KAN operator medical image segmentation method based on diffusion mechanism

Abstract

The invention discloses a KAN operator medical image segmentation method based on a diffusion mechanism. The method comprises the following steps: constructing a diffusion segmentation model based on the diffusion mechanism and a KAN operator; the model comprises a de-noising module, and the de-noising module comprises a de-noising encoder, a semantic encoder, a KAN module and a de-noising decoder. Inputting a to-be-processed original medical image into the model for the following processing: extracting semantic information in the original medical image by using a semantic encoder to obtain a first feature; inputting the output of each layer of the semantic encoder to a corresponding layer of a denoising encoder, and extracting local detail information in the noise image by using the denoising encoder to obtain a second feature; optimizing and extracting the second feature by using a KAN module to obtain a third feature; and fusing the first feature, the second feature and the third feature, inputting the fused features into a denoising decoder, inputting the output of each layer of the denoising encoder into the corresponding layer of the denoising decoder, and outputting a segmentation result. The method can significantly improve the segmentation precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a KAN operator medical image segmentation method based on a diffusion mechanism. Background Art

[0002] Medical image segmentation mainly deals with the segmentation of various images involved in the medical field, such as common CT (Computed Tomography) images, MRI (Magnetic Resonance Imaging) images, etc. Its main task is to segment the region of interest from these medical images or videos, such as specific organ parts, targets of interest (such as brain tumors), etc.

[0003] Traditional medical image segmentation methods mainly rely on manual judgment, which is limited by factors such as human experience, energy, and emotional fluctuations, and the judgment accuracy is unstable. At present, medical image analysis is often achieved with the help of deep learning methods. For example, Patent Document 1 (application number CN202410313700.X) discloses a brain MRI tumor segmentation method based on a GAN-UNet variant network, and proposes an innovative brain MRI tumor segmentation technology. This technology uses the generator and discriminator under the framework of the generative adversarial network (GAN), and realizes the extraction and analysis of semantic features from low-level pixels to high-order dimensions through the encoder and decoder structure, while introducing refined channel attention blocks and multi-scale pyramid fusion blocks to enhance feature extraction accuracy. In addition, the method also includes key data preprocessing steps, such as labeling, Z-score normalization, and resizing of MRI images to improve the accuracy of tumor area segmentation.

[0004] Patent document 2 (application number CN202410023147.6) discloses a method for joint segmentation and registration of brain tumor images, and proposes a technology for improving the accuracy of brain tumor image processing by jointly training segmentation models and registration models. This technology uses multiple seed segmentation modules and evaluates the effect through the Dice score. At the same time, it uses the improved VTN network for registration training and realizes feature fusion through jump connections and dense connections. In addition, this method also specially designs the target loss function, taking into account the loss of image dissimilarity terms and segmentation models, as well as appearance-perceived regularization, to improve segmentation accuracy and registration robustness.

[0005] However, although the method based on the GAN-UNet variant network in Patent Document 1 has made progress in improving segmentation accuracy, the technology has limitations in generalization ability and reasoning time, especially when dealing with tumors of different types or morphologies, and further optimization may be required to improve the adaptability and efficiency of the model. Although the joint segmentation and registration method in Patent Document 2 improves the accuracy of segmentation and the robustness of registration, it is insufficient when dealing with the variable morphology and complex spatial distribution characteristics of tumors. In practical applications, more technical innovations may be needed to further improve the accuracy of segmentation and registration, especially in complex clinical situations.

[0006] In summary, when faced with complex clinical environments that require rapid response, the existing deep learning-based medical image segmentation methods need to be further improved in segmentation accuracy, model robustness, and processing efficiency. Summary of the invention

[0007] In order to solve the above problems existing in the prior art, the present invention provides a KAN operator medical image segmentation method based on diffusion mechanism. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0008] In a first aspect, the present invention proposes a KAN operator medical image segmentation method based on a diffusion mechanism, comprising:

[0009] A diffusion segmentation Diff-KAN model is constructed based on the diffusion mechanism and the KAN operator; the diffusion segmentation Diff-KAN model includes a denoising module using an encoder-decoder structure, and the denoising module includes a semantic encoder, a denoising encoder, a KAN module, and a denoising decoder;

[0010] The original medical image to be processed is input into the diffusion segmentation Diff-KAN model for the following processing:

[0011] Add random Gaussian noise to the label to generate a pure noise image; perform channel splicing on the original medical image and the pure noise image, and input the spliced ​​noise image into the denoising encoder;

[0012] The semantic information in the original medical image is extracted by using a semantic encoder to obtain a first feature; each layer output of the semantic encoder is input into a layer of corresponding size in a denoising encoder, and the denoising encoder is used to extract local detail information in the noisy image to obtain a second feature;

[0013] The KAN module is used to optimize and refine the first feature to obtain the third feature;

[0014] The first feature, the second feature and the third feature are fused, and the fused features are input into the denoising decoder. At the same time, each layer output of the denoising encoder is input into the layer of corresponding size in the denoising decoder through a jump connection, and the segmentation result is output.

