Renal tumor image segmentation method and system based on multi-scale feature extraction

Abstract

The invention relates to a kidney tumor image segmentation method and system based on multi-scale feature extraction, and the method comprises the steps: extracting the multi-scale features of a kidney tumor image to be processed, dividing feature similarity groups, carrying out the optimization through combining the texture and shape, and obtaining similar feature groups; generating a complementary information flow, adaptively enhancing an interaction weight between groups by using an attention mechanism, and obtaining an enhanced complementary feature representation; evaluating and denoising noise, establishing association mapping among groups based on features after denoising, and performing multi-scale feature fusion and boundary refining by adopting a U-Net network to obtain tumor boundary representation; and the boundary fuzzy degree is evaluated, iterative optimization is carried out by adopting a graph convolutional network, optimized collaborative hierarchical representation is obtained, and finally a high-precision kidney tumor segmentation result and a pixel-level contour are output. According to the method, through multi-scale feature cooperation, attention interaction and boundary optimization, the problems of boundary blur and noise sensitivity in kidney tumor segmentation are effectively solved, and the accuracy and robustness of segmentation are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for segmenting kidney tumor images based on multi-scale feature extraction. Background Technology

[0002] Tumor image segmentation, a key branch of medical image analysis, is directly related to the accuracy of clinical diagnosis and the formulation of patient treatment plans. With the advancement of high-precision imaging equipment, the importance of this field has become increasingly prominent, as it helps doctors accurately identify tumor boundaries from complex tissues, thereby guiding surgical or radiotherapy decisions. However, in practical applications, this process often faces the challenge of variable tumor morphology, making the reliability of segmentation results a bottleneck restricting the progress of precision medicine.

[0003] Currently, existing automatic or semi-automatic renal tumor segmentation methods, especially those based on deep learning, still have several limitations in practical applications. First, renal tumors exhibit high heterogeneity in morphology, size, location, and texture features, making it difficult for single-scale feature extraction strategies to simultaneously capture their macroscopic contours and microscopic details. This results in insufficient sensitivity of algorithms to small tumors or regions with blurred boundaries, easily leading to missed or over-segmentation in practical applications, directly affecting the reliability of clinical diagnosis. Second, existing methods often employ simple splicing or addition operations when performing multi-scale feature fusion, lacking effective mechanisms for discriminating between feature complementarity and redundancy. This simple fusion approach fails to fully realize the synergistic enhancement between information at different scales and may even introduce noise and inconsistencies, reducing the accuracy and stability of the segmentation results and weakening the algorithm's generalization ability in complex cases. Furthermore, due to noise, artifacts, and insufficient contrast between tumors and normal tissues during image acquisition, segmentation boundaries often appear blurred, broken, or discontinuous. Traditional post-processing methods have limited optimization capabilities for such complex boundaries, and parameter adjustments rely on experience. This forces doctors to make a lot of tedious manual corrections, which greatly increases the clinical workload, reduces the efficiency of diagnosis and treatment, and becomes a key bottleneck for the technology to be applied in clinical practice.

[0004] Therefore, designing a renal tumor image segmentation method that can adaptively fuse multi-scale features, effectively suppress noise interference, and accurately optimize blurred boundaries has become a pressing technical challenge in this field. This invention aims to address these problems by proposing a renal tumor image segmentation method and system based on multi-scale feature extraction. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for renal tumor image segmentation based on multi-scale feature extraction. Through the synergistic optimization and boundary refinement of multi-scale features, the accuracy and robustness of renal tumor segmentation are significantly improved, making it suitable for the clinical diagnostic needs of complex CT images.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A kidney tumor image segmentation method based on multi-scale feature extraction includes:

[0008] Multi-scale features are extracted from kidney tumor images to be processed using a pyramid structure, and a multi-scale feature set is constructed.

[0009] Clustering algorithms are used to divide the multi-scale feature set into similarity groups, and texture and shape patterns are combined to determine similar feature groups;

[0010] Based on the similar feature groups, complementary information streams are obtained, and the interaction weights between different groups are enhanced through an attention mechanism to obtain enhanced complementary feature representations.

[0011] Based on the enhanced complementary feature representation, the noise interference level is determined and denoising is performed to obtain the denoised multi-scale features.

[0012] The association mapping between groups is established through the denoised multi-scale features, and the tumor boundary representation is determined by using the U-Net architecture for feature fusion and boundary refinement.

