Medical image segmentation method and system based on semantic deviation suppression

Abstract

The invention discloses a medical image segmentation method and system based on semantic deviation suppression, and the method comprises the following steps: inputting two data input streams into two different coding and decoding networks based on two-dimensional convolution respectively, and outputting to obtain two different segmentation prediction distributions with segmentation truth values; respectively dividing the two segmented prediction distributions into a plurality of region blocks with the same size, and taking a difference value between the confidence coefficient of each region block and the confidence coefficient of the surrounding region block as a semantic deviation; performing deviation degree grade division on the semantic deviation to generate a semantic deviation grade; injecting a semantic deviation level into the to-be-segmented multi-organ medical image data, and performing boundary displacement suppression; the boundary displacement suppression comprises the following steps of: in segmentation prediction distribution, acquiring a highest semantic deviation region block and a mark number thereof according to a semantic deviation degree grade; selecting the lowest semantic deviation region block in the segmentation prediction distribution under the other intensity; and covering the highest semantic deviation region block with the lowest semantic deviation region block.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a medical image segmentation method system based on semantic deviation suppression. Background Art

[0002] Medical image segmentation, derived from computed tomography or magnetic resonance imaging, is crucial for a variety of clinical applications. Obtaining large medical datasets with precise annotations to train segmentation models is challenging because reliable annotations can only be provided by experts. This limits the development of medical image segmentation algorithms and poses substantial challenges to further research and implementation.

[0003] To reduce the burden of manual annotation and address these challenges, semi-supervised medical image segmentation is becoming a popular approach, encouraging segmentation models to learn from readily available unlabeled data with limited labeled data. Research in semi-supervised medical image segmentation has also employed consistency learning to ensure that the network model's decision boundary lies within a low-density boundary. By ensuring consistent features or predictions under different perturbations, this strategy has gradually become one of the most effective solutions for learning from unlabeled data. Depending on the perturbation, consistency learning methods can be divided into three categories: 1) Data perturbation, which primarily involves applying different forms of input to the same model. For example, the most common input perturbation for images is strong or weak data augmentation. 2) Model perturbation, which primarily involves applying the same input to different networks. 3) Task perturbation, which involves using various auxiliary tasks to improve the performance of the primary task, including but not limited to regression, lesion repair, reconstruction, and specific knowledge extraction.

[0004] A drawback of existing technologies is that most methods utilize only one of the aforementioned perturbations to ensure robust consistency learning. This limited use of perturbations limits the performance of consistency learning and creates fuzzy decision boundaries, as a specific single perturbation can only handle a limited number of cases. However, utilizing mixed or multiple perturbations can easily cause the consistency learning process to spiral out of control, leading to greater semantic deviations between models and limited learning quality. Summary of the Invention

[0005] In view of the above-mentioned defects of the prior art, the present invention provides a medical image segmentation method and system based on semantic deviation suppression, so as to provide more accurate boundary information results when performing image organ segmentation.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A medical image segmentation method based on semantic deviation suppression includes the following steps:

[0008] S1. Using the organ boundary information obtained from the multi-organ medical image data with segmentation truth values as a boundary prior constraint;

[0009] S2. Performing data augmentation processing of two different strengths on the multi-organ medical image data to be segmented to generate two different data input streams representing two different forms of the multi-organ medical image data; inputting the two data input streams into two different two-dimensional convolution-based encoding and decoding networks, respectively, to output two different segmentation prediction distributions with true segmentation values;

[0010] S3, dividing the two segmentation prediction distributions into a plurality of regions of equal size, and taking the difference between the confidence of each region and the confidence of the region within the preset domain as the semantic deviation; and classifying the semantic deviation into deviation levels to generate a semantic deviation level;

[0011] S4, injecting the semantic deviation level into the multi-organ medical image data to be segmented to suppress boundary displacement;

[0012] The boundary displacement suppression includes:

[0013] In one of the segmentation prediction distributions, the highest semantic deviation region block and its label are obtained according to the semantic deviation degree level; the lowest semantic deviation region block is selected in the segmentation prediction distribution at another strength; and the highest semantic deviation region block is covered with the lowest semantic deviation region block.

[0014] Preferably, in step S2, the data enhancement processing includes weak data enhancement and strong data enhancement; the weak data enhancement includes random rotation, flipping and scaling; and the strong enhancement includes random occlusion or color jittering based on the weak data enhancement.

