] Figure 4 This is a schematic diagram of the structure of a decoder according to an embodiment of this application;

[0025] Figure 5 This is a schematic diagram illustrating the construction of a pixel-level weight map according to an embodiment of this application;

[0026] Figure 6 This is a schematic diagram of comparative experimental results based on embodiments of this application;

[0027] Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0029] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0030] Example 1

[0031] This application provides a method for constructing a segmentation model for CT images of endometrial cancer. By constructing a multi-directional coded structural response operator and generating a coded structural gradient field, it explicitly captures multi-scale structural change information, enhances the representation of weak boundaries and small lesions, and improves the segmentation accuracy of the model. Specifically, refer to... Figure 1 The method includes:

[0032] Multiple endometrial CT images with cancerous areas marked were obtained as a training dataset.

[0033] A segmentation architecture for CT images of endometrial cancer is constructed, wherein the segmentation architecture for CT images of endometrial cancer is a U-Net structure including an encoding unit, a bottleneck layer and a decoding unit;

[0034] Each encoder in the coding unit includes a coding layer and a coding internal geometry field. The coding internal geometry field constructs coding structure response operators corresponding to different directions to perform group convolution on each channel of the coding features to obtain multiple coding structure response features. The multiple coding structure response features are aggregated by the second norm to obtain a coding structure gradient field. The coding structure gradient field is fused with the initial coding features to obtain the encoder output. The coding features are the output features of the coding layer.

[0035] The bottleneck layer has the same structure as the encoder. The output of the bottleneck layer is fused with the output of the encoder at the corresponding scale and then input into the decoding unit.

[0036] Each decoder in the decoding unit includes a decoding layer and a decoding internal geometry field. The decoding internal geometry field constructs decoding structure response operators corresponding to different directions, which convolve the decoding features to obtain multiple decoding structure response features. The multiple decoding structure response features are aggregated using the second norm to obtain a decoding structure gradient field. The decoding structure gradient field is mapped to a gating factor map. The decoding features are gated and modulated using the gating factor map to obtain a gating result. The gating result is concatenated with the multiple decoding structure response features along the channel dimension to obtain the decoder output. The output of the last decoder in the decoding unit is the segmentation result, and the decoding features are the output features of the decoding layer.

[0037] The endometrial cancer CT image segmentation model was obtained by training the training dataset on the endometrial cancer CT image segmentation architecture.

[0038] In this current embodiment, the scheme uses the CT portion of the ECPC-IDS public dataset for endometrial cancer as the training dataset, and divides it into non-overlapping training and testing sets in an 8:2 ratio. To adapt to the imaging characteristics of endometrial cancer lesions with low proportions and weak grayscale contrast, and to improve the learning efficiency of the model, this scheme performs image standardization, spatial normalization, and targeted data augmentation on each CT image in the dataset. Specifically, image standardization normalizes the grayscale values ​​(Henry units) of the CT images to a preset range to eliminate intensity differences caused by different scanning devices and parameters. Spatial normalization resamples all CT images and their corresponding targets to a uniform spatial resolution to ensure consistency in spatial scale. Targeted data augmentation applies random rotation, flipping, and elastic deformation to each CT image to simulate the natural morphological changes of lesions within the cavity and to expand the morphological diversity of the samples.

[0039] In this scheme, the encoder has the following structure: Figure 2 As shown, the encoding layer in the encoder uses multiple convolutional layers of different sizes in parallel to encode the input features to obtain encoded features. Specifically, the encoding layer uses 1×1 convolutional layers, 3×3 convolutional layers and 5×5 convolutional layers in parallel, and the convolution results of the 1×1 convolutional layer, 3×3 convolutional layer and 5×5 convolutional layer are concatenated and input to obtain encoded features.

[0040] In the current embodiment, the coding structure response operator includes a direction index and a channel index. The direction index confirms the convolution direction of the corresponding coding structure response operator, and the channel index is used to guide the coding structure response operator to convolve the corresponding channel of the coding feature. The direction index is generated by pre-setting convolution directions that cover different structural changes such as horizontal, vertical and diagonal directions. For example, typical directions such as horizontal, vertical and ±45° can be selected, but it is not limited to only four directions. The number of directions D can be set or expanded in the implementation according to the task complexity and computational load.

