Urine system stone feature extraction and clinical quantitative evaluation method and system

Abstract

This invention relates to the field of medical image processing technology, and particularly to a method and system for feature extraction and clinical quantitative assessment of urinary system stones. The method includes the following steps: S1: acquiring an input tomographic image and a multi-label mask containing the organ to be detected; S2: constructing five independent stone search regions using the optimized multi-label mask based on anatomical prior relationships; S3: calibrating the density values ​​of the tomographic image; S4: performing morphological closing operations and connected component labeling on the stone candidate voxels; S5: performing edge voxel verification and minimum volume filtering on the stone connected components; and S6: calculating the image features of the final stone region. In this invention, by defining five stone search regions—left kidney, right kidney, bladder, left ureter, and right ureter—the stone retrieval space is limited, eliminating interference from bone and vascular calcification in non-target areas and reducing the false positive rate.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for feature extraction and clinical quantitative assessment of urinary system stones. Background Technology

[0002] Preoperative clinical evaluation of urinary tract stones is based on non-contrast computed tomography (CT) images. Current automated evaluation methods mainly involve fixed density threshold segmentation of the entire abdomen, high-density voxel screening assisted by deep learning masks, or extraction of manually measured features for machine learning prediction. These techniques aim to obtain information on stone distribution, size, and density, providing auxiliary reference data for clinical surgical planning.

[0003] Existing automated methods for detecting urinary system stones employ global processing logic. Due to the heterogeneity of background tissue density in anatomical spaces such as the kidneys, ureters, and bladder, current technologies cannot dynamically establish segmentation thresholds based on local pixel statistical features of the region. This leads to missed detections or false positives when the system encounters low-density lesions or background tissue interference, reducing the accuracy of stone extraction. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a method and system for extracting the characteristics of urinary system stones and for clinical quantitative evaluation, aiming to improve the problem of low accuracy in stone extraction.

[0005] In a first aspect, the present invention provides the following technical solution: a method for extracting the characteristics of urinary system stones and conducting clinical quantitative assessment, including...

[0006] Step S1: Obtain the input tomographic image and a multi-label mask containing the organ to be detected, and perform morphological optimization processing on the multi-label mask;

[0007] Step S2: Construct five independent stone-finding regions based on anatomical prior relationships using optimized multi-label masks;

[0008] Step S3: The density value of the tomographic image is calibrated, and the stone segmentation threshold is dynamically determined in each stone search area to extract stone candidate voxels;

[0009] Step S4: Perform morphological closing operation and connected component labeling on the candidate voxels of the stones to obtain the connected components of the stones in order to achieve stone separation.

[0010] Step S5: Perform edge voxel verification and minimum volume filtering on the connected regions of the stone to obtain the final stone region;

[0011] Step S6: Calculate the imaging features of the final stone area, and fuse the imaging features with the patient's basic information to extract features, outputting stone burden quantitative assessment indicators and surgical approach technical difficulty reference information.

[0012] Preferably, in step S1, the morphological optimization process for the multi-label mask includes:

[0013] Perform morphological closing operations on the left and right kidney masks to fill the cavities in the renal pelvis and calyces.

[0014] Preferably, in step S2, the step of constructing five independent stone-finding zones includes:

[0015] The direction of expansion toward the midline of the body is determined based on the positional relationship between the iliac artery mask and the midline of the body.

[0016] directional expansion is performed on the side of the iliac artery mask facing the midline of the body to obtain the initial expansion area;

[0017] The maximum distance from the low-density voxel within the initial expansion region to the surface of the corresponding iliac artery is calculated, and the expansion radius is determined accordingly, thereby constructing the left ureter search area and the right ureter search area.

[0018] Preferably, in step S2, the step of constructing five independent stone-finding zones further includes:

[0019] The left ureter search area is obtained by subtracting the lumbar spine expansion area from the directional expansion area of ​​the left iliac artery and then subtracting the left iliac artery itself.

[0020] The right ureter search area is obtained by subtracting the lumbar spine expansion area from the right iliac artery directional expansion area and then subtracting the right iliac artery itself area.

[0021] Preferably, in step S3, the steps of calibrating the density values ​​of the input tomographic image and extracting candidate voxels for stones include:

[0022] The adaptive parameter adjustment network dynamically adjusts the standard air density value and preset maximum expansion physical span.

[0023] Linear correction of all pixels in the image is performed using calibration offset;

[0024] Within each search area, the stone segmentation threshold is dynamically determined based on local density statistical characteristics, and voxels with density values ​​higher than the corresponding stone segmentation threshold are selected as candidate stone voxels.

[0025] Preferably, in step S4, the steps of performing morphological closing operations and connected component labeling on the candidate stone voxels include:

[0026] The closing operation is performed using a 26-connected structuring element, wherein the size of the structuring element does not exceed a preset spatial continuity threshold.

[0027] The 26-connected-domain labeling algorithm is used for labeling.

[0028] Preferably, in step S5, the steps of performing edge voxel verification and minimum volume filtering on the stone connectivity region include:

[0029] When the density ratio of the edge voxel to the adjacent non-stone voxel is less than the preset ratio threshold, and the density value continuously and monotonically decreases along the direction from the inside of the stone to the outside background, the adjacent voxel is included in the stone region.

[0030] Among them, the preset ratio threshold and the preset volume threshold are adjusted online through an adaptive parameter adjustment network.

[0031] Preferably, in step S6, the step of outputting the reference information on the technical difficulty of the surgical approach includes:

[0032] Calculate the number of skeletal branches in the final stone region, and determine the stone morphology classification accordingly;

[0033] The image features are combined with the patient's basic information to construct a fused feature vector;

[0034] By retrieving recommended passages from the medical knowledge graph using a search-enhanced generative architecture or page indexing technology, and then inputting them together with fused feature vectors into a multi-task clinical prediction model, the model outputs a surgical approach technical difficulty score and planning text suggestions.

[0035] Preferably, the adaptive parameter adjustment network employs at least one of lightweight neural networks, deep learning models, heuristic algorithms, or reinforcement learning agents to perform online adaptive optimization based on metadata features and global noise distribution of tomographic images.