[0015] Preferably, in step S4, the expression for exchanging the region blocks corresponding to the labels is:

[0016]

[0017]

[0018] in, is the label of the lowest semantic deviation area block under weak enhancement, represents the label of the lowest semantic deviation region block under strong enhancement, S represents the strong enhancement, W represents the weak enhancement, Z represents the label of the semantic region block; X represents the multi-organ medical image data, The multi-organ medical image data region block corresponding to the label of the region block with the highest semantic deviation under strong enhancement, The multi-organ medical image data region block corresponding to the label of the region block under weak enhancement, The multi-organ medical image data region block corresponding to the label of the region block with the highest semantic deviation under weak enhancement, The multi-organ medical image data region block corresponding to the label of the region block under strong enhancement is represented.

[0019] In the second aspect, a medical image segmentation system based on semantic deviation suppression includes:

[0020] A boundary information constraint module is used to extract organ boundary information from multi-organ medical image data with segmentation truth values, and use the organ boundary information as the boundary information prior constraint;

[0021] a data diversity module, configured to perform different preprocessing on the multi-organ medical image data to be segmented to obtain different expression forms of the multi-organ medical image data to be segmented;

[0022] A semantic deviation level division module is used to divide the segmentation prediction distribution to generate the regional blocks, calculate the semantic deviation, and divide the semantic deviation into deviation degree levels to generate the semantic deviation degree levels;

[0023] A dynamic enhancement module, which suppresses boundary displacement by injecting the semantic deviation level into the multi-organ medical image data to be segmented, thereby obtaining boundary information;

[0024] The organ segmentation module obtains a segmentation result according to the boundary information.

[0025] In a third aspect, an electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.

[0026] In a fourth aspect, a computer-readable storage medium stores a computer program thereon, wherein the computer-readable storage medium stores instructions, characterized in that when the instructions are executed on a computer or a processor, the computer or processor executes the steps of the method described in the first aspect.

[0027] Compared with the prior art, the beneficial effects of the present invention are embodied in:

[0028] 1. Different from traditional medical image segmentation methods, this invention, within the framework of a deep two-dimensional convolutional coding and decoding model, considers the semantic smoothness of image pixel blocks. By analyzing the degree of semantic offset between pixel blocks, the model's understanding of organ boundaries in medical images is enhanced. The image's semantic deviation level distribution is then incorporated into the model's segmentation predictions to calibrate the model's ability to recognize boundary information, thereby providing more accurate boundary information results for image organ segmentation.

[0029] 2. This invention extracts segmentation boundary information from existing segmented images and uses it as a priori boundary constraints, ensuring the consistency of boundary information. It also improves the model's ability to understand the boundary features of medical image organs, thereby making the final medical image organ segmentation results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 1 is a flow chart of the method of Example 1. DETAILED DESCRIPTION

[0031] In order to make the technical means, creative features, objectives and effects of the invention easier to understand, the present invention is further described with reference to specific figures. However, the present invention is not limited to the following implementation cases.

[0032] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the efficacy and purpose that can be achieved by the present invention.

[0033] Example 1:

[0034] like Figure 1 A medical image segmentation method based on semantic deviation suppression is shown, comprising the following steps:

[0035] S1. Using the organ boundary information obtained from the multi-organ medical image data with segmentation truth values as a boundary prior constraint;

[0036] S1.1, standardization preprocessing and morphological gradient algorithm to obtain basic organ contour information;

[0037] In this example, grayscale truncation and normalization are performed on a 2D medical image dataset annotated with voxel-level organ segmentation labels to eliminate grayscale deviations caused by differences in imaging equipment. A morphological gradient operator is then used to perform boundary detection on the preprocessed 2D slices. This operator extracts organ edges by performing the difference between dilation and erosion operations, generating initial organ contours with pixel-level accuracy. The 2D boundary data output from this stage provides the underlying spatial features for subsequent 3D anatomical structure separation.