[0015] In a second aspect, the present invention proposes a KAN operator medical image segmentation device based on a diffusion mechanism, which is used to implement the KAN operator medical image segmentation method based on a diffusion mechanism proposed in the first aspect of the present invention, and the device includes:

[0016] A model building module, used for building a diffusion segmentation Diff-KAN model based on a diffusion mechanism and a KAN operator; the diffusion segmentation Diff-KAN model includes a denoising module using an encoder-decoder structure, and the denoising module includes a semantic encoder, a denoising encoder, a KAN module, and a denoising decoder;

[0017] The image processing module is used to input the original medical image to be processed into the diffusion segmentation Diff-KAN model for the following processing:

[0018] Add random Gaussian noise to the label to generate a pure noise image; perform channel splicing on the original medical image and the pure noise image, and input the spliced ​​noise image into the denoising encoder;

[0019] The semantic information in the original medical image is extracted by using a semantic encoder to obtain a first feature; each layer output of the semantic encoder is input into a layer of corresponding size in a denoising encoder, and the denoising encoder is used to extract local detail information in the noisy image to obtain a second feature;

[0020] The KAN module is used to optimize and refine the second feature to obtain the third feature;

[0021] The first feature, the second feature and the third feature are fused, and the fused features are input into the denoising decoder. At the same time, each layer output of the denoising encoder is input into the layer of corresponding size in the denoising decoder through a jump connection, and the segmentation result is output.

[0022] In a third aspect, the present invention provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0023] Memory, used to store computer programs;

[0024] The processor is used to execute the program stored in the memory to implement the method proposed in the first aspect of the present invention.

[0025] In a fourth aspect, the present invention proposes a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the method proposed in the first aspect of the present invention is implemented.

[0026] Beneficial effects of the present invention:

[0027] 1. The present invention constructs a diffusion segmentation Diff-KAN model based on a diffusion mechanism and a KAN operator. The diffusion segmentation Diff-KAN model includes a denoising module using an encoder-decoder structure, and the denoising module includes a semantic encoder, a denoising encoder, a KAN module, and a denoising decoder; wherein the semantic encoder is used to extract semantic information from the original medical image, the denoising encoder is used to extract local detail information from the noise image corresponding to the original medical image, the KAN module is used to further optimize and refine the output features of the denoising encoder, and the denoising decoder processes the fusion features of the output features of the semantic encoder, the denoising encoder, and the KAN module, and outputs the segmentation result. The diffusion segmentation Diff-KAN model constructed by the present invention is based on the encoder-decoder structure, and the KAN operator is cleverly integrated into the diffusion model, so as to achieve more refined and accurate processing of the image. Among them, the diffusion mechanism allows the model to simulate the physical diffusion process during processing, thereby transferring information between different areas of the image and enhancing the global consistency of segmentation, while the KAN operator optimizes the model's ability to capture local features through specific mathematical operations and simplifies the model structure; on the other hand, a unique architecture of denoising encoder and semantic encoder is adopted. The denoising encoder focuses on removing noise from the image and restoring clear features, while the semantic encoder is responsible for extracting semantic information from the image. The collaborative work of these two encoders enables the model to understand and process image content more comprehensively. By combining the diffusion model, KAN operator and deep learning model, the accuracy and robustness of the algorithm are significantly improved.

[0028] 2. The diffusion segmentation Diff-KAN model designed by the present invention adopts a U-shaped encoder-decoder structure, which enables the model to gradually abstract and refine features from the bottom up when processing images. At the same time, it combines the dense skip residual connection technology, which not only enhances the model's ability to segment complex images, but also significantly improves the model's processing efficiency. Residual connections allow the model to directly transmit information in the deep network, reducing information loss and improving the model's learning ability and generalization.