[0036] Secondly, this invention provides the following technical solution: a system for extracting and clinically quantifying the characteristics of urinary system stones, comprising:

[0037] The image masking module is configured to acquire the input tomographic image and a multi-label mask containing the organ to be detected;

[0038] The search area module is configured to construct five independent stone search areas based on anatomical prior relationships, including the left kidney search area, right kidney search area, bladder search area, left ureter search area, and right ureter search area;

[0039] The threshold segmentation module is configured to perform density value calibration on the tomographic image and dynamically determine the stone segmentation threshold in each search area based on local density statistical features, and perform threshold segmentation on the search area to extract stone candidate voxels.

[0040] The stone segmentation module is configured to perform morphological closing operations and connected component labeling on the stone candidate voxels to obtain stone connected components in order to achieve stone segmentation.

[0041] The verification and filtering module is configured to perform edge voxel verification and minimum volume filtering on the stone connectivity region to obtain the final stone region.

[0042] The quantitative assessment module is configured to calculate the imaging features of the final stone region, construct a fusion feature vector containing the imaging features and the patient's basic information, and call the medical knowledge graph and multi-task clinical prediction model to output a surgical procedure technical difficulty score and surgical approach planning text suggestions.

[0043] The present invention has the following beneficial effects:

[0044] 1. In this invention, the ureter search area is constructed by using five stone search areas: left kidney, right kidney, bladder, left ureter and right ureter. The organ mask is used to subtract the expansion area of ​​interfering structures. The iliac artery is directionally expanded to subtract the expansion of the lumbar spine and the iliac artery itself to construct the ureter search area. This limits the stone search space and eliminates the interference of bone and blood vessel calcification in non-target areas, thereby reducing the false positive rate.

[0045] 2. In this invention, by introducing an adaptive parameter adjustment network, the dependence of traditional hard-coded parameters on specific devices and fixed empirical values ​​is changed. Based on deep reinforcement learning agent, the adaptive threshold is optimized in real time according to image quality, which effectively blocks the false expansion of boundaries and false detection of isolated noise points caused by local artifact noise, and improves the detection robustness of the model on multi-center and multi-device data.

[0046] 3. In this invention, the closed-loop integration of a large language model and retrieval-enhanced generation technology upgrades the original single numerical evaluation index into a dual-track auxiliary decision expression of numerical scores and interpretable text suggestions. The surgical approach planning text suggestions output by the emergence and reasoning capabilities of the large language model have the logical support of medical guidelines, which improves the credibility and guiding value of decision support in clinical practice. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method for feature extraction and clinical quantitative assessment of urinary system stones proposed in this invention;

[0048] Figure 2 This is a system module diagram of the urinary system stone feature extraction and clinical quantitative assessment system proposed in this invention. Detailed Implementation

[0049] The technical solutions in 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.

[0050] Example 1

[0051] In the first embodiment of the present invention, the present invention provides a method for feature extraction and clinical quantitative assessment of urinary system stones, such as... Figure 1 As shown, it includes the following steps:

[0052] Step S1: Obtain the input tomographic image and a multi-label mask containing the organ to be detected, and perform morphological optimization processing on the multi-label mask;

[0053] In step S1, the morphological optimization process for the multi-label mask includes:

[0054] Perform morphological closing operations on the left and right kidney masks to fill the cavities in the renal pelvis and calyces regions;

[0055] Specifically, the input tomographic image is acquired. This tomographic image is a plain, non-contrast computed tomography (CT) image, displayed in the computer system as a three-dimensional voxel density matrix. The three-dimensional voxel density matrix is ​​stored in a formal format. The stored value of each voxel is the original pixel value. In this scheme, the rescaling slope in the DICOM file header is utilized. and rescalar intercept The original pixel values ​​are converted into Hounsfield Unit (HU) values, which reflect the physical density of the tissue. The calculation formula is as follows:

[0056] ;

[0057] Obtain the multi-label mask output by the segmentation model. This segmentation model employs a 3D medical image segmentation architecture. In this scheme, the segmentation model uses... As input, perform semantic segmentation prediction and output a spatially aligned 3D multi-label mask matrix. .

[0058] This multi-label mask matrix This includes masks for the left kidney, right kidney, bladder, left iliac artery, right iliac artery, lumbar spine, and hip bone. In this scheme, this multi-label mask matrix... The preferred label value assignment is as follows: 0 for background, 1 for left kidney, 2 for right kidney, 3 for bladder, 4 for left iliac artery, 5 for right iliac artery, 6 for lumbar vertebrae, and 7 for hip bone.

[0059] Based on this multi-label mask matrix Extract the independent binary matrices for each organ. Take the binary matrix of the left kidney as an example. For example, the extraction logic is as follows:

[0060] ;

[0061] Construct the right kidney binary matrix according to the same logic. Bladder binary matrix Left iliac artery binary matrix Right iliac artery binary matrix Lumbar spine binary matrix and the binary matrix of the hip bone .

[0062] Morphological optimization was performed on the multi-label mask. Since the segmentation model is prone to voxel loss or internal voids in low-contrast regions such as the renal pelvis and calyces, optimization is necessary. This optimization includes performing morphological closing operations on the left and right kidney masks to fill voids in the renal pelvis and calyces regions.

[0063] The specific execution process is as follows: using three-dimensional structural elements with preset spatial dimensions. For the binary matrix of the left kidney Perform morphological closing operation to obtain the optimized left kidney mask. The mathematical expression is as follows:

[0064] ;

[0065] in, Represents a three-dimensional morphological dilation operator, whose computational voxels In structural elements The maximum value within the range; This represents a three-dimensional morphological erosion operator, whose computational voxels... In structural elements The minimum value within the range. Similarly, for the binary matrix of the right kidney. Perform the same operation to obtain the optimized right kidney mask. .

[0066] In this morphological closing operation, the three-dimensional structuring element The shape is preferably spherical. This three-dimensional structural element Physical diameter The preferred size is 2mm.

[0067] In this scheme, three-dimensional structural elements discrete voxel radius Determined based on the spatial resolution of the image. If the voxel spacing of the tomographic image is respectively along the three coordinate axes... , , The discrete voxel radius The preferred calculation method is:

[0068] ;

[0069] As a numerical implementation method, when and At that time, the calculation yielded Voxel.

[0070] By performing this morphological closing operation, firstly, a dilation operation is used to merge non-1-valued voxels inside the kidney caused by low-density urine or partial volume effects, achieving spatial connectivity. Then, an erosion operation is used to eliminate unnecessary edge expansion, filling the cavities inside the renal pelvis and calyces while maintaining the macroscopic outline of the kidney. Finally, an optimized multi-label mask set is output.

