Intelligent assistance method for percutaneous nephrolithotomy

Abstract

The present application relates to the technical field of surgical assistance, in particular to an intelligent auxiliary method for percutaneous nephrolithotomy. The method comprises: acquiring a kidney image corresponding to a target kidney; identifying the kidney image to determine the target puncture parameters corresponding to the target kidney. Based on the accurate identification of the image, the quantitative parameters such as the puncture point, angle, depth and obstacle avoidance range are outputted, replacing the traditional empirical puncture scheme and improving the accuracy of the puncture operation. Real-time surgical data of the target kidney during the puncture process is acquired; based on the real-time surgical data and the kidney image, the postoperative estimation result corresponding to the target kidney is determined. The accuracy of the postoperative outcome prediction is improved. High-risk patients are identified in advance to assist doctors in formulating targeted postoperative nursing schemes (such as strengthening anti-infection treatment for high-risk patients). Quantitative basis is provided for doctor-patient communication, so that patients can more clearly understand the expected postoperative recovery and improve the transparency of diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of surgical assistance technology, specifically to an intelligent assistance method for percutaneous nephrolithotomy. Background Technology

[0002] Percutaneous nephrolithotomy is the preferred minimally invasive procedure for treating complex kidney stones. However, the surgical outcome is affected by multiple factors, including the characteristics of the stones, the patient's renal anatomy, and the surgeon's experience. Intraoperative complications such as renal pelvic hypertension, massive bleeding, and postoperative urosepsis and residual stones still limit clinical efficacy.

[0003] The existing technology has obvious limitations: First, the preoperative puncture planning relies heavily on the surgeon's subjective experience, which makes the accuracy of the puncture greatly affected by individual differences in experience; second, the postoperative results usually depend on the surgeon's subjective experience for evaluation, so accurate postoperative evaluation results cannot be obtained, thus making it difficult to do a good job in postoperative rehabilitation.

[0004] Therefore, there is an urgent need to develop an intelligent assistance method for percutaneous nephrolithotomy (PCNL) to solve the core challenges of precision, safety, and individualized management in PCNL surgery. Summary of the Invention

[0005] This invention provides an intelligent assistance method for percutaneous nephrolithotomy (PCNL) to address the issues of precision, safety, and individualized management in PCNL surgery.

[0006] In a first aspect, the present invention provides an intelligent assisted method for percutaneous nephrolithotomy, the method comprising: Acquire kidney imaging images corresponding to the target kidney; identify the kidney imaging images to determine the target puncture parameters corresponding to the target kidney; acquire real-time surgical data of the target kidney during the puncture process; and determine the postoperative prediction results corresponding to the target kidney based on the real-time surgical data and kidney imaging images.

[0007] In one optional implementation, the kidney imaging image is identified to determine the target puncture parameters corresponding to the target kidney, including: identifying the kidney imaging image to determine the region of interest corresponding to the target kidney; the region of interest includes the target stone region, the target kidney parenchyma region, the target blood vessel region, and the target puncture path functional region; identifying the region of interest and extracting the initial features of interest corresponding to the region of interest; and determining the target puncture parameters corresponding to the target kidney based on the initial features of interest.

[0008] In one optional implementation, the region of interest includes the target stone region, the target renal parenchyma region, the target vascular region, and the target puncture path functional region. Identifying the renal image to determine the region of interest corresponding to the target kidney includes: determining the grayscale distribution of the renal image based on the maximum inter-class variance method; determining a grayscale threshold for stone identification based on the grayscale distribution; determining candidate stone regions based on the stone identification grayscale threshold; calculating grayscale co-occurrence matrix parameters corresponding to the candidate stone regions; outputting a three-dimensional mask of the stone based on the grayscale co-occurrence matrix parameters; determining the target stone region based on the three-dimensional mask of the stone; and identifying the renal image to determine the three-dimensional masks of each renal structure, including the renal cortex and renal medulla. The renal pelvis and calyces are labeled to generate the target renal parenchymal region. The Hessian matrix vascular enhancement algorithm is used to identify the renal images and obtain the initial vascular anatomy. Blood flow signals corresponding to each vessel in the renal images are acquired. Based on the blood flow signals, vascular activity is determined. Based on the initial vascular anatomy and vascular activity, the target vascular region is determined. Based on preset principles, a preset number of initial puncture paths are determined. Each initial puncture path is identified, and the basic puncture path functional characteristics and obstructive puncture path functional characteristics corresponding to each initial puncture path are determined. Based on the basic puncture path functional characteristics and obstructive puncture path functional characteristics, the target puncture path functional region is determined within each initial puncture path.

[0009] In one optional implementation, the initial features of interest include initial stone features; the region of interest includes the target stone region and the target renal parenchyma region; the region of interest is identified, and the initial features of interest corresponding to the region of interest are extracted, including: identifying the target stone region and determining the basic morphological features of the stone corresponding to the target stone region; statistically analyzing the mean CT value, standard deviation, coefficient of variation, kurtosis, and skewness corresponding to the CT values ​​within the target stone region; determining the gray-level heterogeneity features corresponding to the target stone region based on the standard deviation, coefficient of variation, kurtosis, and skewness; calculating the gradient mean and gradient variance corresponding to the CT values ​​within the target stone region; determining the stone texture features and composition prediction features corresponding to the target stone region based on the gradient mean and gradient variance; determining the lithotripsy difficulty prediction features based on the CT value difference between the target stone region and the target renal parenchyma region; and generating initial stone features based on the basic morphological features of the stone, gray-level heterogeneity features, stone texture features, composition prediction features, and lithotripsy difficulty prediction features.

[0010] In one optional implementation, the initial features of interest include initial kidney features; the region of interest includes the target kidney parenchyma; identifying the region of interest and extracting the initial features of interest corresponding to the region of interest includes: identifying the target kidney parenchyma and determining the basic anatomical features of the kidney; the basic anatomical features of the kidney include parameters of the renal calyx and fornix and parameters of the renal parenchyma; the parameters of the renal calyx and fornix include the radius of curvature of the renal calyx and fornix, the opening angle of the calyx neck, and the direction vector of the axis of the target renal calyx; identifying the target kidney parenchyma and determining the average CT value and fornix thickness of the fornix; calculating the elasticity index feature of the fornix based on the average CT value and fornix thickness; determining the renal parenchyma tolerance feature based on the CT value and renal parenchyma parameters within the target kidney parenchyma; determining the calyx neck traversability score corresponding to the target kidney parenchyma based on the opening angle of the calyx neck; and generating the initial kidney features based on the basic anatomical features of the kidney, the elasticity index feature of the fornix, the renal parenchyma tolerance feature, and the calyx neck traversability score. In one optional implementation, the initial features of interest include initial vascular features; the region of interest includes a target vascular region; identifying the region of interest and extracting the initial features of interest corresponding to the region of interest includes: identifying the target vascular region and determining the basic vascular distribution features; the basic vascular distribution features include vascular diameter, vascular orientation vector, vascular density in the puncture area, and features of the hypovascular area; calculating the single vascular risk score corresponding to each vascular based on the vascular diameter and the distance between each vascular and the fornix; calculating the regional vascular risk score corresponding to the target vascular region based on each single vascular risk score; calculating the obstacle avoidance distance and obstacle avoidance angle between the preset puncture path and each vascular in the functional area of ​​the target puncture path, and generating obstacle avoidance association features; generating initial vascular features based on the basic vascular distribution features, single vascular risk scores, regional vascular risk scores, and obstacle avoidance association features.

[0011] In one optional implementation, the initial features of interest include initial puncture path features; the region of interest includes the functional region of the target puncture path; identifying the region of interest and extracting the initial features of interest corresponding to the region of interest includes: identifying a preset puncture path in the functional region of the target puncture path to determine basic puncture path features; the basic puncture path features include: path length and the angle between the preset puncture path and the long axis of the kidney; identifying blood vessels, renal parenchymal scar tissue, and renal parenchymal elastic regions in the functional region of the target puncture path to determine obstacle quantification features; identifying the preset puncture path in the functional region of the target puncture path to determine the path efficiency index and operational accessibility score; and generating initial puncture path features based on the basic puncture path features, obstacle quantification features, path efficiency index, and operational accessibility score.

[0012] In one optional implementation, determining the target puncture parameters corresponding to the target kidney based on the initial features of interest includes: inputting the initial features of interest into a preset LASSO regression model to output candidate core features; the preset LASSO regression model uses the target puncture point, target puncture angle, target puncture depth, obstacle avoidance range boundary, and target perfusion pressure threshold as joint target variables; inputting the candidate core features into a preset expert evaluation model to evaluate the candidate core features and output target core features; inputting the target core features into a preset puncture parameter determination model; the preset puncture parameter determination model identifies each sub-feature in the target core features and determines the category corresponding to each sub-feature; the categories include safety category, operation category, and anatomical category; and generating safety category feature groups and operation category groups according to the categories corresponding to each sub-feature. The system extracts feature groups and anatomical category feature groups. Safety category feature groups are input into the first decision tree branch, outputting safety risk features. Operation category feature groups are input into the second feature extraction branch, outputting operation difficulty features. Anatomical category feature groups are input into the third feature extraction branch, outputting anatomical adaptation features. Anatomical adaptation scores characterize the compatibility of the puncture path with renal anatomy. Based on safety risk features, anatomical adaptation features, and stone location features in the basic stone morphology features, the target puncture point is predicted. Based on anatomical adaptation features and operation difficulty features, the target puncture angle is predicted. Based on safety risk features and operation difficulty features, the target puncture depth is predicted. Based on safety risk features and the target vascular region within the region of interest, the obstacle avoidance range boundary is predicted. Based on safety risk features and operation difficulty features, the target perfusion pressure threshold is predicted.

[0013] In one optional implementation, the postoperative prediction result corresponding to the target kidney is determined based on real-time surgical data and kidney imaging images, including: acquiring initial stone features, initial kidney features, and initial vascular features corresponding to the kidney imaging images; spatially overlaying the basic vascular distribution features in the initial vascular features with the puncture trajectory data in the real-time surgical data to generate operation-anatomical fit features; identifying the kidney imaging images to determine the renal pelvis morphological features; constructing pressure-volume correlation features based on the renal pelvis morphological features and real-time renal pelvis pressure data in the real-time surgical data; fusing the initial stone features, initial kidney features, and initial vascular features with the dynamic features in the real-time surgical data to generate dynamic-static fusion features; generating initial postoperative prediction features based on the operation-anatomical fit features, pressure-volume correlation features, and dynamic-static fusion features; and determining the postoperative prediction result corresponding to the target kidney based on the initial postoperative prediction features.

[0014] In one optional implementation, the postoperative prediction result corresponding to the target kidney is determined based on the initial postoperative prediction features, including: filtering the initial postoperative prediction features using L1 regularization penalty of LASSO regression to obtain candidate feature factors; inputting the candidate feature factors into a preset random forest model, ranking them by feature importance score, and selecting target feature factors whose feature importance score is greater than a preset importance score threshold; the target feature factors include: stone feature factors, surgical operation feature factors, anatomical risk feature factors, and intraoperative status feature factors; inputting the target feature factors into a preset postoperative prediction result determination model; the preset postoperative prediction result determination model identifies the target feature factors and determines... The system identifies target static and dynamic features. The static features are input into a feature interaction layer, outputting higher-order interaction features. These higher-order interaction features include interaction features between duration of blood pressure exceeding a preset threshold and degree of hydronephrosis, interaction features between vascular density and number of puncture channels, interaction features between total perfusion volume and renal parenchymal thickness, and interaction features between lithotripsy duration and stone CT value. The dynamic features are input into a temporal feature extraction layer, outputting temporal fusion features. The static features, higher-order interaction features, and temporal fusion features are integrated into a final feature vector. This final feature vector is input into a prediction layer, outputting postoperative prediction results. Postoperative prediction results include at least one of the following: stone clearance rate, probability of complications, and trend of renal function recovery.

[0015] The intelligent assisted method for percutaneous nephrolithotomy provided in this application acquires renal images corresponding to the target kidney, providing anatomical data for all subsequent analyses. It clearly presents the three-dimensional spatial distribution of the kidney, stones, and blood vessels, avoiding risks such as vascular damage and renal parenchymal tearing caused by blind puncture. It eliminates the influence of anatomical differences among patients, enabling personalized treatment plans. The method identifies the renal images and determines the target puncture parameters. Based on accurate image recognition, it outputs quantitative parameters such as puncture point, angle, depth, and obstacle avoidance range, replacing traditional experience-based puncture procedures and improving the accuracy of the puncture operation. It avoids high-risk blood vessels, scar tissue, and other obstacle areas in advance, reducing the incidence of complications such as intraoperative bleeding and extravasation of perfusion fluid, shortening the operation time, and minimizing renal parenchymal damage. It acquires real-time surgical data of the target kidney during the puncture process, capturing dynamic information such as the puncture trajectory, renal pelvis pressure, and bleeding rate in real time, overcoming the limitations of preoperative static image analysis. It provides immediate basis for intraoperative adjustments (such as real-time correction when the puncture trajectory deviates from the planned path), achieving a closed loop of "preoperative planning - intraoperative adaptive adjustment". Accumulating intraoperative dynamic data provides real-world surgical process characteristics to support postoperative prediction. Postoperative predictions are determined based on real-time surgical data and kidney imaging. Integrating preoperative static anatomical features with intraoperative dynamic procedural features improves the accuracy of postoperative outcome prediction. High-risk patients are identified in advance, assisting physicians in developing targeted postoperative care plans (such as intensive anti-infection treatment for high-risk patients). Quantitative data is provided for doctor-patient communication, allowing patients to better understand postoperative recovery expectations and improving transparency in diagnosis and treatment. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the first process of the intelligent assistance method for percutaneous nephrolithotomy according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the second process of the intelligent assistance method for percutaneous nephrolithotomy according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0019] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0020] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] According to an embodiment of the present invention, an embodiment of an intelligent assisted method for percutaneous nephrolithotomy is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0022] This embodiment provides an intelligent assistance method for percutaneous nephrolithotomy, which can be used in electronic devices. Figure 1 This is a flowchart of an intelligent assisted method for percutaneous nephrolithotomy according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain the kidney image corresponding to the target kidney.

[0023] Among them, kidney imaging images include at least one of multi-slice spiral CT images and color Doppler ultrasound images.

[0024] Specifically, the electronic device can receive kidney images corresponding to the target kidney input by the user, or it can receive kidney images corresponding to the target kidney sent by other devices.

[0025] Electronic equipment can perform noise reduction and slice thickness reconstruction on multi-slice spiral CT images, and unify the coordinate system (establishing a three-dimensional coordinate system based on the coronal plane of the human body); it can also perform frame extraction and enhancement processing on color Doppler ultrasound images to highlight the grayscale differences of stones, blood vessels and renal calyx structures.

