A ct and mr fusion method for precise segmentation of urinary system tumors

Abstract

The application discloses a CT and MR fusion method for precise segmentation of urinary system tumors, and relates to the technical field of medical diagnosis and image processing. The CT and MR fusion method for precise segmentation of urinary system tumors comprises the following steps: step 1, image preprocessing, step 2, bidirectional multi-scale registration, step 3, cross-modality attention fusion, step 4, edge enhancement type segmentation, and step 5, anatomical constraint post-processing. The method adopts a multi-scale joint model of "marker point coarse registration + bidirectional B-spline fine registration", initializes global alignment through three key anatomical marker points, and optimizes local deformation by using a hybrid metric function, so that the final registration error is less than or equal to 0.5 mm, and the precision is improved by more than 30% compared with the prior art. The effect directly solves the feature fusion space misplacement problem caused by traditional one-way registration, lays a foundation for subsequent precise extraction of tumor features, and realizes precise matching of the structure positioning advantage of CT and the soft tissue recognition advantage of MRI in space.

Description

Technical Field

[0001] This invention relates to the field of medical diagnostics and image processing technology, and in particular to a CT and MR fusion method for precise segmentation of urinary system tumors. Background Technology

[0002] Accurate diagnosis and staging of urinary system tumors (such as kidney cancer and bladder cancer) are core prerequisites for clinical treatment decisions. CT and MRI, as two core imaging technologies, have irreplaceable complementary value in tumor assessment: CT has excellent anatomical resolution, clearly displaying tumor calcifications, kidney contours, and spatial relationships with surrounding organs, but its ability to identify lesion details such as the extent of tumor infiltration and soft tissue edema is limited; MRI, with its high soft tissue contrast, can accurately distinguish the boundaries between tumors and normal renal parenchyma and bladder mucosa, and is particularly suitable for detecting microinvasive lesions, but it suffers from insufficient spatial localization accuracy and susceptibility to scanning position deviations. Therefore, cross-modal fusion of CT and MRI has become a key technical approach to improve tumor segmentation accuracy.

[0003] However, existing cross-modal image fusion and segmentation methods face a core bottleneck in clinical applications: insufficient closed-loop optimization of cross-modal registration-feature fusion-segmentation verification, which makes it difficult for tumor segmentation accuracy to meet the needs of precise clinical diagnosis and treatment.

[0004] Specifically, traditional fusion methods often employ unidirectional registration models (such as simple affine transformation or local elastic registration), failing to fully consider the complexity of the urinary system's anatomical structure (such as irregular kidney morphology and differences in bladder filling status). Registration errors are generally maintained at 0.8-1.2 mm, directly leading to spatial misalignment in subsequent feature fusion. In the feature fusion stage, most methods use simple weighted summation or channel splicing, failing to dynamically balance the weights of CT structural features and MRI soft tissue features, resulting in redundancy or loss of key information. Clinical data shows that such methods have a Dice similarity coefficient (DSC) of only 0.75-0.82 for segmenting small renal cell carcinomas with a diameter <3 cm, and an accuracy rate of less than 70% for identifying submucosal invasion of bladder cancer. Furthermore, existing segmentation models lack anatomical constraints on the urinary system, making segmentation results susceptible to artifacts such as vascular sections and stones, with a false positive rate as high as 15-20%, leading to overtreatment (such as unnecessary partial nephrectomy) or undertreatment (such as tumor residue) in some patients. Summary of the Invention

[0005] The purpose of this invention is to at least solve one of the technical problems existing in the prior art, and to provide a CT and MR fusion method for precise segmentation of urinary system tumors, which can solve the problem of spatial misalignment in feature fusion.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a CT and MR fusion method for precise segmentation of urinary system tumors, comprising step 1 image preprocessing, step 2 bidirectional multi-scale registration, step 3 cross-modal attention fusion, step 4 edge-enhanced segmentation, and step 5 anatomical constraint postprocessing, wherein step 1 image preprocessing includes step 101 format conversion and cropping, step 102 adaptive noise reduction processing, and step 103 grayscale normalization; Step 2, bidirectional multi-scale registration, includes step 201, coarse registration with marker initialization, step 202, fine registration with bidirectional B-spline, and step 203, registration verification. Step 3, cross-modal attention fusion, includes step 301, dual-branch feature extraction, step 302, cross-modal attention fusion module, and step 303, feature dimensionality reduction. Step 4, edge-enhanced segmentation, includes step 401, tumor edge feature enhancement extraction, step 402, improved U-Net segmentation network, and step 403, loss function optimization. Step 5, anatomical constraint post-processing, includes step 501, morphological optimization, step 502, anatomical constraint verification, and step 503, result output.