[0041] Specifically, the encoded features are grouped by channel and different encoded structure response operators are applied, allowing different channels to independently learn their more suitable local structure response patterns. This avoids the inadequacy of a single shared operator in representing diverse structural morphologies. The formula for grouping and convolving the encoded features based on the encoded structure response operators is expressed as follows:

[0042]

[0043] Where x is any spatial location of the encoded feature, For the first The encoded features of the c-th channel output by the encoder. For the first The encoder's coded structure response operator with direction index d and channel index c. For the first The encoded structural response features of the encoder output at spatial position x, direction d, and channel c.

[0044] Specifically, this scheme learns the structural changes in CT images in different directions by encoding structural response operators. Whether the structural changes are horizontal, vertical or oblique, they will be detected by the encoder, thereby increasing the recognition of cancerous areas.

[0045] Furthermore, before training the segmentation architecture for endometrial cancer CT images, the coding structure response operator within each encoder is initialized using a first-order difference, and each of these coding structure response operators is adaptively updated during the training process. A schematic diagram of obtaining the coding structure gradient field based on the coding structure response operator is shown below. Figure 3 As shown.

[0046] Specifically, the coding structure response operator is not fixed to a preset template but is adaptively updated following the end-to-end training process. Initializing the coding structure response operator in the form of first-order difference can make the coding structure response operator both sensitive to weak boundaries and adaptable to different imaging conditions and lesion morphology.

[0047] Furthermore, the structural response features of multiple encoded structures are aggregated using the L2 norm in the directional dimension to obtain structural gradient features corresponding to different channels. The structural gradient features of each channel are then integrated to obtain the encoded structural gradient field. This encoded structural gradient field represents the structural changes of the encoded features in different channels at the current scale, as expressed by the formula:

[0048]

[0049] in, For the first The encoded structural response features of the encoder outputs at spatial position x, direction d, and channel c, where D is the number of directions. For numerically stable terms, For the first The structure gradient features of the encoded features output by each encoder at spatial position x and channel c are integrated to obtain the encoded structure gradient field.

[0050] Specifically, performing L2 norm aggregation on multiple coding structural response features in the direction dimension makes the internal structural gradient field more robust to changes in direction. When the boundary of the coding feature presents different directions at different locations, the response in a single direction may be unstable, while multi-directional aggregation can more consistently reflect "whether there is a significant structural change", thus providing a more reliable indication of fuzzy transition zones and weak contours.

[0051] The encoded structure gradient field obtained by this scheme has a clear cross-layer meaning. Its data corresponds to the strength of structural changes in space and reflects the structural saliency under the receptive field of the layer in scale. Therefore, it can be output at multiple scales as a traceable internal geometric quantity in the encoding stage.

[0052] Specifically, this scheme strengthens backbone features by constructing a coding structure response operator, enabling the network to maintain sensitivity to the intensity of local structural changes at different scales. This fundamentally alleviates the problem of small lesions being submerged by the background and the difficulty in finely depicting blurred transition zones, providing a structure-aware endogenous driving paradigm for the segmentation of low-contrast, small-target medical images.

[0053] In the current embodiment, the gradient field of the coding structure is fused with the initial coding features to obtain a first fused feature, and the first fused feature is residually connected with the input features of the current encoder to obtain the encoder output.

[0054] Specifically, fusing the gradient field of the coding structure with the initial coding features allows the backbone features to be compensated and enhanced with structural change information before entering downsampling, thereby improving the ability to represent boundaries and weak structures during the coding stage and providing a higher quality and clearer high-resolution feature source for subsequent skip connections.

[0055] Specifically, the first integrated feature is residually connected to the input feature of the current encoder so that the injection of structural information does not destroy the stability of the original semantic representation, thereby maintaining a balance between training convergence and feature expression under complex imaging conditions.

[0056] In the current embodiment, the bottleneck layer is used to continuously retain and enhance global structural information at the deepest layer of the encoding end, providing key support for decoding reconstruction and training optimization.