[0038] S1.2. Multi-organ subvoxel separation using tomographic decoupling and topological reconstruction;

[0039] In this embodiment, the three-dimensional gold standard segmentation atlas is decoupled along the standard anatomical planes (such as the sagittal plane and the coronal plane) to generate a serialized two-dimensional tomographic atlas. Subsequently, a multimodal boundary scanning protocol is adopted to apply the Canny operator to each tomographic image to achieve sub-pixel edge detection, while binding the inter-layer spatial pose information to preserve the three-dimensional coordinate correlation. Finally, an intelligent topological reconstruction algorithm is used to perform nonlinear interpolation on the discrete inter-layer boundaries, integrate the anatomical morphological features between the tomographic layers, eliminate topological contradictions, and reconstruct an organ surface point cloud model that conforms to anatomical continuity in the three-dimensional Euclidean space, thereby completing sub-voxel organ separation.

[0040] In this embodiment, organ boundary information, i.e., organ contours, is obtained through geometric morphological operations; a two-dimensional medical image dataset with voxel-level organ segmentation labels is subjected to standardized preprocessing (including grayscale truncation and normalization operations), and then a morphological gradient operator is used to perform two-dimensional boundary detection.

[0041] In this embodiment, anatomical morphological features are used to achieve subvoxel-level separation of multiple organs. The specific division steps include: based on the gold standard annotation system of the expert-level segmentation atlas, the distribution of the edge features of each anatomical structure is accurately defined through the organ topological mapping algorithm. Specifically, based on the voxel-level spatial coordinate analysis engine, the edge coordinate solution of each organ is performed: in the first stage, digital tomography technology is used to decouple the three-dimensional gold standard annotation volume along the anatomical standard plane to generate a serialized two-dimensional anatomical atlas; in the second stage, a multimodal boundary scanning protocol is used to perform sub-pixel edge detection based on the Canny operator on each tomographic image, and spatial pose metadata is synchronously bound; in the final stage, an intelligent topological reconstruction algorithm is used to perform nonlinear interpolation and fusion of discrete interlayer boundary features in three-dimensional Euclidean space, and finally output an organ surface point cloud model with anatomical rationality.

[0042] S2. Performing data augmentation processing of two different strengths on the multi-organ medical image data to be segmented to generate two different data input streams, representing two different forms of multi-organ medical image data; inputting the two data input streams into two different two-dimensional convolution-based encoding and decoding networks, and outputting two different segmentation prediction distributions with true segmentation values;

[0043] Data augmentation processing includes weak data augmentation and strong data augmentation. Weak data augmentation includes random rotation, flipping, and scaling. Strong augmentation includes random occlusion or color dithering based on weak data augmentation. In this embodiment, weak data augmentation uses random rotation and scaling, while strong data augmentation uses random occlusion based on weak data augmentation.

[0044] The generation of the split prediction distribution includes:

[0045] A U-Net variant architecture based on multi-scale residual connections is constructed, employing a cross-layer feature transfer mechanism to achieve multi-resolution representation learning of anatomical semantics. During the encoding phase, cascaded dilated convolutional modules are used to capture the global context of organ tissues. During the decoding phase, a deconvolutional pyramid is applied to restore spatial semantics from multi-level feature maps. Finally, a softmax classifier is used to generate pixel-level anatomical structure probability distribution maps, enabling organ and tissue topology analysis with submillimeter positioning accuracy. Based on the probability of each voxel belonging to each organ, a normalized activation method and a maximum value method are used to obtain a unique segmentation prediction distribution.

[0046] S3. Divide the two segmentation prediction distributions into multiple regions of equal size, and use the difference between the confidence of each region and the confidence of the region within its preset domain as the semantic deviation; use the semantic deviation values to rank the semantic deviations from high to low, and generate a semantic deviation level.

[0047] In this embodiment, the two segmentation prediction distributions are divided into 4*4 area blocks of the same size.

[0048] The generation of semantic deviation includes:

[0049] For voxel blocks in multi-organ medical imaging data, the region block range is divided and the domain range is set. The segmentation prediction distribution of each region block and the region blocks within its domain range is recorded in turn to construct the semantic deviation.

[0050] In this embodiment, the multi-organ medical image data is set to a region block size of 64*64 pixels, and four adjacent region blocks in both the row and column directions are set as the domain range.

[0051] In this embodiment, the boundary semantic deviation is calculated:

[0052] By calculating the absolute difference in average confidence between each region block and each region block within its domain, the boundary semantic deviation is obtained. The model can better identify the confidence of the boundary information of each region block, help to better improve the erroneous boundary feature information in the boundary prior constraints, help to identify fuzzy or complex organ segmentation boundaries, and improve its judgment ability for unfamiliar image data.