[0071] Step S2: Construct five independent stone-finding regions based on anatomical prior relationships using optimized multi-label masks;

[0072] In step S2, the steps of constructing five independent stone-finding zones include:

[0073] The direction of expansion toward the midline of the body is determined based on the positional relationship between the iliac artery mask and the midline of the body.

[0074] directional expansion is performed on the side of the iliac artery mask facing the midline of the body to obtain the initial expansion area;

[0075] Calculate the maximum distance from the low-density voxel within the initial expansion region to the surface of the corresponding iliac artery, and determine the expansion radius accordingly, thereby constructing the left ureter search area and the right ureter search area;

[0076] In step S2, the step of constructing five independent stone-finding zones also includes:

[0077] The left ureter search area is obtained by subtracting the lumbar spine expansion area from the directional expansion area of ​​the left iliac artery and then subtracting the left iliac artery itself.

[0078] The right ureter search area is obtained by subtracting the lumbar spine expansion area from the right iliac artery directional expansion area and then subtracting the right iliac artery itself area.

[0079] Specifically, to eliminate false positives in stone detection caused by artifacts surrounding large bone structures, the expansion regions of the lumbar spine and hip bone are first established. The binary matrix of the lumbar spine is defined as follows: The binary matrix of the hip bone is A spherical skeleton dilation topology operator is employed. Morphological dilation is performed on the binary matrices of the lumbar spine and hip bone, converting the physical radius into discrete spatial voxel radii based on the voxel spacing along each axis of the image. The mathematical expression corresponding to the dilation process is as follows:

[0080] ;

[0081] ;

[0082] In the formula, Indicates the lumbar spine expansion area. ⊕⊕ represents the expansion region of the hip bone, and ⊕⊕ represents the morphological expansion operator.

[0083] The left kidney search area is obtained by subtracting the lumbar spine expansion region from the morphologically optimized left kidney mask, and the right kidney search area is obtained by subtracting the lumbar spine expansion region from the morphologically optimized right kidney mask. Let the optimized left kidney mask be... The optimized right kidney mask is Left kidney search area Search area with the right kidney The formula for matrix-level Boolean subtraction is as follows:

[0084] ;

[0085] ;

[0086] This Boolean subtraction operation removes high-density interfering voxels caused by overlap between the kidney parenchyma contour edge and the spinal skeleton or by partial volume effects.

[0087] Subtracting the hip bone expansion area from the bladder mask yields the bladder search area. Let the original bladder mask be... Bladder search area The calculation formula is as follows:

[0088] ;

[0089] This step eliminates boundary overlap interference caused by high-density bone around the bladder from the pelvis and hip bones, ensuring that the search area is focused inside the bladder.

[0090] The construction of the left and right ureteral search areas utilizes the anatomical prior relationship that the pelvic segment of the ureter runs closely along the iliac artery. Due to the small diameter of the ureter and its low contrast in unenhanced images, this method uses directional dilation of a high-contrast vascular mask to cover the potential space of the ureter. The left iliac artery mask is obtained. and right iliac artery mask Calculate the coordinates of the midline of the human body in the transverse section (X-axis) of the tomographic image. In this scheme, the coordinates of the human body midline position... Ideally, it should be half the total width of the image's transverse section in pixels. The direction of expansion towards the body midline is determined based on the positional relationship between the left and right iliac artery masks and the body midline. If the abscissa of the vascular anatomical structure center is greater than... Then the direction of expansion towards the midline is determined as the negative direction of the X-axis; if the x-coordinate of the center of the vascular anatomy structure is less than... Then the direction of expansion toward the midline is determined as the positive direction of the X-axis.

[0091] Inflation is performed on the side of the iliac artery mask facing the midline of the body to obtain the initial inflation area. Using the left iliac artery mask... For example, using unidirectional asymmetric structural elements Perform morphological dilation along a defined vector direction toward the midline of the body to obtain the initial dilation region of the left iliac artery. Similarly, the initial expansion region of the right iliac artery is obtained. .

[0092] Low-density voxels with density values ​​within a preset low-density range are selected within the initial expansion region. This preset low-density range corresponds to the HU distribution space of the ureteral wall dilated by urine or hydronephrosis. The initial expansion region of the left iliac artery is traversed. Extract the coordinates of all voxels within the preset low-density range to form a low-density voxel set. .

[0093] Calculate the maximum distance from the low-density voxel to the corresponding iliac artery surface. For the low-density voxel set... Each voxel coordinate in Calculate the voxel coordinates to the left iliac artery mask. outer surface voxel set The minimum three-dimensional Euclidean distance. The formula for calculating the Euclidean distance is as follows:

[0094] ;

[0095] After obtaining the minimum Euclidean distance for all low-density voxels, the maximum value is extracted and determined as the maximum distance. , represented as:

[0096] ;

[0097] The expansion radii of the directional expansion regions of the left and right iliac arteries are determined based on the maximum distance. In this scheme, the target expansion radius... The preferred calculation method is the maximum distance. With preset safety redundancy constant The algebraic sum of is calculated using the following formula:

[0098] ;

[0099] Preset safety redundancy constant Using this dynamically determined target expansion radius Reconstruct the directional structural elements in the corresponding direction for the left iliac artery mask. Perform a unidirectional expansion procedure to obtain the final directional expansion area of ​​the left iliac artery. Similarly, the directional expansion region of the right iliac artery is determined based on the maximum distance calculated from the right iliac artery. .

[0100] The left ureter search area is obtained by subtracting the lumbar spine expansion area and then the left iliac artery's own area from the directional expansion area of ​​the left iliac artery. Similarly, the right ureter search area is obtained by subtracting the lumbar spine expansion area and then the right iliac artery's own area from the directional expansion area of ​​the right iliac artery. (Left ureter search area) and right ureteral search area The mathematical formula for its construction is as follows:

[0101] ;

[0102] ;

[0103] This construction method eliminates artifacts of vascular wall nodules caused by arteriosclerosis and calcification in large blood vessels by subtracting the iliac artery region itself; and eliminates interference from lateral osteophytes of the spine by subtracting the lumbar spine expansion region; ultimately forming a highly specific ureteral stone retrieval range. After the five independent stone search areas are constructed, each region is independently saved as a corresponding three-dimensional binary matrix, which is used as the spatial constraint boundary for subsequent regional adaptive threshold segmentation.