[0057] Specifically, the bottleneck layer has the same structure as the encoder, which can continue the gradient field of the encoded structure in the encoder output. This allows the semantic features of the deepest layer to still carry the intensity of global structural changes. Furthermore, the fact that the bottleneck layer has the same structure as the encoder ensures that the output format of the bottleneck layer is consistent with that of the encoder, which facilitates skip connections in the decoding unit.

[0058] Specifically, the output of the bottleneck layer is fused with the encoder output of the corresponding scale into a conventional skip connection method for the U-Net structure, which will not be elaborated in this scheme.

[0059] In the current embodiment, the structure diagram of the decoder is as follows: Figure 4 As shown, the decoded structural response operator includes a direction index, which determines the convolution direction of the corresponding decoded structural response operator. Multiple decoded structural response features are aggregated using the second norm in the direction dimension to obtain the decoded structural gradient field. The decoded structural gradient field represents the structural change information at each spatial location at the current scale.

[0060] Specifically, the formula for obtaining the response features of the decoded structure is expressed as:

[0061]

[0062] in, For the first The decoded features output by each decoder, where x is any spatial location x in the decoded features. For the first The decoding structure response operator with direction index d in each decoder, where D is the number of directions. For the first The decoding structural response features of the decoder output at spatial position x and direction d.

[0063] Specifically, the construction and adaptive update methods of the decoding structure response operator are the same as those of the encoding structure response operator.

[0064] Specifically, the formula for L2-norm aggregation of multiple decoded structural response features in the direction dimension is expressed as:

[0065]

[0066] in, For the decoded structure gradient field at spatial location x, For the first The decoded structural response features of the decoded features output by each decoder at spatial position x and direction d, where D is the number of directions. It is a numerically stable term, that is to say, This represents the intensity of structural change at the current spatial location x within the current scale of the receptive field.

[0067] In the current embodiment, a 1×1 convolution is used to perform channel mapping on the decoded structure gradient field, and a Sigmoid activation function is used to compress the channel projection results to the [0,1] interval to obtain the gating factor map, as expressed by the formula:

[0068]

[0069] in, For the first The gating factor of the decoded features output by each decoder at the current spatial location x. For 1×1 convolution, Let σ represent the decoded structure gradient field at spatial location x, and let σ denote the Sigmoid activation function.

[0070] In other words, the gating factor map is composed of the gating factors at each spatial location of the decoded feature.

[0071] Specifically, using 1×1 convolution to map the decoded structure gradient field through channels can ensure that the gating is learnable, lightweight, and numerically bounded. Furthermore, the structural strength can be recalibrated in a trainable manner along the channel dimension, allowing the contribution of different scales and channels to the gating to be adaptively adjusted, thus better matching the low-contrast and morphologically variable boundary distribution of endometrial cancer.

[0072] In the current embodiment, the formula for gating and modulating the decoded features is expressed as follows:

[0073]

[0074] in, For the first The gated result of the decoded features output by each decoder at the current position x. For the first The decoded features output by each decoder at spatial location x. For the first The gating factor of the decoded features output by each decoder at the current position x. G is the gain intensity coefficient. As G increases, the significant regions of the structure are amplified by a bounded method, making the boundary transition zone and the weak contour more difficult to smooth in the step-by-step upsampling.

[0075] Specifically, this scheme uses a gain-based modulation method instead of directly multiplying by G to avoid excessive suppression of weaker structural regions and maintain the stability of the semantic backbone.

[0076] In the current embodiment, since amplifying the salient structural region solely through gating modulation may still be insufficient to express the contour orientation and local morphological differences, this scheme concatenates the gating result with multiple decoded structural response features along the channel dimension to obtain the decoder output, expressed by the formula:

[0077]

[0078] in, For decoder output, It is a non-linear activation. This indicates splicing along the channel dimension. For 1×1 convolution, For the first The gated result of the decoded features output by each decoder at the current position x. For the first The decoding structural response features of the decoder output at spatial position x and direction D.

[0079] Specifically, in the process of concatenating the gating results with the response features of multiple decoding structures along the channel dimension, gating is responsible for controlling "where to pay more attention" (spatial selective enhancement), directional branches provide "what kind of change pattern should be emphasized" (contour direction / edge change), and fusion compression unifies the two into a compact reconstructed representation, making the decoding output more boundary discriminative while maintaining semantic consistency, and continuously suppressing the interference of background texture on contour localization during the hierarchical reconstruction process. Among them, 1×1 convolution is used for fusion and compression, thereby avoiding channel expansion and completing information alignment at the same time.

