Medical image segmentation method and system based on implicit neural rendering

Abstract

The invention discloses a medical image segmentation method and system based on implicit neural rendering, and belongs to the technical field of medical image processing, and the method specifically comprises the steps: obtaining medical image data, and carrying out the preprocessing of the image data; the preprocessed image data is input into a lightweight CNN-Transform network, and an initial segmentation result is obtained; inputting the initial segmentation result into an implicit neural rendering network for fine processing to obtain an image segmentation result; the segmentation performance is improved through the combination of a lightweight CNN-Transform network and an implicit neural rendering network. The lightweight CNN-Transform network can be used for solving the problems of target deformation and large data distribution difference in the medical image; meanwhile, the implicit neural rendering network regards a segmentation task as a rendering problem, high-uncertainty points in a preliminary segmentation result are corrected through implicit neural representation, uncertain prediction of selected points is corrected, the segmentation precision of a boundary region is remarkably improved, errors are reduced, and therefore the overall segmentation quality is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular to a medical image segmentation method and system based on implicit neural rendering. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Medical image segmentation aims to divide medical images into several regions with specific attributes. This technology can not only detect abnormalities in human body regions, but also provide important guidance for clinicians. Therefore, it plays a key role in computer-assisted diagnosis and treatment, disease analysis, image-guided surgery, tissue and organ reconstruction, and treatment planning. However, medical images usually face challenges such as higher noise density, lower contrast, and blurred edges. Therefore, how to quickly and accurately segment specific organs and lesions from medical images has always been a major problem in the field of intelligent medicine. In recent years, with the rapid development of computer hardware, researchers have continuously introduced new medical image automatic segmentation algorithms through in-depth theoretical research and a large amount of experimental data analysis. These algorithms can be mainly divided into two categories: methods based on convolutional neural networks (CNNs) and methods based on Transformer networks.

[0004] In the field of medical image segmentation, CNN demonstrates efficient computing capabilities by relying on the scale invariance and inductive bias of its convolution operation. However, due to the fixed nature of the receptive field, CNN is limited in capturing the relationship between distant pixels and lacks the ability to model long-distance features. In order to address the shortcomings of CNN in extracting global features, researchers were inspired by the success of Transformer in the field of natural language processing and introduced it into the field of medical image segmentation. The Transformer network effectively makes up for the shortcomings of CNN by virtue of its ability to establish long-distance dependencies and capture global information. Subsequently, more and more researchers began to explore models that combine CNN and Transformer. This hybrid architecture uses the complementary advantages of CNN and Transformer to achieve local and global information modeling of medical images, thereby achieving better medical image segmentation results.

[0005] Although existing methods have achieved great success in the field of medical image segmentation, these networks still have some problems:

[0006] 1. There are noisy areas such as fuzzy boundaries and heterogeneous textures in medical images, which will produce high gradients during training, distract the model's attention and mislead the optimization, while CNN lacks the ability to selectively capture the features of potential fuzzy boundaries in medical images.

[0007] 2. The convolution operator of CNN is spatially invariant and channel-specific, and has limited ability to change different visual modalities when processing different spatial positions. Due to the limitation of the receptive field, it is difficult to extract features of small targets or targets with blurred edges.

[0008] 3. The application of Transformer in medical image segmentation has problems such as high computational complexity, highly redundant established dependencies, and expensive training costs.

[0009] Existing medical image segmentation models often exhibit inaccurate results when dealing with high-uncertainty areas such as boundaries, resulting in segmentation results that are not precise and reliable enough. Summary of the invention

[0010] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a medical image segmentation method and system based on implicit neural rendering. On the basis of initial segmentation using a lightweight CNN-Transformer network, an implicit neural rendering network is used to correct uncertain rough predictions, thereby achieving fine-grained control of segmentation quality and improving the segmentation effect.

[0011] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0012] In a first aspect, the present invention provides a medical image segmentation method based on implicit neural rendering, comprising:

[0013] Acquiring medical image data and preprocessing the image data;

[0014] Input the preprocessed image data into the lightweight CNN-Transformer network to obtain the initial segmentation result;

[0015] Inputting the initial segmentation result into an implicit neural rendering network for refinement processing to obtain an image segmentation result;

[0016] Among them, the implicit neural rendering network first selects the first N high-uncertainty points from the initial segmentation results according to the point selection mechanism calculated by entropy, encodes the position of each point to transform the position information into a high-dimensional continuous space, and connects the position code with the coarse segmentation features of the point and the features aggregated by the multi-scale decoder to output fine-grained features for optimization.