[0013] The degree of boundary ambiguity is determined based on the tumor boundary representation and optimized to obtain an optimized collaborative hierarchical representation;

[0014] The optimized collaborative hierarchical representation outputs the renal tumor segmentation results, and the renal tumor region contour is generated through pixel-level annotation.

[0015] This invention also provides a system for renal tumor image segmentation based on multi-scale feature extraction, comprising:

[0016] The feature extraction module is used to extract multi-scale features from the kidney tumor image to be processed through a pyramid structure and construct a multi-scale feature set.

[0017] The similarity segmentation module is used to divide the multi-scale feature set into similarity groups using a clustering algorithm, and to determine similar feature groups by combining texture and shape patterns;

[0018] The complementary enhancement module is used to obtain complementary information streams based on the similar feature groups, enhance the interaction weights between different groups through an attention mechanism, and obtain enhanced complementary feature representations.

[0019] The noise reduction optimization module is used to determine the degree of noise interference based on the enhanced complementary feature representation and perform noise reduction to obtain the denoised multi-scale features.

[0020] The feature association module is used to establish an association mapping between groups through the denoised multi-scale features, and to perform feature fusion and boundary refinement using the U-Net architecture to determine the tumor boundary representation.

[0021] The boundary optimization module is used to determine the degree of boundary ambiguity based on the tumor boundary representation and optimize it to obtain the optimized collaborative hierarchical representation;

[0022] The segmentation and annotation module is used to output the renal tumor segmentation result based on the optimized collaborative hierarchical representation, and to generate the renal tumor region contour through pixel-level annotation.

[0023] The beneficial effects of this invention are as follows:

[0024] This invention addresses the challenges of renal tumor segmentation, including significant differences in feature scales, strong noise interference, and blurred boundaries. It constructs a logically coherent solution through multi-scale feature extraction, pyramid structure, clustering, attention mechanisms, denoising, and U-Net architecture optimization. First, a pyramid structure is used to extract coarse and fine-grained features. Clustering algorithms are then used to group similar features, enhancing the complementarity of texture and shape information. An attention mechanism strengthens interaction between groups, and denoising filtering reduces noise interference. U-Net is used for feature fusion and boundary refinement, iteratively optimizing blurred boundaries to ultimately output a precise tumor region contour. This invention significantly improves the accuracy and robustness of renal tumor segmentation through multi-scale feature synergy and boundary refinement, making it suitable for the clinical diagnostic needs of complex CT images. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of a kidney tumor image segmentation method based on multi-scale feature extraction according to an embodiment of the present invention;

[0027] Figure 2 This is a diagram illustrating the pyramid structure and workflow of an embodiment of the present invention;

[0028] Figure 3 This is a diagram illustrating the U-Net structure and workflow of an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of a kidney tumor image segmentation system based on multi-scale feature extraction according to an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] This embodiment provides a kidney tumor image segmentation method based on multi-scale feature extraction, such as... Figure 1 As shown, it includes:

[0033] Multi-scale features are extracted from kidney tumor images to be processed using a pyramid structure, and a multi-scale feature set is constructed.

[0034] Clustering algorithms are used to divide the multi-scale feature set into similarity groups, and texture and shape patterns are combined to determine similar feature groups;

[0035] Based on the similar feature groups, complementary information streams are obtained, and the interaction weights between different groups are enhanced through an attention mechanism to obtain enhanced complementary feature representations.

[0036] Based on the enhanced complementary feature representation, the noise interference level is determined and denoising is performed to obtain the denoised multi-scale features.

[0037] The association mapping between groups is established through the denoised multi-scale features, and the tumor boundary representation is determined by using the U-Net architecture for feature fusion and boundary refinement.

[0038] The degree of boundary ambiguity is determined based on the tumor boundary representation and optimized to obtain an optimized collaborative hierarchical representation;

[0039] The optimized collaborative hierarchical representation outputs the renal tumor segmentation results, and the renal tumor region contour is generated through pixel-level annotation.