[0104] Step S3: The density value of the tomographic image is calibrated, and the stone segmentation threshold is dynamically determined in each stone search area to extract stone candidate voxels;

[0105] In step S3, the steps of density value calibration and stone candidate voxel extraction for the input tomographic image include:

[0106] The adaptive parameter adjustment network dynamically adjusts the standard air density value and preset maximum expansion physical span.

[0107] Linear correction of all pixels in the image is performed using calibration offset;

[0108] Within each search area, the stone segmentation threshold is dynamically determined based on local density statistical characteristics, and voxels with density values ​​higher than the corresponding stone segmentation threshold are selected as stone candidate voxels.

[0109] Specifically, the density value calibration process for this tomographic image begins with parsing the metadata of the DICOM file. This is done by reading the rescaling slope parameter from the metadata. With rescaling intercept parameter The stored pixel values ​​are converted into tissue density values ​​that conform to physical definitions. The conversion logic for tissue density values ​​at each voxel coordinate point is as follows:

[0110] ;

[0111] In the formula, coordinates The stored pixel value, This is the initial tissue density value before calibration.

[0112] To correct for systematic drift caused by CT scans, air feature sampling bounding boxes are established in the four corner regions and the surrounding edge regions of the tomographic image. These sampling bounding boxes extend inwards from the edge of the image matrix by a predetermined pixel thickness. Within the range defined by these bounding boxes, voxels with density values ​​falling within the air distribution range are selected as air region voxels.

[0113] Count the total number of air voxels falling within this air distribution range. And calculate the arithmetic mean of the density values ​​of all voxels in the set to obtain the average air density value. Among them, the standard density value of air The preset value is -1000 HU, which serves as the standard reference for air in CT scans and is connected to an adaptive parameter adjustment network for dynamic bias adjustment. To counteract tissue density and anatomical deformation deviations caused by cross-device scanning, the air standard density value in this embodiment... Furthermore, empirical values ​​such as the preset maximum expansion physical span of unidirectional asymmetric structural elements are not statically fixed, but are dynamically adjusted by an adaptive parameter adjustment network. This adaptive parameter adjustment network uses lightweight neural networks, deep learning models, or multi-center feedback datasets. By inputting metadata features of the current tomographic image and global noise distribution, it performs online adaptive optimization and bias fine-tuning of the air standard density value parameter and the maximum expansion physical span parameter. (Calibration offset) The calculation formula is as follows:

[0114] ;

[0115] Establish a mechanism to verify the effectiveness of calibration offsets. Determine the total number of voxels in the air region. Does it meet the preset threshold of 1000 voxels, and simultaneously determine the calibration offset? Whether the absolute value is within the preset offset threshold range, set to 50HU, and simultaneously determine the calibration offset. Is the absolute value within the preset offset threshold range? Reaching the preset quantity threshold and If the absolute value does not exceed the preset offset threshold, the calibration offset is deemed valid, and linear correction is performed on all pixels in the image. Correction of HU density value. The calculation formula is as follows:

[0116] ;

[0117] Using this corrected HU density value The left kidney search area output in step S2 Right kidney search area Bladder search area Left ureteral search area and right ureteral search area Internally, each component independently performs dynamic decision-making regarding the stone segmentation threshold.

[0118] Within each stone-finding area, soft tissue background is first removed. Specifically, a preset background tissue threshold is set. The threshold is set to 30 HU to extract voxels within the search area whose density values ​​are higher than this background tissue threshold, thus constructing a high-density candidate voxel set. This preset background tissue threshold is used for preliminary filtering of renal parenchyma, urine, and pelvic soft tissue.

[0119] Calculate the statistical distribution characteristics of each high-density candidate voxel set and extract the first quartile. and the third and fourth quartiles .in, This represents the density value at the 25th percentile after the high-density candidate voxel set is sorted in ascending order of density value. This represents the density value at the 75th percentile after sorting by density value in ascending order.

[0120] Calculate the ratio of the third quartile to the first quartile. Based on this ratio Calculate the stone segmentation threshold corresponding to each search area. The continuous mapping function for the stone segmentation threshold is as follows:

[0121] ;

[0122] In the formula, , , These are the preset fitting parameters. The left kidney search area, right kidney search area, bladder search area, left ureter search area, and right ureter search area each correspond to independent fitting parameter values. This mapping function uses ratios... It reflects the degree of dispersion of local density distribution and enables adaptive adjustment of the threshold as tissue characteristics change.

[0123] After obtaining the independent stone segmentation threshold for each search area, threshold segmentation is performed within the corresponding search area space to extract candidate stone voxels. The decision logic is as follows: if the corrected tissue density value of the voxels within the search area... Greater than the stone segmentation threshold corresponding to this region If the voxel is selected, it is identified as a candidate voxel for stones. This operation generates a binary mask for candidate stones in each region, providing target regions for subsequent morphological segmentation and precise boundary verification.

[0124] Step S4: Perform morphological closing operations and connected component labeling on candidate stone voxels to obtain stone connected components in order to separate stones into individual particles;

[0125] In step S4, the steps of performing morphological closing operations and connected component labeling on candidate stone voxels include:

[0126] The 26-connected structuring element is used for the closing operation, and the size of the structuring element does not exceed the preset spatial continuity threshold.

[0127] The 26-connected-component labeling algorithm is used for labeling;

[0128] Specifically, after obtaining the binary mask of the candidate stones in each region, the following steps are performed: morphological closing operation and connected component labeling are performed on the candidate stone voxels to obtain the connected components of the stones in order to separate the stones into individual particles.

[0129] A 26-connected structuring element is used to perform morphological closing operations on candidate voxels of stones, filling the gaps within the same stone. This 26-connected structuring element is represented in three-dimensional discrete space as a symmetric voxel matrix of size 3×3×3 with all internal values ​​of 1. Three-dimensional morphological dilation and erosion operations are then performed sequentially on the binary masks of candidate stones output from each search region using this structuring element. The mathematical calculation formula is as follows:

[0130] ;

[0131] In the formula, The binary masks of candidate stones for each search area extracted in step S3. Corresponding to a 26-connected structure element. For morphological dilation operators, For morphological erosion operators, This corresponds to the updated binary mask matrix of candidate stones after the closing operation. This operation can effectively eliminate internal micro-gaps or fractures caused by uneven density, partial volume effect, or image noise within the stone, making the voxel outline of a single stone tend to have a complete skeleton.

