Method and System for Constructing an Automatic Renal Function Estimation Model Based on Non-Contrast CT

Abstract

The present invention discloses a method and system for constructing an automatic renal function estimation model based on non-contrast CT, including the following steps: Step S1, collecting image data of patients with renal atrophy or hydronephrosis; Step S2, using the renal parenchyma and hydronephrosis regions manually delineated by radiologists to train an automatic segmentation model for renal parenchyma and hydronephrosis regions based on UNETR; Step S3, extracting radiomics features from the images and screening the data; Step S4, constructing renal function index regression models, including a GFR regression model and an SRF regression model; Step S5, verifying the estimated values of the regression models; Step S6, using the renal function index regression models to calculate GFR and SRF and distinguish kidneys in different health states. By using this method, the glomerular filtration rate and split renal function of each kidney can be automatically estimated. For the treatment of patients with renal atrophy or hydronephrosis, a rapid, convenient and low-cost renal function assessment method is provided.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to a method and system for constructing an automatic renal function estimation model based on non-contrast CT. Background Art

[0002] GFR (Glomerular Filtration Rate) is a key indicator for measuring kidney function, while SRF (Split Renal Function) reflects the contribution ratio of a single kidney to the overall renal function. In urological surgical decision-making, such as the treatment of patients with renal atrophy or hydronephrosis, accurate assessment of these parameters is crucial.

[0003] Limitations of the prior art:

[0004] 1. Single Photon Emission Computed Tomography (SPECT): As the gold standard method for measuring unilateral kidney GFR and SRF, SPECT has obvious limitations. First of all, SPECT involves the use of radioactive isotopes, posing a potential radiation risk to patients. Secondly, the operation process of SPECT is complex and requires professional technical personnel for operation and interpretation, which limits its wide application in clinical practice. Moreover, SPECT equipment is expensive and the examination cost is high, which is not conducive to medical environments with limited resources. Finally, the imaging time of SPECT is long, which is not conducive to emergency or situations requiring rapid diagnosis.

[0005] 2. Contrast-Enhanced Computed Tomography (CECT): As another imaging method for evaluating renal function, CECT enhances the images of the kidneys by using iodine contrast agents to estimate GFR and SRF. However, iodine contrast agents may pose an additional risk of nephrotoxicity to patients with renal insufficiency and are not applicable to patients with a history of contrast agent allergy. In addition, CECT also requires specific equipment and professional operators, limiting its clinical application.

[0006] 3. Non-contrast Computed Tomography (non-contrast CT, NCCT, i.e., non-contrast CT): Non-contrast CT (i.e., CT scan without using iodine contrast agents) has certain potential in evaluating renal function because it avoids contrast agent-related side effects and costs. However, the application of traditional non-contrast CT in evaluating renal function is limited by its low accuracy and lack of automated image processing methods: manually segmenting the kidney region is time-consuming and subjective, calculating GFR only based on the renal parenchyma region, the parameters are too simple, and the accuracy is not high, limiting its application in clinical practice. Summary of the Invention

[0007] The object of the present invention is to provide a method and system for constructing an automatic renal function estimation model based on non-contrast CT to solve the problems raised in the above background art.

[0008] To achieve the above object of the invention, one aspect of the present invention provides a method for constructing an automatic renal function estimation model based on non-contrast CT, comprising the following steps:

[0009] Step S1, collecting patient image data with renal atrophy or hydronephrosis and preprocessing the images;

[0010] Step S2, using the renal parenchyma and hydronephrosis regions manually delineated by radiologists to train an automatic segmentation model for the renal parenchyma and hydronephrosis regions based on UNETR (UNet Transformers);

[0011] Step S3, extracting radiomics features from the images of the renal parenchyma and hydronephrosis regions obtained from the automatic segmentation model in step S2 and the original images, and screening the data;

[0012] Step S4, constructing renal function index regression models, including a GFR regression model and an SRF regression model.

[0013] Step S5, verifying the consistency and differences between the estimated values of the regression model and the measured values based on SPECT;

[0014] Step S6, using the renal function index regression model to calculate GFR and SRF and distinguish kidneys in different health states.

[0015] Further, step S1 includes the following steps:

[0016] Step S101, collecting patient data with renal atrophy or hydronephrosis, where the patients have undergone NCCT and SPECT imaging, and obtaining the gold standard G of the patients' GFR and SRF calculated by the SPECT workstation as the dependent variable for regression modeling;

[0017] Step S102, preprocessing the NCCT images, including cropping the range of voxel gray value I to the renal window and normalizing it using the z-score method, and the formula is as follows:

[0018] ,

[0019] where, is to calculate the gray mean value, is to calculate the gray standard deviation, is the normalized voxel gray value.

[0020] Further, step S2 includes the following steps:

[0021] Step S201: Use the renal parenchyma and hydronephrosis regions manually delineated by radiologists as training samples for the deep learning-based automatic kidney segmentation model.

[0022] Step S202: Train an automatic segmentation model for the renal parenchyma and hydronephrosis regions based on UNETR.

[0023] Further, step S3 includes the following steps:

[0024] Step S301: Extract radiomics features from the NCCT images and original images of the renal parenchyma and hydronephrosis regions obtained from the automatic segmentation model in step S2, including shape features, first-order statistical features, texture features, and filtering features.

[0025] Step S302: Screen the data with reference to ICC (Intra class consistency coefficient) and SCC (Spearman correlation coefficients).

[0026] Further, screening the data with reference to ICC is to exclude features with an ICC lower than 0.9 between manual and automatic segmentation to improve the repeatability of radiomics.