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flowchart of a KAN operator medical image segmentation method based on a diffusion mechanism provided by an embodiment of the present invention;

[0031] Figure 2 is a framework diagram of a diffusion segmentation Diff-KAN model provided by an embodiment of the present invention;

[0032] Figure 3 Schematic diagram of the structure of the word segmentation KAN network provided by an embodiment of the present invention;

[0033] Figure 4 is a schematic diagram of the structure of the KAN layer provided in an embodiment of the present invention;

[0034] Figure 5 It is a structural block diagram of a KAN operator medical image segmentation device based on a diffusion mechanism provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] The core idea of ​​the diffusion mechanism is to gradually transform the complex data distribution into a simple distribution (such as the standard normal distribution) by simulating a continuous random process, and then gradually restore the original data distribution from this simple distribution. This process can be regarded as a process of "noising" and "denoising". In the diffusion process, noise is first gradually added to the data until the data becomes sufficiently blurred and can be approximately regarded as sampled from a simple distribution (such as the standard normal distribution). Then, through the reverse process, the noise is gradually removed and the original data is finally restored. By gradually removing the noise in the image, not only the recognition and segmentation accuracy of the key information in the image is significantly improved, but also the image features are clearer by improving the visual effect of the image. The robustness of this model enables it to maintain high segmentation accuracy in the face of noise and image degradation, and its multi-scale feature fusion capability ensures that both global and local information can be considered in the segmentation process, further improving the accuracy of segmentation.

[0037] KAN (Kolmogorov-Arnold Network) is a neural network architecture based on the Kolmogorov-Arnold representation theorem, which is designed to replace the traditional multi-layer perceptron. The KAN operator uses learnable activation functions and parameterized activation functions on the edge as weights, abandons the reliance on linear weight matrices, simplifies the model structure, and maintains or improves performance. This design not only improves the generalization and adaptability of the model, enabling it to flexibly cope with different types of images and noise patterns, but also enhances the interpretability of the model, which is particularly important for medical image analysis.

[0038] Therefore, the first aspect of the present invention proposes a diffusion-based medical image segmentation method based on the diffusion mechanism and the KAN operator.

[0039] See also Figure 1 , Figure 1 : is a flow chart of a diffusion-based medical image segmentation method provided by an embodiment of the present invention. The method mainly includes:

[0040] Step 1, construct a diffusion segmentation Diff-KAN model based on the diffusion mechanism and the KAN operator; the diffusion segmentation Diff-KAN model includes a denoising module using an encoder-decoder structure, and the denoising module includes a semantic encoder, a denoising encoder, a KAN module and a denoising decoder.

[0041] Specifically, see Figure 2 , Figure 2 : This is a framework diagram of the diffusion segmentation Diff-KAN model provided by an embodiment of the present invention. The semantic encoder includes multiple first convolution blocks, each of which includes a convolution layer (Conv), a batch normalization layer (BN), and a ReLU activation function. A maximum pooling layer is added after each first convolution block, and the pooling window size is 2x2. The output of each convolution block can be written as:

[0042] X L =Pool(Conv(X L-1 ));

[0043] Where, X L represents the output feature map of the Lth convolutional block, Pool represents the pooling operation, Conv represents the convolution operation, and X L-1 represents the output feature map of the L-1th convolutional layer, then the final output of the semantic encoder is X L .

[0044] Optionally, as an implementation manner, the semantic encoder in this embodiment may include 3 convolution blocks, ie, L=3, and the kernel size of each convolution block is set to 3x3, the step size is 1, and the padding amount is 1.

[0045] It should be noted that since the diffusion model generates images from a random Gaussian noise, ∈~N(0,1)∈~N(0,1), this process can be achieved by predicting the noise given the noise input, that is:

[0046] ∈ t =DiffKAN(I t ,t);

[0047] In the formula, I t is the label image corrupted by Gaussian noise, and t is the time step I that controls the noise intensity t =N(0,1).

[0048] The diffusion model can inject learnable temporal information into each convolutional block, giving the network the ability to perceive temporal dynamics.

[0049] For further information, please see Figure 2 , wherein the denoising encoder includes a plurality of second convolution blocks, each of which includes a convolution layer, a batch normalization layer, and a ReLU activation function; wherein the number of the second convolution blocks is the same as the number of the first convolution blocks, and each first convolution block is connected to a second convolution block of a corresponding size to fuse features of different scales in the semantic encoder with features of corresponding scales in the denoising encoder.