[0080] Specifically, the gated signal in the decoder is not a static edge map given by external priors or fixed edge operators, but an internal structural strength representation generated self-consistently by the fused reconstructed features. Therefore, its distribution can adaptively match the structural features of lesions of different scales and morphologies as training progresses. By explicitly using this structural strength for gain modulation and orientation fusion, the decoding process can selectively emphasize areas with significant structural changes in space, while suppressing the interference of flat backgrounds and irrelevant textures on contour recovery, making weak boundaries less likely to be smoothed or submerged during progressive upsampling and convolutional reconstruction.

[0081] In the current embodiment, the gated structure and the response features of multiple decoding structures are concatenated along the channel dimension to obtain a second fused feature, and the second fused feature is residually connected with the input features of the current decoder to obtain the decoder output.

[0082] In the current embodiment, the gradient field of the coding structure output by each encoder is upsampled to a first resolution size, and each coding structure gradient field is compressed and mapped to a single-channel image to obtain a corresponding saliency map. All saliency maps are fused to obtain a global saliency field. A pixel-level weight map is constructed based on the global saliency field, and the segmentation error is weighted by the pixel-level weight map to obtain the final loss function. The pixel-level weight map represents the weight of each pixel. In the pixel-level weight map, cancerous regions have high weights and flat regions have low weights. The size of the first resolution is the same as the size of the CT image input to the endometrial cancer CT image segmentation architecture.

[0083] Specifically, a schematic diagram of constructing a pixel-level weight map is shown below. Figure 5 As shown, in order to fuse structural information of different scales in the same spatial dimension, this scheme upsamples the coded structure gradient field output by each encoder to the first resolution size, thereby avoiding positional misalignment caused by scale differences.

[0084] Specifically, the formula for obtaining the saliency plot is:

[0085]

[0086] in, In order to be with the first The saliency maps corresponding to each encoder, where x represents any spatial location. This represents the Sigmoid activation function. This indicates that the single-channel structural strength is obtained by averaging the channel dimensions, where k is the compression slope coefficient. Indicates upsampling, Indicates the first The gradient field of the encoded structure output by each encoder.

[0087] Specifically, the formula for obtaining the global saliency field is expressed as:

[0088]

[0089] in, Let L be the global saliency field, and L be the number of encoders. For encoder index, In order to be with the first The saliency map corresponding to each encoder.

[0090] Specifically, a pixel-level weight map is constructed using weight enhancement coefficients, expressed by the following formula:

[0091]

[0092] in, This is a pixel-level weighted map, where x represents any spatial location, i.e., the pixel position. This is the weight enhancement coefficient. It is a globally salient field.

[0093] Specifically, the final loss function is expressed as follows:

[0094]

[0095] in, For segmentation error, This is the predicted output at spatial location x. Label the truth value at spatial location x. This is a pixel-level weighted map.

[0096] Specifically, the segmentation error in this scheme can be constructed using cross-entropy, Dice, or any combination of their loss functions. Structural saliency fields directly generated within the network During training, more updates are naturally allocated to structurally sensitive regions such as boundary transition zones, small lesions and their neighborhoods, thus maintaining a smaller weight for large flat backgrounds. Furthermore, this weight comes from the fusion result of the gradient fields of multi-scale internal structures, which can simultaneously take into account the common indication of structural location by shallow details and deep semantics, making the error allocation closer to the true distribution of difficult-to-distinguish regions, thereby improving boundary stability and the ability to detect small lesions, and ensuring that the internal geometric quantities formed in the encoding stage are consistently utilized in the optimization stage.

[0097] Furthermore, when training the endometrial cancer CT image segmentation architecture using the training dataset, the optimal model parameters are retained when the final loss function meets the set threshold or the number of iterations is reached to complete the training and obtain the endometrial cancer CT image segmentation model.