[0040] Specifically, this embodiment addresses the challenges of renal tumor segmentation characterized by large differences in feature scales, strong noise interference, and blurred boundaries. It constructs a logically coherent solution through multi-scale feature extraction, pyramid structure, clustering, attention mechanism, denoising, and U-Net architecture optimization. First, a pyramid structure is used to extract coarse and fine-grained features. Then, a clustering algorithm is used to group similar features, enhancing the complementarity of texture and shape information. An attention mechanism strengthens interaction between groups, and denoising filtering reduces noise interference. U-Net is used for feature fusion and boundary refinement, iteratively optimizing blurred boundaries to ultimately output a precise tumor region contour. This embodiment significantly improves the accuracy and robustness of renal tumor segmentation through collaborative optimization of multi-scale features and boundary refinement, making it suitable for the clinical diagnostic needs of complex CT images.

[0041] Furthermore, multi-scale features are extracted from the renal tumor images to be processed using a pyramid structure, and a multi-scale feature set is constructed, including:

[0042] The kidney tumor image to be processed is subjected to grayscale normalization to obtain a standardized image;

[0043] Based on the standardized images, a multi-scale image sequence is constructed using the Gaussian pyramid algorithm to obtain coarse-grained and fine-grained hierarchical images;

[0044] Extract the edge gradient information of the coarse-grained layer image to obtain a coarse-grained edge representation;

[0045] The coarse-grained edge representation is fused into the fine-grained hierarchical image, and spatial features are extracted from the fused representation using a convolutional neural network to obtain a multi-scale spatial feature set.

[0046] Specifically, such as Figure 2 As shown, this embodiment extracts multi-scale features from the input renal tumor CT image. First, the original CT image is preprocessed, assuming an input image resolution of 512×512 pixels and a grayscale value range of 0-4095. Gaussian filtering (standard deviation σ=1.5) is used to smooth the image and reduce noise interference. Simultaneously, histogram equalization is used to enhance image contrast, ensuring a more significant grayscale difference between the tumor region and normal tissue. Next, a four-layer Gaussian pyramid is constructed by extracting coarse-grained and fine-grained hierarchical representations using a pyramid structure. The original image is used as layer 0 (512×512). Downsampling (halving the resolution each time to generate images of 256×256, 128×128, and 64×64 respectively) forms layers 1 to 3. Features are extracted from each layer using a convolutional neural network (CNN) such as ResNet-18, with a kernel size of 3×3 and a stride of 1, generating feature maps at the corresponding scale.

[0047] For example, layer 0 outputs a 64-channel feature map with a size of 512×512; layer 3 outputs a 256-channel feature map with a size of 64×64. To fuse multi-scale features, the feature maps of each layer are upsampled to the original resolution (512×512), and bilinear interpolation is used to ensure smooth feature transitions. Subsequently, through feature concatenation, the feature maps of all layers are merged along the channel dimension to obtain a preliminary multi-scale feature set with dimensions of 512×512×(64+128+192+256). To ensure the effectiveness of the feature set, principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional features, resulting in a feature dimension of approximately 512×512×100.

[0048] Among them, coarse-grained features (low-resolution layers) capture the overall morphology of the tumor, such as the outline of a tumor approximately 3.5 cm in size; fine-grained features (high-resolution layers) highlight local details, such as the gray-scale gradient changes at the tumor edge (gradient values ​​approximately 200-300).

[0049] Furthermore, a clustering algorithm is used to divide the multi-scale feature set into similarity groups, and similar feature groups are determined by combining texture and shape patterns, including:

[0050] The K-means clustering algorithm is used to group the multi-scale feature set to obtain preliminary similar groups;

[0051] The feature similarity within the preliminary similarity group is calculated by texture feature analysis, and the corresponding preliminary similarity groups are merged according to the similarity to obtain the texture adjustment group;

[0052] Based on shape features, extract key points within the texture adjustment group to obtain a set of shape key points, and use principal component analysis to reduce the dimensionality of the set of shape key points to obtain a dimensionality-reduced shape set.

[0053] Determine the matching degree between the dimensionality reduction shape set and the texture adjustment group, and determine the similar feature group based on the matching degree.

[0054] Specifically, in this embodiment, to divide similarity groups, the K-means clustering algorithm is used. After several iterations, five groups are obtained. By analyzing the center vectors of these groups, texture feature analysis is introduced, and the gray-level co-occurrence matrix is ​​used to calculate the average contrast. At the same time, shape feature extraction uses the Hu invariant moment descriptor to calculate the invariant moment value. Based on these quantitative indicators, five similar feature groups are finally determined. The average cosine similarity within each group is above 0.85, ensuring the compactness and separability of the groups.