[0132] To prevent erroneous adhesion and merging of originally independent adjacent stones in space, the size of the 26 connected structural elements does not exceed a preset spatial continuity threshold, which is set to 3 mm. If the discrete spatial physical size of the structural element exceeds this preset value, it will lead to the erroneous connection of adjacent independent stones. In this scheme, the preset spatial continuity threshold is set based on the minimum anatomical spacing of clinically common stones and the spatial resolution of tomographic images. If the discrete spatial physical size of the structural element exceeds the preset spatial continuity threshold, the expansion step will cross the anatomical gap, erroneously connecting two spatially adjacent but unrelated independent stone particles into the same connected region, resulting in a systematic underestimation or deviation in subsequent stone counting and size measurement. Therefore, this scheme limits the use of 3×3×3 structural elements with single-layer voxel expansion to fill the small gaps inside a single stone and maintain the spatial anatomical isolation of adjacent independent stones.

[0133] A 26-connected-component labeling algorithm is used to label the candidate voxels for stones after the closing operation, and each labeling result is determined as a stone connected component. In the three-dimensional discrete space voxel lattice, for any candidate voxel to be detected, the algorithm traverses and detects all 26 adjacent coordinate points of the voxel in the face adjacency direction, edge adjacency direction, and corner adjacency direction in three-dimensional space. If there are voxels that are also activated as stone candidates among these 26 adjacent coordinate points, they are classified into the same connected component set according to the spatial connectivity determination criteria.

[0134] Using this 26-connected-domain labeling algorithm, all stone voxels that are spatially continuous within 26 neighborhoods across the entire graph are assigned the same non-zero integer label value, while spatially disconnected independent stone regions are assigned different non-zero integer label values. The connected-domain labeling result is ultimately represented as a three-dimensional stone label matrix. The formal definition is as follows:

[0135] ;

[0136] In the formula, The unique non-zero positive integer index label is assigned, with a value ranging from 1 to the total number of stones detected. By using the exclusive mapping and allocation of this label matrix, accurate segmentation and discretization counting of all stone targets in the tomographic image are achieved, providing underlying structured data support for subsequent low-threshold edge voxel verification and three-dimensional topological geometric feature calculation for each stone.

[0137] Step S5: Perform edge voxel verification and minimum volume filtering on the connected regions of the stone to obtain the final stone region;

[0138] In step S5, the steps of edge voxel verification and minimum volume filtering for the stone connected domains include:

[0139] When the density ratio of the edge voxel to the adjacent non-stone voxel is less than the preset ratio threshold, and the density value continuously and monotonically decreases along the direction from the inside of the stone to the outside background, the adjacent voxel is included in the stone region.

[0140] Among them, the preset ratio threshold and the preset volume threshold are adjusted online through an adaptive parameter adjustment network;

[0141] Specifically, after completing the discretization and granulation of the stone connected domain, step S5 is executed: perform edge voxel verification and minimum volume filtering of the stone connected domain to obtain the final stone region.

[0142] Edge voxel verification is performed on the connected regions of the stones extracted through low-threshold segmentation. Each acquired connected region of the stone is traversed, and edge voxels within that region are identified. An edge voxel is defined as a voxel point in the 3D stone label matrix that belongs to the current connected region of the stone and whose 26-neighborhood contains at least one background voxel; denoted as […]. The corrected tissue density value corresponding to this edge voxel is denoted as... .

[0143] For each edge voxel Identify the background voxels that are spatially adjacent to it, and denote them as adjacent non-stone voxels. The corrected tissue density value corresponding to the adjacent non-calculous voxel is denoted as... .

[0144] Calculate the ratio between the density value of the edge voxel and the density value of the adjacent non-stone voxel. The mathematical formula for calculating this ratio is as follows:

[0145] ;

[0146] The calculated ratio The comparison is made against a preset ratio threshold, with an initial value set at 1.5. Additionally, a preset volume threshold is set with an initial value set at 10. This is used to filter noise. The preset ratio threshold is not a statically hard-coded fixed empirical value, but rather connected to an adaptive parameter adjustment network. When the network detects a decrease in the global signal-to-noise ratio of the current tomographic image due to low-dose scanning or device heterogeneity, the adaptive parameter adjustment network dynamically adjusts the preset ratio threshold online using feature feedback from lightweight neural networks, deep learning models, heuristic algorithms, or reinforcement learning agents, thereby blocking boundary false expansion caused by local artifact noise. (The final sentence refers to determining the ratio.) Whether it is less than the preset ratio threshold after dynamic fine-tuning, the degree of edge expansion is limited by the density ratio constraint, preventing the stone boundary from spreading disorderly to the surrounding high-density soft tissue area.

[0147] A spatial density gradient change verification mechanism is established to determine the spatial extension direction from the interior of the stone to the external background. This spatial extension direction is then used to determine the edge voxel. A predetermined number of background voxels are continuously extracted from the outside to construct a spatial voxel density sequence. .in Density value of corresponding edge voxels , Density values ​​of adjacent non-stone-forming voxels .

[0148] Determine whether the voxel density values ​​in the spatial voxel density sequence satisfy the constraint of continuous monotonically decreasing. The mathematical inequalities for this constraint are expressed as follows:

[0149] ;

[0150] This criterion is used to verify the physical scattering attenuation law of radiation from the center of the stone outward, and to exclude the interference of discrete high-density voxels caused by isolated calcifications in the blood vessel wall or image artifacts.

[0151] When the ratio If the density value is less than a preset ratio threshold and satisfies the aforementioned continuously monotonically decreasing constraint when extending along the direction from the inside of the stone to the external background, the adjacent non-stone voxel is determined to be a true boundary voxel of the current stone. This adjacent non-stone voxel is then included in the stone region. In the 3D stone label matrix, the label value of this adjacent non-stone voxel is updated from 0 to a non-zero integer label value of the current stone's connected component. This verification process is iterated layer by layer until all boundary voxels no longer meet the inclusion criteria.