[0007] Preferably, step 101 involves format conversion and cropping: DICOM format CT and MRI images (MRI including T1WI enhancement, T2WI or DWI sequences) acquired clinically were uniformly converted to NIfTI format and cropped to 320×320×80 pixels based on the anatomical location of the urinary system (completely covering the target area of ​​kidney-ureter-bladder), and background noise interference from non-target areas was removed.

[0008] Preferably, step 102 involves adaptive noise reduction processing: CT images: Adaptive median filtering is used, and the window size is dynamically adjusted according to the pixel grayscale variance (3×3 window when σ5, 5×5 window when 15≤σ≤25, and 7×7 window when σ>25). The number of filtering iterations is fixed at 3 times to balance the noise reduction effect and the preservation of structural details. MRI images: Non-local mean filtering was used, with a similarity window of 7×7×3 and a search window of 11×11×5. The noise reduction weight factor h=0.8±0.1, focusing on preserving the boundary features between the tumor and the surrounding soft tissue.

[0009] Preferably, step 103 involves grayscale normalization: CT: Map the scan HU value to the clinically effective range of [-100, 400] (covering the density range of renal parenchyma, tumor calcification and surrounding tissues), standardized formula: ICTnorm=400-(-100)ICT-(-100)×255, output grayscale range [0, 255]; MRI: Local Z-score normalization is used, with the formula: IMRInorm=σIMRI-μ, where μ is the mean of a local 3×3 window and σ is the standard deviation of the window. After normalization, the gray levels are uniformly mapped to [0,255] to eliminate signal differences between different scanning devices.

[0010] Preferably, step 201 involves initializing and coarsely registering the marker points: Three key anatomical markers of the urinary system (lower pole of the kidney, beginning of the ureter, and top of the bladder) are automatically extracted, and a global fast alignment is achieved using a Euclidean distance minimization registration algorithm. Number of iterations: 70; convergence threshold: 0.001 pixels; registration error control ≤1.0mm. Transformation matrix: 4×4 affine matrix, containing x / y / z axis translation (≤5mm) and rotation (≤3°) parameters to accommodate slight positional deviations in clinical images.

[0011] Preferably, step 202 involves bidirectional B-spline fine registration; Based on an improved two-way elastic registration model, the alignment of local anatomical structures under deformation is optimized: Transformation model: 3rd order B-spline affine transformation, grid node spacing 10mm, maximum number of iterations 300; Metric function: Hybrid metric: M = 0.6 × MI + 0.4 × GS (MI is mutual information, GS is gradient similarity), which improves the matching accuracy of cross-modal images; Regularization parameter: 0.2±0.1, controls the smoothness of the deformation field and avoids excessive distortion of the anatomical structure.

[0012] Preferably, step 203 involves registration verification: After registration, the normalized mutual information (NMI≥0.85) and mean square error (MSE≤20) of CT and MRI are calculated. If the verification criteria are not met, the fine registration step is re-executed to ensure the reliability of the registration.

[0013] Preferably, step 301 involves dual-branch feature extraction: CT branch (structural features): 3-layer convolutional network (Conv3×3, stride 1, padding 1), number of output channels 64→128→256, activation function ReLU, focusing on extracting tumor calcifications and density difference anatomical features. MRI branch (soft tissue features): Using the same network structure as the CT branch, soft tissue features of tumor edema area and infiltration range are extracted in a targeted manner (DWI sequence focuses on information on water molecule diffusion restriction, T2WI focuses on tissue signal differences). Step 302 Cross-modal attention fusion module: Introducing a channel-space dual attention mechanism to dynamically allocate dual-modal feature weights: Channel attention: Calculates the channel correlation of bi-branch features, with a weighting coefficient: Wc=sigmoid(MLP(GlobalAvgPool(FCT,FMRI))), ranging from 0.3 to 0.7, highlighting key feature channels; Spatial attention: Generates a 2D spatial attention map, formula: Ws=sigmoid(Conv2d(Concatenate(FCT,FMRI))), which accurately locates the tumor region and enhances the feature response; The fusion formula is: Ffusion = Wc × (Ws × FCT + (1-Ws) × FMRI), which achieves adaptive complementarity between structural and soft tissue features; Step 303: Feature dimensionality reduction; By using 1×1 convolution, the number of fused feature channels is reduced to 128, which reduces the computational complexity of the model while fully preserving the core tumor feature information.