[0055] Furthermore, complementary information streams are obtained based on the similar feature groups, and the interaction weights between different groups are enhanced through an attention mechanism to obtain enhanced complementary feature representations, including:

[0056] Complementary information streams are obtained from the similar feature groups, and an attention mechanism is used to calculate the interaction weights between groups.

[0057] The interaction weights are used to enhance inter-group interaction, and complementary information flows are fused based on the enhanced inter-group interaction to obtain enhanced complementary feature representations.

[0058] Specifically, this embodiment uses cosine similarity to calculate complementarity and obtain complementary information streams from similar feature groups. For example, the color texture information of one group is complementary to the distance gradient information of another group. If the complementarity is higher than a threshold, the complementary information stream is extracted as an information difference set, and principal component analysis (PCA) is used for dimensionality reduction to reduce noise. Next, the interaction weights between different groups are enhanced through an attention mechanism. For example, a multi-head self-attention model is used to calculate the attention score for the complementary information stream, which increases the interaction weights. The enhancement process involves LayerNorm and residual connections to stabilize the gradient. Finally, the enhanced attention output is fused with the original similar group features, and a weighted summation is used to enhance the complementary feature representation: α·attention output + (1-α)·original features.

[0059] Furthermore, the noise interference level is determined based on the enhanced complementary feature representation, and denoising is performed to obtain denoised multi-scale features, including:

[0060] The noise interference level is determined based on the enhanced complementary feature representation. If the noise level exceeds a preset threshold, the location of the noise point is determined, and the noise point is processed by Gaussian filtering to obtain a denoised image.

[0061] The denoised image is decomposed by wavelet transform to obtain multi-scale sub-bands, and the enhanced complementary feature representations are fused to obtain the denoised multi-scale features.

[0062] Specifically, this embodiment determines the degree of image noise interference based on the enhanced complementary feature representation, using Peak Signal-to-Noise Ratio (PSNR) as the evaluation metric. Assuming a calculated PSNR of 25dB, the preset threshold is 30dB. A PSNR below the threshold indicates severe noise interference, requiring denoising. The denoising algorithm employs Non-Local Means (NLM), setting the search window to 21x21, the similarity window to 7x7, and the smoothing parameter h=10. Noise points are filtered to generate a denoised feature map. After denoising, multi-scale features are recalculated using wavelet transform (Haar wavelet, decomposition level = 3) to extract low-frequency and high-frequency sub-bands, obtaining a multi-scale feature representation containing edge and texture information. Finally, weighted fusion is used to integrate these features into a single feature representation.

[0063] Furthermore, an association mapping between groups is established using the denoised multi-scale features. A U-Net architecture is employed for feature fusion and boundary refinement to determine the tumor boundary representation, including:

[0064] Based on each scale layer of the denoised multi-scale features, the association mapping between groups is calculated to obtain the association mapping matrix, and highly associated groups are selected according to the association mapping matrix to obtain the selected group set.

[0065] The selected group set is fused using the U-Net architecture. The fused features are then refined through convolutional operations to obtain a refined boundary map and extract the tumor boundary representation.

[0066] Specifically, this embodiment first uses a global attention mechanism (Non-Local Neural Network) based on the denoised multi-scale features to calculate the correlation matrix between pixels in the feature map. The matrix dimension is N×N (N is the total number of pixels), and weights are generated through Softmax normalization to enhance the feature association of the tumor region. Subsequently, as... Figure 3 As shown, a U-Net architecture is used for feature fusion. The encoder contains four convolutional layers, each using a 3×3 kernel, with channel numbers of 64, 128, 256, 512, and 1024 respectively. The decoder concatenates the encoder features through skip connections to fuse multi-scale information. An attention gate mechanism is added to the skip connections, and the weights are normalized to [0,1] using a sigmoid function, increasing the weight of tumor boundary features by approximately 30%. Finally, boundary refinement uses a conditional random field (CRF), taking the probability map output by U-Net as input, setting the Gaussian kernel parameter σ=5, and iterating five times to optimize the boundary. The Dice coefficient of the tumor boundary increases from 0.85 to 0.92.