[0026] Step S102: Identify the kidney image and determine the target puncture parameters corresponding to the target kidney.

[0027] Specifically, electronic devices can identify kidney images, extract core image features, and then determine the target puncture parameters corresponding to the target kidney based on the core image features.

[0028] This step will be explained in detail below.

[0029] Step S103: Obtain real-time surgical data of the target kidney during the puncture process.

[0030] Specifically, electronic devices can use surgical robots or navigation systems to collect real-time coordinates of the puncture needle, needle insertion speed, dynamic changes in puncture depth, and the number of puncture channels. These devices can include PCNL surgical systems with positioning capabilities, holmium laser lithotripters, and infusion pumps (supporting real-time flow and pressure monitoring) to collect data such as lithotripsy duration, cumulative impact energy, infusion flow rate (250-400 ml / min), and total infusion volume.

[0031] The electronic device can also collect renal pelvis pressure through an intrarenal pelvic pressure measuring catheter (inserted retrogradely through the urethra into the renal pelvis and connected to the anesthesia machine pressure measuring module) to distinguish between instantaneous blood pressure higher than a preset threshold (≥30 mmHg and <1 min), transient blood pressure higher than a preset threshold (1-10 min), and sustained blood pressure higher than a preset threshold (>10 min).

[0032] The electronic device can also collect heart rate, blood pressure, and blood oxygen saturation in real time via a monitor. It can also identify peak bleeding rates (>5 ml / min marks high risk), real-time urine white blood cell counts, and extravasation warning signals (widening of the perirenal space in the image) through endoscopic imaging.

[0033] Step S104: Based on real-time surgical data and kidney imaging images, determine the postoperative prediction result corresponding to the target kidney.

[0034] Specifically, the electronic device can identify real-time surgical data and kidney imaging images, then fuse the surgical data features corresponding to the real-time surgical data with the kidney imaging features corresponding to the kidney imaging images to generate target fusion features. Then, based on the target fusion features, the postoperative prediction result corresponding to the target kidney is determined.

[0035] This step will be explained in detail below.

[0036] The intelligent assisted method for percutaneous nephrolithotomy provided in this application acquires renal images corresponding to the target kidney, providing anatomical data for all subsequent analyses. It clearly presents the three-dimensional spatial distribution of the kidney, stones, and blood vessels, avoiding risks such as vascular damage and renal parenchymal tearing caused by blind puncture. It eliminates the influence of anatomical differences among patients, enabling personalized treatment plans. The method identifies the renal images and determines the target puncture parameters. Based on accurate image recognition, it outputs quantitative parameters such as puncture point, angle, depth, and obstacle avoidance range, replacing traditional experience-based puncture procedures and improving the accuracy of the puncture operation. It avoids high-risk blood vessels, scar tissue, and other obstacle areas in advance, reducing the incidence of complications such as intraoperative bleeding and extravasation of perfusion fluid, shortening the operation time, and minimizing renal parenchymal damage. It acquires real-time surgical data of the target kidney during the puncture process, capturing dynamic information such as the puncture trajectory, renal pelvis pressure, and bleeding rate in real time, overcoming the limitations of preoperative static image analysis. It provides immediate basis for intraoperative adjustments (such as real-time correction when the puncture trajectory deviates from the planned path), achieving a closed loop of "preoperative planning - intraoperative adaptive adjustment". Accumulating intraoperative dynamic data provides real-world surgical process characteristics to support postoperative prediction. Postoperative predictions are determined based on real-time surgical data and kidney imaging. Integrating preoperative static anatomical features with intraoperative dynamic procedural features improves the accuracy of postoperative outcome prediction. High-risk patients are identified in advance, assisting physicians in developing targeted postoperative care plans (such as intensive anti-infection treatment for high-risk patients). Quantitative data is provided for doctor-patient communication, allowing patients to better understand postoperative recovery expectations and improving transparency in diagnosis and treatment.

[0037] This embodiment provides an intelligent assistance method for percutaneous nephrolithotomy, which can be used in electronic devices. Figure 2 This is a flowchart of an intelligent assisted method for percutaneous nephrolithotomy according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain the kidney image corresponding to the target kidney.

[0038] Please refer to the above description of step S101 for details on this step, which will not be repeated here.

[0039] Step S202: Identify the kidney image and determine the target puncture parameters corresponding to the target kidney.

[0040] Specifically, step S202 above may include the following steps: Step S2021: Identify the kidney image and determine the region of interest corresponding to the target kidney.

[0041] The region of interest includes the target stone region, the target renal parenchyma region, the target vascular region, and the functional region of the target puncture path; specifically, step S2021 may include the following steps: Step a1: Determine the grayscale distribution of the kidney image based on the Otsu's method.

[0042] Specifically, the electronic device can use the Otsu algorithm (maximum inter-class variance method) to iterate through all possible grayscale values ​​and calculate the brightness difference (inter-class variance) between the "suspected stone area" (foreground) and the "kidney parenchyma / urine area" (background) when each grayscale value is used as a boundary. It then finds the grayscale value with the largest difference to divide the image into bright and dark areas. The electronic device can first select the area where the kidney is located through anatomical localization, perform grayscale statistics on the kidney area, and output a grayscale histogram (visually displaying the pixel proportions of different brightness levels) and the grayscale interval corresponding to the optimal segmentation.

[0043] It should be noted that high-density stones have significantly higher gray values ​​than kidney parenchyma and urine, laying the foundation for determining a specific "brightness standard" for stones.

[0044] Step a2: Determine the grayscale threshold for stone identification based on grayscale distribution.

[0045] Specifically, the electronic device can dynamically adjust the final grayscale threshold for stone identification based on the optimal grayscale value found by the Otsu algorithm and with reference to the clinically recognized preset range of CT values ​​for stones (280-320 HU). Only grayscale values ​​within this range will be initially identified as stones.

[0046] Optionally, if calcifications are present in the kidney imaging (their grayscale values ​​are similar to those of stones, making them easily confused), the electronic device can utilize the "bimodal characteristic" of grayscale distribution (stones and calcifications will form two brightness peaks) to supplement an "exclusion threshold" (the grayscale value of calcifications is usually >350 HU). This ultimately forms a dual screening mechanism: a "primary stone identification grayscale threshold (280-320 HU) + an exclusion stone identification grayscale threshold (>350 HU)," preventing calcifications from being misidentified as stones.

[0047] Step a3: Based on the grayscale threshold for stone identification, determine the candidate stone region.

[0048] Specifically, the electronic device can mark pixels in a kidney image whose grayscale values ​​are within the grayscale threshold range for stone identification as white (foreground) and the rest as black (background), thus obtaining a clear black-and-white binary image. The white areas are the candidate stone areas that are suspected to be stones.

[0049] Electronic devices can denoise initial binary images. They can perform a "closing operation" using 5×5 rectangular structuring elements (to fill tiny pores within the stone area), and then perform an "opening operation" using 3×3 rectangular structuring elements (to remove small noise points from the image). Finally, only images with an area ≥5mm² are retained. 2 The connected regions (roughly the size of a sesame seed) are used to exclude interference from small-volume artifacts.

[0050] Step a4: Calculate the gray-level co-occurrence matrix parameters corresponding to the candidate stone regions.

[0051] Specifically, the electronic device can traverse each pixel (x, y) within the candidate stone region, and starting from that pixel, determine the adjacent pixels (x', y') according to a set angle θ and distance d. If the adjacent pixel (x', y') is still within the candidate region, the frequency of occurrence of grayscale value combinations (i=I(x, y), j=I(x', y')) is counted and filled into the corresponding positions in the GLCM matrix. The above process is repeated to construct grayscale co-occurrence matrices corresponding to the four angles (0°, 45°, 90°, 135°).

[0052] To eliminate the influence of candidate stone region size on matrix element values, the gray-level co-occurrence matrix for each angle needs to be normalized (converting element values ​​into probabilities). The sum S of all elements in a single gray-level co-occurrence matrix is ​​calculated. Each element GLCM(i,j) in the gray-level co-occurrence matrix is ​​divided by S to obtain the normalized element P(i,j) = GLCM(i,j) / S, satisfying that the sum of all P(i,j) is 1.

[0053] Then, based on the normalized gray-level co-occurrence matrices of the four angles, the energy value, entropy value, and contrast value of each angle are calculated respectively, and the average value of the four angles is taken as the final gray-level co-occurrence matrix parameters (to ensure the isotropic nature of the features and avoid randomness in a single direction).

[0054] The energy value reflects the uniformity of grayscale distribution in the candidate stone area. The more uniform the texture (e.g., a stone), the higher the energy value; the looser the texture (e.g., a blood clot), the lower the energy value. The calculation formula is: Among them, the energy value of gallstones is ≥0.15, and the energy value of blood clots is <0.1.

[0055] Entropy reflects the complexity of the grayscale distribution in a candidate stone region. The more complex the texture (e.g., a stone with irregular internal crystal arrangement), the higher the entropy value; the simpler the texture (e.g., calcifications with uniform structure), the lower the entropy value. Calculation formula: ( (To minimize the value, avoid log(0) being meaningless). Among them, the entropy value of the stone is ≥2.0, and the entropy value of the calcification is <2.0.

[0056] Contrast reflects the degree of grayscale difference between adjacent pixels within a candidate stone region. A greater grayscale difference (e.g., a clear boundary between the stone and surrounding tissue) results in higher contrast; a smaller difference (e.g., urine or renal parenchyma) results in lower contrast. Calculation formula: Among them, the stone contrast ratio was >50, and the urine and renal parenchyma contrast ratio was <30.

[0057] The electronic device can calculate the arithmetic mean of the energy value, entropy value, and contrast value from four angles to obtain the final gray-level co-occurrence matrix parameters of the candidate stone region (e.g., "energy value = 0.18, entropy value = 2.3, contrast value = 58").

[0058] Step a5: Output the 3D mask of the stone based on the gray-level co-occurrence matrix parameters.

[0059] Specifically, the electronic device can perform a secondary screening of candidate stone regions based on the parameter thresholds corresponding to the gray-level co-occurrence matrix parameters. For example, it can eliminate "fake" stones such as calcifications with an entropy value < 2.0 and blood clots with an energy value < 0.1, retaining only the true stone regions. Then, it performs inter-slice connectivity analysis on the screened two-dimensional stone regions (stones in each layer of CT image), using interpolation algorithms to fill in the gaps between layers, generating a continuous three-dimensional "stone model" (three-dimensional mask). In the mask, stone regions are marked with 1 and non-stone regions are marked with 0, ensuring that the stone model closely matches the actual stone morphology by ≥ 95%.

[0060] Step a6: Determine the target stone region based on the three-dimensional mask of the stone.

[0061] Specifically, the electronic device calculates the three-dimensional boundary coordinates (Xmin / Xmax, Ymin / Ymax, Zmin / Zmax) of the mask, clearly delineating the three-dimensional extent of the stone within the kidney, thereby identifying the target stone area. Furthermore, the electronic device can also identify two key locations—the main protrusion of the stone (the endpoint of the stone's maximum diameter near the renal calyx in three-dimensional space, the core target point during lithotripsy) and the interface between the stone and the renal calyx (the mask boundary area with a distance of <1mm, which needs to be carefully covered during puncture).

[0062] Step a7: Identify the kidney image and determine the three-dimensional mask of each kidney structure.

[0063] The structures include the renal cortex, renal medulla, renal pelvis, and renal calyces.

[0064] Specifically, the electronic device can input kidney images into a pre-trained U-Net attention mechanism model. The U-Net model encoder extracts low-dimensional to high-dimensional features (such as edges, textures, and morphological features) progressively through convolutional and pooling layers. A spatial attention module is embedded at the skip connection between the encoder and decoder. This spatial attention module automatically assigns higher feature extraction weights (30%-50% higher than other regions) to the renal calyces and fornix by calculating the importance score of each pixel in the feature map. Based on clinical anatomical priorities, the renal calyces and fornix are key areas for PCNL surgery puncture and require priority in ensuring segmentation accuracy. The attention module enhances the feature response of this region, suppresses interference from other areas of the renal parenchyma, and ensures that the structural details of the fornix are not missed.

[0065] The U-Net model's attention mechanism decoder upsamples the high-level features extracted by the encoder through transposed convolutional layers, gradually restoring image resolution. Simultaneously, it fuses the corresponding layer feature maps from the encoder output to supplement detailed information. A boundary detection branch is added to the decoder output, which enhances the boundary features of each structure (such as the abrupt grayscale gradient regions between the renal cortex and medulla) by calculating gradient values. Morphological dilation-erosion operations are used to refine the boundary contours, preventing jagged or broken segmentation results.

[0066] Next, connectivity analysis was performed on the preliminary segmentation results to remove isolated small pixel clusters (area < 3mm). 2 This ensures the continuity of tubular / cavitary structures such as renal calyces and renal pelvis (e.g., the connectivity between renal calyces and renal pelvis, and the complete encapsulation of the renal cortex).

[0067] Finally, the attention mechanism U-Net model outputs a two-dimensional segmentation map corresponding to the kidney image. The classification probability of each pixel is mapped to the most likely structural category (renal cortex, renal medulla, renal pelvis, renal calyces) using the Softmax function. Then, specific label values ​​are assigned to different structures according to preset rules, i.e., renal cortex=1, renal medulla=2, renal pelvis=3, renal calyces=4, generating a two-dimensional mask (the mask value is the corresponding structural label, and the background value is 0). This determines the three-dimensional mask of each continuous kidney structure, ensuring seamless connection between layers (interlayer structural offset ≤0.1mm).

[0068] Step a8: Annotate each structure to generate the target renal parenchyma region.

[0069] Specifically, the electronic device can determine the target renal parenchymal region based on the three-dimensional masks of various structures. The target renal parenchymal region = renal cortex mask + renal medulla mask (both are the core parenchymal tissues of the kidney and the main areas of action for the puncture procedure). The electronic device removes the following non-parenchymal structures from the overall kidney mask: renal pelvis and calyces masks (which are hollow structures with no parenchymal function and must be avoided during puncture); background masks (non-renal tissue with no clinical significance); and pathological masks such as calcifications and cysts (if such areas were identified earlier, they need to be additionally marked as "parenchymal abnormalities").

[0070] Then, the electronic device can use pixel-level logical OR operations to fuse the renal cortex mask and the renal medulla mask to generate a preliminary renal parenchyma mask. Preliminary parenchyma mask = renal cortex mask ∪ renal medulla mask. After the operation, regions with a pixel value of 1 in the mask represent renal parenchyma, and regions with a pixel value of 0 represent non-parenchymal structures.