[0098] Specifically, when the model training is completed, the trained endometrial cancer CT image segmentation model is validated using a validation set. The model's performance is evaluated by region overlap accuracy, boundary alignment, and lesion detection capability. Region overlap accuracy is evaluated by calculating the Dice coefficient and intersection-over-union ratio (IoU). Boundary alignment is measured by the 95% Hausdorff distance (HD95) as the main indicator, which measures the maximum local deviation between the segmented contour and the true boundary. Lesion detection capability is evaluated by indicators such as accuracy and recall to assess the model's overall lesion recognition efficiency.

[0099] Alternatively, based on the evaluation results on the validation set, the network structure parameters, training hyperparameters, and weight coefficients related to the internal structural gradient field can be iteratively adjusted and optimized. This process is repeated until the model achieves stable and superior overall performance on the validation set. Finally, a one-time final performance report is generated on the test set to confirm that the obtained model configuration meets the clinical practical needs of automatic segmentation of CT images of endometrial cancer.

[0100] In this scheme, the segmentation architecture for endometrial cancer CT images can more effectively transform regions with significant structural changes into continuous and clear contour representations at each level of high-resolution restoration, reducing boundary breaks and morphological shifts. Furthermore, it continuously protects small-volume lesions and their surrounding details during the progressive upsampling process, thereby achieving more stable and refined segmentation results even in the complex boundaries and low-contrast scenarios of endometrial cancer CT. Consequently, the coded structural gradient field generated by the coding unit and propagated with the feature flow is consistently reused in the three stages of coding representation enhancement, decoding reconstruction regulation, and training error redistribution, forming an end-to-end closed-loop constraint on the internal geometric quantities.

[0101] To verify the effectiveness of this scheme, ablation experiments were conducted on different combinations of technical features proposed in this invention under the same dataset and consistent training conditions to analyze the contribution of each component module to the segmentation performance. The results of the ablation experiments are shown in Table 1.

[0102] Table 1 Ablation Experiment Results

[0103]

[0104] Under the complete model conditions, the model achieves optimal performance on key metrics such as Dice, IoU, HD95, and Recall (Dice=0.8743, IoU=0.7767, HD95=3.7725, Recall=0.9025), indicating that the construction of the coding structure gradient field at the encoder end, the gradient gating modulation at the decoder end, and the weighting of the loss function based on the pixel-level weight map can form an effective synergy: on the one hand, it improves the overall region overlap accuracy, and on the other hand, it achieves smaller deviations in the boundary distance metric, while maintaining a high lesion detection capability.

[0105] In group G1, where the weighting of the loss function by the pixel-level weight map was removed, all model metrics decreased (Dice=0.8642, IoU=0.7612, HD95=3.9291, Recall=0.8675). Compared to G0, Dice decreased by 0.0101, IoU decreased by 0.0155, HD95 increased by 0.1566, and Recall decreased by 0.0350. This indicates that without structural gradient-guided adaptive error redistribution, the optimization effort for boundary transition zones and small-volume lesion regions during training was insufficient, leading to increased boundary localization errors, decreased overall segmentation consistency, and impacted lesion detection rates. This result validates the positive effect of pixel-level weight map weighting of the loss function on error focusing in structurally sensitive regions and improving training stability during the optimization phase.

[0106] In group G2, the construction of the coding structure gradient field and the weighting of the loss function by the pixel-level weight map were removed, but the gated modulation at the decoder was retained, resulting in a further decrease in model performance (Dice=0.8612, IoU=0.7563, HD95=4.0476, Recall=0.8724). Compared to G0, Dice decreased by 0.0131, IoU decreased by 0.0204, HD95 increased by 0.2751, and Recall decreased by 0.0301, indicating that the construction and propagation of the internal structure gradient field at the encoder has a fundamental contribution to the overall performance: when the multi-scale internal gradient geometry information provided by the construction of the coding structure gradient field is lacking, even if the decoder still has gated reconstruction capabilities, it is difficult to obtain the same boundary stability and region overlap accuracy as the complete model. It is worth noting that G2 shows a slight increase in Recall compared to G1 (0.8724 vs 0.8675), but at the same time, HD95 increases further. This reflects that when there is a lack of coding structure gradient field construction and error redistribution mechanism, the model is more likely to produce the phenomenon of "detection but unstable boundary", that is, some lesion areas can still be covered, but the contour deviation and boundary continuity are more significantly affected.