[0067] Furthermore, based on the tumor boundary representation, the degree of boundary ambiguity is determined and optimized to obtain an optimized collaborative hierarchical representation, including:

[0068] Edge pixel sequences are extracted from the tumor boundary representation and processed using a Gaussian filter to obtain smooth edge sequences;

[0069] The gradient magnitude value is calculated based on the smooth edge sequence to obtain a set of low gradient positions. If the number of consecutive positions in the set of low gradient positions exceeds a preset consecutive threshold, it is marked as a fuzzy candidate region.

[0070] Calculate the pixel variation coefficient within the blurred candidate region, and add the blurred candidate regions with variation coefficients higher than the preset variation threshold to the blurred region set.

[0071] Based on the set of fuzzy regions, the boundary points of the regions are obtained. The random forest algorithm is used to optimize the association mapping of the boundary points of the regions. The positions of the fuzzy candidate regions in the smooth edge sequence are iteratively adjusted to obtain the adjusted edge sequence.

[0072] Based on the adjusted edge sequence, a multi-layer node graph representation is constructed. The input graph convolutional network is processed to obtain a preliminary collaborative hierarchical representation. If the proportion of fuzzy regions in the preliminary collaborative hierarchical representation is higher than a preset proportion threshold, the boundary point association mapping is repeatedly optimized to obtain an optimized collaborative hierarchical representation.

[0073] Specifically, this embodiment, based on a refined tumor boundary representation, determines whether a boundary ambiguity problem exists. Fourier transform analysis of the spectral energy distribution is used to calculate the proportion of high-frequency component energy. If the proportion is below 0.3, it is considered clear; otherwise, ambiguity exists. The analysis process involves performing discrete Fourier transform on the boundary pixel values, extracting the amplitude spectrum, and integrating the high-frequency region. The result shows a ambiguity index of 0.42, confirming the existence of a ambiguity problem. Subsequently, the proportion of the ambiguity region is evaluated by using a threshold segmentation method to identify uncertain pixels, i.e., the transition region with gray values ​​between 0.4 and 0.6. The area of ​​this region is calculated to be 0.28 of the total boundary area. If it is higher than a preset threshold of 0.25, iterative optimization of the association mapping is triggered. A graph neural network (GNN) is used to construct a pixel-level association graph, where nodes are boundary pixels, and edge weights are calculated using a Gaussian similarity function. The initial iteration uses an Adam optimizer with a learning rate of 0.001 to update node embeddings. After three iterations, convergence is achieved. The optimization process analyzes the improvement of the cosine similarity of the embedding vectors, reducing ambiguity association errors. Finally, by fusing the GNN output with the original features through a multilayer perceptron, an optimized collaborative hierarchical representation is output, which includes low-level edge details and high-level semantic context.

[0074] Furthermore, based on the optimized collaborative hierarchical representation, the renal tumor segmentation result is output, and the renal tumor region contour is generated through pixel-level annotation, including:

[0075] A first segmentation result is generated based on the optimized collaborative hierarchical representation, and a second segmentation result is determined by using pixel-level annotation technology in the first segmentation result.

[0076] The contour of the renal tumor region is generated from the second segmentation result using a contour extraction algorithm. If the pixel-level accuracy of the renal tumor region contour reaches a preset standard, the contour data of the renal tumor region is output.

[0077] Specifically, this embodiment outputs the renal tumor segmentation result under the optimized collaborative hierarchical representation. A probability map is generated through pixel-level softmax classification, and a binary mask is obtained by binarization. The precise contour is extracted by Canny edge detection, generating a closed contour path with a length of approximately 1250 pixels and a coordinate sequence such as [(100.5,200.3), (101.2,201.1), ...]. Verification shows that the contour Jaccard index is 0.91 and the boundary Hausdorff distance is 2.3 pixels, ensuring accurate clinical annotation of the tumor region.

[0078] To verify the effectiveness of the proposed method, experiments were conducted on the publicly available kidney tumor segmentation dataset KiTS2019. The method was compared with current mainstream methods, and the results show that the average Dice coefficient of the proposed method reaches 0.92±0.03, which is approximately 8.2% higher than the benchmark U-Net method (0.85±0.05). In terms of the Hausdorff distance metric, which measures boundary segmentation accuracy, the proposed method achieves a distance of 2.3 mm, significantly better than U-Net's 3.5 mm. These data fully demonstrate the significant effect of the proposed method in improving the accuracy and robustness of kidney tumor segmentation.