[0071] The electronic device annotates the preliminary parenchymal tissue mask with clinical attributes and adds structured labels to facilitate subsequent puncture path planning. The electronic device also annotates the spatial coordinate range of the preliminary parenchymal tissue mask region and outputs the 3D bounding box (x, y) of the preliminary parenchymal tissue mask region. min ,y min ,z min ,x max ,y max ,z max ), clearly define the location of the actual area in the image; The electronic device calculates the thickness of the renal parenchyma at different levels (such as the coronal and sagittal planes), marking the "effective parenchyma area" (safe puncture area) with a thickness ≥3mm and the "weak parenchyma area" (risk puncture area) with a thickness <3mm. The electronic device marks the junction of the renal parenchyma with the renal pelvis and calyces as the "restricted area boundary" of the puncture path.

[0072] Finally, the electronic device can use morphological closing operations (dilation followed by erosion) to process the initial parenchymal mask, eliminating burrs and micro-voids at the mask boundaries, making the contour of the parenchymal region more closely match the anatomical shape of the kidney. Specifically, the electronic device can select 3×3×3 spherical structural elements (adapting to the three-dimensional curved surface of the kidney) to avoid excessive smoothing that could distort the parenchymal region. Abnormal regions in the mask (such as small cysts and calcifications within the parenchyma) are then secondary-marked: the pixel value of the abnormal region is corrected from 1 to 2, and it is labeled as a "parenchymal abnormality area". In the final output, "normal parenchymal region" (pixel 1) and "parenchymal abnormality area" (pixel 2) are distinguished, providing a basis for avoiding abnormal regions in the puncture path.

[0073] Step a9: The Hessian matrix vascular enhancement algorithm is used to identify the kidney image to obtain the initial vascular anatomy.

[0074] Specifically, the electronic device can calculate the Hessian matrix for each pixel in a kidney image, identify tubular structures by analyzing the matrix's eigenvalues, and set a threshold for the proportion of eigenvalues ​​(λ2 / λ1 < 0.3, where λ1 is the largest eigenvalue and λ2 is the second largest eigenvalue) to filter out candidate vascular regions. Then, the electronic device can perform connectivity analysis on the candidate vascular regions, removing fragments shorter than 5 mm (to avoid mistaking vascular fragments for complete vessels), obtaining a two-dimensional initial vascular anatomy, which is subsequently used to generate a three-dimensional structure through interlayer reconstruction.

[0075] Step a10: Obtain the blood flow signals corresponding to each blood vessel in the kidney imaging image.

[0076] Specifically, the electronic device can use the ultrasound blood flow signal thresholding method, setting a systolic blood flow velocity >10cm / s as an effective blood flow signal, and by identifying red or blue blood flow areas in the kidney imaging image, mark the blood vessels with blood flow perfusion and generate a two-dimensional mask of blood flow signals.

[0077] Step a11: Determine vascular activity based on blood flow signals.

[0078] Specifically, electronic devices can determine whether blood vessels are physiologically active by measuring the intensity, continuity, and speed of blood flow signals, and exclude "ineffective blood vessels" without blood flow perfusion, such as calcified vessels and image afterimages.

[0079] For example, the criteria are as follows: Highly active vessels: continuous blood flow signal, systolic velocity > 20 cm / s, marked as Grade 1 (must be avoided during surgery); Moderately active vessels: relatively continuous blood flow signal, systolic velocity 10-20 cm / s, marked as Grade 2 (must be avoided with caution); Inactive vessels: no blood flow signal or systolic velocity < 10 cm / s, marked as Grade 0 (will be removed later, no need to avoid).

[0080] Step a12: Based on the initial vascular anatomy and vascular activity, determine the target vascular region.

[0081] Specifically, the electronic device can perform a "logical AND operation" between the initial vascular anatomy mask and the vascular activity mask, retaining only vascular regions with an activity level ≥1 and eliminating inactive vessels.

[0082] Then, the electronic device can use an ellipse fitting algorithm to calculate the vessel diameter (major axis diameter, accurate to 0.1 mm), extract the vessel orientation vector, and statistically analyze the vessel density in the functional area of ​​the target puncture path (between the posterior axillary line and the scapular line, in the 11th intercostal space / below the 12th rib). Based on the vessel diameter and the distance from the renal calyx fornix, the first-risk vessels (diameter > 2 mm and distance < 3 mm) and the second-risk vessels (diameter 1-2 mm and distance 3-5 mm) are marked, and a 3 mm avoidance zone for the first-risk vessel (which must never be crossed during puncture) is delineated. Finally, a three-dimensional mask and risk labeling map of the target vessel area are generated.

[0083] Step a13: Based on preset principles, determine a preset number of initial puncture paths.

[0084] Specifically, the electronic device can automatically generate 3-5 initial puncture paths following the MARS principle (body, acute angle, dome, short distance) recognized in PCNL surgery. Each path is a cylindrical three-dimensional area with a diameter of 5mm (similar to a "precision channel").

[0085] For example, the generation rules are as follows: Main body principle: The endpoint of the path must cover the main protrusion of the stone to ensure that the main part of the stone can be reached during lithotripsy; Dome principle: The endpoint of the path is selected in the first priority area of ​​the renal calyx dome (high elasticity + low vascular risk) to reduce puncture damage; Short distance principle: The starting point of the path is selected in the pre-set puncture area on the body surface (between the 11th intercostal space / 12th intercostal space, the posterior axillary line and the scapular line) to ensure that the distance from the skin to the renal calyx is the shortest and to reduce trauma; Acute angle principle: The angle between the path and the axis of the renal calyx is controlled between 30° and 60° to make it easier for doctors to operate.

[0086] Step a14: Identify each initial puncture path and determine the basic puncture path functional characteristics and obstacle puncture path functional characteristics corresponding to each initial puncture path.

[0087] Specifically, the electronic device can identify each initial puncture path and determine the path length (the straight-line distance from the origin on the body surface to the end point in the fornix, accurate to 0.1 cm), the angle between the initial puncture path and the long axis of the kidney (accurate to 0.1°), and the proportion of the elastic zone in the initial puncture path (the length of the path passing through the highly elastic zone / the total path length, the higher the better). Then, based on the path length of the initial puncture path, the angle between the initial puncture path and the long axis of the kidney, and the proportion of the elastic zone in the initial puncture path, the electronic device determines the functional characteristics of the basic puncture path.

[0088] The electronic device can also determine the number of first-risk vessels crossed (ideally 0), the proportion of renal parenchymal scarring (length of the initial puncture path through the scar area / total path length, ≤5% is preferred), and the distance to the second-risk vessel (minimum distance between the initial puncture path and the second-risk vessel, ≥2mm is preferred) for each initial puncture path. Then, based on the number of first-risk vessels crossed, the proportion of renal parenchymal scarring, and the distance to the second-risk vessel for each initial puncture path, the electronic device determines the functional characteristics of the obstacle puncture path corresponding to each initial puncture path.

[0089] Step a15: Based on the functional characteristics of the basic puncture path and the functional characteristics of the obstacle puncture path, the functional area of ​​the target puncture path is determined in each initial puncture path.

[0090] Specifically, electronic devices can calculate the path efficiency index based on the following formula: Path efficiency index = (elastic zone percentage × 0.4 + angle fit × 0.3) ÷ (number of first-risk blood vessels crossed × 0.5 + scar percentage × 0.2), where angle fit = 1 - |actual angle - 45°| / 45° (45° is the ideal angle, the closer it is to the ideal angle, the higher the score). Then, the electronic devices are sorted in descending order of path efficiency index, and the candidate puncture path with the largest path efficiency index is selected from each initial puncture path. Based on the candidate puncture path, the functional area of ​​the target puncture path is determined.

[0091] Step S2022: Identify the region of interest and extract the initial features of interest corresponding to the region of interest.

[0092] Specifically, the initial features of interest include initial stone features; the region of interest includes the target stone region and the target renal parenchyma region; step S2022 above may include the following steps: Step b1: Identify the target stone region and determine the basic morphological characteristics of the stone corresponding to the target stone region; Specifically, the electronic device can identify the target stone region and determine the location characteristics of the stone within that region. The three-dimensional space of a kidney imaging image is composed of discrete voxels, each with a fixed volume. The electronic device can calculate the stone volume by counting the number of voxels within the target stone region. The calculation formula is: Voxel volume = slice thickness × pixel pitch × pixel pitch (for example, when the slice thickness is 0.625 mm and the pixel pitch is 0.5 mm, the voxel volume = 0.625 × 0.5 × 0.5 = 0.15625 mm). 3 The electronic device can count the total number of voxels N (the number of voxels with a mask value of 1) within the target stone area. The electronic device calculates the stone volume V = N × voxel volume, with the result accurate to 0.1 cm. 3 (1cm) 3 =1000mm 3).

[0093] Specifically, the electronic device can traverse the three-dimensional coordinates (x, y, z) of all voxels within the target stone area. i ,y i ,z i Then, calculate the Euclidean distance between any two voxel points. Then, the electronic device selects the maximum value among all Euclidean distances as the maximum diameter of the stone, accurate to 0.1 mm.

[0094] Sphericity measures how closely the stone's shape approximates an ideal sphere; the closer the sphericity is to 1, the easier it is to remove the stone through the puncture channel. Electronic devices can extract the stone's surface mesh using 3D reconstruction technology and calculate the surface area; then, substituting... Calculations yielded The result range is 0-1.

[0095] The main protrusion is the endpoint of the stone's maximum diameter closest to the renal calyx, and it is the core target point for lithotripsy. The electronic device can filter out the coordinates of two endpoints corresponding to the stone's maximum diameter, P1(x1,y1,z1) and P2(x2,y2,z2). Combining this with a three-dimensional mask of the renal calyx, the electronic device determines which endpoint is closer to the renal calyx wall and selects that endpoint as the main protrusion; the three-dimensional coordinates of the protrusion are output, accurate to 0.1 mm.

[0096] The contact area between the stone and the renal calyx reflects the degree of adhesion between the stone and the renal calyx wall. A larger area indicates that the stone is more easily fixed and the risk of movement after puncture is lower. Electronic equipment can perform three-dimensional surface reconstruction of the stone and renal calyx; it traverses each vertex of the stone surface and calculates its shortest distance to the renal calyx wall; it then calculates the surface area covered by vertices with a distance <1mm to obtain the contact area between the stone and the renal calyx, accurate to 1mm. 2 .

[0097] The straight-line distance from the center of the stone to the fornix reflects the spatial relationship between the stone target and the critical puncture area (renal calyx fornix). The shorter the distance, the higher the precision required for puncture. Electronic equipment can calculate the geometric center coordinates of the stone: the mean of all voxel coordinates. The electronic device extracts the three-dimensional coordinates P0(x0,y0,z0) of the apex of the renal calyx fornix. The electronic device also calculates the Euclidean distance from the center of the stone to the fornix. Accurate to 0.1mm.

[0098] Step b2: Calculate the mean CT value, standard deviation, coefficient of variation, kurtosis, and skewness corresponding to the CT values ​​within the target stone area.

[0099] Specifically, the electronic device can extract the CT values ​​of all voxels within the target stone area, forming a CT value dataset CT={CT1,CT2,...,CT...} NThen, the electronic device calculates the following statistical parameters: mean CT value: This reflects the average density of the stones; standard deviation: This reflects the dispersion of CT values; coefficient of variation: To eliminate the influence of the average CT value on the degree of dispersion, it is convenient to compare stones of different densities; Kurtosis: describes the steepness of the CT value distribution curve. The higher the kurtosis value, the steeper the curve, indicating that the stone density distribution is more concentrated (such as the higher kurtosis of infected stones); Skewness: describes the symmetry of the CT value distribution curve. Skewness > 0 indicates a right-skewed distribution (high proportion of high-density voxels), and skewness < 0 indicates a left-skewed distribution (high proportion of low-density voxels).

[0100] Step b3: Based on standard deviation, coefficient of variation, kurtosis, and skewness, determine the gray-level heterogeneity characteristics corresponding to the target stone region.

[0101] Specifically, electronic devices can stitch together standard deviation, coefficient of variation, kurtosis, and skewness to obtain the grayscale heterogeneity characteristics corresponding to the target stone region.

[0102] Step b4: Calculate the gradient mean and gradient variance corresponding to the CT values ​​within the target stone region.

[0103] Specifically, the electronic device can perform gradient calculations on the three-dimensional CT data within the target stone area, and use the Sobel operator to calculate the gradient values ​​G in the x, y, and z directions respectively. x G y G z Then, the total gradient value of the voxels is calculated. .

[0104] The electronic device collects the total gradient values ​​of all voxels and calculates the mean gradient. and gradient variance .

[0105] Step b5: Based on the gradient mean and gradient variance, determine the stone texture features and composition prediction features corresponding to the target stone region.

[0106] Specifically, electronic devices can compress the CT value within the target stone area to level L (usually level 32), using the following formula: CT min CT scan max The minimum and maximum CT values ​​are defined within the target stone area.

[0107] The electronic device can divide the total gradient value calculated in step b4 above into M levels (usually 8 levels), as shown in the formula: .

[0108] The electronic device counts the number of voxels with gray level g and gradient level t within the target stone region, and after normalization, obtains the probability matrix P(g,t), which satisfies... .

[0109] Then, the electronic device calculates the gradient mean. It reflects the overall roughness of the texture, and the calculation formula is: The higher the gradient mean, the more drastic the change in stone density and the coarser the texture.

[0110] Electronic devices calculate gradient variance, which reflects the uniformity of texture roughness. The calculation formula is as follows: The larger the gradient variance, the greater the difference in density variation intensity in different regions, and the more uneven the texture.

[0111] Electronic devices can extract Local Binary Pattern (LBP) histogram features. This feature focuses on the microscopic texture pattern of the stone surface, characterizing the gray-level distribution of local voxels through binary encoding, supplementing detailed features that gradient parameters cannot cover. The core logic takes each voxel as the center, compares its gray-level value with the surrounding neighboring voxels, and generates a binary code. The distribution pattern of the code directly reflects the roughness, smoothness, and regularity of the local texture.

[0112] Specifically, the electronic device performs surface reconstruction on the three-dimensional mask of the target stone area, acquiring continuous two-dimensional slices (slice thickness consistent with CT slice thickness, typically 0.625~1mm), focusing on surface texture analysis (the area directly interacting with the instruments during lithotripsy). Then, a 3×3 neighborhood window is selected, with the gray value I of the central voxel as the criterion. c The threshold is set; if the gray value I of the neighboring voxels is... i ≥I c If the code bit is 1, then the code bit is 1; otherwise, it is 0. Combine the code bits of the 8 neighboring domains in clockwise / counterclockwise order to generate an 8-bit binary number, which is then converted into a decimal value (range 0~255), which is the LBP value of the central voxel.

