Medical image segmentation method based on functional query to improve cross-center generalization performance

Abstract

This invention proposes a medical image segmentation method to improve cross-center generalization performance, addressing issues such as edge blurring, false positive prediction, and sample discrepancies in multi-center medical image segmentation. The method employs an encoder-decoder and query-guided dual-path deformer-decoder structure, enhancing functionality through semantically driven queries. First, at the encoder output, a coarse masking prediction network generates semantic query hints for samples. An additional cross-attention layer fuses the semantic query and sample hints, adding edge enhancement and false positive queries for each semantic category. By generating semantic edge labels and dynamic false positive labels, combined with masking attention and cross-attention mechanisms, efficient prediction of edge enhancement and false positive suppression is achieved. Inspired by knowledge distillation and multi-granularity segmentation, this invention enhances the adaptability of queries to sample discrepancies, explores various query optimization methods, and significantly improves the cross-center performance of medical image segmentation models.

Description

Technical Field

[0001] This application belongs to the field of image processing technology, and relates to a medical image segmentation method and system, specifically a medical image segmentation method and system based on functional queries to improve cross-center generalization performance. Background Technology

[0002] Over the past few decades, thanks to the development of deep learning, especially deep convolutional neural networks (CNNs), medical image segmentation has made tremendous progress. The U-Net network, initially used for medical image segmentation, introduced skip connections, effectively fusing shallow texture information and deep semantic information. The U-Net structure takes into account the data distribution and structural characteristics of medical images, achieving good segmentation results in most cases, and has gained widespread use and recognition.

[0003] Since the advent of DETR, query-based deformer decoders have driven the unification of development across multiple computer vision tasks due to their powerful relational and attribute modeling capabilities. Queries, as learnable multidimensional vectors, effectively capture various prior information from training data. Specifically, the interaction between these queries and image representations in the deformer decoder enables dynamic adaptation to each sample, even in unfamiliar domains, which aligns perfectly with the dynamic network requirements of DG. During training, the parameters in each query serve as implicit and flexible prototypes corresponding to fixed semantics, capturing knowledge learned from the source domain. During deployment, their interaction with individual samples ensures knowledge preservation and dynamic adjustment.

[0004] The collection, annotation, and organization of medical images often have multi-center characteristics. Differences between centers stem from differences in equipment, annotators' preferences, and other factors. These differences have fostered cross-center research on medical image generalization segmentation. Researchers are dedicated to exploring various methods to mitigate the performance degradation caused by data differences during the training and model deployment processes.

[0005] Recent studies have found that individual differences exist between samples, leading to fluctuations in model performance. Furthermore, segmentation results for cross-center medical images often exhibit low edge discrimination and frequent false positives. For example, some methods address these issues by adding extra small convolutional networks that predict domain label regression data for the image and dynamically output a small number of convolutional parameters. During deployment, these parameters are better adapted to the target sample through domain regression analysis. Other methods add additional supervision information to edges or enhance edge-related information in the image representation. Still others add multiple semantic categories to the segmentation process to mitigate the performance impact of false positives.

[0006] Inspired by previous work on query-based deformer decoders in cross-center medical image generalization, this invention starts from the same network design to solve or mitigate the three common problems of cross-center generalization (false positives, blurred edges, and large sample variability). Based on the query design, it improves the cross-center medical image segmentation performance of deformer decoders by using methods such as cue learning, supervised learning, and attention mechanisms. Summary of the Invention

[0007] The purpose of this invention is to provide a medical image segmentation method and system based on functional queries to improve cross-center generalization performance. This method can effectively improve problems such as false positives, blurred edges, and large sample variability in domain generalization studies, and can significantly improve the generalization ability of the network.

[0008] The technical solution adopted by the method of the present invention is: a medical image segmentation method based on functional query to improve cross-center generalization performance, which uses a query-guided medical image cross-center generalization segmentation network to segment medical images;

[0009] The query-guided medical image cross-center generalized segmentation network includes an encoder-decoder structure based on a deep residual network, a query-based deformer decoder module, and a segmentation mask prediction module.

[0010] The encoder-decoder structure based on a deep residual network consists of an encoder composed of several residual blocks, each followed by a pooling layer for spatial downsampling. The decoder uses skip connections to pass the context information captured by the encoder to the corresponding decoder layer to assist in the recovery of high-resolution feature maps. Each layer of the decoder consists of upsampling operations and residual blocks to gradually recover image details.