[0079] This embodiment also provides a renal tumor image segmentation system based on multi-scale feature extraction, such as... Figure 4 As shown, it includes:

[0080] The feature extraction module is used to extract multi-scale features from the kidney tumor image to be processed through a pyramid structure and construct a multi-scale feature set.

[0081] The similarity segmentation module is used to divide the multi-scale feature set into similarity groups using a clustering algorithm, and to determine similar feature groups by combining texture and shape patterns;

[0082] The complementary enhancement module is used to obtain complementary information streams based on the similar feature groups, enhance the interaction weights between different groups through an attention mechanism, and obtain enhanced complementary feature representations.

[0083] The noise reduction optimization module is used to determine the degree of noise interference based on the enhanced complementary feature representation and perform noise reduction to obtain the denoised multi-scale features.

[0084] The feature association module is used to establish an association mapping between groups through the denoised multi-scale features, and to perform feature fusion and boundary refinement using the U-Net architecture to determine the tumor boundary representation.

[0085] The boundary optimization module is used to determine the degree of boundary ambiguity based on the tumor boundary representation and optimize it to obtain the optimized collaborative hierarchical representation;

[0086] The segmentation and annotation module is used to output the renal tumor segmentation result based on the optimized collaborative hierarchical representation, and to generate the renal tumor region contour through pixel-level annotation.

[0087] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for segmenting renal tumor images based on multi-scale feature extraction, characterized in that, include: Multi-scale features are extracted from kidney tumor images to be processed using a pyramid structure, and a multi-scale feature set is constructed. Clustering algorithms are used to divide the multi-scale feature set into similarity groups, and texture and shape patterns are combined to determine similar feature groups; Based on the similar feature groups, complementary information streams are obtained, and the interaction weights between different groups are enhanced through an attention mechanism to obtain enhanced complementary feature representations. Based on the enhanced complementary feature representation, the noise interference level is determined and denoising is performed to obtain the denoised multi-scale features. The association mapping between groups is established through the denoised multi-scale features, and the tumor boundary representation is determined by using the U-Net architecture for feature fusion and boundary refinement. The degree of boundary ambiguity is determined based on the tumor boundary representation and optimized to obtain an optimized collaborative hierarchical representation; The optimized collaborative hierarchical representation outputs the renal tumor segmentation results, and the renal tumor region contour is generated through pixel-level annotation.

2. The renal tumor image segmentation method based on multi-scale feature extraction according to claim 1, characterized in that, Multi-scale features were extracted from the renal tumor images to be processed using a pyramid structure, and a multi-scale feature set was constructed, including: The kidney tumor image to be processed is subjected to grayscale normalization to obtain a standardized image; Based on the standardized images, a multi-scale image sequence is constructed using the Gaussian pyramid algorithm to obtain coarse-grained and fine-grained hierarchical images; Extract the edge gradient information of the coarse-grained layer image to obtain a coarse-grained edge representation; The coarse-grained edge representation is fused into the fine-grained hierarchical image, and spatial features are extracted from the fused representation using a convolutional neural network to obtain a multi-scale spatial feature set.

3. The renal tumor image segmentation method based on multi-scale feature extraction according to claim 1, characterized in that, Clustering algorithms are used to divide the multi-scale feature set into similarity groups, and texture and shape patterns are combined to determine similar feature groups, including: The K-means clustering algorithm is used to group the multi-scale feature set to obtain preliminary similar groups; The feature similarity within the preliminary similarity group is calculated by texture feature analysis, and the corresponding preliminary similarity groups are merged according to the similarity to obtain the texture adjustment group; Based on shape features, extract key points within the texture adjustment group to obtain a set of shape key points, and use principal component analysis to reduce the dimensionality of the set of shape key points to obtain a dimensionality-reduced shape set. Determine the matching degree between the dimensionality reduction shape set and the texture adjustment group, and determine the similar feature group based on the matching degree.

4. The renal tumor image segmentation method based on multi-scale feature extraction according to claim 1, characterized in that, Based on the aforementioned similar feature groups, complementary information streams are obtained. An attention mechanism is used to enhance the interaction weights between different groups, resulting in enhanced complementary feature representations, including: Complementary information streams are obtained from the similar feature groups, and an attention mechanism is used to calculate the interaction weights between groups. The interaction weights are used to enhance inter-group interaction, and complementary information flows are fused based on the enhanced inter-group interaction to obtain enhanced complementary feature representations.