[0014] Preferably, step 401 involves enhanced extraction of tumor edge features: Embedding the Roberts edge enhancement module accurately captures tumor boundary information: Operator kernel: and The edge detection threshold is 70±20, and a tumor edge feature map Fedge is generated. Feature fusion: combining the edge feature map Fedge with the fused feature map Ffusion; The concatenation (channel count 128+1=129) is input into the segmentation network to improve boundary recognition accuracy; Step 402: Improve the U-Net segmentation network; Encoder: 4 layers of convolution + max pooling (stride 2) to extract multi-scale tumor features in layers; Decoder: 4 layers of transposed convolutions + skip connections, fusing semantic and detail features from the corresponding layers of the encoder; Segmentation head: 2-channel output (tumor / background), Sigmoid activation function, to achieve binary classification segmentation; Step 403 Loss Function Optimization: A hybrid loss function is used to balance segmentation accuracy and training stability: L = 0.3 × DiceLoss + 0.7 × CrossEntropyLoss. The optimizer is Adam, the initial learning rate is 1e-4, the iteration is 300 rounds, and the batch size is 4.

[0015] Preferably, step 501 involves morphological optimization: Closure operation: Using a 5×5×3 spherical structuring element, fill the internal area of ​​the tumor in mm.2 Tiny pores; Small connected regions are removed: regions with a volume <50mm are deleted. 3 Isolated connected regions effectively eliminate interference from vascular cross-sections and stone artifacts; Step 502 Anatomical constraint verification: Renal cell carcinoma: The segmentation result must be strictly within the kidney ROI automatically extracted during the preprocessing stage. Areas outside the renal parenchyma are considered false positives and are removed. Bladder cancer: The segmentation results should be close to the bladder wall (the normal clinical bladder wall thickness ranges from 2 to 5 mm). Isolated areas >3 mm away from the bladder wall are considered artifacts and should be deleted. Step 503 output: Output DICOM format tumor segmentation mask (compatible with clinical imaging workstations), and use voxel counting method (voxel size = 1mm×1mm×1mm) to calculate tumor volume, maximum diameter and three-dimensional boundary coordinates, providing quantitative indicators for clinical diagnosis and treatment.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This CT and MR fusion method for precise segmentation of urinary system tumors significantly improves cross-modal registration accuracy and solves the problem of spatial misalignment: This invention adopts a multi-scale joint model of "coarse registration of marker points + fine registration of bidirectional B-splines". Global alignment is initialized through three key anatomical markers, and local deformation is optimized with a hybrid metric function (0.6×MI+0.4×GS). The final registration error is ≤0.5mm, which is more than 30% more accurate than the existing technology (0.8-1.2mm). This effect directly solves the problem of spatial misalignment in feature fusion caused by traditional unidirectional registration, laying the foundation for the accurate extraction of tumor features. It enables the structural localization advantage of CT and the soft tissue recognition advantage of MRI to achieve precise spatial matching.

[0017] 2. This CT and MR fusion method for precise segmentation of urinary system tumors optimizes feature fusion efficiency and enhances the extraction of core tumor information: It innovatively introduces a channel-space dual attention mechanism, dynamically allocating the weights of CT structural features and MRI soft tissue features (channel weight 0.3-0.7, spatial attention map for precise localization of tumor regions), avoiding the information redundancy problem of traditional simple weighted fusion. Actual clinical validation shows that this fusion strategy improves the tumor segmentation DSC value to 0.90-0.92, which is 8-10% higher than the traditional method (0.75-0.82). In particular, it significantly enhances the ability to distinguish between renal cell carcinoma calcifications and surrounding edema areas, and to identify bladder cancer mucosal thickening and tumor infiltration, solving the problem of insufficient feature extraction for small lesions (diameter <3cm).