[0107] In group G3, removing only the gated modulation module in the decoding stage resulted in the most significant degradation across all metrics (Dice=0.8393, IoU=0.7234, HD95=4.4020, Recall=0.8523). Compared to G0, Dice decreased by 0.0350, IoU decreased by 0.0533, HD95 increased by 0.6295, and Recall decreased by 0.0502. This indicates that gated modulation plays a crucial role in transforming the structural gradient information at the encoding stage into an effective boundary representation in the high-resolution reconstruction stage. When the gradient-gated reconstruction and structural response fusion mechanism is lacking, the decoding path struggles to continuously enhance the contour transition band during upsampling, leading to a significant increase in boundary error and a decline in lesion detection capability.

[0108] The experimental data above demonstrates that the various technical features of this invention have a clear division of labor and synergistic relationship in function: the coding structure gradient field is responsible for explicitly constructing and transmitting the internal structure gradient field during the encoding stage, providing a unified geometric basis for subsequent reconstruction and optimization; the gated modulation module utilizes this geometric basis during the decoding stage to perform position-related reconstruction modulation and edge information fusion, thereby ensuring boundary continuity and fine-grained structure recovery; the pixel-level weight map further adaptively reweights the errors in structurally significant regions during training, enhancing the learning efficiency of boundary transition zones and small-volume lesions. The combined effect of these three features enables the endometrial cancer CT image segmentation model to simultaneously achieve higher region overlap accuracy, lower boundary distance error, and stronger lesion detection capability in endometrial cancer CT image segmentation tasks, showcasing the overall technical advantages and effectiveness of the solution.

[0109] In this study, the endometrial cancer CT image segmentation model was compared with current classic and cutting-edge endometrial cancer CT image segmentation networks, including Unet (2015), Unet++ (2018), TransUnet (2021), MFE-Unet (2024) and AFC-Unet (2025). The results of the comparison experiment are shown in Table 2.

[0110] Table 2 Comparison of experimental results

[0111]

[0112] Comparison of experimental figures as follows Figure 6 As shown in the experimental results, the endometrial cancer CT image segmentation model outperforms other methods in multiple evaluation metrics, including segmentation accuracy, boundary restoration, and lesion identification, demonstrating its significant advantages under complex imaging conditions. Specifically, the endometrial cancer CT image segmentation model achieves a Dice coefficient of 0.8743, a significant improvement compared to Unet's 0.8281, Unet++'s 0.8359, and TransUnet's 0.7609. It also surpasses MFE-Unet's 0.8492 and AFC-Unet's 0.8499, indicating its superior overall segmentation accuracy and refined lesion region restoration.

[0113] In terms of the intersection-over-union (IoU) ratio, the segmentation model for CT images of endometrial cancer achieved 0.7767, which is higher than Unet (0.6703), Unet++ (0.7182) and TransUnet (0.6156), and better than MFE-Unet (0.7382) and AFC-Unet (0.7394). This indicates that the model significantly improves the spatial overlap between the predicted region and the actual annotation, and can more accurately capture the complex morphology and boundary structure of the lesion.

[0114] The HD95 boundary accuracy evaluation index showed that the segmentation model of endometrial cancer CT image was 3.7725, which was lower than Unet (4.5214), Unet++ (4.4582) and TransUnet (5.7011), but better than MFE-Unet (4.6588) and AFC-Unet (4.1545). This indicates that the method has significant advantages in fine-grained boundary reconstruction and boundary continuity maintenance, and can effectively cope with the challenges of blurred and complex morphology of endometrial cancer lesions.

[0115] In terms of recall, the CT image segmentation model for endometrial cancer achieved a score of 0.9025, which is higher than Unet (0.8218), Unet++ (0.8191) and TransUnet (0.7781), and also better than MFE-Unet (0.8490) and AFC-Unet (0.8378). This indicates that the model performs well in detecting small-volume and low-contrast lesions and can effectively reduce the risk of missed detection.