[0053] In this embodiment, by injecting boundary semantic deviation into the segmentation prediction, the accuracy of boundary prediction and the completeness of boundary information are enhanced, so that when processing unfamiliar or blurred medical images, the model can determine the erroneous boundary information predicted by the boundary prior constraints and the boundary information implicit in the internal area of the erroneously predicted organ, thereby obtaining more accurate and complete boundary structure details.

[0054] S4, injecting semantic deviation level into the segmented multi-organ medical image data to suppress boundary displacement;

[0055] Boundary displacement suppression includes:

[0056] In one segmentation prediction distribution, the highest semantic deviation region block and its label are obtained according to the semantic deviation degree level; the lowest semantic deviation region block is selected in the segmentation prediction distribution under another strength; and the highest semantic deviation region block is covered with the lowest semantic deviation region block.

[0057] In step S4, the expression for the area block corresponding to the exchange number is:

[0058]

[0059]

[0060] in, is the label of the highest semantic deviation area block under weak enhancement, Indicates the label of the highest semantic deviation region block under strong enhancement, S indicates strong enhancement, W indicates weak enhancement, and Z indicates the label of the semantic deviation semantic region block; X is multi-organ medical imaging data, It is represented as the multi-organ medical imaging data region block corresponding to the label of the highest semantic deviation region block under strong enhancement, It represents the multi-organ medical image data region block corresponding to the label of the semantic deviation region block under weak enhancement, It is represented as the multi-organ medical imaging data region block corresponding to the label of the highest semantic deviation region block under weak enhancement, Indicates the multi-organ medical image data region blocks corresponding to the labels of the semantic deviation region blocks under strong enhancement.

[0061] In this embodiment, we first use a pre-labeled multi-organ medical imaging dataset to construct an organ contour topology map through a boundary geometric feature extraction module, which serves as a spatial constraint condition during deep learning network training, effectively enhancing the model's ability to represent and learn the boundary areas of anatomical structures; secondly, we use a multimodal data variation processing engine to perform elastic deformation, grayscale perturbation and other diversified enhancement operations on the target image, and use the dynamic expansion of the input data to improve the adaptability of the segmentation results to fluctuations in imaging conditions; then, under the framework of a deep two-dimensional convolutional coding and decoding model, we extract the boundary semantic features of the same sample data in different forms and analyze the boundary semantic deviations generated by the model under data diversity. , enhance the model's understanding of organ boundary features, be able to identify and fully understand the structure of organ boundaries in images, and show higher robustness and completeness; then inject boundary semantic deviation into the organ segmentation prediction distribution to suppress boundary displacement. The boundary displacement suppression strategy enables the model to not only capture the missed boundary semantic information, but also correct the misunderstanding of the prior constraints of boundary information when performing image segmentation tasks, thereby providing a more complete and accurate organ boundary prediction distribution when performing image segmentation; by comprehensively considering data consistency, the model can effectively reduce errors and improve segmentation accuracy when processing image data with fuzzy and sparse semantic features.

[0062] Example 2:

[0063] A medical image segmentation system based on semantic deviation suppression, comprising:

[0064] A boundary information constraint module is used to extract organ boundary information from multi-organ medical imaging data with segmentation truth values, and use the organ boundary information as a priori constraint on boundary information;

[0065] A data diversity module is used to perform different preprocessing on the multi-organ medical image data to be segmented to obtain different expression forms of the multi-organ medical image data to be segmented;

[0066] The semantic deviation level classification module is used to divide the segmentation prediction distribution, generate regional blocks, calculate the semantic deviation, and classify the semantic deviation into deviation levels to generate semantic deviation levels;

[0067] The dynamic enhancement module injects semantic deviation levels into the multi-organ medical image data to be segmented, eliminating the boundary displacement caused by the prior constraints of boundary information under different expression forms, and obtains boundary information;

[0068] The organ segmentation module uses the boundary information prior constraints of the encoding and decoding network to generate boundary displacement information for data input in different forms of expression, suppresses boundary displacement based on the semantic deviation level distribution, eliminates erroneous boundaries that have occurred boundary displacement, and obtains accurate segmentation results.