[0011] The query-based deformer decoder module includes a query hint enhancement module and a deformer decoder module;

[0012] The query suggestion enhancement module includes a query suggestion acquisition module and a suggestion cross-attention module. The query suggestion acquisition module first uses a network composed of two convolutional layers as a coarse mask segmentation head network to obtain coarse mask prediction. It then obtains image features corresponding to different semantics through dot product of the coarse mask and encoder features. Finally, it maps the image features to query suggestion information corresponding to specific semantics through global pooling and a fully connected layer. The suggestion cross-attention module is a cross-attention layer.

[0013] The deformer decoder module includes a semantically quantity-dependent learnable embedding vector, a first deformer block, and a second deformer block. The semantically quantity-dependent learnable embedding vector is used to expand the query to the number of samples. The first deformer block is a single-path deformer decoder design, including a self-attention layer, a cross-attention layer for updating the image query, and a fully connected layer. The second deformer block is a dual-path deformer decoder design, which includes an additional cross-attention layer for updating the query image after the fully connected layer. Each layer of the first and second deformer blocks has a residual design with direct input-output connection and a normalization operation for the output. The second deformer block receives the updated image representation and query from the first deformer block.

[0014] The mask prediction module includes a projection network consisting of two fully connected layers connected by an activation layer, which is added after removing the traditional segmentation head network. This projection network projects the query output by the query-based deformer decoder module and the image representation output by the encoder-decoder structure into the same dimensional space. The image representation output by the encoder-decoder structure and the projected query are multiplied by a dot product, and then a sigmoid function is performed to obtain a semantic mask prediction based on a threshold.

[0015] Preferably, the encoder-decoder structure built based on the deep residual network is a pre-trained network; during training, the loss function used consists of the binarized mask prediction dice_loss and the pixel-level cross_entropy_loss;

[0016]

[0017]

[0018] Where, p i Let g represent the predicted probability of the i-th pixel. i This represents the true label of the i-th pixel, and N represents the total number of pixels in the image.

[0019] Preferably, the query-based deformer decoder module is a trained network; during training, the loss function used consists of the binarized mask prediction dice_loss and the pixel-level cross_entropy_loss.

[0020]

[0021]

[0022] Where, p i Let g represent the predicted probability of the i-th pixel. iThis represents the true label of the i-th pixel, and N represents the total number of pixels in the image.

[0023] Preferably, the mask prediction module is a pre-trained network; during training, the loss function used consists of the binarized mask prediction die_loss and the pixel-level cross_entropy_loss.

[0024]

[0025]

[0026] Where, p i Let g represent the predicted probability of the i-th pixel. i This represents the true label of the i-th pixel, and N represents the total number of pixels in the image.

[0027] Preferably, the query-guided medical image cross-center generalization segmentation network is a pre-trained network;

[0028] During training, morphological operations and Gaussian filtering are performed on the binary mask of the mask annotation data provided in the segmentation training dataset to obtain additional edge supervision information. Additional semantic edge enhancement queries are added for each semantic class. Edge prediction is obtained after pixel-level calculation of the edge query and image representation output. The edge prediction is supervised based on the edge supervision information.

[0029] The mask prediction after each forward propagation is binarized and superimposed with the binary mask corresponding to the label as a false positive suppression label. This label changes dynamically during the training process. At the same time, a corresponding false positive suppression query is added for each semantic class. The query and image representation output false positive prediction. The false positive prediction is supervised based on the dynamic false positive suppression label.

[0030] When the two work together, each type of language corresponds to three types of queries. The semantic mask prediction in the forward propagation process (which is also the dynamic supervision information corresponding to the mask prediction in the false positive prediction label) is generated by superimposing the mask corresponding to the edge reinforcement query and the original semantic query.

[0031] This invention also provides a medical image segmentation system based on functional queries to improve cross-center generalization performance, comprising:

[0032] One or more processors;

[0033] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the query-guided cross-center generalized segmentation method for medical images.

[0034] The present invention also provides a non-volatile computer-readable storage medium containing a computer program that, when executed by one or more processors, causes the processors to perform the medical image segmentation method based on functional queries to improve cross-center generalization performance.

[0035] The present invention also provides a computer program product, including computer program instructions, which, when executed on a computer, cause the computer to execute the medical image segmentation method based on functional queries to improve cross-center generalization performance.