[0050] Optionally, in this embodiment, the denoising encoder also includes three second convolution blocks and adopts the same convolution processing technology as the semantic encoder, that is, each second convolution block has three key components: convolution layer (Conv), batch normalization layer (BN) and ReLU activation function. This structure not only deepens the network's ability to capture local features, but also ensures the stability and efficiency of network training.

[0051] Specifically, the semantic encoder receives the raw unprocessed medical images as input, extracts and generates multi-scale features, which contain rich semantic information of the original data. Furthermore, by cleverly performing feature fusion on these different scales with the corresponding size features generated by the denoising encoder, that is, through the splicing operation, a comprehensive multi-scale feature set can be obtained. This fusion process not only enhances the model's ability to capture image details, but also significantly improves the recognition and segmentation accuracy of complex structures such as brain tumors by integrating information at different levels.

[0052] This embodiment adopts a unique architecture of denoising encoder and semantic encoder. The denoising encoder focuses on removing noise in the image and restoring clear features, while the semantic encoder is responsible for extracting semantic information in the image. The collaborative work of these two encoders enables the model to understand and process image content more comprehensively, especially performing well in multi-scale feature fusion.

[0053] Furthermore, the KAN module includes a plurality of word segmentation KAN networks, which are connected in series in sequence, and the output end of the first word segmentation KAN network is connected to the input end of the last word segmentation KAN network; wherein,

[0054] The input of the first word segmentation KAN network is connected to the output of the denoising encoder as the input of the entire KAN module;

[0055] The output of the last word segmentation KAN network is connected to the input of the denoising decoder as the output of the entire KAN module.

[0056] Optionally, as an implementation, Figure 2 As shown, this embodiment designs a KAN module including four word segmentation KAN networks.

[0057] See also Figure 3 , Figure 3 is a schematic diagram of the structure of the word segmentation KAN network provided by an embodiment of the present invention. Specifically, each word segmentation KAN network includes a word segmentation module, a mapping module, a plurality of KAN layers and layer normalization connected in sequence; wherein,

[0058] The word segmentation module is used to reshape the input features into a 2D image block sequence through a word segmentation algorithm;

[0059] The mapping module is used to map the 2D image block sequence to the D-dimensional embedding space to obtain high-dimensional features and pass them to multiple KAN layers;

[0060] Multiple KAN layers are connected in series to optimize and refine the high-dimensional features of the input;

[0061] Layer normalization is used to process the output features of the last KAN layer to obtain the output of the current word segmentation KAN network.

[0062] For example, for the first word segmentation KAN network, its input is the output X of the denoising encoder L , through the word segmentation algorithm, X L Reshape into a sequence of flat 2D image patches Among them, the size of each image block is P×P, and the number of image blocks generated is H L For X L The length, WL For X L The vectorized image blocks are then transformed using a trainable linear projection E (i.e., a convolutional layer with a kernel size of 3). Mapped to a potential D-dimensional embedding space, namely:

[0063]

[0064] Where Z0 is the E high-dimensional feature obtained after mapping.

[0065] Next, the high-dimensional features obtained above are transferred to multiple KAN layers for processing. In this embodiment, three KAN layers are preferred. The structure of each KAN layer is as follows: Figure 4 shown.

[0066] It should be noted that each of the multiple KAN layers also includes a Batch normalization layer and a ReLU activation function.

[0067] Finally, layer normalization (LN) is applied and the output features are passed to the next block.

[0068] The principle of the entire word segmentation KAN network can be expressed as follows:

[0069] Z k =LN(KAN(Z k-1 ));

[0070] In the formula, Z k and Z k-1 They represent the output feature maps of the kth and k-1th word segmentation KAN networks respectively.

[0071] The word segmentation KAN network designed in the present invention utilizes feature word segmentation and high-dimensional mapping technology to convert the features output by the convolution block into a flat 2D image block sequence through word segmentation technology, and applies trainable linear projection to map these image blocks to a high-dimensional embedding space (D dimension), which greatly enriches the expressive power of the features. Finally, these high-dimensional features are passed to the KAN layer. This step further refines the features, so that the model can more accurately capture and respond to the key information of brain tumor images.

[0072] For further information, please see Figure 2, where the denoising decoder maintains consistency and symmetry with the denoising encoder in structure, and also includes three convolution blocks. This structure cleverly utilizes the previous word segmentation KAN block to process features. At the same time, each decoder block has the ability to double the resolution of the input features. This design not only accelerates the upsampling process of features, but also effectively enhances the model's ability to reconstruct the fine structure of the original data by gradually restoring the details of the image. In this way, the denoising decoder not only improves the image resolution, but also ensures the richness and completeness of the feature information, providing a solid foundation for achieving high-precision brain tumor segmentation.