[0113] The distribution frequency of LBP values ​​of all voxels in electronic devices is statistically analyzed, and the value range of 0 to 255 is divided into 16 intervals (bins) to construct an LBP histogram.

[0114] The electronic device extracts three core parameters from the LBP histogram: Energy. (p) k (where is the frequency of the k-th bin), reflecting the regularity of the texture; the lower the energy, the more irregular the texture; entropy: This reflects the complexity of the texture; the higher the entropy, the more complex the texture. Contrast: ( (The average value of bins) reflects the difference in brightness of the texture; the higher the contrast, the more obvious the texture's unevenness.

[0115] Electronic devices combine the core parameters of the gray-level-gradient co-occurrence matrix with the LBP histogram parameters to form a structured texture feature vector, in the following format: Texture Feature Vector = in: The gradient mean, Let E be the gradient variance, E be the LBP energy, En be the LBP entropy, and C be the LBP contrast.

[0116] Next, the electronic device acquires the average CT value (calculated in step b2) and texture feature vector (calculated as described above in this step) of the target stone region. Then, the electronic device can determine the component prediction features based on the prediction rule table. The prediction rule table is established based on the statistical regularities of clinical cases and is used to achieve accurate matching between parameters and components. For example, Table 1 is the prediction rule table.

[0117] Table 1 Prediction Rules

[0118] The electronic device outputs the combined average CT value and key texture parameters to generate the final component prediction feature vector, in the following format: Component Prediction Feature Vector = [Average CT Value, Gradient Mean, LBP Energy, Predicted Component, Matching Confidence]. The matching confidence is calculated based on the degree of agreement between the parameters and the rules (maximum score 100%). For example, the confidence is 90%~100% when the parameters perfectly match the rules, and 60%~89% when there is a partial match.

[0119] Step b6: Based on the CT value difference between the target stone area and the target renal parenchyma area, determine the predictive features of lithotripsy difficulty.

[0120] Specifically, the electronic device can calculate the difference in CT values ​​between the kidney stone and the renal parenchyma. .in, The average CT value corresponding to the target stone area. The average CT value corresponding to the target renal parenchymal region.

[0121] The electronic device obtains the LBP entropy calculated in step b5, and then calculates the rockfall difficulty prediction feature based on the CT value difference and LBP entropy. Rockfall difficulty prediction feature = (ΔCT / 100×0.6+LBP entropy value×0.4)×10.

[0122] Step b7: Based on the basic morphological characteristics, grayscale heterogeneity characteristics, stone texture characteristics, composition prediction characteristics, and stone fragmentation difficulty prediction characteristics, generate initial stone characteristics.

[0123] Specifically, the electronic device integrates the basic morphological features, grayscale heterogeneity features, stone texture features, composition prediction features, and stone fragmentation difficulty prediction features to generate initial stone features. This encompasses five dimensions: morphology, density, texture, composition, and stone fragmentation difficulty. The basic morphological feature subset includes: volume, maximum diameter, sphericity, coordinates of the main convex point, contact area, and distance from the center to the dome; the grayscale heterogeneity feature subset includes: standard deviation, coefficient of variation, kurtosis, and skewness; the texture and composition prediction feature subset includes: gradient mean, gradient variance, LBP histogram parameters, and composition prediction vector; and the stone fragmentation difficulty prediction feature subset includes: CT value difference, LBP entropy value, stone fragmentation difficulty score, and pressure concentration coefficient.

[0124] Specifically, the initial features of interest also include initial kidney features; the region of interest includes the target kidney parenchyma region; step S2022 above may also include the following steps: Step b8: Identify the target renal parenchyma region and determine the basic anatomical features of the kidney.

[0125] The basic anatomical features of the kidney include parameters of the renal calyx and fornix, as well as parameters of the renal parenchyma. The parameters of the renal calyx and fornix include the radius of curvature of the renal calyx and fornix, the opening angle of the calyx neck, and the direction vector of the target renal calyx axis.

[0126] Specifically, the electronic device extracts the three-dimensional coordinates {(x) of all voxels on the surface of the dome. i ,y i ,z i )|i=1,2,...,N}. Then, the least squares method is used to fit the equation of the sphere: Where (a,b,c) are the coordinates of the sphere's center, and R is the radius of the sphere (i.e., the radius of curvature of the dome). The electronic device solves the equation to obtain the radius of the sphere R, with the result accurate to 0.1mm.

[0127] The electronic device extracts the coordinates of the two endpoints of the neck opening, P1(x1,y1,z1) and P2(x2,y2,z2), as well as the coordinates of the midpoint of the neck, P0(x0,y0,z0). Then, it calculates the vector. and The electronic device calculates the opening angle θ of the cup neck using the vector angle formula: Morphological classification coding: θ>120° is U-shaped neck (code 0), θ≤120° is O-shaped neck (code 1).

[0128] Electronic equipment extracts the dome vertex coordinates P q (x q ,y q ,z q Furthermore, the coordinates of all voxels in the plane of the renal calyx neck are extracted, and the midpoint P of the plane is calculated. m (x m,y m ,z m ); Calculate the connecting vector (i.e., the axis direction vector): Output the normalized three-dimensional directional coordinates (x, y, z) (vector magnitude is 1).

[0129] The electronic device identifies the renal cortex and medulla regions one voxel along a predetermined candidate puncture path (a straight line from the skin target to the fornix target), and measures the cortical thickness Tc and medullary thickness Tm respectively; the total thickness T = Tc + Tm, with the result accurate to 0.1 mm.

[0130] In addition, the electronic device extracts the coordinates of all voxels in the renal parenchyma region and uses principal component analysis (PCA) to extract the first principal component (i.e., the direction vector of the kidney's long axis). The normal vector of the body's coronal plane is (The coronal plane is perpendicular to the x-axis, and the normal vector is along the x-axis direction); Calculate the angle α between the two vectors: The results are accurate to 0.5°.

[0131] Step b9: Identify the target renal parenchyma region and determine the average CT value and thickness of the fornix.

[0132] Specifically, the electronic device identifies the target renal parenchyma region and determines the Region of Interest (ROI) in the fornix of the target renal parenchyma (containing only fornix tissue, excluding urine, stones, etc. within the renal calyces). The electronic device calculates the mean CT value of all voxels in the ROI region of the fornix, i.e., the mean CT value of the fornix. The electronic device measures the tissue thickness along the normal direction of the dome, i.e., the dome thickness (D). q (Accurate to 0.1mm).

[0133] Step b10: Calculate the elastic index characteristics of the dome based on the average CT value and thickness of the dome.

[0134] Specifically, the electronic device calculates the elastic index characteristics of the dome based on the average CT value and thickness of the dome, using the following formula: Among them, the elasticity index of the fornix is ​​>5: good elasticity, low risk of puncture damage; 3≤fornix elasticity index≤5: moderate elasticity, force needs to be controlled during puncture; elasticity index of the fornix <3: poor elasticity, easy to puncture rupture, it is recommended to change the target point.

[0135] Step b11: Determine the renal parenchymal tolerance characteristics based on CT values ​​and renal parenchymal parameters within the target renal parenchymal region.

[0136] Among them, the renal parenchymal tolerance characteristics reflect the tolerance of the renal parenchyma in the puncture path area to puncture injury, and are divided into two dimensions: "density uniformity" and "respiratory mobility".

[0137] Specifically, the electronic device extracts all voxel CT values ​​within the target renal parenchyma region. The electronic device calculates the standard deviation S corresponding to all voxel CT values ​​in the renal parenchyma region. r : in This represents the average CT value of the target renal parenchymal region.

[0138] The renal respiratory mobility reflects the displacement of the kidney during respiration. The higher the predicted value, the more necessary it is to wait for the "resting period" of the respiratory cycle (such as the end of expiration) during the puncture, otherwise the puncture is prone to deviation.

[0139] The electronic device calculates the predicted renal respiratory mobility value, which is calculated as: BMI × 0.1 + renal long axis angle × 0.05. A predicted renal respiratory mobility value > 3 indicates high mobility, requiring respiratory gating or paused puncture; a predicted renal respiratory mobility value 1 ≤ predicted renal respiratory mobility value ≤ 3 indicates moderate mobility, requiring puncture after observing respiratory rhythm intraoperatively; a predicted renal respiratory mobility value < 1 indicates low mobility, requiring routine puncture.

[0140] The electronic device is based on the standard deviation S of all voxel CT values ​​in the renal parenchyma region. r The renal respiratory mobility predictor value was used to determine the renal parenchymal tolerance characteristics.

[0141] Step b12: Based on the calyx opening angle, determine the calyx duct ...

[0142] Specifically, the electronic device can first determine the calyx type based on the calyx opening angle, and then determine the calyx duct accessibility score corresponding to the target renal parenchyma region based on the calyx type and the calyx opening angle.

[0143] For example, a U-shaped neck (opening angle > 120°) has a base score of 6 points, and an O-shaped neck (≤ 120°) has a base score of 4 points; for every 10° increase in opening angle, 0.5 points are added (for example, a U-shaped neck with an opening angle of 140° would have 2 × 0.5 = 1 point added, for a total score of 6 + 1 = 7 points).

[0144] Step b13 generates initial kidney features based on basic kidney anatomical features, fornix elasticity index features, renal parenchyma tolerance features, and calyx duct ...

[0145] Specifically, the electronic device integrates the basic anatomical features of the kidney, the elasticity index features of the fornix, the tolerance features of the renal parenchyma, and the calyx duct patency scores calculated in steps b8-b12 to form a structured initial kidney feature set, covering four dimensions: anatomical morphology, elasticity, tolerance, and patency, providing a complete basis for kidney feature analysis for puncture path planning and surgical procedure formulation.

[0146] Specifically, the initial features of interest also include initial vascular features; the region of interest includes the target vascular region; step S2022 above may also include the following steps: Step b14: Identify the target vascular region and determine the basic distribution characteristics of the blood vessels.

[0147] Among them, the basic distribution characteristics of blood vessels include blood vessel diameter, blood vessel orientation vector, blood vessel density in the puncture area, and characteristics of the hypovascular area.

[0148] Specifically, the electronic device can identify the target blood vessel region and determine the target blood vessel within that region. Then, the electronic device performs three-dimensional segmentation of the target blood vessel, extracting continuous cross-sections of each vessel. The electronic device then performs ellipse fitting (least squares method) on the pixels of each cross-section to obtain the major axis D of the ellipse. max short axis D min The electronic device uses the major axis diameter of the ellipse as the diameter of the blood vessel, accurate to 0.1 mm. If the blood vessel is branched, the diameters of the main trunk and each branch are calculated separately.

[0149] Electronic equipment extracts the central axis of target vessels surrounding the renal calyx fornix region (using a vascular skeleton extraction algorithm), obtaining the coordinates of key nodes of the axis {(x1,y1,z1),(x2,y2,z2),...,(x n ,y n ,z n Then, the three-dimensional direction vectors of the axes are calculated using linear fitting or principal component analysis (PCA): The electronic device normalizes the vector (with a magnitude of 1) and outputs the final three-dimensional coordinates (x, y, z).

[0150] Vascular density quantifies the density of blood vessel distribution within the functional area of ​​the target puncture path. The higher the density, the greater the probability of hitting a blood vessel during puncture, and the higher the risk of bleeding.

[0151] The electronic device calculates the vascular density in the puncture area using the following formula: .in: The sum of the cross-sectional areas of all blood vessels within the target puncture path functional area (between the posterior axillary line and the scapular line, in the 11th intercostal space / below the 12th rib) (unit: mm2); L is the puncture path length (the straight-line distance from the skin target point to the fornix target point, unit: mm). The electronic device can mark all regions that meet the conditions within the functional area of ​​the target puncture path with a blood vessel density of <0.1 mm2 / mm as the threshold; extract the three-dimensional coordinate boundary (vertex coordinates of the minimum bounding box) of each hypovascular region; calculate the surface area (accurate to 1 mm2) and three-dimensional coordinates (xc, yc, zc) of the geometric center of the hypovascular region; sort them by area from largest to smallest, and preferentially recommend hypovascular regions with an area >50 mm2 as puncture targets.

[0152] Step b15: Calculate the single vessel risk score for each vessel based on the vessel diameter and the distance between each vessel and the fornix.

[0153] Specifically, the electronic device can add the vessel diameter to the distance between the vessel and the fornix to calculate the individual vessel risk score. Based on the individual vessel risk score, the electronic device can determine the risk level of each vessel. When the individual vessel risk score is greater than a first preset vessel score threshold, the vessel's risk level is the first risk level; when the individual vessel risk score is less than or equal to the first preset vessel score threshold but greater than a second preset vessel score threshold, the vessel's risk level is the second risk level; when the individual vessel risk score is less than or equal to the second preset vessel score threshold, the vessel's risk level is the third risk level. The risk level corresponding to the first risk level is greater than the second risk level, and the risk level corresponding to the second risk level is greater than the third risk level.

[0154] Step b16: Calculate the regional vascular risk score corresponding to the target vascular region based on the risk scores of each individual vascular vessel.

[0155] Specifically, the electronic device adds up the individual vessel risk scores corresponding to each vessel within the target vascular region to obtain the regional vascular risk score corresponding to the target vascular region.

[0156] Step b17: Calculate the obstacle avoidance distance and obstacle avoidance angle between the preset puncture path and each blood vessel in the target puncture path functional area, and generate obstacle avoidance association features.

[0157] Specifically, the electronic device can extract the linear equation Lpuncture of the preset puncture path and the central axis equation Lvesse of the blood vessel of the first risk level in the target puncture path functional area.

[0158] Then, the shortest distance d between the two is calculated using the formula for the shortest distance between straight lines in space. min (i.e., obstacle avoidance distance) in, The direction vector corresponding to the preset puncture path. Let be the vector from the preset puncture path to the central axis of the blood vessel of the first risk level. Where d min≥A preset safe distance threshold (e.g., 3mm) can achieve safe avoidance; d min If the preset safe distance threshold is not met, the electronic device needs to adjust the preset puncture path.

[0159] The electronic device can extract the direction vector of the preset puncture path in the target puncture path functional area. Second risk level vascular orientation vector Then, calculate the intersection angle β (i.e., the obstacle avoidance angle) using the vector angle formula: If the obstacle avoidance angle is greater than the preset obstacle avoidance angle threshold, the obstacle avoidance angle is determined to be low risk; if the obstacle avoidance angle is less than or equal to the preset obstacle avoidance angle threshold, the obstacle avoidance angle is determined to be high risk, and the puncture angle needs to be adjusted.