[0116] The endometrial cancer CT image segmentation model achieves fine segmentation of lesions in endometrial cancer CT images through the synergistic effect of internal structural gradient modeling, gradient-gated high-resolution reconstruction, and weighted loss function based on the internal structural gradient field and pixel-level weight map. The encoded structural gradient field explicitly generates and propagates multi-scale internal structural gradient fields in the intermediate feature domain using learnable multi-directional first-order structural responses, preserving the intensity of local structural changes as a computable and traceable geometric quantity along with the feature flow. Gated modulation consistently calculates the internal directional gradient response and its magnitude representation based on the reconstructed features at the current scale, mapping the magnitude representation to a spatial gating factor to perform pixel-by-pixel gain modulation on the reconstructed features, while fusing directional gradient information to stably recover boundary transition zones and weak contours. The pixel-level weight map uses the internal structural gradient field to generate pixel-level saliency weights for weighted loss function, adaptively allocating errors to structurally sensitive regions. Through the above mechanism, the CT image segmentation model for endometrial cancer outperforms existing methods in terms of Dice, IoU, HD95, and Recall, effectively improving boundary clarity and the ability to detect small lesions, and providing reliable technical support for automatic clinical segmentation.

[0117] Example 2

[0118] Based on the same concept, this application also proposes an application method for a segmentation model of CT images of endometrial cancer, including:

[0119] Acquire the CT image of the endometrium to be segmented, and input the CT image of the endometrium to be segmented into the endometrial cancer CT image segmentation model trained in Example 1 to obtain the segmentation result of the cancerous region.

[0120] Example 3

[0121] This embodiment also provides an electronic device, see reference. Figure 7 It includes a memory 404 and a processor 402, the memory 404 storing a computer program and the processor 402 being configured to run the computer program to perform the steps in any of the above method embodiments.

[0122] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0123] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0124] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0125] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the methods for constructing a CT image segmentation model for endometrial cancer in the above embodiments.

[0126] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0127] Transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0128] The input / output device 408 is used to input or output information. In this embodiment, the input information may be multiple endometrial CT images with cancerous areas marked, and the output information may be the optimal parameters of the endometrial cancer CT image segmentation architecture, segmentation results, etc.

[0129] Optionally, in this embodiment, the processor 402 can be configured to perform the following steps via a computer program:

[0130] Multiple endometrial CT images with cancerous areas marked were obtained as a training dataset.

[0131] A segmentation architecture for CT images of endometrial cancer is constructed, wherein the segmentation architecture for CT images of endometrial cancer is a U-Net structure including an encoding unit, a bottleneck layer and a decoding unit;

[0132] Each encoder in the coding unit includes a coding layer and a coding internal geometry field. The coding internal geometry field constructs coding structure response operators corresponding to different directions to perform group convolution on each channel of the coding features to obtain multiple coding structure response features. The multiple coding structure response features are aggregated by the second norm to obtain a coding structure gradient field. The coding structure gradient field is fused with the initial coding features to obtain the encoder output. The coding features are the output features of the coding layer.

[0133] The bottleneck layer has the same structure as the encoder. The output of the bottleneck layer is fused with the output of the encoder at the corresponding scale and then input into the decoding unit.

[0134] Each decoder in the decoding unit includes a decoding layer and a decoding internal geometry field. The decoding internal geometry field constructs decoding structure response operators corresponding to different directions, which convolve the decoding features to obtain multiple decoding structure response features. The multiple decoding structure response features are aggregated using the second norm to obtain a decoding structure gradient field. The decoding structure gradient field is mapped to a gating factor map. The decoding features are gated and modulated using the gating factor map to obtain a gating result. The gating result is concatenated with the multiple decoding structure response features along the channel dimension to obtain the decoder output. The output of the last decoder in the decoding unit is the segmentation result, and the decoding features are the output features of the decoding layer.

[0135] The endometrial cancer CT image segmentation model was obtained by training the training dataset on the endometrial cancer CT image segmentation architecture.