[0018] 3. This CT and MR fusion method for precise segmentation of urinary system tumors improves the accuracy of segmentation boundaries and clinical fit, reducing the risk of misdiagnosis and misjudgment: By enhancing tumor boundary features through the Roberts edge enhancement module and combining it with anatomical constraints of the urinary system (renal cancer is confined to the kidney ROI, and bladder cancer is attached to the bladder wall by 2-5 mm) to form a dual verification system, the matching degree of segmentation boundaries is improved by 15%, and the false positive rate is reduced to below 5% (compared to 15-20% in existing technologies). In clinical examples, the relative error between the segmentation results of renal cancer and the measured values ​​of pathological specimens is only 3.1%, and the judgment of the invasion depth of bladder cancer is completely consistent with the postoperative pathology. This effectively avoids errors in surgical planning (such as tumor residue or excessive resection of normal tissue) caused by segmentation deviations, and provides a reliable quantitative basis for surgical planning and prognostic assessment.

[0019] 4. This CT and MR fusion method for precise segmentation of urinary system tumors has strong clinical adaptability and is compatible with existing diagnostic and treatment workflows: This method outputs a DICOM format segmentation mask, which can be directly connected to clinical imaging workstations. Moreover, all parameters (such as filter window, number of iterations, and anatomical constraint threshold) are validated based on 50 clinical samples and are compatible with scanning equipment in different hospitals (CT tube voltage 100-120kV, MRI different sequence parameters). This effect solves the problem of the disconnect between traditional fusion methods and clinical diagnostic and treatment workflows. It can be applied without additional equipment modification, significantly reducing the threshold for clinical promotion. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the process modules of the present invention. Detailed Implementation