[0036] Compared with the prior art, the beneficial effects of the present invention include:

[0037] (1) When segmenting medical images across centers, individual distribution differences are obvious. The original cross-attention mechanism is insufficient to make the knowledge in the query correspond fully with the current image. This invention is inspired by cue learning. At the encoder output, a coarse mask is predicted based on image features. The coarse semantic region average pooling features are obtained based on the coarse mask and image features as query cue for each image. Before the query input is based on the query deformer decoder, an additional cross-attention layer is added to mix semantic query and query cue.

[0038] (2) The edges of cross-center medical image segmentation masks are often unclear. To enhance the mask edges, and since the edge information of similar semantics in medical images is similar, this invention generates edge supervision corresponding to each semantic class based on the original supervision information, and adds additional edge enhancement queries for each semantic class, using edge labels to supervise edge enhancement prediction. The final semantic mask is composed of the mask obtained from the original semantic query and the edge enhancement semantic query.

[0039] (3) Cross-center medical image segmentation masks often produce false positives. To suppress false positives, inspired by the query-based multi-granularity interactive basic segmentation model "segment everything," this invention adds an additional query to predict information containing certain false positives, making the original semantic query prediction more accurate. Therefore, in each forward propagation process, false positive mask labels are generated based on semantic mask prediction and labels as the prediction target of the false positive query, making the prediction mask of the original semantic query more accurate.

[0040] (4) In order to enhance the functionality of the query and enable the self-attention mechanism to be used smoothly in the original query, edge query, and false positive query within each semantic class, and to enable the self-attention mechanism to be applied smoothly in the query interaction between different semantic classes, this invention removes a part of the cross attention layer, so that the query update process only interacts with the encoder output representation, and uses a mask attention mechanism in the self-attention layer of the query.

[0041] (5) The design around the query simultaneously addresses common problems in three types of medical image segmentation tasks: large sample variability, false positives, and unclear edges. Attached Figure Description

[0042] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0043] Figure 1 This is a diagram of the query-guided medical image cross-center generalized segmentation network structure proposed in this application embodiment. Detailed Implementation

[0044] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0045] Please see Figure 1 This embodiment provides a medical image segmentation method based on functional queries to improve cross-center generalization performance, which uses a query-guided medical image cross-center generalization segmentation network for medical image segmentation.

[0046] The query-guided medical image cross-center generalized segmentation network includes an encoder-decoder structure built based on the deep residual network ResNet-34, a query-based deformer decoder module, and a segmentation mask prediction module.

[0047] The encoder-decoder structure, built upon the deep residual network ResNet-34, consists of an encoder composed of multiple residual blocks. Each residual block is followed by a pooling layer for spatial downsampling, thereby extracting feature maps rich in semantic information. During decoding, the module utilizes skip connections to pass the contextual information captured by the encoder to the corresponding decoder layer, aiding in the recovery of high-resolution feature maps. Each layer of the decoder comprises upsampling operations and residual blocks, progressively restoring image details.

[0048] The query-based deformer decoder module includes a query hint enhancement module and a deformer decoder module based on multi-functional queries.

[0049] The query suggestion enhancement module includes a query suggestion acquisition module and a suggestion cross-attention module. The query suggestion acquisition module first uses a network composed of two convolutional layers as a coarse mask segmentation head network to obtain coarse mask prediction. It obtains image features corresponding to different semantics by multiplying the coarse mask and encoder features. Then, it maps the image features to query suggestion information corresponding to specific semantics through global pooling and fully connected layers. The suggestion cross-attention module is a cross-attention layer, where Q is the original semantic query and K and V are the previously obtained query suggestion information.

[0050] The deformer decoder module based on multi-functional queries contains semantically number-related learnable embedding vectors, each with a length of 512, corresponding to the channel size of the encoder's output image representation. Each semantic class contains three queries. Within the module, the semantically guided learnable embedding vectors are first expanded to include the number of samples. The module contains two deformer blocks: the first is a single-path deformer decoder design, containing a self-attention layer, a cross-attention layer for image update queries, and a fully connected layer; the second is a dual-path deformer decoder design, adding an additional cross-attention layer for query-updated images after the fully connected layer. Each layer has a residual design with direct input-output connection and a normalization operation for the output. The self-attention layer computes all queries of the sample as Q, K, V; the cross-attention layer for updating the image query uses the query as Q, the encoder output representation as K, V, and outputs the updated query. Furthermore, a masked attention design is used for interactions between different semantic queries, allowing the original semantic query to interact with edge enhancement and false positive suppression queries, while edge queries and false positive queries do not interact directly. Within semantics, false positive queries interact, but edge queries do not interact. The fully connected layer processes the updated query; the cross-attention layer for updating the image uses the query as K, V and the image representation as Q. The second deformer block accepts the updated image representation and query from the previous deformer block.