[0136] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0137] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0138] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 7Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0139] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing a CT image segmentation model for endometrial cancer, characterized in that, Includes the following steps: Multiple endometrial CT images with cancerous areas marked were obtained as a training dataset. A segmentation architecture for CT images of endometrial cancer is constructed, wherein the segmentation architecture for CT images of endometrial cancer is a U-Net structure including an encoding unit, a bottleneck layer and a decoding unit; Each encoder in the coding unit includes a coding layer and a coding internal geometry field. The coding internal geometry field constructs coding structure response operators corresponding to different directions to perform group convolution on each channel of the coding features to obtain multiple coding structure response features. The multiple coding structure response features are aggregated by the second norm to obtain a coding structure gradient field. The coding structure gradient field is fused with the initial coding features to obtain the encoder output. The coding features are the output features of the coding layer. The bottleneck layer has the same structure as the encoder. The output of the bottleneck layer is fused with the output of the encoder at the corresponding scale and then input into the decoding unit. Each decoder in the decoding unit includes a decoding layer and a decoding internal geometry field. The decoding internal geometry field constructs decoding structure response operators corresponding to different directions, which convolve the decoding features to obtain multiple decoding structure response features. The multiple decoding structure response features are aggregated using the second norm to obtain a decoding structure gradient field. The decoding structure gradient field is mapped to a gating factor map. The decoding features are gated and modulated using the gating factor map to obtain a gating result. The gating result is concatenated with the multiple decoding structure response features along the channel dimension to obtain the decoder output. The output of the last decoder in the decoding unit is the segmentation result, and the decoding features are the output features of the decoding layer. The endometrial cancer CT image segmentation model was obtained by training the training dataset on the endometrial cancer CT image segmentation architecture.

2. The method for constructing a CT image segmentation model for endometrial cancer according to claim 1, characterized in that, The coding structure response operator includes a direction index and a channel index. The direction index determines the convolution direction of the corresponding coding structure response operator, and the channel index is used to guide the coding structure response operator to convolve the corresponding channel of the coding feature.

3. The method for constructing a CT image segmentation model for endometrial cancer according to claim 1, characterized in that, Before training the segmentation architecture for CT images of endometrial cancer, the coding structure response operator in each encoder is initialized in the form of a first-order difference, and each of the coding structure response operators is adaptively updated during the training process.

4. The method for constructing a CT image segmentation model for endometrial cancer according to claim 1, characterized in that, Multiple coding structure response features are aggregated using the second norm in the directional dimension to obtain structural gradient features corresponding to different channels. The structural gradient features under each channel are integrated to obtain the coding structure gradient field, which represents the structural change information of coding features under different channels at the current scale.

5. The method for constructing a CT image segmentation model for endometrial cancer according to claim 1, characterized in that, The gradient field of the coding structure is fused with the initial coding features to obtain the first fused feature. The first fused feature is then residually concatenated with the input features of the current encoder to obtain the encoder output.

6. The method for constructing a CT image segmentation model for endometrial cancer according to claim 1, characterized in that, The decoded structural response operator includes a direction index, which determines the convolution direction of the corresponding decoded structural response operator. Multiple decoded structural response features are aggregated using the second norm in the direction dimension to obtain the decoded structural gradient field, which represents the structural change information at each spatial location at the current scale.

7. The method for constructing a CT image segmentation model for endometrial cancer according to claim 1, characterized in that, The formula for gating and modulating the decoded features is expressed as follows: in, For the first The gated result of the decoded features output by each decoder at the current position x. For the first The decoded features output by each decoder at spatial location x. For the first The gating factor of the decoded features output by each decoder at the current position x. This is the gain intensity coefficient.

8. The method for constructing a CT image segmentation model for endometrial cancer according to claim 1, characterized in that, The gradient field of the coding structure output by each encoder is upsampled to the size of the first resolution. Each coding structure gradient field is compressed and mapped to a single-channel image to obtain the corresponding saliency map. All saliency maps are fused to obtain the global saliency field. A pixel-level weight map is constructed based on the global saliency field. The segmentation error is weighted by the pixel-level weight map to obtain the final loss function. The pixel-level weight map represents the weight of each pixel. In the pixel-level weight map, cancerous areas have high weights and flat areas have low weights. The size of the first resolution is the same as the size of the CT image input to the endometrial cancer CT image segmentation architecture.

9. A method for applying a CT image segmentation model for endometrial cancer, characterized in that, include: Obtain the CT image of the endometrium to be segmented, and input the CT image of the endometrium to be segmented into the endometrial cancer CT image segmentation model trained by any one of the methods described in claims 1-8 to obtain the segmentation result of the cancerous region.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform a method for constructing a CT image segmentation model for endometrial cancer as described in any one of claims 1-8.

Images