[0017] Furthermore, in the lightweight CNN-Transformer network, the input image passes through the encoder and the lightweight Transformer module in sequence to extract features, and the decoder generates an initial segmentation result based on the features.

[0018] Furthermore, before performing the self-attention mechanism operation, the lightweight Transformer module processes each candidate query image block through a learnable gating function, and only retains query image blocks with low gating scores to obtain a small self-attention matrix.

[0019] Furthermore, the encoder and decoder include a four-layer architecture, and each layer of the architecture adopts the same operation.

[0020] Furthermore, each layer architecture of the encoder includes an uncertainty-guided feature recalibration module, a downsampling operation, a dynamic deformable convolution, a BN batch normalization operation and a ReLU activation function in sequence.

[0021] Furthermore, each layer architecture of the decoder sequentially includes an upsampling operation, a dynamic deformable convolution, a BN batch normalization operation, a ReLU activation function, and an uncertainty-guided feature recalibration module.

[0022] Furthermore, in the uncertainty-guided feature recalibration module, the uncertainty of each position of the input feature map is calculated through a normalized entropy operation, and the calculated low-uncertainty perception feature map is applied to the input feature map, and the eigenvalue is adjusted through element-level multiplication operations to obtain a reweighted feature map to replace the input feature map and pass it to subsequent network layers.

[0023] In a second aspect, the present invention provides a medical image segmentation system based on implicit neural rendering, comprising:

[0024] A data acquisition module, which is configured to: acquire medical image data and pre-process the image data;

[0025] An initial segmentation module, which is configured to: input the preprocessed image data into a lightweight CNN-Transformer network to obtain an initial segmentation result;

[0026] A fine segmentation module is configured to: input the initial segmentation result into an implicit neural rendering network for fine processing to obtain an image segmentation result;

[0027] Among them, the implicit neural rendering network first selects the first N high-uncertainty points from the initial segmentation results according to the point selection mechanism calculated by entropy, encodes the position of each point to transform the position information into a high-dimensional continuous space, and connects the position code with the coarse segmentation features of the point and the features aggregated by the multi-scale decoder to output fine-grained features for optimization.

[0028] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the medical image segmentation method based on implicit neural rendering as described in the first aspect.

[0029] In a fourth aspect, the present invention provides a computer 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 program, the steps in the medical image segmentation method based on implicit neural rendering as described in the first aspect are implemented.

[0030] One or more of the above technical solutions have the following beneficial effects:

[0031] In terms of segmentation effect, this paper proposes for the first time a medical image segmentation model based on implicit neural rendering and lightweight CNN-Transformer. Through the uncertainty-guided feature recalibration module at the encoder and decoder layers, low-uncertainty features are reweighted to further optimize the feature extraction process, and dynamic deformable convolution is used to adaptively learn the kernel deformation offset and weight coefficient distribution according to specific tasks and data, effectively extract the features of small targets or targets with blurred edges, and use a lightweight Transformer module to establish truly effective long-range dependencies to improve network performance. Finally, an implicit neural rendering module is used to optimize the segmentation results, especially to refine boundaries and process uncertain areas, which improves segmentation accuracy and boundary processing capabilities compared to traditional methods.

[0032] In terms of practicality and scalability, the method of the present invention has a simple and flexible structure. The modular design of the model makes it easy to add new feature extraction modules or optimization modules. When the task objectives need to be expanded or adjusted, there is no need to change the model architecture on a large scale. It has high practicality and scalability and can adapt to a variety of task requirements. The model can support multiple types of medical image segmentation tasks, including but not limited to the segmentation of organs such as the heart, lungs, and liver, and has broad application potential.