[0073] In addition, for the diffusion segmentation Diff-KAN model constructed by the present invention, its denoising module specifically adopts a U-shaped structure and an encoder-decoder structure with dense jump connections, and the denoising encoder, KAN module and denoising decoder form a classic U-shaped structure as a whole.

[0074] The model constructed in the present invention adopts a classic U-shaped structure, supplemented by dense jump connection technology, to carefully construct the Diff-KAN model. This structural design draws on the hierarchical processing mechanism of the biological visual system, cleverly bypasses the information loss problem in the traditional deep network through jump connections, and effectively transmits low-level detail information directly to the high level of the model. Such a design not only retains the key details in the image, but also realizes the in-depth extraction of high-level semantic information through the collaborative work of the encoder-decoder structure. The encoder part is responsible for capturing and encoding the essential features of the input data, while the decoder part focuses on converting these features into accurate segmentation results. The use of this structure enables the Diff-KAN model to achieve more refined and accurate segmentation effects when processing complex brain tumor images, significantly improving the performance and application value of the model.

[0075] Specifically, given the skip connection feature Z of the kth network in the word segmentation KAN network k and the characteristic Z' of the last loading block k+1 , then the input feature Z' of the kth upsampling block k for:

[0076] Z' k =Cat(Z' k+1 ,(Z k ));

[0077] In the formula, Cat() represents the feature concatenation operation. Similarly, given the skip feature X of the lth layer in the convolution part of the encoder, l , and the feature X of the last upsampling block l ' +1 , then the input feature X of the lth upsampling block l 'for:

[0078] X l '=Cat(X l ' +1 ,(X l )).

[0079] At this point, the construction of the diffusion segmentation Diff-KAN model has been completed.

[0080] The present invention proposes a diffusion segmentation Diff-KAN model based on the diffusion mechanism and the KAN operator. By introducing the diffusion mechanism and the KAN operator, a more sophisticated and accurate image processing is achieved. The diffusion mechanism allows the model to simulate the physical diffusion process during the processing, thereby transferring information between different areas of the image and enhancing the global consistency of the segmentation. The KAN operator optimizes the model's ability to capture local features through specific mathematical operations.

[0081] Furthermore, in terms of model design, the present invention adopts a unique architecture of a denoising encoder and a semantic encoder. The design of these two encoders enables the model to process both local details and global semantic information of the image. The denoising encoder focuses on removing noise from the image and restoring clear features, while the semantic encoder is responsible for extracting semantic information from the image. The collaborative work of these two encoders enables the model to understand and process image content more comprehensively, especially in multi-scale feature fusion.

[0082] In addition, this model adopts a U-shaped structure design, which imitates the processing flow of the human visual system, allowing the model to gradually abstract and refine features from the bottom up when processing images. The U-shaped structure combined with residual connection technology not only enhances the model's ability to segment complex images, such as brain tumor images, but also significantly improves the model's processing efficiency. Residual connections allow the model to directly transmit information in deep networks, reducing information loss and improving the model's learning ability and generalization.

[0083] The present invention combines the diffusion mechanism, the KAN operator and the U-shaped structure, so that the obtained diffusion segmentation Diff-KAN model shows excellent performance in segmentation accuracy, model robustness and processing speed.

[0084] Step 2: Input the original image of the medical image to be processed into the diffusion segmentation Diff-KAN model for the following processing:

[0085] Add random Gaussian noise to the label to generate a pure noise image; perform channel splicing on the original medical image and the pure noise image, and input the spliced ​​noise image into the denoising encoder;

[0086] The semantic information in the original medical image is extracted by using a semantic encoder to obtain a first feature; each layer output of the semantic encoder is input into a layer of corresponding size in a denoising encoder, and the denoising encoder is used to extract local detail information in the noisy image to obtain a second feature;

[0087] The KAN module is used to optimize and refine the second feature to obtain the third feature;

[0088] The first feature, the second feature and the third feature are fused, and the fused features are input into the denoising decoder. At the same time, each layer output of the denoising encoder is input into the layer of corresponding size in the denoising decoder through a jump connection, and the segmentation result is output.

[0089] It should be noted that before the original medical image to be processed is input into the diffusion segmentation Diff-KAN model for processing, the diffusion segmentation Diff-KAN model is trained using a training set.