[0021] This section will describe in detail specific embodiments of the present invention. Preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the drawings is to supplement the textual description with graphics, so that people can intuitively and vividly understand each technical feature and overall technical solution of the present invention, but they should not be construed as limiting the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0023] In the description of this invention, terms such as greater than, less than, and exceeding are understood to exclude the stated number, while terms such as above, below, and within are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0024] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0025] Please see Figure 1 The present invention provides a technical solution: a CT and MR fusion method for precise segmentation of urinary system tumors, comprising the following steps: Step 1 Image preprocessing: modality-specific noise reduction and standardization; Step 101: Format Conversion and Cropping DICOM format CT and MRI images (MRI including T1WI enhancement, T2WI or DWI sequences) acquired clinically were uniformly converted to NIfTI format and cropped to 320×320×80 pixels based on the anatomical location of the urinary system (completely covering the target area of ​​kidney-ureter-bladder), and background noise interference from non-target areas was removed. Step 102 Adaptive noise reduction processing: CT images: Adaptive median filtering is used, and the window size is dynamically adjusted according to the pixel grayscale variance (3×3 window when σ5, 5×5 window when 15≤σ≤25, and 7×7 window when σ>25). The number of filtering iterations is fixed at 3 times to balance the noise reduction effect and the preservation of structural details. MRI images: Non-local mean filtering was used, with a similarity window of 7×7×3 and a search window of 11×11×5. The noise reduction weight factor h=0.8±0.1, focusing on preserving the boundary features between the tumor and the surrounding soft tissue. Step 103 Grayscale Standardization: CT: Map the scan HU value to the clinically effective range of [-100, 400] (covering the density range of renal parenchyma, tumor calcification and surrounding tissues), standardized formula: ICTnorm=400-(-100)ICT-(-100)×255, output grayscale range [0, 255]; MRI: Local Z-score normalization is used, with the formula: IMRInorm=σIMRI-μ, where μ is the mean of the local 3×3 window and σ is the standard deviation of the window. After normalization, the gray levels are uniformly mapped to [0,255] to eliminate signal differences between different scanning devices. Step 2: Bidirectional Multi-Scale Registration: Global-Local Joint Alignment Step 201: Initialize coarse registration of marker points: Three key anatomical markers of the urinary system (lower pole of the kidney, beginning of the ureter, and top of the bladder) are automatically extracted, and a global fast alignment is achieved using a Euclidean distance minimization registration algorithm. Number of iterations: 70; convergence threshold: 0.001 pixels; registration error control ≤1.0mm. Transformation matrix: 4×4 affine matrix, containing x / y / z axis translation (≤5mm) and rotation (≤3°) parameters to accommodate slight positional deviations in clinical images; Step 202: Two-way B-spline fine registration; Based on an improved two-way elastic registration model, the alignment of local anatomical structures under deformation is optimized: Transformation model: 3rd order B-spline affine transformation, grid node spacing 10mm, maximum number of iterations 300; Metric function: Hybrid metric: M = 0.6 × MI + 0.4 × GS (MI is mutual information, GS is gradient similarity), which improves the matching accuracy of cross-modal images; Regularization parameter: 0.2±0.1, controls the smoothness of the deformation field and avoids excessive distortion of the anatomical structure; Step 203 Registration Verification: After registration, the normalized mutual information (NMI≥0.85) and mean square error (MSE≤20) of CT and MRI are calculated. If the verification criteria are not met, the fine registration step is repeated to ensure the reliability of the registration. Step 3: Cross-modal attention fusion: Adaptive and precise aggregation of dual-branch features Step 301: Dual-branch feature extraction CT branch (structural features): 3-layer convolutional network (Conv3×3, stride 1, padding 1), number of output channels 64→128→256, activation function ReLU, focusing on extracting tumor calcifications and density difference anatomical features. MRI branch (soft tissue features): Using the same network structure as the CT branch, soft tissue features of tumor edema area and infiltration range are extracted in a targeted manner (DWI sequence focuses on information on water molecule diffusion restriction, T2WI focuses on tissue signal differences). Step 302 Cross-modal attention fusion module: Introducing a channel-space dual attention mechanism to dynamically allocate dual-modal feature weights: Channel attention: Calculates the channel correlation of bi-branch features, with a weighting coefficient: Wc=sigmoid(MLP(GlobalAvgPool(FCT,FMRI))), ranging from 0.3 to 0.7, highlighting key feature channels; Spatial attention: Generates a 2D spatial attention map, formula: Ws=sigmoid(Conv2d(Concatenate(FCT,FMRI))), which accurately locates the tumor region and enhances the feature response; The fusion formula is: Ffusion = Wc × (Ws × FCT + (1-Ws) × FMRI), which achieves adaptive complementarity between structural and soft tissue features; Step 303: Feature dimensionality reduction; By using 1×1 convolution, the number of fused feature channels is reduced to 128, which reduces the computational complexity of the model while fully preserving the core feature information of the tumor. Step 4: Edge-enhanced segmentation: Automated and precise extraction of tumor regions Step 401: Enhanced extraction of tumor edge features: Embedding the Roberts edge enhancement module accurately captures tumor boundary information: Operator kernel: and The edge detection threshold is 70±20, and a tumor edge feature map Fedge is generated. Feature fusion: combining the edge feature map Fedge with the fused feature map Ffusion; The concatenation (channel count 128+1=129) is input into the segmentation network to improve boundary recognition accuracy; Step 402: Improve the U-Net segmentation network; Encoder: 4 layers of convolution + max pooling (stride 2) to extract multi-scale tumor features in layers; Decoder: 4 layers of transposed convolutions + skip connections, fusing semantic and detail features from the corresponding layers of the encoder; Segmentation head: 2-channel output (tumor / background), Sigmoid activation function, to achieve binary classification segmentation; Step 403 Loss Function Optimization: A hybrid loss function is used to balance segmentation accuracy and training stability: L = 0.3 × Dice Loss + 0.7 × CrossEntropy Loss. The optimizer is Adam, the initial learning rate is 1e-4, the iteration is 300 rounds, and the batch size is 4. Step 5: Post-processing of anatomical constraints: artifact removal and anatomical accuracy verification; Step 501 Morphological optimization: Closure operation: Using a 5×5×3 spherical structuring element, fill the internal area of ​​the tumor in mm. 2 Tiny pores; Small connected regions are removed: regions with a volume <50mm are deleted. 