[0033] In terms of computational efficiency, the present invention effectively improves computational efficiency by introducing a lightweight Transformer module in model design. Traditional Transformer models usually require a large amount of computing resources to process the global self-attention mechanism. The lightweight Transformer of the present invention reduces computational complexity by designing a lighter and more efficient self-attention mechanism and reducing redundant calculations. In addition, the model effectively reduces unnecessary calculations by adaptively selecting the feature area of ​​interest. Therefore, while maintaining high-precision segmentation performance, the present invention significantly improves the computational efficiency of the model, and is suitable for computing environments with limited resources and real-time applications that require rapid response. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] Figure 1 is a flow chart of a medical image segmentation method according to an embodiment of the present invention;

[0036] Figure 2 is a structural diagram of a medical image segmentation model of a medical image segmentation method according to an embodiment of the present invention;

[0037] Figure 3 is a schematic diagram of an uncertainty-guided feature recalibration module in a medical image segmentation model according to an embodiment of the present invention;

[0038] Figure 4 is a schematic diagram of a dynamic deformable convolution module in a medical image segmentation model according to an embodiment of the present invention;

[0039] Figure 5 is a schematic diagram of a lightweight Transformer module in a medical image segmentation model according to an embodiment of the present invention;

[0040] Figure 6 This is an example diagram of the medical image segmentation result according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0042] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0043] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0044] The method of the present invention improves the segmentation performance by combining a lightweight CNN-Transformer network (LCTNet) and an implicit neural rendering network (INRN). The lightweight CNN-Transformer network adopts a U-shaped structure, integrates uncertainty-guided feature recalibration modules and dynamic deformable convolutions in each layer of the encoder and decoder to optimize feature extraction and deal with the problems of large target deformation and data distribution differences in medical images. The lightweight Transformer module is located at the bottleneck position of the encoder and decoder, and achieves efficient remote dependency modeling by reducing redundant calculations and dependencies; the implicit neural rendering network regards the segmentation task as a rendering problem, and uses implicit neural representation to correct the high uncertainty points in the preliminary segmentation results to correct the uncertain prediction of the selected points, significantly improve the segmentation accuracy of the boundary area, reduce errors, and thus improve the overall segmentation quality.

[0045] Embodiment 1

[0046] like Figure 1 As shown, this embodiment discloses a medical image segmentation method based on implicit neural rendering, which includes the following steps:

[0047] S1: Acquire medical image data and preprocess the image data;

[0048] In this embodiment, an experiment is conducted using a dataset publicly available on the Internet. The dataset comes from the ACDC Challenge (Automated Cardiac Diagnosis Challenge, cardiac image segmentation), including the segmentation of organs such as the left ventricle (LV), right ventricle (RV) and myocardium (Myo). The acquired medical image dataset is preprocessed and data enhancement is performed, and the data enhancement operation uses methods such as random rotation, random scaling, cropping, contrast adjustment and gamma enhancement.

[0049] Specifically, several slice-level short-axis cardiac MRI images and corresponding left ventricle (LV), right ventricle (RV) and myocardium (Myo) categories and pixel-level annotations of corresponding positions are obtained. Each slice-level short-axis cardiac MRI image is preprocessed and data augmented, and the data set is divided into a training set, a validation set, and a test set.

[0050] S2: Input the preprocessed image data into the lightweight CNN-Transformer network to obtain the initial segmentation result;

[0051] This embodiment method proposes a medical image segmentation model based on implicit neural rendering and lightweight CNN-Transformer, including a lightweight CNN-Transformer network and an implicit neural rendering network. The two networks form a unified architecture and can be trained in an end-to-end learning manner. The lightweight CNN-Transformer network extracts multi-scale features of the input medical image through an uncertainty-guided feature recalibration module and a dynamic deformable convolution encoder, respectively. The lightweight Transformer module at the bottleneck position further models global context information, and the corresponding decoder combines these features to generate an initial segmentation result (initial segmentation map); on this basis, the implicit neural rendering network refines the boundaries and other areas with high uncertainty in the initial segmentation result, and improves the segmentation quality by correcting the uncertain rough segmentation prediction.

[0052] In this embodiment, if Figure 2 As shown in the figure, the lightweight CNN-Transformer network (LCTNet) consists of an encoder, a lightweight Transformer module, and a decoder. The input image is sequentially extracted through the encoder and the lightweight Transformer module, and the decoder generates an initial segmentation result based on the features. The encoder module is used to extract feature information. There are four layers of architecture, each layer takes the same operation, including the uncertainty-guided feature recalibration module, downsampling operation (Max Pooling), dynamic deformable convolution and BN batch normalization operation (Batch Norm) and ReLU activation function. The lightweight Transformer module is at the bottleneck position connecting the encoder and the decoder, and its input is the output of the last layer of the encoder; the decoder module combines multi-scale feature information through skip connections for feature fusion. There are four layers of architecture, each layer takes the same operation, including upsampling operation (Upsampling), dynamic deformable convolution and BN batch normalization operation and ReLU activation function, and uncertainty-guided feature recalibration module. The last layer of the decoder output maps the number of feature channels to the number of categories through a 1*1 convolution kernel, and outputs the initial segmentation map.