[0090] In this embodiment, the training set is constructed using the following method:

[0091] Collect existing medical imaging data, analyze the effective attributes in the medical imaging data based on principal component analysis, and generate a usable data set;

[0092] The data set is preprocessed and divided into training set, validation set and test set; the training set is used to train the diffusion segmentation Diff-KAN model, and the validation set and test set are used to verify and evaluate the generalization ability of the diffusion segmentation Diff-KAN model.

[0093] Optionally, as an implementation method, first, we extensively collect raw medical imaging data of brain tumors, including key attributes such as scan date, image type, tumor location, size, pathological type, patient age and gender, etc. Through close cooperation with domain experts and the use of advanced data analysis methods such as principal component analysis, we filter out the most effective attributes from the raw data and build a high-quality dataset that can be used for deep learning model training.

[0094] Then, a series of preprocessing operations were performed on the raw medical imaging data in the dataset to ensure the quality of the data. The preprocessing operations included standardizing the images and performing data augmentation operations using techniques such as rotation, flipping, and random cropping, thereby increasing the diversity and robustness of model training.

[0095] Specifically, the brain tumor area and its adjacent tissues are defined as the main regions of interest and marked as foreground objects. The pixel intensities of these regions are carefully analyzed to obtain their mean and standard deviations. On this basis, the entire dataset is standardized by centering and scaling the pixel intensities to eliminate the intensity differences between different images.

[0096] A variety of enhancement techniques were implemented on the standardized dataset, including but not limited to image rotation, flipping, random cropping, etc. These operations are designed to simulate various variations that may be encountered during clinical image acquisition, thereby enhancing the model's ability to recognize these variations.

[0097] Finally, the processed data set is precisely divided into training set, test set, and validation set according to a certain ratio. This strategy is crucial for developing a high-performance segmentation model. Through this division, not only the overfitting problem caused by the lack of verification link is avoided, but also the model can be iterated continuously by repeatedly adjusting and optimizing the attribute information to achieve the best segmentation effect. The rigor and foresight of this methodology provide a solid foundation for achieving high-precision segmentation of brain tumor images.

[0098] The obtained training set is input into the constructed diffusion segmentation Diff-KAN model for training. It should be noted that in the task context of medical brain tumor segmentation, the final segmentation map can be obtained from the output feature map, that is, Among them, C Y is the number of semantic categories, and Y is the ground-truth of the segmentation. Therefore, in order to measure the accuracy of the segmentation result, this embodiment designs a comprehensive loss function, which consists of three parts: Dice loss, binary cross entropy loss, and mean square error loss, namely:

[0099]

[0100] Where, L total Represents the total loss function, L dice (DiffKAN(I),Y) represents Dice loss, L bce (DiffKAN(I),Y) represents the binary cross entropy loss, L mse (DiffKAN(I),Y) represents the mean square error loss.

[0101] This loss function comprehensively considers the segmentation contour consistency, category balance, and pixel-level accuracy, ensuring that the model is fully optimized during training, thereby achieving higher accuracy and reliability in brain tumor segmentation tasks.

[0102] After the training is completed, the original medical images can be input into the trained diffusion segmentation Diff-KAN model for processing.

[0103] First, random Gaussian noise is added to the label to generate a pure noise image. Then, the original medical image and the pure noise image are channel-joined to generate a noise image.

[0104] Then, the original medical image is used as the input of the semantic encoder, and the semantic information in the original medical image is extracted through the convolution operation to obtain the first feature. At the same time, each layer output of the semantic encoder is input to the corresponding size level in the denoising encoder, and the concatenated noisy image is used as the input of the denoising encoder, and the local detail information in the noisy image is extracted through the convolution operation to obtain the second feature.

[0105] Next, the second feature output by the denoising encoder is input into the KAN module for optimization and refinement to obtain the third feature.

[0106] Finally, the first feature, the second feature and the third feature are fused, and the fused features are input into the denoising decoder. At the same time, each layer output of the denoising encoder is input into the layer of corresponding size in the denoising decoder through a jump connection, and the segmentation result is output.

[0107] It should be noted that before the original medical images are input into the trained diffusion segmentation Diff-KAN model for processing, they also need to be subjected to data enhancement operations such as rotation, flipping, and random cropping.

