Intelligent segmentation system for kidney tumor in CT (Computed Tomography) image

Abstract

The invention relates to the technical field of medical image processing, and particularly discloses an intelligent segmentation system for a kidney tumor in a CT image, and the system comprises an image preprocessing module, an ROI automatic detection module, a kidney tumor segmentation module, a segmentation result post-processing module, a visual interaction module, and a model updating module. The system improves image quality through normalization and filtering, uses a target detection network to position a kidney area, realizes precise tumor segmentation based on an integrated channel attention mechanism and a U-Net network guided by structure priori, optimizes a boundary in combination with a conditional random field, adopts a multi-plane fusion strategy to improve three-dimensional consistency, and calculates tumor volume. The system supports federal learning update, improves the cross-mechanism generalization ability, has the advantages of high segmentation precision, strong boundary reducibility, good structural rationality, strong clinical deployment and the like, and is suitable for kidney tumor auxiliary diagnosis and quantitative analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a CT image kidney tumor intelligent segmentation system. BACKGROUND

[0002] Kidney tumor is a common clinical entity of urinary system malignancy, which usually lacks obvious symptoms in early stage, thus often relies on imaging methods for discovery and diagnosis. Computed tomography (CT) is the main imaging method for current clinical evaluation of kidney anatomical structure and tumor morphology, which has important value in tumor location, boundary, size judgment and preoperative planning.

[0003] In clinical practice, doctors usually observe CT images layer by layer, and manually delineate or semi-automatically label tumor regions, but this process has the following problems: first, the subjectivity of doctors is strong, and the consistency between different operators is poor; second, in the face of large amounts of image data, manual segmentation is inefficient and repetitive; third, the tumor boundary often presents a fuzzy transition, which is easy to be confused with the surrounding kidney tissue, and traditional methods based on threshold, region growing and other methods are difficult to accurately segment. In addition, the gray scale distribution difference and anatomical structure variation between different devices and patients also lead to insufficient generalization ability of existing rule-based segmentation algorithms.

[0004] In recent years, deep learning technology has developed rapidly, and models such as convolutional neural network (CNN) and U-Net structure have made significant progress in medical image segmentation, which can automatically learn feature distribution, identify boundaries and lesion regions from training data. However, there are still several technical bottlenecks that need to be solved: first, the model only relies on image texture learning, lacks structural prior guidance, and is prone to anatomically unreasonable segmentation results; second, the model is not sensitive to low-contrast regions and small-volume tumors; third, the training data is limited, the model has insufficient cross-institutional generalization ability, and there are obstacles to data privacy transmission.

[0005] In summary, the current intelligent segmentation of kidney tumors in CT images still faces multiple challenges such as low accuracy, insufficient generalization, lack of structural constraints and lack of practical deployment feasibility.

[0006] Therefore, there is an urgent need to design a CT image kidney tumor intelligent segmentation system to solve the above problems. SUMMARY

[0007] The present application aims to provide a CT image kidney tumor intelligent segmentation system to solve the problems raised in the background.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0009] The application discloses a CT image kidney tumor intelligent segmentation system, which comprises an image preprocessing module, an ROI automatic detection module, a kidney tumor segmentation module, a segmentation result post-processing module, a visual interaction module and a model updating module.

[0010] The image preprocessing module is used for converting an original CT image I raw into a normalized image I norm , and further filtering to generate a denoising image I denoise .

[0011] The ROI automatic detection module is used for outputting kidney region rectangular frame coordinates (x min , y min , x max , y max ) and extracting an image block.

[0012] The kidney tumor segmentation module is based on a multi-scale U-Net network and is trained by using the following joint loss function:

[0013] L = alpha * L Dice + beta * L BCE + gamma * L SP

[0014] wherein L represents a total loss function, L Dice represents a Dice coefficient loss, L BCE represents a binary cross-entropy loss, L SP represents a structure prior loss, alpha, beta and gamma represent loss weight coefficients, alpha + beta + gamma = 1 and alpha, beta and gamma are in the range of [0, 1].

[0015] As a preferred technical scheme of the application, the Dice coefficient loss L Dice is expressed as:

[0016]

[0017] wherein i represents a pixel index, 1 <= i <= N, N represents the total number of pixels in the image, p i represents the value of the i-th pixel in a model prediction result, and ranges from 0 to 1, g i represents the true label value of the i-th pixel, and is 0 or 1, wherein 1 represents a tumor and 0 represents a non-tumor.