[0053] Specifically, Figure 3As shown in the figure, the uncertainty-guided feature recalibration module (UGFR) aims to suppress high uncertainty areas, strengthen low uncertainty areas, and optimize the feature extraction process by re-weighting low uncertainty features. For an input feature map, the uncertainty of each position is calculated by normalized entropy operation. The calculated low uncertainty perception feature map U is applied to the original feature map Q. The feature value is adjusted by element-level multiplication operation to obtain a re-weighted feature map to replace the original feature map and pass it to the subsequent network layer. It should be noted that this module does not require additional supervision data, can effectively reduce the gradient interference caused by high uncertainty areas during training, and improve the optimization effect of the model.

[0054] Specifically, Figure 4 As shown in FIG. 1 , the dynamic deformable convolution module (DDCM) realizes the dynamic adjustment of the convolution kernel shape and weight during the network training process, which is more conducive to extracting the features of irregular targets in medical images. H×W×C First, the initial convolution kernel K is used for sampling, and then the deformation offset Δγ of the convolution kernel is obtained through network learning m = {m = 1, 2, 3, ..., N}, the offset is used to adjust the position of the initial convolution kernel sampling, and the output feature map Y is at the coordinate γ n The value at can be expressed as:

[0055]

[0056] Among them, γ m Represents the relative position coordinates of the sampling points in the convolution kernel, γ n Indicates the coordinate point of the current output feature map.

[0057] By linearly combining different weight matrices through learnable parameters, dynamic adjustment of convolution kernel weights is achieved to overcome the defect that different inputs share the same convolution kernel parameters. The output feature map can be expressed as:

[0058]

[0059] in, represents the output feature map generated by the dynamic deformable convolution module, σ represents the activation function, which is used to introduce nonlinear characteristics; ω n represents the weight matrix, x represents the input feature map, n is the number of weight coefficients, α nis a weight coefficient with learnable parameters. Dynamic deformable convolution can adaptively adjust the shape and weight of the convolution kernel according to specific tasks and data distribution, thereby enhancing the feature extraction capability of complex and irregular targets in the encoder and decoder modules.

[0060] Specifically, Figure 5 As shown, the Lightweight Transformer Module (LTM) is located at the bottleneck position connecting the encoder and the decoder, and its input is the output of the last layer of encoder. While reducing the computational complexity, it builds a truly effective long-range dependency relationship for it to pass to the decoder. Before the self-attention mechanism is operated, since it is meaningless to build a long-range dependency relationship for the background image patch and redundant calculations will be generated, each candidate query image patch is processed by a learnable gating function, and only query image patches with low gating scores, i.e., those that are not easily identifiable as background, are retained to obtain a smaller self-attention matrix. The candidate query image patch refers to the initial query patch that has not been screened by the gating function. Patches that are easily identifiable as background will be eliminated before being projected to the query. It is worthless to build long-range dependencies for them. After screening by the gating function, only query patches with low gating scores (i.e., excluding easily identifiable background patches) will be retained to establish dependencies with the key patch.

[0061] The self-attention matrix can be expressed as:

[0062]

[0063] in, represents the patch representation after the gating function excludes the patches that are easily identified as background patches, E q represents the projection matrix of query patches, F p represents the patch representation of the input feature map, E k represents the projection matrix of key patches, d m represents the feature dimension, and P represents positional encoding.

[0064] After obtaining the initially established dependency relationship, the redundant dependency of each query image block is eliminated to reduce the perceptual redundancy generated by the key image block. Specifically, an adaptive threshold is generated for each query image block according to the dependency distribution and internal features, and the threshold is used to select A. p Eliminate redundant dependencies to obtain truly useful dependencies Then perform the remaining attention operations and convert the final self-attention matrix The point multiplication operation is performed with the value matrix, and it is applied to the feedforward neural network layer to obtain the final output to be passed to the decoder. It should be noted that the lightweight Transformer module designed in this way not only effectively reduces the amount of calculation while ensuring the global information modeling capability, but also improves the performance and lightweightness of the model.

[0065] S3: Inputting the initial segmentation result into an implicit neural rendering network for refinement processing to obtain an image segmentation result.