[0108] The present invention proposes a diffusion-based medical image segmentation method, which aims to overcome the limitations of the existing technology in the accuracy of brain tumor segmentation and the generalization ability of the model. The method first pre-processes the original medical image in detail, including unifying the slice interval, standardizing the pixel intensity and applying data enhancement technology to reduce the differences between different images and improve the generalization of the model. Subsequently, a Diff-KAN model containing an encoder-decoder structure and a diffusion mechanism is constructed, which further enhances the recognition and segmentation accuracy of the brain tumor area by introducing the KAN operator. The model is systematically trained using the training set data, and the model performance is strictly evaluated using the validation set and the test set, ensuring the robustness and accuracy of the model on different data. Finally, a high-precision brain tumor segmentation image is obtained by post-processing the segmentation results output by the model. The present invention not only significantly improves the segmentation accuracy, but also optimizes the learning efficiency and processing speed of the model, providing clinicians with an efficient and accurate brain tumor image analysis tool, which has important clinical application value and broad market prospects.

[0109] Based on the same inventive concept, the second aspect of the present invention also provides a KAN operator medical image segmentation device based on a diffusion mechanism. Figure 5 , Figure 5 : is a structural block diagram of a KAN operator medical image segmentation device based on a diffusion mechanism provided by an embodiment of the present invention, and the device specifically includes:

[0110] A model building module, used for building a diffusion segmentation Diff-KAN model based on a diffusion mechanism and a KAN operator; the diffusion segmentation Diff-KAN model includes a denoising module using an encoder-decoder structure, and the denoising module includes a semantic encoder, a denoising encoder, a KAN module, and a denoising decoder;

[0111] The image processing module is used to input the original medical image to be processed into the diffusion segmentation Diff-KAN model for the following processing:

[0112] Add random Gaussian noise to the label to generate a pure noise image; perform channel splicing on the original medical image and the pure noise image, and input the spliced ​​noise image into the denoising encoder;

[0113] The semantic information in the original medical image is extracted by using a semantic encoder to obtain a first feature; each layer output of the semantic encoder is input into a layer of corresponding size in a denoising encoder, and the denoising encoder is used to extract local detail information in the noisy image to obtain a second feature;

[0114] The KAN module is used to optimize and refine the second feature to obtain the third feature;

[0115] The first feature, the second feature and the third feature are fused, and the fused features are input into the denoising decoder. At the same time, each layer output of the denoising encoder is input into the layer of corresponding size in the denoising decoder through a jump connection, and the segmentation result is output.

[0116] Based on the same inventive concept, the third aspect of the present invention further provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0117] Memory, used to store computer programs;

[0118] The processor is used to execute the program stored in the memory to implement the method proposed in the first aspect of the present invention.

[0119] Based on the same inventive concept, the third aspect of the present invention further provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the method proposed in the first aspect of the present invention is implemented.

[0120] As for the device provided in the second aspect, the electronic device provided in the third aspect, and the storage medium embodiment provided in the fourth aspect of the present invention, since they are basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0121] It should be noted that the device, electronic device and storage medium of the embodiments of the present invention are respectively a device, electronic device and storage medium that apply the above-mentioned diffusion-based medical image segmentation method. All embodiments of the above-mentioned XX method are applicable to the device, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0122] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, devices (equipment), or computer program products. Therefore, the present application may adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware, which are collectively referred to as "modules" or "systems" herein. Moreover, the present application may adopt the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program codes. The computer program is stored / distributed in a suitable medium, provided together with other hardware or as a part of hardware, or may adopt other distribution forms, such as by Internet or other wired or wireless telecommunication systems.

[0123] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A KAN operator medical image segmentation method based on diffusion mechanism, characterized in that: include: A diffusion segmentation Diff-KAN model is constructed based on a diffusion mechanism and a KAN operator; the diffusion segmentation Diff-KAN model includes a denoising module using an encoder-decoder structure, and the denoising module includes a semantic encoder, a denoising encoder, a KAN module, and a denoising decoder; The original medical image to be processed is input into the diffusion segmentation Diff-KAN model for the following processing: Adding random Gaussian noise to the label to generate a pure noise image; performing channel splicing on the original medical image and the pure noise image, and inputting the spliced ​​noise image into the denoising encoder; Extracting semantic information from the original medical image using the semantic encoder to obtain a first feature; Inputting each layer output of the semantic encoder into a layer of corresponding size in the denoising encoder, and using the denoising encoder to extract local detail information in the noisy image to obtain a second feature; Utilizing the KAN module to optimize and refine the second feature to obtain a third feature; The first feature, the second feature and the third feature are fused, and the fused features are input into the denoising decoder. At the same time, each layer output of the denoising encoder is input into the layer of corresponding size in the denoising decoder through a jump connection, and the segmentation result is output.