[0160] Specifically, the obstacle avoidance association feature integrates the above two dimensions, and the format is: Obstacle Avoidance Association Feature = [Shortest Avoidance Distance of First Risk Vessel, Crossing Angle of Second Risk Vessel, Avoidance Judgment Result]. The "Avoidance Judgment Result" is "Safe / Needs Adjustment", for example: [4.2mm, 60°, Safe] or [2.5mm, 30°, Needs Adjustment].

[0161] Step b18: Generate initial vascular features based on basic vascular distribution characteristics, single vascular risk scores, regional vascular risk scores, and obstacle avoidance association features.

[0162] Specifically, the electronic device integrates all vascular-related parameters from steps b14-b17 to form a structured initial vascular feature set, covering four dimensions: "distribution morphology, single vascular risk, regional risk, and obstacle avoidance relationship," providing a complete vascular feature basis for the safe optimization of the puncture path.

[0163] Specifically, the initial features of interest also include initial puncture path features; the region of interest includes the functional region of the target puncture path; step S2022 above may also include the following steps: Step b19: Identify the preset puncture path in the target puncture path functional area and determine the basic puncture path characteristics.

[0164] The basic puncture path features include: path length and the angle between the preset puncture path and the long axis of the kidney.

[0165] Specifically, path length refers to the straight-line distance from the skin puncture point to the target point in the renal calyx fornix, which directly reflects the tissue depth that the puncture needle needs to penetrate. The shorter the length, the less trauma the puncture and the more convenient the operation.

[0166] The electronic device can identify the preset puncture path in the target puncture path functional area and determine the marked preset skin puncture point P. s (x s ,ys ,z s ) and the target point P in the renal calyx fornix t (x t ,y t ,z t Then, the path length Lp is calculated using the three-dimensional Euclidean distance formula: .

[0167] The long axis of the kidney is the main anatomical axis of the kidney. The angle between the pre-set puncture path and the long axis of the kidney determines the degree of conformity between the puncture direction and the anatomical structure of the kidney. When the angle is between 30° and 60°, the puncture path is closer to the physiological direction of the kidney and causes less damage to the renal parenchyma.

[0168] Electronic devices can extract the direction vector of the kidney's long axis through principal component analysis (PCA). (Step b8 has been calculated). Then, calculate the direction vector of the preset puncture path. Next, the electronic device can use the vector angle formula to calculate the angle γ between the preset puncture path and the long axis of the kidney: Finally, the two parameters are combined into a feature vector, in the following format: Basic puncture path feature = [L p ,γ].

[0169] Step b20 involves identifying the blood vessels, renal parenchymal scar tissue, and renal parenchymal elastic regions within the functional area of ​​the target puncture path to determine the quantitative characteristics of the obstruction.

[0170] Specifically, the electronic device can construct a cylindrical ROI region with a radius of 3mm centered on a preset puncture path (covering the area that may be disturbed by the puncture needle). Then, the number Nv of all first-risk vessels within this ROI is counted. If Nv=0, it is optimal (no first-risk vessel obstruction); if Nv≥1, the puncture path needs to be adjusted.

[0171] Kidney parenchymal scar tissue is mostly fibrotic areas caused by previous inflammation, surgery, or stone stimulation. It is hard and has poor blood supply, which can easily lead to needle tract tearing and difficulty in healing during puncture. The higher the proportion of scar tissue, the greater the puncture risk.

[0172] Within the preset puncture path, the electronic device can identify scar tissue based on CT values ​​and texture features. Typical characteristics of scar tissue include a decreased CT value (10-20 HU lower than normal renal parenchyma) and uneven texture (large grayscale variance). The electronic device can measure the total length Ls of scar tissue within the preset puncture path; then, it calculates the scar percentage Rs. If R sA scar percentage < the first preset scar percentage threshold (e.g., 10%) is considered excellent; a scar percentage ≤ Rs ≤ the second preset scar percentage threshold (e.g., 30%) is considered medium risk; and a scar percentage > the second preset scar percentage threshold requires avoiding the preset puncture path.

[0173] The highly elastic area of ​​the renal parenchyma (elasticity index ≥ 0.8) has a soft texture and rich blood supply, resulting in less damage and faster healing during puncture. The higher the proportion of the elastic area, the better the safety and tolerability of the puncture route.

[0174] The electronic device can mark areas with an elasticity index ≥ 0.8 within a preset puncture path based on the renal parenchyma elasticity index calculated in step b10; measure the total length Le of the high-elasticity area within the path; and calculate the proportion Re of the elastic area. If R e > The first preset elastic zone percentage threshold (e.g., 80%) is preferred, and the second preset elastic zone percentage threshold (e.g., 50%) is ≤ R. e A percentage ≤ the first preset elastic zone threshold (e.g., 30%) indicates medium risk. e <The second preset elastic zone percentage threshold must be avoided by avoiding this preset puncture path.

[0175] Electronic devices combine three metrics into a feature vector, such as: obstacle quantification feature = [Nv, Rs, Re].

[0176] Step b21: Identify the preset puncture path in the target puncture path functional area, and determine the path efficiency index and operation accessibility score.

[0177] Among them, the path efficiency index represents the cost-effectiveness of the preset puncture path in terms of "stone removal effect-risk". The higher the index, the higher the probability of stone removal under the premise of ensuring safety.

[0178] The calculation formula is: Path Efficiency Index = Predicted Stone-Free Rate / Path Risk Score; where the predicted stone-free rate is a joint prediction based on the basic morphological characteristics of the stone and the initial puncture path characteristics. For example, the predicted stone-free rate = 0.4 × stone volume coefficient + 0.3 × sphericity coefficient + 0.3 × path angle coefficient. The range is 0-1, with higher values ​​indicating more thorough stone removal (e.g., volume < 1 cm). 3 The predicted rate of stone-free spherical calcium oxalate stones can reach 0.9%.

[0179] The path risk score is a weighted score formula that integrates the quantitative characteristics of obstacles: Path Risk Score = Nv × 5 + Rs × 0.1 (the proportion of the elastic zone is a positive indicator and is not included in the risk score), ranging from 0 to 10 points, with higher values ​​indicating greater risk. An efficiency index > 0.2 indicates a highly efficient path, 0.1-0.2 indicates a moderately efficient path, and < 0.1 indicates an inefficient path.

[0180] The accessibility score comprehensively considers the rationality of the puncture path angle, the convenience of the puncture path length, and the degree of obstacle interference, quantifying the ease of operation for the surgeon. The higher the score, the more convenient the operation and the shorter the operation time. For example, Table 2 shows the scoring rules (total score 1-10, weighted sum of 3 dimensions).

[0181] Table 2. Scoring Rules (Illustrated)

[0182] The accessibility score is obtained by adding the scores for puncture path angle, puncture path length, and puncture path obstruction. An accessibility score of 8-10 indicates excellent accessibility (good operator field of vision, minimal movement, recommended as the first choice); a score of 5-7 indicates good accessibility (moderate difficulty, can be considered as an alternative path); and a score of 1-4 indicates poor accessibility (high difficulty, prone to complications, not recommended).

[0183] Step a22: Generate initial puncture path features based on basic puncture path features, obstacle quantification features, path efficiency index, and operational accessibility score.

[0184] Specifically, the electronic device integrates all the core parameters of steps b19-b21 to form a structured initial puncture path feature set, covering three dimensions: basic morphology, obstacle risk, and efficiency accessibility, providing a complete basis for comparing and selecting the best option among multiple paths.

[0185] Step S2023: Based on the initial features of interest, determine the target puncture parameters corresponding to the target kidney.

[0186] Specifically, step S2023 above may include the following steps: Step c1: Input the initial features of interest into the preset LASSO regression model and output candidate core features.

[0187] The pre-defined LASSO regression model uses the target puncture point, target puncture angle, target puncture depth, obstacle avoidance range boundary, and target infusion pressure threshold as joint target variables.

[0188] Specifically, LASSO (Minimum Absolute Shrinkage and Selection Operator) regression is a feature selection and dimensionality reduction algorithm. By applying L1 regularization to the regression coefficients, it compresses the coefficients of unimportant features to 0, thereby selecting the candidate core features that have the greatest influence on the target variable and avoiding the curse of dimensionality and overfitting.

[0189] The electronic device can input the initial features of interest extracted in the previous steps, covering a full-dimensional dataset including initial stone features, initial kidney features, initial vascular features, and initial puncture path features, into a pre-trained pre-defined LASSO regression model. The pre-defined LASSO regression model outputs a set of features with non-zero coefficients, namely candidate core features (such as vascular density, renal parenchymal tolerance features, lithotripsy difficulty score, and operative accessibility score). The training process of the pre-defined LASSO regression model and its processing of the initial features of interest to output candidate core features are existing technologies and will not be elaborated upon here.

[0190] The joint target variable set includes {target puncture point, target puncture angle, target puncture depth, obstacle avoidance range boundary, and target infusion pressure threshold}.

[0191] Step c2: Input the candidate core features into the preset expert evaluation model, evaluate the candidate core features, and output the target core features.

[0192] Specifically, the electronic device can input candidate core features and their correlation coefficients with the target variable into a pre-defined expert evaluation model. The pre-defined expert evaluation model can evaluate each candidate core feature from the following perspectives: whether it directly affects the safety or effectiveness of the puncture (e.g., vascular density and obstacle avoidance distance are mandatory features); whether the measurement accuracy of the candidate core features meets clinical requirements (e.g., path length accuracy must be ≤0.1mm); and whether the candidate core features are independent, eliminating redundant features (e.g., "path length" is highly correlated with "puncture depth," retaining the more critical "path length"). Based on the evaluation results, it outputs a target core feature set that has passed expert consensus verification, providing input for subsequent puncture parameter prediction.

[0193] Step c3: Input the core features of the target into the preset puncture parameter determination model.

[0194] Specifically, the electronic device can input the target core features determined in step c2 into the feature input layer of the puncture parameter determination model through a preset interface, and complete the data format standardization (such as unified units and normalization processing) to prepare for subsequent feature classification and branch calculation.

[0195] Step c4: The preset puncture parameters determine the model to identify each sub-feature in the core features of the target and determine the category corresponding to each sub-feature.

[0196] Specifically, the categories include safety category, operation category, and dissection category.

[0197] The preset puncture parameter determination model identifies each sub-feature in the core features of the target and divides the core features of the target into three categories: safety category, operation category, and anatomy category. Each category of features corresponds to a core target (safety, efficiency, and adaptation) predicted by the puncture parameters, thus achieving "classification processing and accurate modeling".

[0198] For example, the safety category is used to assess risks such as puncture bleeding and renal pelvis hypertension, and may include related features such as vascular density, obstacle avoidance distance, renal parenchymal tolerance characteristics, regional vascular risk score, renal pelvis pressure prediction value, and contact area between the stone and renal calyx; the operation category is used to assess the difficulty and ease of operation of lithotripsy, and may include related features such as stone volume, lithotripsy difficulty prediction characteristics, path length, operation accessibility score, and standard deviation of stone CT value; the anatomy category is used to assess the fit between the path and the renal anatomy, and may include related features such as the radius of curvature of the renal calyx fornix, the opening angle of the calyx neck, the direction vector of the renal calyx axis, and the angle of the renal long axis.

[0199] Step c5: Generate safety category feature group, operation category feature group and anatomy category feature group according to the category corresponding to each sub-feature.

[0200] Specifically, the electronic device can integrate the sub-features classified in step c4 into categories to generate safety category feature groups, operation category feature groups, and anatomy category feature groups. Each feature group serves as an independent input and enters the corresponding branch model for in-depth feature mining.

[0201] Among them, the safety category feature group = [vascular density, obstacle avoidance distance, standard deviation of renal parenchyma CT value, regional vascular risk score, renal pelvis pressure prediction value, contact area between stone and renal calyx]; the operation category feature group = [stone volume, lithotripsy difficulty score, path length, operation accessibility score, standard deviation of stone CT value]; the anatomy category feature group = [radius of curvature of the fornix, calyx neck opening angle, renal calyx axis direction vector, angle of the long axis of the kidney].

[0202] Step c6: Input the security category feature group into the first decision tree branch and output the security risk features.

[0203] Specifically, the first decision tree branch uses a two-layer decision tree ensemble model to perform in-depth analysis of the safety category feature group, quantify the safety risks during the puncture process (such as bleeding and renal pelvis hypertension), and output a continuous risk score of 0-10.

[0204] Specifically, the first-layer decision tree uses blood vessel density and obstacle avoidance distance as core splitting nodes to initially divide the region into "high / medium / low blood vessel risk" areas, outputting a basic blood vessel risk score (0-4 points). The splitting rule is: blood vessel density > 0.3mm. 2 / mm and obstacle avoidance distance <2mm → First risk (4 points); otherwise → Low risk (0 points).

[0205] The deep decision tree incorporates the standard deviation of renal parenchyma CT values ​​and the predicted value of renal pelvis pressure to construct feature interaction nodes (such as the standard deviation of renal parenchyma vascular density) and quantify the risk of puncture bleeding. At the same time, it combines the contact area between the stone and the renal calyx to correct the risk weight of renal pelvis hypertension (the larger the contact area, the higher the risk of renal pelvis hypertension).

[0206] Finally, the scores of the two decision trees are weighted and summed to obtain a safety risk characteristic value of 0-10. If the safety risk characteristic value is ≥6, it is considered the highest risk, requiring enhanced obstacle avoidance and stress control; if the safety risk characteristic value is <3, it is considered low risk, and a conventional puncture strategy can be adopted.

[0207] Step c7: Input the operation category feature group into the second feature extraction branch and output the operation difficulty feature.

[0208] Specifically, the second feature extraction branch analyzes the operation category feature groups based on the integrated learning logic of multi-scene segmentation and feature interaction mining, quantifies the complexity of the surgical operation, and outputs a continuous difficulty score of 0-10.

[0209] Specifically, the electronic device uses stone volume and path length as splitting nodes to divide basic scenarios such as "short path small stones (easy to operate)" and "long path large stones (difficult to operate)". Then, key interaction nodes are constructed, such as lithotripsy difficulty score × stone value standard deviation (reflecting the combined effect of stone hardness and texture complexity) and path length × operation accessibility score (reflecting the constraint of path length on operation convenience). Then, based on clinical operation time data, the feature scores are mapped to three levels of difficulty: 0-3 points: low difficulty (short path + small volume stone, ultrasonic lithotripsy is sufficient); 3-7 points: medium difficulty (conventional holmium laser lithotripsy); 7-10 points: high difficulty (long path + high hardness stone, requiring high-power holmium laser + ultrasonic aspiration).

[0210] Step c8: Input the anatomical category feature group into the third feature extraction branch and output the anatomical adaptation features.