[0066] In this embodiment, an implicit neural rendering network (INRN) is used to correct uncertain rough predictions to improve segmentation quality and achieve refinement of boundary areas.

[0067] First, the first N high-uncertainty points are selected from the initial segmentation result (i.e., the rough segmentation result) according to the point selection mechanism of entropy calculation, and the position encoding is performed on each point to transform the position information into a high-dimensional continuous space, and the position encoding is connected with the rough segmentation feature of the point and the features aggregated by the multi-scale decoder to output fine-grained features for optimization. Specifically, in order to optimize the multi-scale pixel-level features in parallel, the output features [F1, F2, ..., F M ] Perform weighted aggregation to obtain multi-scale decoder aggregation features Where W i =[Softmax(Conv([F1,F2,...,F N ]))].

[0068] The resulting fine-grained features are fed into a rendering head consisting of an implicit neural representation. The goal of the rendering head is to correct the uncertain predictions about these selected points. A 3-layer MLP design is used to correct the class labels of the selected points. The standard cross entropy loss L is used. Rend The model is supervised and trained in an end-to-end manner. Ultimately, through this implicit neural rendering process, the segmentation accuracy of the boundary area can be significantly improved, the error can be reduced, and the overall segmentation quality can be improved.

[0069] The total loss function of the model is expressed as:

[0070] L Total =L Dice +L Ce +λL Rend

[0071] Among them, L Dice is the Dice loss, LCe is the cross entropy loss of the model, L Rend To supervise the cross entropy loss part of the rendering head, λ is used to adjust the weights.

[0072] When training the lightweight CNN-Transformer network and the implicit neural rendering network, set the data path for model loading, the model storage path, etc., as well as the training parameters such as the learning rate, optimization algorithm, number of iterations, and data enhancement method to implement the training of the medical image segmentation model based on implicit neural rendering and lightweight CNN-Transformer. The model loads the best model weights saved during the training process, inputs the medical images in the test set into the model to generate segmentation results, and displays the test results.

[0073] Specifically, the training set is used for training, and an efficient dynamic encoder is used to extract multi-scale features of the input medical image. The lightweight Transformer module at the bottleneck position further models the global context information. The corresponding decoder generates segmentation results based on these features, and the implicit neural rendering module further refines the segmentation results, especially the highly uncertain areas such as boundaries. During the test process, the best model weights saved during the training process are loaded, and the medical images in the test set are input into the model to generate segmentation results. The segmentation results are compared with the true labels, and the evaluation indicators are calculated to analyze the model performance.

[0074] Therefore, the Lightweight CNN-Transformer Network (LCTNet) is composed of a simple U-shaped structure. It uses uncertainty-guided feature recalibration modules in each layer of the encoder and decoder to reweight low-uncertainty features, further optimize the feature extraction process, and adaptively learn the deformation offset and weight coefficient of the convolution kernel through dynamic deformable convolution to effectively deal with the problems of large data distribution differences and large target deformation in medical image segmentation. The lightweight Transformer module is at the bottleneck position connecting the encoder and decoder. Before the self-attention mechanism is operated, it significantly reduces redundant calculations by deleting query patches with high background confidence, and eliminates redundant dependencies for each query patch to reduce the perceptual redundancy of each key patch, so as to establish more effective long-range dependencies while reducing computational complexity.

[0075] Implicit Neural Rendering Network (INRN): It treats medical image segmentation as an end-to-end rendering problem, and estimates the uncertainty of the sampled points in the preliminary segmentation results to generate uncertain coordinate point features. These features reflect the uncertainty and detail information of the anatomical structure at each location in the image. The features of the sampled uncertain coordinate points are adaptively combined and the features of the coordinate points are continuously encoded to correct the uncertain anatomical regions using implicit neural representations. Specifically, the rendering head consists of 3 layers of MLP to correct the class labels of the selected points. This process is supervised by the cross entropy loss function to correct the uncertain rough predictions and achieve fine-grained control of the segmentation quality to improve the segmentation effect. The loss function is used to constrain the training of the network. From Figure 6 It can be seen that the technical solution of the present invention ensures the accuracy of medical image segmentation.