[0018] As a preferred technical scheme of the application, the structure prior loss L SP is defined as:

[0019]

[0020] wherein P s(i): the prior probability value of the i-th pixel in the structure template belonging to the kidney tissue, ranging from 0 to 1; Q s (i): the probability value of the i-th pixel in the model prediction belonging to the tumor region, ranging from 0 to 1; N: the total number of pixels in the image; log(·): natural logarithm function; L SP : Kullback-Leibler divergence as a structure prior guided loss.

[0021] As a preferred technical solution of the application, in the channel attention mechanism of the kidney tumor segmentation module, the weight w c of the c-th channel is calculated as:

[0022] w c =σ(MLP(GAP(F c )));

[0023] Where: F c : the c-th feature channel tensor, GAP(·): global average pooling function, output channel average value scalar; MLP(·): multi-layer perceptron function (containing at least two layers of fully connected neural network); σ(·): Sigmoid activation function, output value range [0, 1]; w c : the attention weight of the c-th channel.

[0024] As a preferred technical solution of the application, the image preprocessing module performs the following processing: gray scale normalization:

[0025]

[0026] Where: I raw (x,y,z): original voxel gray value; μ: image gray mean; σ: image gray standard deviation; (x,y,z): voxel coordinates.

[0027] Two-dimensional Gaussian wave:

[0028]

[0029] Where: G(i,j,l): the weight of the Gaussian kernel function at the (i,j,l) point, satisfying ∑G(i,j,l) = 1; k: filter radius; I denoise (x,y,z): filtered voxel value.

[0030] As a preferred technical solution of the application, the ROI automatic detection module uses the YOLOv5 target detection network to output the rectangular frame coordinates: x min ,y min : upper left corner coordinates, x max ,y max: lower right corner coordinates, the above-mentioned coordinates are used to extract an image block containing a kidney structure for subsequent segmentation processing.

[0031] As a preferred technical solution of the present application, the segmentation result post-processing module adopts a conditional random field model, and a cost function is defined as:

[0032]

[0033] Wherein: x i : label of the i th pixel, 0 or 1; ψ u (x i ): cost function of pixel i, related to model prediction confidence; ψ p (x i , x j ): smoothing penalty term of pixel pair (i, j), considering spatial distance and pixel gray difference.

[0034] As a preferred technical solution of the present application, the system performs weighted fusion on three view segmentation results, and the formula is:

[0035] M final = λ 1 M axial + λ 2 M coronal + λ 3 M sagittal

[0036] Wherein: M axial , M coronal , M sagittal : segmentation masks in axial, coronal and sagittal directions respectively; λ 1, λ 2, λ 3: corresponding fusion weights, satisfying λ 1 + λ 2 + λ 3 = 1, each λ i ∈ [0, 1]; M final : three-dimensional tumor mask graph finally obtained by fusion.

[0037] As a preferred technical solution of the present application, the visualization interaction module is used to automatically calculate the tumor volume V, and the calculation formula is:

[0038]

[0039] Wherein: v voxel : physical volume of each voxel, unit: mm 3 ; m i : segmentation label of the i th voxel, 1 represents tumor, and 0 represents non-tumor; N: number of all voxels.

[0040] As a preferred technical solution of the present application, the model updating module uses a federal average algorithm FedAvq to aggregate multiple client models, and the calculation formula is:

[0041]

[0042] wherein: the model weight obtained by the kth client in the tth round of training; w t+1 : the aggregated model weight of the t+1th round; n k : the amount of data used by the kth client; K: the total number of clients; the total number of global samples.

[0043] Compared with the prior art, the beneficial effects of the present application are:

[0044] 1. The present application significantly improves the automatic segmentation accuracy and stability of the renal tumor region by integrating image preprocessing, kidney target detection, attention-enhanced multi-scale segmentation network and structure prior guidance mechanism. Compared with the traditional image segmentation method based on gray threshold or region growing, the present system is more robust in identifying low-contrast boundaries and small-volume lesions, can realize pixel-level accurate delineation, reduce the degree of artificial dependence and alleviate the burden of doctors.