5. The renal tumor image segmentation method based on multi-scale feature extraction according to claim 1, characterized in that, The noise interference level is determined based on the enhanced complementary feature representation, and denoising is performed to obtain denoised multi-scale features, including: The noise interference level is determined based on the enhanced complementary feature representation. If the noise level exceeds a preset threshold, the location of the noise point is determined, and the noise point is processed by a Gaussian filtering algorithm to obtain a denoised image. The denoised image is decomposed by wavelet transform to obtain multi-scale sub-bands, and the enhanced complementary feature representations are fused to obtain the denoised multi-scale features.

6. The renal tumor image segmentation method based on multi-scale feature extraction according to claim 1, characterized in that, The association mapping between groups is established using the denoised multi-scale features. A U-Net architecture is used for feature fusion and boundary refinement to determine the tumor boundary representation, including: Based on each scale layer of the denoised multi-scale features, the association mapping between groups is calculated to obtain the association mapping matrix, and highly associated groups are selected according to the association mapping matrix to obtain the selected group set. The selected group set is fused using the U-Net architecture. The fused features are then refined through convolutional operations to obtain a refined boundary map and extract the tumor boundary representation.

7. The renal tumor image segmentation method based on multi-scale feature extraction according to claim 1, characterized in that, The degree of boundary ambiguity is determined based on the tumor boundary representation and optimized to obtain an optimized collaborative hierarchical representation, including: Edge pixel sequences are extracted from the tumor boundary representation and processed using a Gaussian filter to obtain smooth edge sequences; The gradient magnitude value is calculated based on the smooth edge sequence to obtain a set of low gradient positions. If the number of consecutive positions in the set of low gradient positions exceeds a preset consecutive threshold, it is marked as a fuzzy candidate region. Calculate the pixel variation coefficient within the blurred candidate region, and add the blurred candidate regions with variation coefficients higher than the preset variation threshold to the blurred region set. Based on the set of fuzzy regions, the boundary points of the regions are obtained. The random forest algorithm is used to optimize the association mapping of the boundary points of the regions. The positions of the fuzzy candidate regions in the smooth edge sequence are iteratively adjusted to obtain the adjusted edge sequence. Based on the adjusted edge sequence, a multi-layer node graph representation is constructed. The input graph convolutional network is processed to obtain a preliminary collaborative hierarchical representation. If the proportion of fuzzy regions in the preliminary collaborative hierarchical representation is higher than a preset proportion threshold, the boundary point association mapping is repeatedly optimized to obtain an optimized collaborative hierarchical representation.

8. The renal tumor image segmentation method based on multi-scale feature extraction according to claim 1, characterized in that, The optimized collaborative hierarchical representation outputs the renal tumor segmentation result, and the renal tumor region contour is generated through pixel-level annotation, including: A first segmentation result is generated based on the optimized collaborative hierarchical representation, and a second segmentation result is determined by using pixel-level annotation technology in the first segmentation result. The contour of the renal tumor region is generated from the second segmentation result using a contour extraction algorithm. If the pixel-level accuracy of the renal tumor region contour reaches a preset standard, the contour data of the renal tumor region is output.

9. A renal tumor image segmentation system based on multi-scale feature extraction, characterized in that, include: The feature extraction module is used to extract multi-scale features from the kidney tumor image to be processed through a pyramid structure and construct a multi-scale feature set. The similarity segmentation module is used to divide the multi-scale feature set into similarity groups using a clustering algorithm, and to determine similar feature groups by combining texture and shape patterns; The complementary enhancement module is used to obtain complementary information streams based on the similar feature groups, enhance the interaction weights between different groups through an attention mechanism, and obtain enhanced complementary feature representations. The noise reduction optimization module is used to determine the degree of noise interference based on the enhanced complementary feature representation and perform noise reduction to obtain the denoised multi-scale features. The feature association module is used to establish an association mapping between groups through the denoised multi-scale features, and to perform feature fusion and boundary refinement using the U-Net architecture to determine the tumor boundary representation. The boundary optimization module is used to determine the degree of boundary ambiguity based on the tumor boundary representation and optimize it to obtain the optimized collaborative hierarchical representation; The segmentation and annotation module is used to output the renal tumor segmentation result based on the optimized collaborative hierarchical representation, and to generate the renal tumor region contour through pixel-level annotation.

Images