[0211] Among them, the anatomical fit score is used to characterize the fit between the puncture path and the kidney anatomy; Specifically, the third feature extraction branch can use the renal calyx axis direction vector and the calyx neck opening angle as the core to determine whether the path is along the renal calyx axis direction (optimal anatomical path). Then, interactive nodes are constructed, such as the radius of curvature of the fornix × the angle of the kidney's long axis (reflecting the requirements of the fornix morphology on the puncture angle) and the calyx neck opening angle × the thickness of the renal parenchyma (reflecting the influence of the calyx neck morphology on stone expulsion). Then, based on clinical anatomical adaptation standards, the feature scores are converted into 0-10 points: 8-10 points: excellent (path along the renal calyx axis, minimal damage); 5-7 points: good (minor angle adjustment required); <5 points: poor (path replanning required).

[0212] Step c9: Based on the safety risk characteristics, anatomical adaptation characteristics, and stone location characteristics in the basic morphological characteristics of the stone, predict the target puncture point.

[0213] Specifically, the electronic device can use the geometric center of the fornix in the anatomical adaptation features as the initial target point (the fornix is ​​the preferred anatomical location for puncture with minimal damage). Then, a vascular distribution mask based on safety risk features is superimposed. If the initial target point is located in the first-risk vascular region (vascular density > 0.3 mm),... 2 If the target puncture point is located at a distance of 1.5 mm (the diameter of the blood vessel), the point is shifted towards the area with poor blood vessels, with the shift distance equal to 1.5 times the blood vessel diameter (to ensure obstacle avoidance safety). Next, considering the relative position of the stone and the renal calyx, if the stone is close to the calyx neck, the target point is slightly adjusted 5-10 mm towards the calyx neck to improve the stone removal efficiency after fragmentation. Finally, the three-dimensional coordinates (X, Y, Z) of the target puncture point are obtained with an accuracy of ≤0.1 mm, and the risk level (low / medium / high) of the target point is marked.

[0214] Step c10: Based on anatomical adaptation features and operational difficulty features, predict the target puncture angle.

[0215] Specifically, the electronic device can use the renal calyx axis direction vector in the anatomical adaptation features as a reference, with an initial angle set at 30°-60° (conforming to the MARS acute angle puncture principle to reduce renal parenchymal damage). If the calyx neck opening angle is >120° (U-shaped calyx neck): the angle is increased by 5° to widen the stone passage; if the calyx neck opening angle is ≤120° (O-shaped calyx neck): the angle is decreased by 5° to reduce the risk of calyx neck tearing. If the operation difficulty feature score is ≥7 points (high complexity scenario), the angle is adjusted to 35°-55° (the operator's comfort range to improve stability). The vascular distribution features are superimposed; if the intersection angle between the path and the second risk vessel is <45°, the angle is offset by 3°-5° to ensure an obstacle avoidance distance ≥2mm. The final target puncture angle is obtained, with the angle value retained to one decimal place and an accuracy ≤0.5°.

[0216] Step c11: Based on safety risk characteristics and operational difficulty characteristics, predict the target puncture depth.

[0217] Specifically, the prediction of the target puncture depth needs to balance effective puncture (breaking through the fornix) and safe puncture (avoiding excessive puncture and damaging the renal pelvis). Through a three-layer logic of "basic calculation - safety calibration - stone adhesion correction", the accurate depth value and safe fluctuation range are output.

[0218] The electronic device can calculate the initial puncture depth based on the path length in the operation difficulty characteristics. The initial puncture depth = puncture path length + 0.5cm (ensuring the puncture needle penetrates the fornix and enters the renal calyx). If the safety risk characteristic score is ≥6 (first risk): the initial puncture depth is reduced by 0.2cm, using a "shallow puncture + slow needle advance" strategy to reduce the risk of bleeding; if the fornix thickness is <3mm: the initial puncture depth is adjusted to path length + 0.3cm (to avoid excessive depth and damage to the renal pelvis); if the fornix thickness is >5mm: the initial puncture depth is adjusted to path length + 0.7cm (to ensure effective penetration of the fornix). If the contact area between the stone and the renal calyx is >30mm... 2 (High risk of adhesion) The initial puncture depth is further reduced by 0.2cm to avoid stone displacement caused by puncture. Finally, the adjusted target puncture depth is obtained. The depth value is retained to one decimal place with an accuracy of ≤0.1cm, and a safe fluctuation range (±0.2cm) is output.

[0219] Step c12: Based on the safety risk characteristics and the target blood vessel region in the region of interest, predict the boundary of the obstacle avoidance range.

[0220] Specifically, predicting the boundary of the obstacle avoidance range is key to preventing massive intraoperative bleeding. Through the logic of "risk area identification - three-dimensional boundary generation - dynamic expansion", the prohibited crossing areas of the puncture path are clearly defined.

[0221] The electronic device can automatically identify the first-risk blood vessel, the second-risk blood vessel, and renal parenchymal scar tissue by combining safety risk features and semantic segmentation models, generating a two-dimensional risk mask (marking dangerous areas). Then, the two-dimensional mask is mapped to three-dimensional space, and the avoidance range is calculated according to the blood vessel risk level: First-risk blood vessel: avoidance range = vessel diameter × 1.5 + 3 mm; Second-risk blood vessel: avoidance range = vessel diameter × 1.2 + 2 mm; Scar tissue: avoidance range = scar area radius + 2 mm. If the safety risk feature score is ≥ 7 (extremely high risk), the overall avoidance range is increased by 1 mm, and a "vascular variation warning zone" (such as the area of ​​an arcuate artery crossing the fornix) is marked. Finally, the boundary of the avoidance range is obtained, which can include the boundary of the first-risk area and the boundary of the second-risk area, clearly defining the restricted area of ​​the puncture path.

[0222] Step c13: Based on safety risk characteristics and operational difficulty characteristics, predict the target perfusion pressure threshold.

[0223] Specifically, predicting the perfusion pressure threshold is key to preventing complications of renal pelvic hypertension. It is necessary to balance the clarity of the lithotripsy field of vision with the safe upper limit of renal pelvic pressure, and output the threshold through the logic of "basic threshold setting - multi-factor adjustment - dynamic feedback".

[0224] The electronic device can set the initial perfusion pressure threshold to ≤200 mmHg based on clinical renal pelvis pressure control standards (ensuring a safe upper limit of <30 mmHg for intrarenal pelvis pressure). If the safety risk characteristic score is ≥6 (first risk): the threshold is reduced to 180 mmHg (reducing the risk of reflux caused by renal pelvis hypertension); if the lithotripsy difficulty score is ≥8 (high-hardness stones): the threshold is further reduced to 160 mmHg (requiring high-power lithotripsy to reduce perfusion fluid volume); if the standard deviation of the renal parenchyma CT value is >10 HU (low renal parenchyma tolerance): the threshold is further reduced by 20 mmHg. The electronic device, combined with the intraoperative renal pelvis pressure prediction value, automatically lowers the threshold by 10 mmHg if the predicted blood pressure is higher than the preset threshold for >1 minute. Finally, the target perfusion pressure threshold is obtained, with the threshold adjusted in 5 mmHg increments, with an accuracy ≤5 mmHg, and a pressure warning line (e.g., 180 mmHg) is output.

[0225] Step S203: Obtain real-time surgical data of the target kidney during the puncture process.

[0226] Please refer to the above description of step S103 for details on this step, which will not be repeated here.

[0227] Step S204: Based on real-time surgical data and kidney imaging images, determine the postoperative prediction result corresponding to the target kidney.

[0228] Specifically, step S204 above can include the following steps: Step S2041: Obtain the initial stone features, initial kidney features, and initial vascular features corresponding to the kidney imaging image.

[0229] Specifically, please refer to the above description of step S2022 for details on this step, which will not be repeated here.

[0230] Step S2042: Spatial overlay of the basic vascular distribution features in the initial vascular features with the puncture trajectory data in the real-time surgical data to generate operation-anatomy fit features.

[0231] Specifically, the electronic device can spatially register the three-dimensional coordinates of the basic vascular distribution in the initial vascular features with the real-time puncture trajectory, and use the Euclidean distance algorithm to calculate the actual distance deviation between the puncture path and the first-risk vessel (safety risk score ≥ 6 points): The result was accurate to 0.1 mm.

[0232] Then, the electronic device extracts the dynamic change curve of the renal parenchyma penetration thickness during the puncture process, calculates the penetration thickness fluctuation value ΔT (the absolute value of the difference between the real-time thickness and the preoperative planned thickness), and the average penetration thickness T. 平均 .

[0233] Next, the electronic device can use a weighted summation formula to generate the dissection fit (values ​​from 0 to 1, with values ​​closer to 1 indicating better fit), the formula is as follows: .

[0234] The safety distance deviation threshold can be set to 3mm (clinical vascular obstacle avoidance safety distance). If d 偏差 >3mm, then A value of 0 indicates extremely poor compatibility and a very high risk of vascular damage.

[0235] Among them, the operation-anatomy fit characteristic ≥0.8: excellent (the puncture trajectory fits the safe path and no adjustment is needed); 0.5-0.8: medium (slight deviation, the puncture angle / depth needs to be fine-tuned); operation-anatomy fit characteristic <0.5: poor (severe deviation, the needle insertion needs to be stopped immediately and the path replanned).

[0236] Step S2043: Identify the kidney imaging image and determine the morphological characteristics of the renal pelvis.

[0237] Specifically, electronic devices can segment the renal pelvis region from kidney images using a three-dimensional semantic segmentation algorithm and extract the morphological features of the renal pelvis. These morphological features include: renal pelvis volume V (unit: mL, accurate to 0.1 mL), calculated using voxel counting to determine the three-dimensional volume of the renal pelvis lumen; and renal pelvis wall thickness T. 壁 (Unit: mm, accurate to 0.1 mm) Take 10 sampling points along the circumference of the renal pelvis wall and calculate the average thickness; Renal pelvis morphology classification: divided into intrarenal type (renal pelvis is located in the renal parenchyma) and extrarenal type (renal pelvis protrudes from the renal parenchyma), with intrarenal type having a higher risk of extravasation.

[0238] Finally, the renal pelvis morphological feature set is obtained as [renal pelvis volume, renal pelvis wall thickness, morphological classification].

[0239] Step S2044: Based on the morphological characteristics of the renal pelvis and the real-time renal pelvis pressure data in the real-time surgical data, construct pressure-volume correlation features.

[0240] Specifically, the electronic device can use renal pelvis volume V as the independent variable and real-time renal pelvis pressure P as the dependent variable, and employ a quadratic polynomial regression to fit the basic correlation: P = aV 2 +bV+c (a, b, and c are fitting coefficients, which are iteratively updated using pressure-volume data collected in real time during the operation).

[0241] Then, a renal pelvis wall thickness correction term is introduced to optimize the model: (where k is the correction factor, T) 正常 The normal reference value for renal pelvis wall thickness is 2 mm; T 壁 The smaller the value, the higher the corrected pressure value, and the greater the risk of leakage.

[0242] Finally, the electronic device calculates the extravasation risk factor of the infusion fluid (value 0-1), using the following formula: (P) 阈值 The safe upper limit for renal pelvis pressure is 30 mmHg; the closer the extravasation risk coefficient is to 1, the higher the risk of extravasation, and an intraoperative warning is triggered when it is ≥0.8.

[0243] Then, the risk coefficient of extravasation was determined as a pressure-volume correlation characteristic.

[0244] Step S2045: The initial stone features, initial kidney features, and initial vascular features are fused with the dynamic features in the real-time surgical data to generate dynamic-static fusion features.

[0245] The dynamic features in real-time surgical data include puncture trajectory data, real-time renal pelvis pressure, bleeding rate, lithotripsy duration, and irrigation fluid volume.

[0246] Specifically, the electronic device can acquire the weight information corresponding to the initial stone features, initial kidney features, initial vascular features, and various dynamic features in the real-time surgical data. Then, it fuses the weight information corresponding to the initial stone features, initial kidney features, initial vascular features, and dynamic features in the real-time surgical data to generate dynamic-static fusion features.

[0247] For example, among the dynamic features, the duration of sustained blood pressure above a preset threshold (renal pelvis pressure ≥30 mmHg) has a weight of 0.3; the peak bleeding rate has a weight of 0.2; among the static features, the stone clearance rate-related features (volume, sphericity) have a weight of 0.25; the regional vascular risk score has a weight of 0.15; and other features, such as the operability score and renal pelvis volume, have a weight of 0.1.

[0248] After normalizing all features to the 0-1 range, the electronic device sums them according to their weights to generate a dynamic-static fusion feature matrix with uniform dimensions (example dimension: 1×50). Each element in the matrix corresponds to a weighted feature value.

[0249] Step S2046: Based on the operation-anatomical fit characteristics, pressure-volume correlation characteristics, and dynamic-static fusion characteristics, generate initial postoperative prediction characteristics.

[0250] Specifically, the electronic device integrates the operation-anatomical fit of step S2042, the pressure-volume correlation features of S2044, and the dynamic-static fusion features of S2045 to form an initial postoperative prediction feature set for predicting postoperative outcomes, covering three dimensions: "operational safety, perfusion risk, and overall surgical effect".

[0251] Step S2047: Based on the initial postoperative prediction features, determine the postoperative prediction result corresponding to the target kidney.

[0252] The postoperative expected results include at least one of the following: stone clearance rate, probability of complication, and trend of renal function recovery.

[0253] Specifically, step S2047 above may include the following steps: Step d1 involves using L1 regularization penalty of LASSO regression to screen the initial postoperative predicted features and obtain candidate feature factors.

[0254] Specifically, the electronic device inputs the initial postoperative predictive features generated in step S2046 (covering manipulation-anatomical fit, pressure-volume correlation, dynamic-static fusion features, etc., typically 50-100 dimensions) into the LASSO regression model. The L1 regularization of the LASSO regression compresses the weights of weakly correlated features by adding an L1 penalty term to the loss function, as shown in the formula: , where y i This represents the actual postoperative outcome value (such as the actual stone clearance rate). The model predicts the value; λ: regularization coefficient (the optimal value is determined through 5-fold cross-validation). The larger λ is, the more weights are assigned to weakly correlated features. j Compressed to 0; w j The weight is the weight of the j-th feature; features with a weight of 0 are directly removed.

[0255] Finally, strong correlation features such as retained stone volume, duration of sustained blood pressure above the preset threshold, and vascular density were obtained, while meaningless features such as image noise-derived features were removed, resulting in 30-40 dimensional candidate risk factors.

[0256] Step d2: Input the candidate feature factors into the preset random forest model, sort them by feature importance score, and filter the target feature factors whose feature importance score is greater than the preset importance score threshold.