[0076] Embodiment 2

[0077] This embodiment provides a medical image segmentation system based on implicit neural rendering, including:

[0078] A data acquisition module, which is configured to: acquire medical image data and pre-process the image data;

[0079] An initial segmentation module, which is configured to: input the preprocessed image data into a lightweight CNN-Transformer network to obtain an initial segmentation result;

[0080] A fine segmentation module is configured to: input the initial segmentation result into an implicit neural rendering network for fine processing to obtain an image segmentation result;

[0081] Among them, the implicit neural rendering network first selects the first N high-uncertainty points from the initial segmentation results according to the point selection mechanism calculated by entropy, encodes the position of each point to transform the position information into a high-dimensional continuous space, and connects the position code with the coarse segmentation features of the point and the features aggregated by the multi-scale decoder to output fine-grained features for optimization.

[0082] In one embodiment, the medical image segmentation system further comprises an output module, wherein the output module comprises file storage and visual display. The generated file is stored in the memory of the computer device, and the generated image segmentation result is visually displayed.

[0083] Embodiment 3

[0084] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method steps of the first embodiment when executing the program.

[0085] Embodiment 4

[0086] The purpose of this embodiment is to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the steps of the method of embodiment 1 are performed.

[0087] The steps involved in the apparatus of the above embodiments 3 and 4 correspond to the method embodiment 1, and the specific implementation method can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0088] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0090] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A medical image segmentation method based on implicit neural rendering, characterized in that: include: Acquiring medical image data and preprocessing the image data; Input the preprocessed image data into the lightweight CNN-Transformer network to obtain the initial segmentation result; Inputting the initial segmentation result into an implicit neural rendering network for refinement processing to obtain an image segmentation result; Among them, the implicit neural rendering network first selects the first N high-uncertainty points from the initial segmentation results according to the point selection mechanism calculated by entropy, encodes the position of each point to transform the position information into a high-dimensional continuous space, and connects the position code with the coarse segmentation features of the point and the features aggregated by the multi-scale decoder to output fine-grained features for optimization.

2. The medical image segmentation method based on implicit neural rendering according to claim 1, characterized in that: The input image in the lightweight CNN-Transformer network passes through the encoder and the lightweight Transformer module in sequence to extract features, and the decoder generates an initial segmentation result based on the features.

3. The medical image segmentation method based on implicit neural rendering according to claim 2, characterized in that: The lightweight Transformer module processes each candidate query image block through a learnable gating function before performing the self-attention mechanism operation, and only retains the query image blocks with low gating scores to obtain a small self-attention matrix.

4. The medical image segmentation method based on implicit neural rendering according to claim 2, characterized in that: The encoder and decoder include a four-layer architecture, and each layer of the architecture performs the same operation.

5. The medical image segmentation method based on implicit neural rendering according to claim 4, characterized in that: Each layer of the encoder architecture sequentially includes an uncertainty-guided feature recalibration module, a downsampling operation, a dynamic deformable convolution, a BN batch normalization operation, and a ReLU activation function.

6. The medical image segmentation method based on implicit neural rendering according to claim 4, characterized in that: Each layer of the decoder architecture includes an upsampling operation, a dynamic deformable convolution, a BN batch normalization operation, a ReLU activation function, and an uncertainty-guided feature recalibration module.

7. The medical image segmentation method based on implicit neural rendering according to claim 6, characterized in that: In the uncertainty-guided feature recalibration module, the uncertainty of each position of the input feature map is calculated through a normalized entropy operation, and the calculated low-uncertainty perception feature map is applied to the input feature map, and the feature value is adjusted through an element-level multiplication operation to obtain a reweighted feature map to replace the input feature map and pass it to subsequent network layers.

8. A medical image segmentation system based on implicit neural rendering, characterized in that: include: A data acquisition module, which is configured to: acquire medical image data and pre-process the image data; An initial segmentation module, which is configured to: input the preprocessed image data into a lightweight CNN-Transformer network to obtain an initial segmentation result; A fine segmentation module is configured to: input the initial segmentation result into an implicit neural rendering network for fine processing to obtain an image segmentation result; Among them, the implicit neural rendering network first selects the first N high-uncertainty points from the initial segmentation results according to the point selection mechanism calculated by entropy, encodes the position of each point to transform the position information into a high-dimensional continuous space, and connects the position code with the coarse segmentation features of the point and the features aggregated by the multi-scale decoder to output fine-grained features for optimization.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the medical image segmentation method based on implicit neural rendering as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the medical image segmentation method based on implicit neural rendering as described in any one of claims 1 to 7 are implemented.

Images