3 Isolated connected regions effectively eliminate interference from vascular cross-sections and stone artifacts; Step 502 Anatomical constraint verification: Renal cell carcinoma: The segmentation result must be strictly within the kidney ROI automatically extracted during the preprocessing stage. Areas outside the renal parenchyma are considered false positives and are removed. Bladder cancer: The segmentation results should be close to the bladder wall (the normal clinical bladder wall thickness ranges from 2 to 5 mm). Isolated areas >3 mm away from the bladder wall are considered artifacts and should be deleted. Step 503 output: Output DICOM format tumor segmentation mask (compatible with clinical imaging workstations), and use voxel counting method (voxel size = 1mm×1mm×1mm) to calculate tumor volume, maximum diameter and three-dimensional boundary coordinates, providing quantitative indicators for clinical diagnosis and treatment; Clinical Example: Example 1: CT-MRI fusion segmentation of renal cell carcinoma (clear cell carcinoma); Case information: A 55-year-old male was found to have a mass in his right kidney during a physical examination one week ago. The clinical suspicion is clear cell carcinoma. He has no obvious clinical symptoms. Image data: CT: Tube voltage 120kV, tube current 250mA, slice thickness 1mm, pitch 1.0, scanning HU value coverage range -200~800; MRI: T1WI enhancement (TR 500ms, TE 10ms), T2WI (TR 3000ms, TE 80ms), slice thickness 1mm, matrix 256×256, scan range covering the entire kidney; Processing result: Registration accuracy: Normalized mutual information (NMI) = 0.89, mean square error (MSE) = 15.3, registration error 0.3mm; Segmentation performance: Dice similarity coefficient (DSC) = 0.92, intersection-over-union ratio (IoU) = 0.86, boundary matching degree (BD) = 0.2mm; Tumor quantitative indicators: Tumor volume 18.6 cm 3 The value was 19.2 cm, compared with the gross pathological specimen measurement after surgery. 3 The relative error is 3.1%, which meets the needs of clinical quantitative diagnosis. Example 2: CT-MRI fusion segmentation of bladder cancer (urothelial carcinoma); Case information: A 62-year-old female presented with intermittent gross hematuria for 2 months, accompanied by urinary frequency and urgency. The clinical diagnosis was a bladder wall mass with local mucosal thickening. Image data: CT enhancement: tube voltage 100kV, tube current 200mA, slice thickness 1mm, pitch 1.0, arterial phase scan, clearly showing the blood supply to the bladder wall; MRI: DWI (b value 800s / mm)2 T2WI (TR 4000ms, TE 80ms), T2WI (TR 3500ms, TE 90ms), layer thickness 1mm, matrix 256×256; Processing result: Registration accuracy: NMI=0.87, MSE=17.8, registration error 0.4mm; Segmentation performance: DSC=0.90, IoU=0.83, BD=0.3mm; Clinical fit: Tumor volume 7.8cm 3 Postoperative pathology confirmed that the tumor infiltration depth was T2 stage, which was completely consistent with the infiltration range indicated by the segmentation results, providing a precise reference for the formulation of the surgical plan; Furthermore, cross-modal registration accuracy is significantly improved, solving the problem of spatial misalignment: This invention adopts a multi-scale joint model of "coarse registration of marker points + fine registration of bidirectional B-splines". Global alignment is initialized through three key anatomical marker points, and local deformation is optimized with a hybrid metric function (0.6×MI+0.4×GS). The final registration error is ≤0.5mm, which is more than 30% more accurate than the existing technology (0.8-1.2mm). This effect directly solves the problem of spatial misalignment of feature fusion caused by traditional unidirectional registration, laying the foundation for the accurate extraction of tumor features. It enables the structural localization advantage of CT and the soft tissue recognition advantage of MRI to achieve accurate spatial matching. Furthermore, the efficiency of feature fusion was optimized to enhance the extraction of core tumor information: an innovative channel-space dual attention mechanism was introduced to dynamically allocate the weights of CT structural features and MRI soft tissue features (channel weight 0.3-0.7, spatial attention map accurately locates the tumor region), avoiding the information redundancy problem of traditional simple weighted fusion. Actual clinical validation showed that this fusion strategy improved the tumor segmentation DSC value to 0.90-0.92, which is 8-10% higher than the traditional method (0.75-0.82). In particular, it significantly enhanced the ability to distinguish between renal cell carcinoma calcifications and surrounding edema areas, and to identify bladder cancer mucosal thickening and tumor infiltration, solving the problem of insufficient feature extraction of small lesions (diameter <3cm). Furthermore, the accuracy and clinical fit of the segmentation boundary are improved, reducing the risk of misdiagnosis and misjudgment: The tumor boundary features are enhanced by the Roberts edge enhancement module, and a dual verification system is formed by combining the anatomical constraints of the urinary system (renal cancer is limited to the kidney ROI, and bladder cancer is attached to the bladder wall by 2-5mm), which improves the matching degree of the segmentation boundary by 15% and reduces the false positive rate to below 5% (15-20% in the prior art). In the clinical embodiment, the relative error between the segmentation result of renal cancer and the measurement value of the pathological specimen is only 3.1%, and the judgment of the invasion depth of bladder cancer is completely consistent with the postoperative pathology. This effectively avoids the error in surgical planning (such as tumor residue, excessive resection of normal tissue) caused by segmentation deviation, and provides a reliable quantitative basis for surgical planning and prognostic assessment. Furthermore, this method exhibits strong clinical adaptability and compatibility with existing diagnostic and treatment workflows: it outputs a DICOM format segmentation mask that can be directly integrated with clinical imaging workstations. All parameters (such as filter window, number of iterations, and anatomical constraint thresholds) are validated based on 50 clinical samples, making it compatible with scanning equipment in different hospitals (CT tube voltage 100-120kV, MRI different sequence parameters). This effect solves the problem of traditional fusion methods being disconnected from clinical diagnostic and treatment workflows, allowing for practical application without additional equipment modifications, and significantly lowering the threshold for clinical promotion. The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A CT and MR fusion method for precise segmentation of urinary system tumors, comprising step 1 image preprocessing, step 2 bidirectional multi-scale registration, step 3 cross-modal attention fusion, step 4 edge-enhanced segmentation, and step 5 anatomical constraint postprocessing, characterized in that: The image preprocessing step 1 includes step 101 format conversion and cropping, step 102 adaptive noise reduction processing, and step 103 grayscale normalization. Step 2, bidirectional multi-scale registration, includes step 201, coarse registration with marker initialization, step 202, fine registration with bidirectional B-spline, and step 203, registration verification. Step 3, cross-modal attention fusion, includes step 301, dual-branch feature extraction, step 302, cross-modal attention fusion module, and step 303, feature dimensionality reduction. Step 4, edge-enhanced segmentation, includes step 401, tumor edge feature enhancement extraction, step 402, improved U-Net segmentation network, and step 403, loss function optimization. Step 5, anatomical constraint post-processing, includes step 501, morphological optimization, step 502, anatomical constraint verification, and step 503, result output.

2. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 101: Format conversion and cropping: DICOM format CT and MRI images (MRI including T1WI enhancement, T2WI or DWI sequences) acquired clinically were uniformly converted to NIfTI format and cropped to 320×320×80 pixels based on the anatomical location of the urinary system (completely covering the target area of ​​kidney-ureter-bladder), and background noise interference from non-target areas was removed.

3. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 102, adaptive noise reduction processing: CT images: Adaptive median filtering is used, and the window size is dynamically adjusted according to the pixel grayscale variance (3×3 window when σ5, 5×5 window when 15≤σ≤25, and 7×7 window when σ>25). The number of filtering iterations is fixed at 3 times to balance the noise reduction effect and the preservation of structural details. MRI images: Non-local mean filtering was used, with a similarity window of 7×7×3 and a search window of 11×11×5. The noise reduction weight factor h=0.8±0.1, focusing on preserving the boundary features between the tumor and the surrounding soft tissue.

4. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 103: Gray-scale normalization: CT: Map the scan HU value to the clinically effective range of [-100, 400] (covering the density range of renal parenchyma, tumor calcification and surrounding tissues), standardized formula: ICTnorm=400-(-100)ICT-(-100)×255, output grayscale range [0, 255]; MRI: Local Z-score normalization is used, with the formula: IMRInorm=σIMRI-μ, where μ is the mean of a local 3×3 window and σ is the standard deviation of the window. After normalization, the gray levels are uniformly mapped to [0,255] to eliminate signal differences between different scanning devices.

5. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 201: Initialization and coarse registration of marker points. Three key anatomical markers of the urinary system (lower pole of the kidney, beginning of the ureter, and top of the bladder) are automatically extracted, and a global fast alignment is achieved using a Euclidean distance minimization registration algorithm. Number of iterations: 70; convergence threshold: 0.001 pixels; registration error control ≤1.0mm. Transformation matrix: 4×4 affine matrix, containing x / y / z axis translation (≤5mm) and rotation (≤3°) parameters to accommodate slight positional deviations in clinical images.

6. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 202 involves bidirectional B-spline fine registration. Based on an improved two-way elastic registration model, the alignment of local anatomical structures under deformation is optimized: Transformation model: 3rd order B-spline affine transformation, grid node spacing 10mm, maximum number of iterations 300; Metric function: Hybrid metric: M = 0.6 × MI + 0.4 × GS (MI is mutual information, GS is gradient similarity), which improves the matching accuracy of cross-modal images; Regularization parameter: 0.2±0.1, controls the smoothness of the deformation field and avoids excessive distortion of the anatomical structure.

7. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 203: Registration verification After registration, the normalized mutual information (NMI≥0.85) and mean square error (MSE≤20) of CT and MRI are calculated. If the verification criteria are not met, the fine registration step is re-executed to ensure the reliability of the registration.

8. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 301, dual-branch feature extraction: CT branch (structural features): 3-layer convolutional network (Conv3×3, stride 1, padding 1), number of output channels 64→128→256, activation function ReLU, focusing on extracting tumor calcifications and density difference anatomical features. MRI branch (soft tissue features): Using the same network structure as the CT branch, soft tissue features of tumor edema area and infiltration range are extracted in a targeted manner (DWI sequence focuses on information on water molecule diffusion restriction, T2WI focuses on tissue signal differences). Step 302 Cross-modal attention fusion module: Introducing a channel-space dual attention mechanism to dynamically allocate dual-modal feature weights: Channel attention: Calculates the channel correlation of bi-branch features, with a weighting coefficient: Wc=sigmoid(MLP(GlobalAvgPool(FCT,FMRI))), ranging from 0.3 to 0.7, highlighting key feature channels; Spatial attention: Generates a 2D spatial attention map, formula: Ws=sigmoid(Conv2d(Concatenate(FCT,FMRI))), which accurately locates the tumor region and enhances the feature response; The fusion formula is: Ffusion = Wc × (Ws × FCT + (1-Ws) × FMRI), which achieves adaptive complementarity between structural and soft tissue features; Step 303: Feature dimensionality reduction; By using 1×1 convolution, the number of fused feature channels is reduced to 128, which reduces the computational complexity of the model while fully preserving the core tumor feature information.

9. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 401, tumor edge feature enhancement extraction: Embedding the Roberts edge enhancement module accurately captures tumor boundary information: Operator kernel: and The edge detection threshold is 70±20, and a tumor edge feature map Fedge is generated. feature Fusion: Combines the edge feature map Fedge with the fusion feature map Ffusion; The concatenation (channel count 128+1=129) is input into the segmentation network to improve boundary recognition accuracy; Step 402: Improve the U-Net segmentation network; Encoder: 4 layers of convolution + max pooling (stride 2) to extract multi-scale tumor features in layers; Decoder: 4 layers of transposed convolutions + skip connections, fusing semantic and detail features from the corresponding layers of the encoder; Segmentation head: 2-channel output (tumor / background), Sigmoid activation function, to achieve binary classification segmentation; Step 403 Loss Function Optimization: A hybrid loss function is used to balance segmentation accuracy and training stability: L = 0.3 × DiceLoss + 0.7 × CrossEntropyLoss. The optimizer is Adam, the initial learning rate is 1e-4, the iteration is 300 rounds, and the batch size is 4.

10. The CT and MR fusion method for precise segmentation of urinary system tumors according to claim 1, characterized in that: Step 501, morphological optimization: Closure operation: Using a 5×5×3 spherical structuring element, fill the internal area of ​​the tumor in mm. 2 Tiny pores; Small connected regions are removed: regions with a volume <50mm are deleted. 3 Isolated connected regions effectively eliminate interference from vascular cross-sections and stone artifacts; Step 502 Anatomical constraint verification: Renal cell carcinoma: The segmentation result must be strictly within the kidney ROI automatically extracted during the preprocessing stage. Areas outside the renal parenchyma are considered false positives and are removed. Bladder cancer: The segmentation results should be close to the bladder wall (the normal clinical bladder wall thickness ranges from 2 to 5 mm). Isolated areas >3 mm away from the bladder wall are considered artifacts and should be deleted. Step 503 output: Output DICOM format tumor segmentation mask (compatible with clinical imaging workstations), and use voxel counting method (voxel size = 1mm×1mm×1mm) to calculate tumor volume, maximum diameter and three-dimensional boundary coordinates, providing quantitative indicators for clinical diagnosis and treatment.

Images