[0045] 2. The present application effectively improves the consistency and boundary continuity of three-dimensional segmentation results by introducing a multi-plane perspective fusion strategy and CRF boundary optimization mechanism. The present system solves the problem of artifacts or discontinuity in single-direction segmentation. The three-way fusion method can adaptively integrate different section information and constrain the anatomical structure in combination with structure priors, so that the segmentation result is more consistent with the actual clinical anatomy. In addition, the system supports automatic calculation of tumor volume at the voxel level, and the output result can directly generate a structured report, improving the efficiency of clinical data visualization and management.

[0046] 3. The present system also integrates a model updating module based on federated learning, which supports cross-institutional model optimization under the premise of protecting patient data privacy, has good engineering deployment capability and generalization performance. The overall architecture of the system is complete and the process is closed loop, which is suitable for deployment in hospital PACS environment and large-scale image data processing, and has wide application prospect and promotion value in scenes such as renal tumor screening, pre-treatment evaluation and postoperative monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce and describe the drawings needed in the description of the embodiments of the present application or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any inventive labor.

[0048] Figure 1 is a system block diagram of a CT image renal tumor intelligent segmentation system of the present application;

[0049] Figure 2 is a segmentation result image of a kidney tumor intelligent segmentation system in a CT image. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the following described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. In order to more clearly explain and illustrate the technical solutions and implementation modes of the present application, the following introduces the specific examples preferred for implementing the technical solutions of the present application.

[0051] In order to realize high-precision and automatic segmentation of the kidney tumor region in the CT image, the present application provides a kidney tumor intelligent segmentation system in a CT image, which integrates multiple function modules such as image preprocessing, target detection, deep learning segmentation, boundary optimization, three-dimensional visualization and model self-updating, and constructs a complete technical process from image input to structured output, which has strong adaptability and clinical practical value. In actual application, the system first pre-processes the input original CT image.

[0052] The original image I raw is subjected to gray scale normalization processing by the mean μ and the standard deviation σ to obtain the normalized image Then, noise suppression is performed by a three-dimensional Gaussian filter, and random high-frequency information is filtered out by using the kernel function G(i,j,l) to obtain a smoothed image I denoise This process enhances the texture continuity of the kidney and tumor region in the image, which is beneficial to subsequent target detection and segmentation accuracy improvement. The pre-processed image will be sent frame by frame into the ROI automatic detection module. This module uses the YOLOv5 target detection network to identify the kidney region of each CT slice image. The detection result is expressed in the form of a rectangular frame coordinate, which contains the upper left corner coordinate (x min ,y min ) and the lower right corner coordinate (x max ,y max ). According to this, the region of interest (ROI) is cropped for input of the segmentation model. This strategy can significantly reduce the network attention area, reduce background interference and improve the model training efficiency.

[0053] The core module of the system is the kidney tumor segmentation module, which uses an improved multi-scale U-Net deep network structure. The network integrates a channel attention mechanism in the encoding stage, and the feature map F cFirstly, global average pooling (GAP) is used to extract statistical features, then multi-layer perceptron (MLP) is used for nonlinear mapping, and the channel weight w is output through the Sigmoid activation function c = σ(MLP(GAP(F c ))) to realize the weighted enhancement of key semantic channels, and the segmentation model outputs the prediction probability map Q s (i), and the joint loss function L = αL Dice + βL BCE + γL SP is used for supervised learning, where the Dice loss L Dice is used to measure the overlap between the predicted mask and the real label, which is defined as:

[0054]

[0055] where N is the total number of pixels, p i is the predicted value of the i-th pixel, g i is the corresponding real value, and the structure prior guidance loss L SP is introduced to improve the model's ability to capture structural information, which is defined by Kullback-Leibler divergence as follows:

[0056]

[0057] where P s (i) is the kidney probability of the i-th pixel in the structure template, and Q s (i) is the output value of the corresponding pixel of the model, which can guide the network to converge in a reasonable structure, and the weights α, β, γ of each loss term satisfy the normalization constraint α + β + γ = 1.

[0058] To improve the edge segmentation quality, the system introduces a post-processing module based on conditional random field (CRF) model to refine the boundary of the segmentation mask. The CRF cost function is represented as:

[0059] E(x) = ∑ i ψ u (x i ) + ∑ i<j ψ p (x i , x j );

[0060] where ψ u (x i ) is the pixel classification confidence cost, and ψ p (x i , x j) is a smooth constraint term between adjacent pixels, considering the spatial position and intensity similarity between pixels, which can effectively repair misjudged edges and hole regions.