[0152] After boundary expansion verification is completed, minimum volume filtering is performed on the updated stone-connected regions. The total number of voxels contained in each independent stone-connected region is counted. The physical volume of a single voxel is calculated based on the spatial resolution parameters of the image. The calculation formula is as follows:

[0153] ;

[0154] In the formula, , These represent the pixel spacing at the transverse section of the tomographic image. This represents the interslice spacing of the tomographic image.

[0155] Calculate the actual physical volume of each independent stone-connected domain. The calculation formula is as follows:

[0156] ;

[0157] The actual physical volume of each stone's connected region The size is compared with a preset volume threshold. This preset volume threshold is defined as the spatial three-dimensional continuous envelope volume. To combat the interference of different scan slice thicknesses and noise levels on the identification of small lesions, this preset volume threshold is also adaptively fine-tuned online by an adaptive parameter adjustment network. The adaptive parameter adjustment network dynamically scales the spatial three-dimensional continuous envelope volume based on the multi-center feedback dataset and the current tomographic image quality. If the actual physical volume of a certain stone connected region... If the volume is smaller than the preset volume threshold after dynamic optimization, the connected region of the stone is identified as local noise or a minor artifact. The label values ​​of all corresponding voxels in the 3D stone label matrix are reset to 0, and filtering is performed to remove them. Regions with actual physical volumes greater than or equal to the preset volume threshold are retained to obtain the final stone region, providing a reliable target mask for subsequent accurate quantification calculations of the stone's 3D geometric features.

[0158] Step S6: Calculate the imaging features of the final stone area, and fuse the imaging features with the patient's basic information to extract features, outputting stone burden quantitative assessment indicators and reference information on the technical difficulty of surgical approach;

[0159] In step S6, the step of outputting reference information on the technical difficulty of the surgical approach includes:

[0160] Calculate the number of skeletal branches in the final stone region and determine the stone morphology classification accordingly;

[0161] Construct a fused feature vector by combining image features with basic patient information;

[0162] By retrieving recommended passages from the medical knowledge graph using a search-enhanced generative architecture or page indexing technology, and inputting them together with the fused feature vectors into a multi-task clinical prediction model, the model outputs a surgical approach technical difficulty score and planning text suggestions.

[0163] Specifically, the image feature calculation process for the final stone region includes extracting anatomical location, volume, maximum diameter, width, average density value, maximum density value, minimum density value, sphericity, and number of skeletal branches. The anatomical location is determined based on the overlap state of the spatial mask matrix of the specific stone-finding area where the final stone region is located. The volume... The maximum diameter is obtained by multiplying the total number of voxels contained within the current stone region by the physical volume of a single voxel. The width is obtained by calculating the maximum three-dimensional Euclidean distance between any two voxel coordinates within the current stone region. The dimensions of the second major axis of the cuboid perpendicular to the direction of the maximum diameter vector are determined by constructing a minimum circumscribed cuboid parallel to the direction of the maximum diameter vector.

[0164] The formula for calculating the surgical approach technical difficulty score S is as follows:

[0165] ;

[0166] In the formula, This represents the physical volume of the final stone-bearing area; Indicates the average density value; This indicates the anatomical location weighting coefficient, which is determined based on the anatomical location of the stone. For example, the weighting coefficient for stones in the lower calyx of the kidney is set higher than that for stones in the upper calyx of the kidney. Indicates the patient's body mass index; Indicates the number of branches in the skeleton; to Preset weighting coefficients for each feature component; This is the model bias correction term. This mapping function achieves an objective quantitative assessment of the technical challenge of the surgical procedure by comprehensively considering the physical volume, texture, spatial morphology complexity, anatomical location accessibility, and patient's physical characteristics of the stone.

[0167] The average density value Maximum density value and minimum density value By statistically analyzing the corrected HU density values ​​of all voxels within the final stone region... The values ​​in the matrix are obtained by performing arithmetic mean calculation, maximum extraction, and minimum extraction, respectively. This sphericity... The formula used to characterize the degree to which the spatial geometry of a gallstone approximates a sphere is as follows:

[0168] ;

[0169] In the formula, This indicates the physical volume of the final stone-bearing area. This represents the three-dimensional surface area of ​​the final stone region. This three-dimensional surface area is obtained by extracting the outer boundary voxels of the current stone region and calculating the sum of the areas of each boundary mesh using an isosurface mesh extraction algorithm.

[0170] The number of skeleton branches was obtained through 3D skeletonization and statistical analysis of 26 neighborhood bifurcation points. The specific process was as follows: a 3D topological refinement algorithm was used to peel away layers of the binary matrix of the final stone region, extracting the 3D skeleton centerline with single-cell thickness that maintains topological connectivity. Each skeleton voxel on this 3D skeleton centerline was traversed, and the number of neighboring voxels belonging to the skeleton centerline within its 26 neighborhood in discrete space was counted. When the number of neighboring voxels in the 26 neighborhood of a skeleton voxel was greater than or equal to 3, the skeleton voxel was identified as a 26-neighborhood bifurcation point. By summing the total number of 26-neighborhood bifurcation points in the final stone region and combining this with topological branch mapping rules, the number of skeleton branches for the stone was determined.

[0171] Automatic stone morphology classification is performed based on the calculated number of skeleton branches. When the number of skeleton branches is greater than or equal to three, antler-shaped morphology classification feature description information is generated for the final stone region; when the number of skeleton branches is less than three, ordinary stone morphology classification feature description information is generated. When the three-dimensional skeletonization process fails due to the stone volume being too small or topological discontinuities, stone morphology is determined based on sphericity. The specific determination logic is as follows: a preset sphericity threshold is set, and when the sphericity... When the sphericity is less than the preset sphericity threshold, the morphology of the final stone region is determined to be staghorn calculus; when the sphericity... If the sphericity is greater than or equal to the preset sphericity threshold, the morphology of the final stone region is determined to be a common stone. This dual-track determination mechanism ensures the accuracy of morphology classification.

[0172] After acquiring the aforementioned image features, these features were fused with the patient's basic information for feature extraction. This basic information included age, gender, body mass index, and history of urological surgery. Anatomical location, volume, maximum diameter, width, average density value, maximum density value, minimum density value, sphericity, number of skeletal branches, and the determined stone morphology classification were then matrix-level concatenated with the patient's basic information and aligned with multimodal features to construct a fused feature vector.