[0257] Among them, the target characteristic factors include: stone characteristic factors, surgical operation characteristic factors, anatomical risk characteristic factors, and intraoperative status characteristic factors; The core parameters of the preset random forest include: number of decision trees (n_estimators): 100-200 (balancing model accuracy and computational efficiency); feature sampling method: randomly select √p features for each tree (p is the dimension of the candidate features); node splitting criterion: Gini coefficient (i.e., "Gini impurity").

[0258] Electronic devices can randomly sample multiple subsets from clinical datasets (including candidate features and postoperative outcome labels) and assign one subset to each decision tree. Each decision tree splits nodes based on the Gini coefficient: features that minimize the "impurity" of nodes are preferentially selected as split nodes (e.g., "duration of sustained blood pressure above a preset threshold" can more accurately distinguish "occurrence / non-occurrence of complications" and will be preferentially selected as split nodes). After all decision trees are trained, a complete preset random forest model is formed.

[0259] The electronic device inputs candidate feature factors into a pre-defined random forest model, which calculates the Gini coefficient for each candidate feature factor. The Gini coefficient measures "how much a feature reduces the uncertainty of data classification / prediction," and the formula is: Where: D: the dataset of the current node; K: the number of categories of the target variable (e.g., K=2 for complication "occurrence / non-occurrence"); p k Let be the proportion of the k-th class of samples in dataset D.

[0260] For each feature, calculate the "total decrease in Gini coefficient when a node splits" across all decision trees. This value is the feature's Gini importance score. For each split node in a single decision tree, calculate the "weighted sum of Gini coefficients of child nodes after splitting" to obtain the Gini decrease for that node. If the splitting feature of that node is feature j, add this decrease to the total score of feature j. Traverse all decision trees to obtain the total Gini importance score for each candidate feature factor. Among them, high-importance features (high Gini coefficient) contribute significantly to postoperative outcome prediction, such as "duration of sustained blood pressure above the preset threshold" (directly associated with infection risk), "peak bleeding rate" (directly associated with bleeding complications), and "stone volume" (directly associated with clearance rate). Low-importance features (low Gini coefficient) contribute little to prediction, such as "renal calyx morphology fit" (only indirectly affects clearance rate and has strong substitutability).

[0261] Then, the electronic device calculates the Gini importance score of all candidate feature factors and sorts them from high to low. The lowest score among the first 15-20 dimensions is taken as the importance score threshold (e.g., if the lowest score among the first 20 dimensions is 0.02, then the importance score threshold is 0.02). Features with Gini coefficients greater than the importance score threshold are retained, while features with scores less than or equal to the importance score threshold are removed (these features have negligible contribution to the prediction; their removal does not affect model accuracy and reduces dimensionality), thus obtaining the target feature factors.

[0262] The selected 15-20 target feature factors are divided into four categories based on "clinical attributes + influencing dimensions." The clinical significance and selection criteria for each category are as follows: Stone characteristic factors, including stone volume, average CT value, and estimated percentage of residual stones, are selected based on the fact that a high Gini coefficient directly determines the difficulty of stone removal and is a core influencing factor on the removal rate, thus predicting the stone removal rate and residual risk. Surgical operation characteristic factors, including the number of puncture channels, cumulative duration of sustained blood pressure above the preset threshold (≥30 mmHg), and total perfusion volume, are selected based on the fact that a high Gini coefficient directly reflects the intensity of the surgical operation and is associated with complications / renal function impairment. The anatomical risk characteristics include renal parenchymal thickness, vascular density in the puncture area, and degree of hydronephrosis. The selection criteria (due to a high Gini coefficient) reflect the patient's basic anatomical risk, amplify the adverse effects of the procedure, and thus predict the risk of bleeding, extravasation, and renal function repair capacity. The intraoperative status characteristics include peak bleeding rate, real-time urine white blood cell count, extravasation warning signal of perfusion fluid, renal calyx morphology fit, and bleeding influence coefficient. The selection criteria (due to a high Gini coefficient) reflect the intraoperative risk status in real time and serve as an immediate warning indicator for complications, thereby predicting the probability of intraoperative / postoperative complications.

[0263] Step d3: Input the target feature factors into the preset postoperative prediction result determination model; Specifically, electronic devices can input characteristic factors into a pre-defined postoperative prediction result determination model.

[0264] Step d4: The model for determining postoperative prediction results identifies target feature factors and determines the target static and dynamic features. Specifically, the model for determining the postoperative prediction results can split the target feature factors according to time attributes and data types, and adopt differentiated processing strategies for different types of features to maximize the mining of feature value.

[0265] The target static features, which do not change over time, are derived from preoperative imaging / intraoperative fixed records and may include stone volume, average CT value, renal parenchymal thickness, vascular density in the puncture area, preoperative eGFR, and degree of hydronephrosis. Static features reflect basic risks and are suitable for interactive analysis. The target dynamic features change dynamically with the surgical progress and are derived from real-time intraoperative monitoring data. These may include renal pelvic pressure change curves, dynamic changes in bleeding rate, cumulative duration of sustained blood pressure exceeding a preset threshold, urinary leukocyte sequence, and extravasation warning time series. Dynamic features reflect intraoperative trends and are suitable for time-series modeling. Step d5 inputs the target static features into the feature interaction layer and outputs higher-order interactive features.

[0266] Among them, the higher-order interaction features include the interaction features of duration of blood pressure above the preset threshold and degree of hydronephrosis, interaction features of vascular density and number of puncture channels, interaction features of total perfusion volume and renal parenchymal thickness, and interaction features of lithotripsy duration and stone dT value.

[0267] Specifically, different static features have significantly different degrees of influence on postoperative outcomes. The feature interaction layer can automatically calculate the importance score of the target static features through the attention mechanism, assign high weights to core features (to amplify their influence), and assign low weights to secondary features (to weaken their interference), ensuring that the model focuses on key clinical factors.

[0268] Specifically, the electronic device can calculate the correlation coefficient between target static features and postoperative outcomes (e.g., the correlation coefficient r=0.8 between the duration of sustained blood pressure above a preset threshold and infectious complications, which is much higher than r=0.2 for renal calyx morphology fit). Then, it combines the preset baseline weights from clinical guidelines (e.g., vascular density is directly related to bleeding risk, with a baseline weight set at 0.25). The electronic device can calculate the gradient contribution of features to the prediction results through backpropagation (the larger the gradient, the higher the importance). For example, among the core features, the duration of sustained blood pressure above a preset threshold, with a weight of 0.25-0.3, directly determines the cumulative effect of renal pelvis pressure and is associated with the risk of infection / extravasation; vascular density, with a weight of 0.2-0.25, directly determines the risk of puncture bleeding and is a core predictor of bleeding complications; lithotripsy time, with a weight of 0.2-0.25, reflects the duration of surgical trauma and is associated with the degree of renal parenchymal damage; among the secondary features, renal calyx morphology fit, with a weight of 0.05-0.1, only indirectly affects the stone clearance rate and can be replaced by "stone volume" and "CT value"; the degree of hydronephrosis (non-interactive), with a weight of 0.08-0.1, has limited impact on its own and only shows value after interacting with the duration of blood pressure above a preset threshold.

[0269] Then, the electronic device performs weighted scaling on the original values ​​of each target static feature, using the formula: Weighted feature value = Original feature value × Attention weight.

[0270] Since a single feature cannot reflect the "cumulative risks" of complex clinical scenarios (such as only looking at the duration of blood pressure above a preset threshold or only looking at the degree of hydronephrosis, it is impossible to accurately predict the risk of extravasation), it is necessary to construct "feature interaction items" based on prior clinical knowledge to quantify the synergistic effect between features.

[0271] Before constructing all interaction features, the electronic device can first normalize the features involved in the interaction (within the range of 0-1) to eliminate differences in units of measurement (such as the unit of duration in minutes, the unit of volume in centimeters). 3 This ensures that the interaction results are comparable. Then, based on the normalized target static features, higher-order interaction features are constructed.

[0272] For example, the interaction feature of duration of sustained blood pressure above the preset threshold and degree of hydronephrosis is: duration of sustained blood pressure above the preset threshold × degree of hydronephrosis. When there is hydronephrosis, the renal pelvis wall is in a dilated state, and its ability to tolerate high pressure is significantly reduced. The longer the blood pressure is above the preset threshold, the risk of extravasation of the hydronephrotic renal pelvis increases exponentially (rather than linearly). For example, degree of hydronephrosis 0.8 (severe) + duration of blood pressure above the preset threshold 0.9 (18 min) → interaction value = 0.72, and the risk of extravasation is more than twice that of a single feature. The interaction feature between vascular density and the number of puncture channels is: vascular density feature × number of puncture channels feature. Single-channel puncture has limited damage to high vascular density areas, but multi-channel puncture will disturb the area multiple times, and the bleeding risk will be "multiplied" with the increase of the number of channels. For example, vascular density 0.9 (high risk) + number of channels 0.8 (3 channels) → interaction value = 0.72, bleeding probability ≥40% (single feature only 20%). The interaction characteristic between total perfusion volume and renal parenchymal thickness is: total perfusion volume characteristic 1 × (1 - renal parenchymal thickness characteristic); the thinner the renal parenchymal thickness (the smaller the characteristic 2), the easier it is for the perfusion fluid to penetrate the renal parenchyma and cause extravasation; the higher the perfusion volume, the more significant this risk; for example, perfusion volume 0.9 (3000ml) + renal parenchymal thickness 0.2 (thin) → interaction value = 0.9 × 0.8 = 0.72, extravasation probability ≥ 30%; The interaction characteristic between lithotripsy time and stone CT value is: lithotripsy time characteristic × stone CT value characteristic. The higher the stone CT value (the greater the hardness), the more energy / time is required for lithotripsy, and the more severe the cumulative damage to the renal parenchyma caused by the lithotripsy energy. For example, CT value 0.9 (1200 HU) + lithotripsy time 0.8 (60 min) → interaction value = 0.72, and the renal function damage coefficient increases by 0.15.

[0273] Step d6: Input the target dynamic features into the temporal feature extraction layer and output the temporal fusion features.

[0274] The target dynamic features may include dynamic time-series data such as renal pelvis pressure curve (1Hz sampling), bleeding rate sequence, and cumulative duration of sustained blood pressure above a preset threshold.

[0275] Specifically, the electronic device can input the target dynamic features into the temporal feature extraction layer. The temporal feature extraction layer learns long-term dependencies (such as the impact of "blood pressure exceeding a preset threshold for more than 10 minutes" on infection) through input gates, forget gates, and output gates; filters out noise such as instantaneous pressure fluctuations, retains key clinical trends, and outputs temporal fusion features that cover the trend change patterns of dynamic features (such as the slope of pressure rise and the time of peak bleeding rate).

[0276] Step d7: Integrate the target static features, high-order interaction features, and temporal fusion features into the final feature vector; Specifically, electronic devices can concatenate target static features, high-order interactive features, and temporal fusion features to generate a final feature vector.

[0277] Step d8: Input the final feature vector into the prediction layer and output the postoperative prediction result.

[0278] The postoperative expected results include at least one of the following: stone clearance rate, probability of complication, and trend of renal function recovery.

[0279] Specifically, the electronic device inputs the 32-dimensional final feature vector generated by the preoperative-intraoperative end-link integration into the prediction layer. This vector contains three types of core features: target static features (approximately 10 dimensions such as stone volume, renal parenchymal thickness, and vascular density); high-order interactive features (4 dimensions such as duration of sustained blood pressure above the preset threshold - degree of hydronephrosis and vascular density - number of puncture channels); and temporal fusion features (16 dimensions such as renal pelvis pressure change trend and bleeding rate dynamic features).

[0280] Based on different postoperative prediction results, four independent but interconnected predictive sub-models are constructed. Each sub-model corresponds to a prediction target and shares the input data of the final feature vector to achieve accurate prediction across multiple dimensions. For example, the core input feature dimensions for the stone clearance rate prediction sub-model include stone volume, average CT value, lithotripsy time, estimated percentage of residual stones, and renal calyx morphology fit. This sub-model can be a piecewise linear regression model with an accuracy ≥88%. The core input feature dimensions for the complication probability prediction sub-model include the cumulative duration of sustained blood pressure above a preset threshold, vascular density, peak bleeding rate, urine white blood cell count, and extravasation risk coefficient. This sub-model can be a multi-class logistic regression + Sigmoid mapping model with a classification accuracy ≥92%. The core input feature dimensions for the renal function recovery trend prediction sub-model include preoperative eGFR, renal parenchymal thickness, total perfusion volume, duration of sustained blood pressure above a preset threshold, and bleeding impact coefficient. This sub-model can be a linear fitting model with a prediction error ≤10%.

[0281] Specifically, the stone clearance rate prediction sub-model can extract stone-related features (volume, CT value, residual percentage) and operation-related features (lithotripsy time, renal calyx morphology fit) from the final feature vector. Then, based on the stone volume, it divides the stone into three intervals and substitutes them into the corresponding formulas to calculate the clearance rate: small volume stones (<1cm) 3 Stone removal rate = Initial basic removal rate - Residual percentage × 0.8 - 5%, where the initial basic removal rate can be set to 90%; medium-volume stones (1-2.5cm) 3Stone removal rate = First basic removal rate - Estimated percentage of residual stones × 1.0 - 3%, where the second basic removal rate can be set at 88%; large stones (>2.5cm) 3 Stone removal rate = Third basic removal rate - Estimated percentage of residual stones × 1.0 - 3%, where the third basic removal rate can be set to 80%. Electronic devices can classify risk levels based on the stone removal rate results (low risk ≥85%, second risk 65%-84%, first risk <65%).

[0282] Specifically, the complication probability prediction sub-model can extract core features from the final feature vector, such as the duration of sustained blood pressure above a preset threshold, vascular density, and peak bleeding rate. Then, based on a risk function, it calculates the infection complication risk index: Infection complication risk index = Duration of sustained blood pressure above a preset threshold × Urine white blood cell count. If the duration of sustained blood pressure above the preset threshold is >10 minutes, the risk index doubles, and is mapped to an occurrence probability of 0-100% using a Sigmoid function. Then, based on vascular density, the probability of bleeding complications is calculated using the following formula: .