[0061] To improve the three-dimensional visual effect and the accuracy of clinical analysis, the system supports independent segmentation of the axial, coronal and sagittal slices, and fusion to obtain the final three-dimensional mask M final , the fusion method is:

[0062] M final =λ1M axial +λ2M coronal +λ3M sagittal ;

[0063] where M axial , M coronal , M sagittal are the segmentation results of the three directions, λ1, λ2, λ3∈[0,1], satisfying λ1+λ2+λ3=1.

[0064] The fusion result is superimposed with the original image through the visualization module, and the user can perform three-dimensional rotation, layer-by-layer browsing and other interactive operations.

[0065] In addition, the visualization module also provides automatic tumor volume calculation function, and its volume V is determined in the following way:

[0066]

[0067] where v voxel is the single voxel volume (unit: mm 3 ), m i is whether the ith voxel belongs to the tumor (1 or 0), and N is the total number of voxels. The calculation result is automatically output to the structured report (DICOM-SR format) for doctor's diagnosis reference.

[0068] To improve the generalization ability of the system and protect data privacy, the system also includes a model updating module. This module uses a federated learning strategy to perform distributed model updating without transmitting original image data. Its aggregation process uses the FedAvg algorithm, and the aggregation formula is as follows:

[0069]

[0070] where, is the model parameter of the kth client in the tth round, n k is the number of its samples, is the total number of samples, and w t+1 is the new global parameter.

[0071] This way supports inter-hospital model co-evolution and is suitable for multi-center deployment environment.

[0072] The embodiment has good engineering feasibility, and has completed verification experiments on several CT image data sets of kidney tumors. The results show that the system has significant advantages in segmentation accuracy, boundary restoration degree, processing speed and the like compared with traditional methods, and has practical popularization and application value.

[0073] Embodiment: The embodiment provides a kidney tumor intelligent segmentation and volume calculation process based on a CT image, and specifically as follows:

[0074] The embodiment is tested and verified based on an abdominal enhanced CT image of a kidney tumor patient from a certain third-grade class-A hospital. The CT image size is 512x512x120, the layer thickness is 5mm, the spatial resolution is 0.75mmx0.75mm, and the data format is DICOM. The system aims to automatically identify the kidney region, segment the tumor, optimize the boundary, fuse the perspective, calculate the volume, and generate a structured report. The system first performs gray scale normalization processing on the original CT image I raw (x,y,z).

[0075] The statistical full-image gray mean μ=58.7 and standard deviation σ=21.4, and the image normalization formula is:

[0076]

[0077] For example, the gray value of a voxel is I raw (x,y,z)=95.3, and after normalization, it is:

[0078]

[0079] Then, three-dimensional Gaussian filtering is performed for denoising, the kernel radius k=1, the center value of the three-dimensional Gaussian kernel G(i,j,l) is 0.15, and the average value of the remaining 26 neighbors is about 0.015, and the sum satisfies ∑G=1. Then, the value of a certain voxel after filtering is:

[0080] I denoise ≈0.15·1.71+0.015·104.2≈0.257+1.563=1.82;

[0081] The denoised image is input into the YOLOv5 target detection module, and the kidney region is identified for each layer slice.

[0082] The detection result outputs the rectangular frame coordinates as (x min ,y min ) = (122, 84), (x max ,y max) = (402, 360), from which the ROI image block with the size of 280x276x120 is cropped, and the cropped image is sent to the kidney tumor segmentation module.

[0083] The module uses a multi-scale U-Net network structure, and integrates a channel attention mechanism in each down-sampling path. Taking channel 3 as an example, the feature map mean value is F3 = 0.64, the MLP output is z = 0.83, and the attention weight is:

[0084]

[0085] The model output probability map Q s (i) and is trained based on a joint loss function:

[0086] L = a · L Dice + β · L BCE + γ · L SP ;

[0087] Wherein a = 0.5, β = 0.3, and γ = 0.2. The loss is calculated as follows:

[0088] Dice loss (prediction of a slice):

[0089] ∑p i g i = 13400, Total pixel number N = 262144:

[0090]

[0091] Structural prior loss (5 pixel examples):

[0092]

[0093]

[0094] The segmentation result is optimized for the edge after CRF post-processing, and the energy function is as follows:

[0095] E(x) = ∑ i ψ u (x i )+ ∑ i<j ψ p (x i ,x j );

[0096] Wherein ψ u is based on the prediction confidence, and ψ p is based on the calculation of the adjacent pixel gray scale and spatial position. The system processes three-direction slices in parallel, and outputs three-view masks: M axial = 0.84, Mcoronal = 0.72, M sagittal = 0.66; weights λ1 = 0.4, λ2 = 0.3, λ3 = 0.3.