[0051] The mask prediction module involves removing the traditional segmentation head network and adding a projection network containing two fully connected layers connected by an activation layer. This projection network projects the query output from the deformer decoder and the image representation output from the original network's decoder into the same 256-dimensional space. The original network's decoder output image representation and the projected query are multiplied by a dot product, and then a sigmoid function is performed to obtain a semantic mask prediction based on a 0.5 threshold. The final actual mask prediction is obtained by superimposing the masks predicted by the edge enhancement query and the original semantic query.

[0052] In one implementation, the encoder-decoder structure built on the deep residual network ResNet-34 is a pre-trained network; during training, the loss function used consists of the binarized mask prediction dice_loss and the pixel-level cross_entropy_loss.

[0053]

[0054]

[0055] p i : Represents the predicted probability of the i-th pixel, that is, the probability that the model predicts that the pixel belongs to the foreground (usually represented as the target or object of interest). In the binary segmentation problem, the model outputs a probability value for each pixel, which can be determined as foreground (1) or background (0) after thresholding.

[0056] g i : Represents the true label of the i-th pixel, that is, the value of this pixel in the true segmentation mask. In binary segmentation problems, the true label is usually 1 or 0, where 1 indicates that the pixel belongs to the foreground and 0 indicates that it belongs to the background.

[0057] N represents the total number of pixels in the image. When calculating dice_loss and cross_entropy_loss, all pixels in the image are traversed for calculation, so N refers to the total number of pixels considered in the entire image.

[0058] In one implementation, the query-based deformer decoder module is a trained network; during training, the loss function used consists of binarized mask prediction dice_loss and pixel-level cross_entropy_loss.

[0059]

[0060]

[0061] p i : Represents the predicted probability of the i-th pixel, that is, the probability that the model predicts that the pixel belongs to the foreground (usually represented as the target or object of interest). In the binary segmentation problem, the model outputs a probability value for each pixel, which can be determined as foreground (1) or background (0) after thresholding.

[0062] g i: Represents the true label of the i-th pixel, that is, the value of this pixel in the true segmentation mask. In binary segmentation problems, the true label is usually 1 or 0, where 1 indicates that the pixel belongs to the foreground and 0 indicates that it belongs to the background.

[0063] N represents the total number of pixels in the image. When calculating dice_loss and cross_entropy_loss, all pixels in the image are traversed for calculation, so N refers to the total number of pixels considered in the entire image.

[0064] In one implementation, the mask prediction module is a trained network; during training, the loss function used consists of the binarized mask prediction die_loss and the pixel-level cross_entropy_loss.

[0065]

[0066]

[0067] p i : Represents the predicted probability of the i-th pixel, that is, the probability that the model predicts that the pixel belongs to the foreground (usually represented as the target or object of interest). In the binary segmentation problem, the model outputs a probability value for each pixel, which can be determined as foreground (1) or background (0) after thresholding.

[0068] g i : Represents the true label of the i-th pixel, that is, the value of this pixel in the true segmentation mask. In binary segmentation problems, the true label is usually 1 or 0, where 1 indicates that the pixel belongs to the foreground and 0 indicates that it belongs to the background.

[0069] N represents the total number of pixels in the image. When calculating dice_loss and cross_entropy_loss, all pixels in the image are traversed for calculation, so N refers to the total number of pixels considered in the entire image.

[0070] In one implementation, the query-guided medical image cross-center generalization segmentation network is a trained network;

[0071] During training, morphological operations and Gaussian filtering are performed on the binary mask of the mask annotation data provided in the segmentation training dataset to obtain additional edge supervision information. Additional semantic edge enhancement queries are added for each semantic class. Edge prediction is obtained after pixel-level calculation of the edge query and image representation output. The edge prediction is supervised based on the edge supervision information.

[0072] The mask prediction after each forward propagation is binarized and superimposed with the binary mask corresponding to the label as a false positive suppression label. This label changes dynamically during the training process. At the same time, a corresponding false positive suppression query is added for each semantic class. The query and image representation output false positive prediction. The false positive prediction is supervised based on the dynamic false positive suppression label.