[0069] Specifically, the organ segmentation module uses deep three-dimensional convolutional layers and activation functions to construct organ semantic segmentation predictions;

[0070] The segmentation prediction of each organ at each voxel point is calculated by the organ semantic segmentation prediction head and boundary displacement suppression, and the organ classification is achieved by taking the maximum value. The expression is as follows:

[0071]

[0072]

[0073] in, and These are the predicted boundary segmentation results of multi-organ medical imaging data under two different expression situations. and Represented as two different organ semantic segmentation prediction heads.

[0074] In this embodiment, two different sets of organ semantic segmentation prediction heads are used to calculate the probability distribution of each voxel corresponding to each organ. This ensures the stability of the segmentation prediction heads during the segmentation process and reduces segmentation errors. Segmentation accuracy is significantly improved, particularly in cases where organ features are similar or organ boundaries are blurred. Furthermore, by operating on the segmentation prediction distributions through boundary displacement suppression, the entire segmentation process is made more accurate, achieving more efficient segmentation performance.

Claims

1. A medical image segmentation method based on semantic deviation suppression, characterized in that: The following steps are involved: S1. Using the organ boundary information obtained from the multi-organ medical image data with segmentation truth values as a boundary prior constraint; S2. Performing data augmentation processing of two different strengths on the multi-organ medical image data to be segmented to generate two different data input streams representing two different forms of the multi-organ medical image data; inputting the two data input streams into two different two-dimensional convolution-based encoding and decoding networks, respectively, to output two different segmentation prediction distributions with true segmentation values; S3, dividing the two segmentation prediction distributions into a plurality of regions of equal size, and taking the difference between the confidence of each region and the confidence of the region within the preset domain as the semantic deviation; and classifying the semantic deviation into deviation levels to generate a semantic deviation level; S4, injecting the semantic deviation level into the multi-organ medical image data to be segmented to suppress boundary displacement; The boundary displacement suppression includes: In one of the segmentation prediction distributions, obtaining a highest semantic deviation region block and its label according to the semantic deviation degree level; Selecting a region block with the lowest semantic deviation from the segmentation prediction distribution at another strength; and covering the region block with the highest semantic deviation with the region block with the lowest semantic deviation.

2. The medical image segmentation method based on semantic deviation suppression according to claim 1, characterized in that: In step S2, the data enhancement process includes weak data enhancement and strong data enhancement; the weak data enhancement includes random rotation, flipping and scaling; The strong enhancement includes performing random occlusion or color jittering on the basis of the weak data enhancement.

3. The medical image segmentation method based on semantic deviation suppression according to claim 2, characterized in that: In step S4, the expression for exchanging the region blocks corresponding to the labels is: ； ； in, is the label of the lowest semantic deviation area block under weak enhancement, represents the label of the lowest semantic deviation region block under strong enhancement, S represents the strong enhancement, W represents the weak enhancement, Z represents the label of the region block; X represents the multi-organ medical image data, The multi-organ medical image data region block corresponding to the label of the highest semantic deviation region block under strong enhancement is represented, The multi-organ medical image data region block corresponding to the label of the region block under weak enhancement, The multi-organ medical image data region block corresponding to the label of the highest semantic deviation region block under weak enhancement, The multi-organ medical image data region block corresponding to the label of the region block under strong enhancement is represented.

4. A medical image segmentation system based on semantic deviation suppression, characterized in that: include: A boundary information constraint module is used to extract organ boundary information from multi-organ medical image data with segmentation truth values, and use the organ boundary information as the boundary information prior constraint; a data diversity module, configured to perform different preprocessing on the multi-organ medical image data to be segmented to obtain different expression forms of the multi-organ medical image data to be segmented; A semantic deviation level division module is used to divide the segmentation prediction distribution to generate the regional blocks, calculate the semantic deviation, and divide the semantic deviation into deviation degree levels to generate the semantic deviation degree levels; a dynamic enhancement module, which suppresses boundary displacement by injecting the semantic deviation level into the multi-organ medical image data to be segmented, thereby obtaining boundary information; The organ segmentation module obtains a segmentation result according to the boundary information.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to claim 1 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, wherein the computer-readable storage medium stores instructions, characterized in that: When the instructions are executed on a computer or a processor, the computer or the processor is caused to perform the steps of the method according to claim 1 .

Images