[0097] The segmentation probability of this position after fusion is:

[0098] M final = 0.4 * 0.84 + 0.3 * 0.72 + 0.3 * 0.66 = 0.75;

[0099] After completing the three-dimensional segmentation, the system automatically counts the number of tumor voxels N tumor = 4280, and the volume of each voxel is:

[0100] v voxel = 0.75 * 0.75 * 5 = 2.8125 mm 3 ;

[0101] The total volume of the tumor is calculated as:

[0102] V = v voxel * N tumor = 2.8125 * 4280 = 12033.75 mm 3 ≈ 12.03 cm 3 ;

[0103] The results are automatically written into the hospital PACS system report in the DICOM SR structured format.

[0104] The system deployment supports multi-center federated learning updates.

[0105] Taking four hospitals as an example, the data amounts are n1 = 500, n2 = 600, n3 = 400, and n4 = 700, and the last round of model weights are: w (1) = 0.80, w (2) = 0.76, w (3) = 0.82, and w (4) = 0.79; the total sample number n = 2200, and the aggregated model is:

[0106]

[0107] The final aggregated model deployment update is completed, realizing cross-institutional model optimization.

[0108] Figure 2 The image segmentation result schematic diagram of the kidney tumor intelligent segmentation system in the CT image is shown, which includes the visualization comparison of the original CT image and the automatic segmentation result of the system.

[0109] In the figure, the upper left and lower left two images are the original transverse images of the patient's abdominal enhanced CT, respectively corresponding to different slice levels;

[0110] The upper right and lower right images in the figure are the automatic segmentation results of the kidney tumor after the corresponding slices are processed by the system, and the red highlighted area in the figure is the kidney tumor area recognized by the system.

[0111] The segmentation result is obtained by the system described in the application through the following process: first, the original image is normalized and Gaussian denoised by the image preprocessing module, and the kidney region is extracted by the ROI automatic detection module. Subsequently, the region image is sent to the U-Net segmentation network integrated with attention mechanism and structure prior, and combined with the Dice loss, cross-entropy loss and KL divergence structure prior in the embodiment, the joint training is outputed. The pixel-level tumor probability map. By setting the threshold and performing conditional random field (CRF) post-processing, the refinement of the segmentation boundary is realized.

[0112] Taking the upper right image in the figure as an example, in the 58th slice, the system successfully identifies two relatively symmetrical tumor foci in the left and right kidney regions, and the red area accurately envelopes the lesion edge, avoiding misjudgment of the normal kidney parenchyma; in the lower right image, the system identifies a single tumor block located in the middle and posterior segment of the left kidney, and the boundary is smooth and highly consistent with the true contour, showing the robustness of the system to low-contrast tumor regions.

[0113] Further combined with the embodiment, the three-dimensional reconstruction result corresponding to the image participates in the volume calculation process. The system automatically counts the number of voxels in the segmentation result as 4280, and combines the voxel volume v voxel = 0.75 x 0.75 x 5 = 2.8125 mm 3 , the tumor volume is calculated as:

[0114] V = 2.8125 x 4280 = 12033.75 mm 3 ≈ 12.03 cm 3 ;

[0115] The result has been automatically written into a structured report and used for doctor evaluation.

[0116] In summary, Figure 2 The image shows the intelligent recognition, pixel-level segmentation and result visualization ability of the system for kidney tumors in different anatomical sections, further verifying the segmentation accuracy and clinical applicability of the system in real clinical images.

[0117] The content not described in detail in the specification belongs to the prior art known to those skilled in the art, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A system for intelligent segmentation of renal tumors in CT images, characterized in that, Comprise: An image preprocessing module, an ROI automatic detection module, a renal tumor segmentation module, a segmentation result post-processing module, a visualization interaction module and a model updating module, wherein: The image pre-processing module is used to convert the original CT image I raw into a normalized image I norm , and further filter processing to generate a denoised image I denoise ; The ROI automatic detection module is configured to output the kidney region rectangular frame coordinates (x min ,y min ,x max ,y max ) and extract the image block. The renal tumor segmentation module is based on a multi-scale U-Net network and trained by using the following joint loss function: L = a · L Dice + β · L BCE + γ · L SP ; Where: L: total loss function; L Dice : Dice coefficient loss; L BCE : Binary cross-entropy loss; L SP : Structure prior loss; α, β, γ: loss weight coefficients, satisfying α + β + γ = 1 and α, β, γ ∈ [0, 1].