[0073] When the two work together, each type of language corresponds to three types of queries. The semantic mask prediction in the forward propagation process (which is also the mask prediction in the false positive prediction label) is generated by superimposing the mask corresponding to the edge enhancement query and the original semantic query.

[0074] This embodiment also provides a medical image segmentation system based on functional queries to improve cross-center generalization performance, including:

[0075] One or more processors;

[0076] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the query-guided cross-center generalized segmentation method for medical images.

[0077] This embodiment also provides a non-volatile computer-readable storage medium containing a computer program that, when executed by one or more processors, causes the processors to perform the medical image segmentation method based on functional queries to improve cross-center generalization performance.

[0078] This embodiment also provides a computer program product, including computer program instructions, which, when executed on a computer, cause the computer to perform the medical image segmentation method based on functional queries to improve cross-center generalization performance.

[0079] The invention will be further illustrated below through specific experiments.

[0080] Compared to the baseline, this invention only increases memory usage by about 30% (from 5G to 7G on the prostate multicenter dataset and from 9G to 12G on the fundus optic cup and optic disc multicenter dataset). This is because this invention uses only a lightweight deformer decoder and eliminates the traditional segmentation head network. Furthermore, the training and inference processes of this invention are not significantly improved. Therefore, the increased time overhead and computational cost of the entire segmentation process are relatively small, making it well-suited for assisting clinicians in diagnosis.

[0081] In the experimental phase, this invention uses the Dice score (Dice), a commonly used evaluation metric in medical image segmentation, and obtains the generalization performance on the isolated target center by training the model on all other center datasets. This invention was validated on two common multi-center medical segmentation tasks. Experimental results demonstrate that this invention outperforms all other methods. Its excellent performance is attributed to the query-based deformer decoder and the auxiliary tasks designed for representation learning enhancement and generalization. Compared to the baseline, this invention achieves a Dice segmentation accuracy of 88.21% on the fundus optic cup / optic disc cross-center generalized segmentation dataset, a 3.01% improvement. On the prostate MRI modality multi-center generalized segmentation dataset, the invention achieves a Dice segmentation accuracy of 89.74%, a 2.66% improvement over the baseline. On the prostate multicenter dataset, data domain 1 (I2CVB) and data domain 3 (BIDMC) are more prone to false positive segmentation or difficulty in segmenting the target due to greater morphological differences. Compared with the baseline, the present invention significantly suppresses false positives. At the same time, since the final prediction result is based on the original semantic query and the edge enhancement query, the present invention significantly enhances the edge of the target and suppresses the difficulty in segmenting the target.

[0082] This invention achieves cross-center segmentation in the context of variable test data distribution and limited training data. Addressing the specific problems of edge blurring, false positive predictions, and discrepancies between samples often caused by the multi-center nature of multi-center medical image segmentation, it enhances the model's cross-center performance by incorporating additional functional enhancements around semantically driven queries in an encoder-decoder architecture with an embedded query-guided dual-path deformer decoder. First, semantic query hints are calculated for samples at the encoder output, and the semantic query and sample semantic query hints are fused. Semantic edge enhancement queries and false positive queries are added for each semantic category. Edge labels for each semantic category are obtained based on the original labels, and dynamic false positive labels are generated based on each prediction and label for each semantic category. Mask attention and cross-attention mechanisms are fully utilized, and the final output mask prediction is obtained based on both the original semantic and edge enhancement queries. This invention enhances the model's performance on weak segmentation edges when facing cross-center segmentation tasks; inspired by knowledge distillation and multi-granularity segmentation, this invention adds queries with false positive suppression capabilities; this invention adds an additional coarse mask prediction network and cross-attention layer to strengthen the original semantic query, making the sample differences of the query more obvious; this invention explores the possibility of multiple queries working together for each semantic class and considers various regularization methods such as mask attention mechanism.