[0173] This embodiment further integrates a decision generation flow based on a large language model and retrieval-enhanced generation technology. The system pre-imports authoritative clinical guidelines such as the European Association of Urology (EAUR) guidelines and the American Urological Association (AUR) guidelines, constructing a structured medical knowledge graph of urological clinical guidelines through entity extraction and relation definition. After generating a fused feature vector, the quantitative evaluation module converts this fused feature vector into input prompts containing rich contextual information. Utilizing the retrieval-enhanced generation architecture combined with efficient page indexing retrieval technology, a deep semantic search is performed within the medical knowledge graph using these input prompts as the retrieval core, quickly recalling guideline recommendations that match the current stone characteristics and patient anatomical conditions.

[0174] Subsequently, the recalled guideline recommendations and the original input prompts are concatenated at the text level and input into a pre-trained multi-task clinical prediction model. This model's multi-branch output architecture enables joint quantitative evaluation and text interpretation reasoning: its numerical prediction branch outputs the total stone volume, zoned stone load, surgical approach technical difficulty score, and postoperative clearance rate reference range; its text interpretation branch, relying on the logical reasoning capabilities of the large language model, outputs surgical approach planning text suggestions with high medical logical interpretability. This forms an intelligent closed-loop output from precise image feature extraction and dynamic guideline text retrieval to macro-level clinical decision support.

[0175] The adaptive parameter adjustment network employs at least one of lightweight neural networks, deep learning models, heuristic algorithms, or reinforcement learning agents to perform online adaptive optimization based on metadata features and global noise distribution of tomographic images.

[0176] Specifically, the network input feature vector include:

[0177] Metadata features: tube voltage (kVp), tube current (mAs), reconstruction kernel type, slice thickness (mm), pixel pitch (mm), and patient body type;

[0178] Global noise statistical features include: mean, standard deviation, skewness, and kurtosis of air region density; mean, standard deviation, and estimated signal-to-noise ratio of the entire image gradient. These features are normalized to form a 16- to 32-dimensional input feature vector. .

[0179] Network Structure: The lightweight MLP network includes an input layer, a first fully connected layer (64 neurons), a ReLU activation layer, a second fully connected layer (32 neurons), another ReLU activation layer, and an output layer. Its output format is as follows:

[0180] ;

[0181] in, This is the weight matrix. For bias parameters, It is a linear output activation function.

[0182] Network output and adaptive compensation: The network output includes an offset of the standard air density value. and correction amount for the maximum expansion physical span of the ureteral search area Therefore, the adaptive calibration parameters are obtained:

[0183] ;

[0184] ;

[0185] In the formula, To preset the standard air density value, -1000HU is preferred. The maximum physical span for expansion is preset.

[0186] Training and Inference Process: The network is trained using a pre-collected multi-center CT image dataset. Training samples include original DICOM / NIfTI images, scan parameters, image quality assessment results, and manually annotated optimal air calibration parameters and optimal expansion span parameters that achieve the highest accuracy in stone detection, as selected through experiments. The network uses the sum of the mean squared error (MSE) between the predicted parameters and the optimal annotated parameters as the loss function.

[0187]

[0188] in, and Loss weighting coefficient (preferred) During online inference, the system reads the metadata and noise statistics of the image to be processed in real time, and the trained network outputs parameter biases to achieve online adaptive compensation for the heterogeneity of different scanning protocols and devices.

[0189] Example 2:

[0190] In existing computer-aided stone detection methods, low contrast in plain non-enhanced tomographic images leads to voids in the generated organ masks of the renal pelvis and calyces, radial artifacts around skeletal structures, low contrast in the ureteral pelvic segment, and tissue density deviations when scanning across different devices. This makes it impossible to achieve independent segmentation and three-dimensional morphological quantification of stones in multiple organs using existing segmentation models or globally fixed thresholds. To address these issues, this invention provides a system for feature extraction and clinical quantitative assessment of urinary system stones, such as... Figure 2 As shown. The specific implementation process of this system is as follows:

[0191] The image masking module is configured to acquire the input tomographic image and a multi-label mask containing the organ to be detected;

[0192] The search area module is configured to construct five independent stone search areas based on anatomical prior relationships. The search areas include the left kidney search area, the right kidney search area, the bladder search area, the left ureter search area, and the right ureter search area.

[0193] The threshold segmentation module is configured to perform density value calibration on the tomographic image and dynamically determine the stone segmentation threshold in each search area based on local density statistical features, and perform threshold segmentation on the search area to extract stone candidate voxels.

[0194] The stone segmentation module is configured to perform morphological closing operations and connected component labeling on candidate stone voxels to obtain connected components of the stones in order to achieve stone segmentation.

[0195] The verification and filtering module is configured to perform edge voxel verification and minimum volume filtering on the connected domains of the stone to obtain the final stone region.

[0196] The quantitative assessment module is configured to calculate the imaging features of the final stone area, construct a fusion feature vector containing imaging features and basic patient information, and call upon the medical knowledge graph and multi-task clinical prediction model to output a surgical procedure technical difficulty score and surgical approach planning text suggestions.

[0197] Specifically, the image masking module calls the 3D image reading engine to parse the rescale slope and rescale intercept in the DICOM file header, converting the original pixel values ​​into the physically defined HU value space. This module performs morphological processing on the multi-label mask, using a spherical structuring element with a diameter of 2mm for closing operations to fill low-density artifact holes in the renal pelvis and calyces, outputting a geometrically complete target mask.

[0198] Search Area Module: The search space is constructed based on anatomical priors. First, the lumbar vertebrae and hip bone masks are expanded by a radius of 5mm to generate a skeletal masking matrix. Overlapping skeletal portions in the kidney and bladder masks are removed using set subtraction. For the ureter, based on the iliac artery location information, the search space is locked by asymmetric directional expansion along the midline of the body with a maximum span of 30mm, and the vessel itself is automatically subtracted to eliminate interference from vascular calcification.

[0199] Thresholding Segmentation Module: Establishes a background-adaptive calibration mechanism. By analyzing the atmospheric features extending 5 pixels from the image's edges, the MLP network architecture is adjusted using adaptive parameters, and the output bias parameters are used for linear correction of the entire image. Within each search region, quartiles are calculated based on the distribution characteristics of local high-density samples, and corresponding stone segmentation thresholds are generated to achieve real-time removal of soft tissue background.