2. The KAN operator medical image segmentation method based on diffusion mechanism according to claim 1, characterized in that: The semantic encoder includes multiple first convolution blocks, each of which includes a convolution layer, a batch normalization layer and a ReLU activation function.

3. The KAN operator medical image segmentation method based on diffusion mechanism according to claim 2, characterized in that: The denoising encoder includes multiple second convolution blocks, each of which includes a convolution layer, a batch normalization layer and a ReLU activation function; wherein the number of the second convolution blocks is the same as the number of the first convolution blocks, and each of the second convolution blocks is connected to the first convolution block of the corresponding size to fuse features of different scales in the semantic encoder with features of corresponding scales in the denoising encoder.

4. The KAN operator medical image segmentation method based on diffusion mechanism according to claim 1, characterized in that: The KAN module includes a plurality of word segmentation KAN networks, which are connected in series in sequence, and the output end of the first word segmentation KAN network is connected to the input end of the last word segmentation KAN network; wherein, The input end of the first word segmentation KAN network is connected to the output of the denoising encoder as the input of the entire KAN module; The output end of the last word segmentation KAN network is connected to the input of the denoising decoder as the output of the entire KAN module.

5. The KAN operator medical image segmentation method based on diffusion mechanism according to claim 4 is characterized in that: The word segmentation KAN network includes a word segmentation module, a mapping module, a plurality of KAN layers and layer normalization connected in sequence; The word segmentation module is used to reshape the input features into a 2D image block sequence through a word segmentation algorithm; The mapping module is used to map the 2D image block sequence to a D-dimensional embedding space to obtain high-dimensional features and pass them to the multiple KAN layers; The multiple KAN layers are connected in series in sequence to optimize and refine the input high-dimensional features; The layer normalization is used to process the output features of the last KAN layer to obtain the output of the current word segmentation KAN network; Each of the multiple KAN layers further includes a Batch normalization layer and a ReLU activation function.

6. The KAN operator medical image segmentation method based on diffusion mechanism according to claim 1, characterized in that: The denoising module specifically adopts a U-shaped structure and an encoder-decoder structure with dense jump connections, and the denoising decoder adopts a structure symmetrical to the denoising encoder.

7. The KAN operator medical image segmentation method based on diffusion mechanism according to claim 1, characterized in that: Before the original medical image to be processed is input into the diffusion segmentation Diff-KAN model for processing, the method further includes: Using the training set to train the diffusion segmentation Diff-KAN model; The training set is constructed using the following method: Collect existing medical image data, analyze effective attributes in the medical image data based on principal component analysis, and generate a usable data set; Preprocessing the data set, and dividing the preprocessed data set into a training set, a validation set, and a test set; the training set is used to train the diffusion segmentation Diff-KAN model, and the validation set and the test set are used to verify and evaluate the generalization ability of the diffusion segmentation Diff-KAN model; The loss function used in training is a comprehensive loss function consisting of Dice loss, binary cross entropy loss and mean square error loss.

8. A KAN operator medical image segmentation device based on a diffusion mechanism, used to implement the KAN operator medical image segmentation method based on a diffusion mechanism according to any one of claims 1 to 7, characterized in that: The device comprises: A model building module, used for building a diffusion segmentation Diff-KAN model based on a diffusion mechanism and a KAN operator; the diffusion segmentation Diff-KAN model includes a denoising module using an encoder-decoder structure, and the denoising module includes a semantic encoder, a noise encoder, a KAN module, and a denoising decoder; The image processing module is used to input the original medical image to be processed into the diffusion segmentation Diff-KAN model for the following processing: Adding random Gaussian noise to the label to generate a pure noise image; performing channel splicing on the original medical image and the pure noise image, and inputting the spliced ​​noise image into the denoising encoder; The semantic encoder is used to extract semantic information in the original medical image to obtain a first feature; each layer output of the semantic encoder is input into a layer of corresponding size in the denoising encoder, and the denoising encoder is used to extract local detail information in the noisy image to obtain a second feature; Utilizing the KAN module to optimize and refine the first feature to obtain a third feature; The first feature, the second feature and the third feature are fused, and the fused features are input into the denoising decoder. At the same time, each layer output of the denoising encoder is input into the layer of corresponding size in the denoising decoder through a jump connection, and the segmentation result is output.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory, used to store computer programs; A processor, configured to execute a program stored in a memory to implement the method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 can be implemented.

Images