[0283] Specifically, the renal function recovery trend prediction sub-model extracts features from the final feature vector, including preoperative baseline renal function, duration of sustained blood pressure above a preset threshold, renal parenchymal thickness, and total perfusion. Damage and repair coefficient calculation: The renal parenchymal injury coefficient is calculated as follows: duration of sustained blood pressure above a preset threshold × 0.01 + total perfusion volume × 0.0001 + hemorrhage impact coefficient × 0.2; the renal repair coefficient is calculated as follows: renal parenchymal thickness × 0.05. Then, based on the renal parenchymal injury coefficient and the renal repair coefficient, the baseline renal function at 1 week post-operation is calculated as follows: baseline renal function at 1 week post-operation = baseline renal function at 1 week post-operation × (1 - renal parenchymal injury coefficient) + renal repair coefficient; the baseline renal function at 1 month post-operation is calculated as follows: baseline renal function at 1 week post-operation = baseline renal function at 1 month post-operation × 1.3. The electronic device can then determine the trend of renal function recovery based on the baseline renal function at 1 week post-operation and the baseline renal function at 1 month post-operation.

[0284] The intelligent assisted method for percutaneous nephrolithotomy provided in this application determines the grayscale distribution of kidney images based on the maximum inter-class variance method, outputs a three-dimensional mask of the stone, and identifies the target stone region. It automatically calculates the grayscale threshold for stone identification, avoiding the subjectivity of manual threshold selection and accurately segmenting candidate stone regions. By combining grayscale co-occurrence matrix parameters to extract stone texture features, it generates a high-precision three-dimensional mask of the stone, achieving precise localization and boundary delineation of the stone region, providing a basis for predicting the difficulty of subsequent lithotripsy. It identifies the three-dimensional masks of various kidney structures and annotates the target renal parenchyma region. It accurately segments anatomical structures such as the renal cortex, renal medulla, renal pelvis, and renal calyces, clarifying the extent and boundaries of the renal parenchyma. This provides anatomical reference for puncture path planning, avoiding puncture damage to key structures such as the renal pelvis and renal cortex, and improving the anatomical adaptability of the puncture. Then, based on Hessian matrix vascular enhancement and combined with blood flow signals, it determines the target vascular region. It enhances the contrast of vascular images, clearly presenting the initial vascular anatomical morphology. By combining blood flow signals to determine vascular activity, distinguishing between high-risk and low-risk vessels, and accurately delineating the target vessel area, a core basis for puncture avoidance is provided, reducing the risk of intraoperative bleeding.

[0285] Next, the initial puncture path is determined, and the functional area of ​​the target puncture path is selected. Multiple initial paths are generated based on preset principles, balancing the dual requirements of "direct access to the stone" and "obstacle avoidance." By analyzing basic functional characteristics (such as path length and angle) and obstacle functional characteristics (such as distance from blood vessels), the optimal functional area of ​​the target puncture path is selected to improve puncture efficiency and safety. Then, features of the target stone / kidney / blood vessel / puncture path are extracted to generate an initial feature set. Features such as stone morphology, texture, composition, and lithotripsy difficulty are comprehensively extracted; kidney anatomy and tolerance characteristics are extracted; blood vessel distribution and risk score characteristics are extracted; and puncture path efficiency and accessibility characteristics are extracted. This forms a standardized, multi-dimensional initial feature set, providing high-quality input data for subsequent intelligent prediction of puncture parameters. Feature classification is performed, generating three feature groups, outputting safety risk / operational difficulty / anatomical adaptation features. Thus, features are categorized into safety, operation, and anatomy categories according to clinical significance, achieving refined feature management. Different branch models are used to quantify safety risks (bleeding, high pressure), operational difficulty (lithotomy, pathway), and anatomical fit (adaptability to renal calyces / cortex), providing a quantitative basis for puncture parameter prediction. Finally, puncture point / angle / depth / obstacle avoidance range / perfusion pressure thresholds are predicted based on three types of features. Multi-dimensional feature fusion prediction replaces traditional empirical puncture parameter formulation, improving parameter accuracy and personalization. Targeted avoidance of high-risk vessels, optimization of operational angle and depth, and control of perfusion pressure comprehensively reduce the probability of intraoperative complications and shorten operation time.

[0286] Finally, based on the fusion of intraoperative dynamic and static features, operation-anatomical fit / pressure-volume correlation / dynamic-static fusion features are generated. Preoperative static imaging features are overlaid with real-time intraoperative data (puncture trajectory, renal pelvis pressure) to verify deviations between the puncture path and the preoperative plan in real time, enabling dynamic adjustments during the operation. Pressure-volume correlation features are constructed to accurately quantify the risk of extravasation of the perfusion fluid, providing a basis for intraoperative pressure control. Initial postoperative prediction features are generated to determine the postoperative prediction results. Integrating features across the entire preoperative-intraoperative timeline comprehensively covers key factors affecting postoperative outcomes. Accurate predictions of stone clearance rate, complication probability, and renal function recovery trends are made, assisting in the development of personalized postoperative care plans and improving doctor-patient communication efficiency and transparency in diagnosis and treatment.

[0287] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A percutaneous nephrolithotomy intelligent assistance method, characterized in that, The method includes: Obtain the kidney image corresponding to the target kidney; The kidney image is identified to determine the target puncture parameters corresponding to the target kidney; Acquire real-time surgical data of the target kidney during the puncture process; Based on the real-time surgical data and the kidney imaging images, the postoperative prediction result corresponding to the target kidney is determined.

2. The method of claim 1, wherein, The step of identifying the kidney image and determining the target puncture parameters corresponding to the target kidney includes: The kidney image is identified to determine the region of interest corresponding to the target kidney. The region of interest is identified, and the initial interest features corresponding to the region of interest are extracted; Based on the initial features of interest, the target puncture parameters corresponding to the target kidney are determined.

3. The method of claim 2, wherein, The region of interest includes the target stone region, the target renal parenchyma region, the target vascular region, and the functional region of the target puncture path; the step of identifying the renal image and determining the region of interest corresponding to the target kidney includes: The grayscale distribution of the kidney image was determined based on the Otsu's method. Based on the grayscale distribution, determine the grayscale threshold for stone identification; Based on the grayscale threshold for stone identification, candidate stone regions are determined; Calculate the gray-level co-occurrence matrix parameters corresponding to the candidate stone region; Based on the gray-level co-occurrence matrix parameters, output a three-dimensional mask of the stone; Based on the three-dimensional mask of the stone, the target stone region is determined; The kidney images are identified to determine the three-dimensional masks of various kidney structures, including the renal cortex, renal medulla, renal pelvis, and renal calyces. Each structure is labeled to generate the target renal parenchyma region; The Hessian matrix vascular enhancement algorithm was used to identify the kidney images to obtain the initial vascular anatomy. Obtain the blood flow signals corresponding to each blood vessel in the kidney imaging image; Based on the blood flow signals, vascular activity is determined; The target vascular region is determined based on the initial vascular anatomy and vascular activity. Based on preset principles, a preset number of initial puncture paths are determined; Each initial puncture path is identified, and the basic puncture path functional characteristics and obstacle puncture path functional characteristics corresponding to each initial puncture path are determined. Based on the functional characteristics of the basic puncture path and the functional characteristics of the obstacle puncture path, the functional region of the target puncture path is determined in each of the initial puncture paths.

4. The method of claim 2, wherein, The initial features of interest include initial stone features; the region of interest includes the target stone region and the target renal parenchyma region. The step of identifying the region of interest and extracting the initial interest features corresponding to the region of interest includes: Identify the target stone region and determine the basic morphological characteristics of the stone corresponding to the target stone region; The mean CT value, standard deviation, coefficient of variation, kurtosis, and skewness corresponding to the CT values ​​within the target stone region were statistically analyzed. Based on the standard deviation, the coefficient of variation, the kurtosis, and the skewness, the gray-level heterogeneity characteristics corresponding to the target stone region are determined; Calculate the gradient mean and gradient variance corresponding to the CT values ​​within the target stone region; Based on the gradient mean and the gradient variance, the stone texture features and composition prediction features corresponding to the target stone region are determined. Based on the CT value difference between the target stone region and the target renal parenchyma region, the characteristics for predicting the difficulty of lithotripsy are determined. The initial stone features are generated based on the basic morphological features of the stone, the grayscale heterogeneity features, the stone texture features, the composition prediction features, and the stone fragmentation difficulty prediction features.

5. The method of claim 2, wherein, The initial features of interest include initial kidney features; the region of interest includes the target kidney parenchyma region; The step of identifying the region of interest and extracting the initial interest features corresponding to the region of interest includes: The target renal parenchyma region is identified to determine the basic anatomical features of the kidney; the basic anatomical features of the kidney include parameters of the renal calyx fornix and parameters of the renal parenchyma; the parameters of the renal calyx fornix include the radius of curvature of the renal calyx fornix, the opening angle of the calyx neck, and the direction vector of the axis of the target renal calyx; The target renal parenchymal region is identified, and the average CT value and thickness of the fornix are determined. Based on the average CT value of the dome and the thickness of the dome, the elastic index characteristics of the dome are calculated; Based on the CT values ​​within the target renal parenchyma region and the renal parenchyma parameters, the renal parenchyma tolerance characteristics are determined; Based on the calyx opening angle, the calyx duct accessibility score corresponding to the target renal parenchymal region is determined; The initial kidney features are generated based on the basic anatomical features of the kidney, the elasticity index features of the fornix, the tolerance features of the renal parenchyma, and the calyx duct patency score.

6. The method of claim 2, wherein, The initial features of interest include initial vascular features; the region of interest includes a target vascular region; the process of identifying the region of interest and extracting the initial features of interest corresponding to the region of interest includes: The target vascular region is identified to determine the basic vascular distribution characteristics; the basic vascular distribution characteristics include vascular diameter, vascular orientation vector, vascular density in the puncture area, and characteristics of the hypovascular area. Based on the diameter of the blood vessel and the distance between each blood vessel and the fornix, calculate the single blood vessel risk score corresponding to each blood vessel. Based on the individual vessel risk scores, the regional vessel risk score corresponding to the target vessel region is calculated. Calculate the obstacle avoidance distance and obstacle avoidance angle between the preset puncture path and each blood vessel in the target puncture path functional area, and generate obstacle avoidance association features; The initial vascular features are generated based on the basic vascular distribution characteristics, the single vascular risk score, the regional vascular risk score, and the obstacle avoidance association features.

7. The method of claim 2, wherein, The initial features of interest include initial puncture path features; the region of interest includes the target puncture path functional region; the process of identifying the region of interest and extracting the initial features of interest corresponding to the region of interest includes: The preset puncture path in the target puncture path functional area is identified to determine the basic puncture path characteristics; the basic puncture path characteristics include: path length and the angle between the preset puncture path and the long axis of the kidney; Identify the blood vessels, renal parenchymal scar tissue, and renal parenchymal elastic regions in the functional area of ​​the target puncture path to determine the quantitative characteristics of the obstacles; The preset puncture path in the target puncture path functional area is identified, and the path efficiency index and operation accessibility score are determined. The initial puncture path features are generated based on the basic puncture path features, the obstacle quantification features, the path efficiency index, and the operation accessibility score.

8. The method according to claim 2, characterized in that, The step of determining the target puncture parameters corresponding to the target kidney based on the initial feature of interest includes: The initial features of interest are input into a preset LASSO regression model, which outputs candidate core features. The preset LASSO regression model uses the target puncture point, target puncture angle, target puncture depth, obstacle avoidance range boundary, and target perfusion pressure threshold as joint target variables. The candidate core features are input into a preset expert evaluation model to evaluate the candidate core features and output the target core features. The core features of the target are input into a preset puncture parameter determination model; The preset puncture parameter determination model identifies each sub-feature in the core features of the target and determines the category corresponding to each sub-feature; the category includes safety category, operation category, and dissection category; Based on the category corresponding to each of the sub-features, generate a safety category feature group, an operation category feature group, and an anatomy category feature group; The security category feature group is input into the first decision tree branch, and the security risk features are output. The operation category feature group is input into the second feature extraction branch, and the operation difficulty feature is output. The anatomical category feature group is input into the third feature extraction branch, and the anatomical adaptation feature is output; the anatomical adaptation score is used to characterize the puncture path and the kidney anatomy. Based on the safety risk characteristics, the anatomical adaptation characteristics, and the stone location characteristics in the basic morphological characteristics of the stone, the target puncture point is predicted. Based on the anatomical adaptation features and the operational difficulty features, the target puncture angle is predicted; Based on the aforementioned safety risk characteristics and operational difficulty characteristics, the target puncture depth is predicted; Based on the safety risk characteristics and the target blood vessel region in the region of interest, predict the boundary of the obstacle avoidance range; Based on the safety risk characteristics and the operational difficulty characteristics, the target perfusion pressure threshold is predicted.

9. The method according to claim 1, characterized in that, The step of determining the postoperative prediction result corresponding to the target kidney based on the real-time surgical data and the kidney imaging image includes: Acquire the initial stone features, initial kidney features, and initial vascular features corresponding to the kidney imaging image; The basic vascular distribution features in the initial vascular features are spatially superimposed with the puncture trajectory data in the real-time surgical data to generate operation-anatomy fit features. The kidney images are identified to determine the morphological characteristics of the renal pelvis; Based on the renal pelvis morphological features and the real-time renal pelvis pressure data in the real-time surgical data, a pressure-volume correlation feature is constructed. The initial stone features, initial kidney features, and initial vascular features are fused with the dynamic features in the real-time surgical data to generate dynamic-static fusion features. Based on the operation-anatomical fit characteristics, the pressure-volume correlation characteristics, and the dynamic-static fusion characteristics, initial postoperative prediction characteristics are generated. Based on the initial postoperative prediction features, the postoperative prediction result corresponding to the target kidney is determined.

10. The method according to claim 9, characterized in that, The determination of the postoperative prediction result corresponding to the target kidney based on the initial postoperative prediction features includes: The initial postoperative predicted features were screened by L1 regularization penalty of LASSO regression to obtain candidate feature factors. The candidate feature factors are input into a preset random forest model, sorted by feature importance score, and target feature factors with a feature importance score greater than a preset importance score threshold are selected; the target feature factors include: stone feature factors, surgical operation feature factors, anatomical risk feature factors, and intraoperative status feature factors; The target feature factors are input into a preset postoperative prediction result determination model; The preset postoperative prediction result determination model identifies the target feature factors and determines the target static features and target dynamic features; The target static features are input into the feature interaction layer, and higher-order interaction features are output. The higher-order interaction features include the interaction features of duration of blood pressure above a preset threshold and degree of hydronephrosis, interaction features of vascular density and number of puncture channels, interaction features of total perfusion volume and renal parenchymal thickness, and interaction features of lithotripsy duration and stone CT value. The target dynamic features are input into the temporal feature extraction layer, and the temporal fusion features are output. The target static features, the higher-order interaction features, and the temporal fusion features are integrated into a final feature vector; The final feature vector is input into the prediction layer, and the postoperative prediction result is output; the postoperative prediction result includes at least one of the following: stone clearance rate, probability of complication, and trend of renal function recovery.

Images