[0200] Stone segmentation module: Performs 3D morphological decoupling. Using structuring elements of a preset size (3×3×3), morphological closing operations are performed on the candidate voxel set to fill the internal micro-gaps. A 26-connected-domain labeling algorithm is used to traverse the 3D space, assigning unique index labels to separate discrete stone targets into independent stone connected domains.

[0201] Verification and Filtering Module: Performs boundary refinement correction. Density gradient analysis is performed on the edge voxels of each stone's connected domain. The density ratio between the stone's interior and adjacent background voxels is compared to a threshold of 1.5, and boundary voxels that meet the monotonically decreasing gradient constraint are selected. Simultaneously, the physical volume of the stone is statistically analyzed and compared with a dynamically adjusted volume threshold of 10. The system compares and automatically identifies and removes noise artifacts.

[0202] Quantitative Assessment Module: Performs multimodal feature fusion. It calculates the geometric features of the stone, such as volume, sphericity threshold of 0.7, and number of skeletal branches, and constructs a fused feature vector with the patient's basic information. This module links to a medical knowledge graph, using enhanced generation architecture matching clinical guidelines through retrieval. It inputs this data into a multi-task clinical prediction model and outputs a surgical approach technical difficulty score and planning suggestion text.

[0203] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting characteristics and conducting clinical quantitative assessment of urinary system stones, characterized in that, include: Step S1: Obtain the input tomographic image and a multi-label mask containing the organ to be detected, and perform morphological optimization processing on the multi-label mask; Step S2: Construct five independent stone-finding regions based on anatomical prior relationships using optimized multi-label masks; Step S3: The density value of the tomographic image is calibrated, and the stone segmentation threshold is dynamically determined in each stone search area to extract stone candidate voxels; Step S4: Perform morphological closing operation and connected component labeling on the candidate voxels of the stones to obtain the connected components of the stones in order to achieve stone separation. Step S5: Perform edge voxel verification and minimum volume filtering on the connected regions of the stone to obtain the final stone region; Step S6: Calculate the imaging features of the final stone area, and fuse the imaging features with the patient's basic information to extract features, outputting stone burden quantitative assessment indicators and surgical approach technical difficulty reference information.

2. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, In step S1, the morphological optimization process for the multi-label mask includes: Perform morphological closing operations on the left and right kidney masks to fill the cavities in the renal pelvis and calyces.

3. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, In step S2, the step of constructing five independent stone-finding zones includes: The direction of expansion toward the midline of the body is determined based on the positional relationship between the iliac artery mask and the midline of the body. directional expansion is performed on the side of the iliac artery mask facing the midline of the body to obtain the initial expansion area; The maximum distance from the low-density voxel within the initial expansion region to the surface of the corresponding iliac artery is calculated, and the expansion radius is determined accordingly, thereby constructing the left ureter search area and the right ureter search area.

4. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, In step S2, the step of constructing five independent stone-finding zones further includes: The left ureter search area is obtained by subtracting the lumbar spine expansion area from the directional expansion area of ​​the left iliac artery and then subtracting the area of ​​the left iliac artery itself. The right ureter search area is obtained by subtracting the lumbar spine expansion area from the right iliac artery directional expansion area and then subtracting the right iliac artery itself area.

5. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, In step S3, the steps of calibrating the density values ​​of the input tomographic image and extracting candidate voxels for stones include: The adaptive parameter adjustment network dynamically adjusts the standard air density value and preset maximum expansion physical span. Linear correction of all pixels in the image is performed using calibration offset; Within each search area, the stone segmentation threshold is dynamically determined based on local density statistical characteristics, and voxels with density values ​​higher than the corresponding stone segmentation threshold are selected as candidate stone voxels.

6. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, In step S4, the steps of performing morphological closing operations and connected component labeling on the candidate stone voxels include: The closing operation is performed using a 26-connected structuring element, wherein the size of the structuring element does not exceed a preset spatial continuity threshold. The 26-connected-domain labeling algorithm is used for labeling.

7. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, In step S5, the steps of performing edge voxel verification and minimum volume filtering on the stone connectivity region include: When the density ratio of the edge voxel to the adjacent non-stone voxel is less than the preset ratio threshold, and the density value continuously and monotonically decreases along the direction from the inside of the stone to the background, the adjacent voxel is included in the stone region. Among them, the preset ratio threshold and the preset volume threshold are adjusted online through an adaptive parameter adjustment network.

8. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, In step S6, the step of outputting the reference information on the technical difficulty of the surgical approach includes: Calculate the number of skeletal branches in the final stone region, and determine the stone morphology classification accordingly; The image features are combined with the patient's basic information to construct a fused feature vector; By retrieving recommended passages from the medical knowledge graph using a search-enhanced generative architecture or page indexing technology, and then inputting them together with fused feature vectors into a multi-task clinical prediction model, the model outputs a surgical approach technical difficulty score and planning text suggestions.

9. The method for extracting features and conducting clinical quantitative assessment of urinary system stones according to claim 1, characterized in that, The adaptive parameter adjustment network employs at least one of lightweight neural networks, deep learning models, heuristic algorithms, or reinforcement learning agents to perform online adaptive optimization based on metadata features and global noise distribution of tomographic images.

10. A system for extracting characteristics and conducting clinical quantitative assessment of urinary system stones, characterized in that, The method for feature extraction and clinical quantitative assessment of urinary system stones according to any one of claims 1-9 includes: The image masking module is configured to acquire the input tomographic image and a multi-label mask containing the organ to be detected; The search area module is configured to construct five independent stone search areas based on anatomical prior relationships, including the left kidney search area, right kidney search area, bladder search area, left ureter search area, and right ureter search area; The threshold segmentation module is configured to perform density value calibration on the tomographic image and dynamically determine the stone segmentation threshold in each search area based on local density statistical features, and perform threshold segmentation on the search area to extract stone candidate voxels. The stone segmentation module is configured to perform morphological closing operations and connected component labeling on the stone candidate voxels to obtain stone connected components in order to achieve stone segmentation. The verification and filtering module is configured to perform edge voxel verification and minimum volume filtering on the stone connectivity region to obtain the final stone region. The quantitative assessment module is configured to calculate the imaging features of the final stone region, construct a fusion feature vector containing the imaging features and the patient's basic information, and call the medical knowledge graph and multi-task clinical prediction model to output a surgical procedure technical difficulty score and surgical approach planning text suggestions.

Images