[0083] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0084] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A medical image segmentation method based on functional queries to improve cross-center generalization performance, characterized in that: A query-guided cross-center generalized segmentation network for medical images is used for medical image segmentation. The query-guided medical image cross-center generalized segmentation network includes an encoder-decoder structure based on a deep residual network, a query-based deformer decoder module, and a segmentation mask prediction module. The encoder-decoder structure based on deep residual networks consists of an encoder composed of several residual blocks, and each residual block is followed by a pooling layer for spatial downsampling. The decoder uses skip connections to pass the context information captured by the encoder to the corresponding decoder layer to assist in the recovery of high-resolution feature maps. Each layer of the decoder consists of upsampling operations and residual blocks to gradually recover image details. The query-based deformer decoder module includes a query hint enhancement module and a deformer decoder module; The query suggestion enhancement module includes a query suggestion acquisition module and a suggestion cross-attention module; The query hint acquisition module first uses a network composed of two convolutional layers as a coarse mask segmentation head network to obtain coarse mask prediction. It obtains image features corresponding to different semantics by multiplying specific coarse masks and encoder features. Then, it maps the image features to query hint information corresponding to specific semantics through global pooling and fully connected layers. The hint cross-attention module is a cross-attention layer. The deformer decoder module includes a semantically quantity-related learnable embedding vector, a first deformer block, and a second deformer block. The semantically quantity-related learnable embedding vector is used to expand the query to the number of samples. The first deformer block is a single-path deformer decoder design, including a self-attention layer, a cross-attention layer for image update queries, and a fully connected layer. The second deformer block is a dual-path deformer decoder design, which includes an additional cross-attention layer for query update images after the fully connected layer. The first deformer block and the second deformer block each have a residual design with direct input to output and a normalization operation for the output; the second deformer block accepts the updated image representation and query from the first deformer block. The mask prediction module includes removing the traditional segmentation head network and adding a projection network containing two fully connected layers connected by an activation layer in the middle. This projection network projects the query output by the query-based deformer decoder module and the image representation output by the encoder-decoder structure into the same dimensional space. The image representation output by the encoder-decoder structure and the projected query are multiplied by a dot, and then sigmoid calculation is performed to obtain semantic mask prediction based on a threshold. The query-guided medical image cross-center generalization segmentation network is a pre-trained network; During training, morphological operations and Gaussian filtering are performed on the binary mask of the mask annotation data provided in the segmentation training dataset to obtain additional edge supervision information. Additional semantic edge enhancement queries are added for each semantic class. Edge prediction is obtained after pixel-level calculation of the edge query and image representation output. The edge prediction is supervised based on the edge supervision information. The mask prediction after each forward propagation is binarized and superimposed with the binary mask corresponding to the label as a false positive suppression label. This label changes dynamically during the training process. At the same time, a corresponding false positive suppression query is added for each semantic class. The query and image representation output false positive prediction. The false positive prediction is supervised based on the dynamic false positive suppression label. When the two work together, each type of language corresponds to three types of queries. The semantic mask prediction in the forward propagation process is generated by superimposing the masks corresponding to the edge-enhanced query and the original semantic query.

2. The medical image segmentation method based on functional queries to improve cross-center generalization performance according to claim 1, characterized in that: The encoder-decoder structure built based on the deep residual network is a trained network; During training, the loss function used is predicted by the binarized mask. and pixel level composition; ； ； in, This represents the predicted probability of the i-th pixel. Let N represent the true label of the i-th pixel, and N represent the total number of pixels in the image.

3. The medical image segmentation method based on functional queries to improve cross-center generalization performance according to claim 1, characterized in that: The query-based deformer decoder module is a trained network; During training, the loss function used is predicted by the binarized mask. and pixel level composition; ； ； in, This represents the predicted probability of the i-th pixel. Let N represent the true label of the i-th pixel, and N represent the total number of pixels in the image.

4. The medical image segmentation method based on functional queries to improve cross-center generalization performance according to claim 1, characterized in that: The mask prediction module is a trained network; During training, the loss function used is predicted by the binarized mask. and pixel level composition; ； ； in, This represents the predicted probability of the i-th pixel. Let N represent the true label of the i-th pixel, and N represent the total number of pixels in the image.

5. A medical image segmentation system based on functional queries to improve cross-center generalization performance, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the medical image segmentation method based on functional queries to improve cross-center generalization performance as described in any one of claims 1 to 4.

6. A non-volatile computer-readable storage medium containing a computer program, characterized in that: When the computer program is executed by one or more processors, the processors perform the medical image segmentation method based on functional queries to improve cross-center generalization performance, as described in any one of claims 1 to 4.

7. A computer program product comprising computer program instructions, characterized in that: When the computer program instructions are executed on a computer, the computer performs the medical image segmentation method based on functional queries to improve cross-center generalization performance as described in any one of claims 1 to 4.

Images