2. The system of claim 1, wherein, The Dice coefficient loss L Dice The expression is: Wherein: i: pixel index, value range 1≤i≤N; N: total number of pixels in the image; p i : the value of the i-th pixel in the model prediction result, ranging from [0, 1]; g i : the true label value of the i-th pixel, taking the value of 0 or 1, 1 representing tumor and 0 representing non-tumor. 3.The CT image kidney tumor intelligent segmentation system according to claim 1, characterized in that, The structure prior loss L SP is defined as: where: P s (i): the prior probability value of the i-th pixel in the structure template belonging to kidney tissue, with a value range of [0, 1]; Q s (i): the probability value of the i-th pixel in the model prediction belonging to the tumor region, with a value range of [0, 1]; N: the total number of pixels in the image; log(·): the natural logarithm function; L SP : Kullback-Leibler divergence as the structure prior guided loss. 4.The CT image kidney tumor intelligent segmentation system according to claim 1, characterized in that, In the channel attention mechanism of the kidney tumor segmentation module, the weight w of the cth channel c The calculation is: w c = σ(MLP(GAP(F c ))) ; where: F c : the c-th feature channel tensor, GAP(·): global average pooling function, output channel average value scalar; MLP(·): multi-layer perception function; σ(·): Sigmoid activation function, output value range is [0, 1]; w c : the c-th channel attention weight. 5.The CT image kidney tumor intelligent segmentation system according to claim 1, wherein, The image preprocessing module performs the following processing: gray scale normalization: where: I raw (x, y, z): original voxel gray value; μ: image gray mean value; σ: image gray standard deviation; (x, y, z): voxel coordinates. Two-dimensional Gaussian wave: where: G(i,j,l): the weight of the Gaussian kernel function at the point (i,j,l), satisfying ∑G(i,j,l)=1; k: filter radius; I denoise (x,y,z): the filtered voxel value. 6.The CT image kidney tumor intelligent segmentation system according to claim 1, characterized in that, The ROI automatic detection module uses a YOLOv5 target detection network to output rectangular frame coordinates: x min ,y min : upper left corner coordinates, x max ,y max : right lower corner coordinates, which are used to extract an image block containing the kidney structure for subsequent segmentation processing. 7.The CT image kidney tumor intelligent segmentation system according to claim 1, wherein, The segmentation result post-processing module adopts a conditional random field model, and the cost function is defined as: E(x) = ∑ i ψ u (x i )+∑ i<j ψ p (x i ,x j ) where: x i : label of the i-th pixel, 0 or 1; ψ u (x i ): cost function of pixel i, related to model prediction confidence; ψ p (x i , x j ): smoothness penalty term for pixel pair (i, j), considering spatial distance and pixel intensity difference. 8.The CT image kidney tumor intelligent segmentation system according to claim 1, wherein, The system also comprises a three-view angle fusion module, which performs weighted fusion on three-view angle segmentation results, and the formula is: M final = λ1M axial + λ2M coronal + λ3M sagittal ; Wherein: M axial ,M coronal ,M sagittal : the segmentation mask in axial, coronal and sagittal directions, respectively; λ1, λ2, λ3: corresponding fusion weights, satisfying λ1+λ2+λ3=1, each λ i ∈[0,1]; M final : the final three-dimensional tumor mask map obtained by fusion. 9.The CT image kidney tumor intelligent segmentation system according to claim 1, wherein, The visualization interaction module is used for automatically calculating the tumor volume V, and the calculation formula is: where: v voxel : physical volume of each voxel, in mm 3 ; m i : segmentation label of the i-th voxel, 1 indicates tumor and 0 indicates non-tumor; N: number of all voxels. 10.The CT image kidney tumor intelligent segmentation system according to claim 1, wherein, The model updating module uses a federated average algorithm FedAvq to aggregate multiple client models, and the calculation formula is: wherein: the model weight obtained by the kth client in the tth round of training; w t+1 : the aggregated model weight of the t+1th round; n k : the amount of data used by the kth client; K: the total number of clients; the total number of global samples.

Images