[0027] Further, screening the data with reference to SCC is to calculate the SCC between one feature and all other features in each pair of highly correlated features, and features with SCC > 0.9 are excluded.

[0028] Further, the method for constructing the GFR regression model includes the following steps:

[0029] Step S411: Use the ElasticNet model to select the most representative features from the screened features and determine the feature weights to estimate rGFR (radiomics based GFR estimation), and the remaining features are input into ElasticNet to obtain rGFR (mL / min×1.73m 2 ) The calculation formula is:

[0030]

[0031] Among them, p_original_glszm_LargeAreaHighGrayLevelEmphasis is the large area high gray level emphasis value based on the gray level size area matrix in the renal parenchyma region, p_original_glszm_SizeZoneNonUniformity is the regional size non-uniformity based on the gray level size area matrix in the renal parenchyma region, p_gradient_ngtdm_Busyness is the busyness based on the gradient neighborhood gray level difference matrix in the renal parenchyma region, p_square_firstorder_TotalEnergy is the total energy in the renal parenchyma region, and h_square_ngtdm_Busyness is the busyness based on the gradient neighborhood gray level difference matrix in the hydronephrosis region;

[0032] Step S412: Combine rGFR with clinical features and use multivariable linear regression to obtain rcGFR (radiomics-clinical based GFR estimation, the GFR estimation value combining radiomics features and clinical features). The formula for rcGFR (mL / min×1.73m 2 ) is as follows:

[0033] rcGFR = age × (-0.147) + rGFR × (1.224) - 1.303,

[0034] where rcGFR is the GFR estimation result.

[0035] Furthermore, the method for constructing the SRF regression model includes the following steps:

[0036] Step S421: Calculate the relative contribution ratio rcSRF (radiomics-clinical based SRF estimation, the SRF estimation value combining radiomics features and clinical features) of the rcGFR of a single kidney to the overall rcGFR, which is expressed by the following formula:

[0037] rcSRFl = rcGFRl / (rcGFRl + rcGFRr),

[0038] rcSRFr = 1 - rcSRFl,

[0039] where rcGFRl and rcGFRr are the rcGFR of the left and right kidneys respectively, and rcSRFl and rcSRFr are the rcSRF of the left and right kidneys respectively;

[0040] Step S422: Calculate the ratio RPV_P of the volume of a single renal parenchyma RPV to the total volume of the renal parenchyma. The calculation formula is:

[0041] RPV_Pl = RPVl / (RPVl + RPVr),

[0042] RPV_Pr = 1 - RPV_Pl,

[0043] Wherein, RPVl and RPVr are the RPVs of the left and right kidneys respectively, and RPV_Pl and RPV_Pr are the RPV_Ps of the left and right kidneys respectively;

[0044] Using linear regression to obtain the pSRF, the estimated value of SRF based on the renal parenchyma. The calculation formula is:

[0045] pSRF = 0.991×RPV_P + 0.5%,

[0046] Where RPV_P is RPV_Pl or RPV_Pr, corresponding to the calculated pSRF of the left or right kidney;

[0047] Step S423, calculate the proportion RHV_P of the unilateral renal hydronephrosis volume RHV to the total renal hydronephrosis volume. The calculation formula is:

[0048] RHV_Pl = RHVl / (RHVl + RHVr),

[0049] RHV_Pr = 1 - RHV_Pl,

[0050] Wherein, RHVl and RHVr are the RHV of the left and right kidneys respectively, and RHV_Pl and RHV_Pr are the RPV_P of the left and right kidneys respectively;

[0051] Using linear regression to obtain the hSRF, the estimated value of SRF based on renal hydronephrosis. The calculation formula is:

[0052] hSRF = (-0.223)×RHV_P + 61.1%,

[0053] Where RHV_P is RHV_Pl or RHV_Pr, corresponding to the calculated hSRF of the left or right kidney;

[0054] Step S424, using linear regression to obtain the final SRF estimated value rcphSRF of the left or right kidney by combining rcSRF, pSRF and hSRF. The calculation formula is:

[0055] rcphSRF = rcSRF×0.416 + pSRF×0.564 + hSRF×0.535 - 25.8%.

[0056] Furthermore, in step S6, the kidneys in different health states are distinguished according to the following rules:

[0057] The damaged kidney is a single kidney with GFR ≤ 30 mL / min × 1.73 m 2 and the normal kidney is a single kidney with GFR > 30 mL / min × 1.73 m 2 ;

[0058] The non-functional kidney is a single kidney with GFR < 10 mL / min × 1.73 m 2 and the functional kidney is a single kidney with GFR ≥ 10 mL / min × 1.73 m 2 ;

[0059] Suitable for nephrectomy is SRF < 15%, and not suitable for nephrectomy is SRF ≥ 15%;

[0060] Renal function decline is SRF < 45%, and no decline is SRF ≥ 45%;

[0061] The non-dominant kidney is SRF ≤ 60%, and the dominant kidney is SRF > 60%.

[0062] Another aspect of the present invention provides a system for constructing an automatic renal function estimation model based on non-contrast enhanced CT, including a collection module, a segmentation module, an extraction module, a modeling module, a verification module, and an application module, where:

[0063] The collection module is used to collect image data of patients with renal atrophy or hydronephrosis and preprocess the images;

[0064] The segmentation module uses the manually delineated renal parenchyma and hydronephrosis regions by radiologists to train an automatic segmentation model for renal parenchyma and hydronephrosis regions based on UNETR;

[0065] The extraction module extracts radiomics features from the images of renal parenchyma and hydronephrosis regions obtained from the automatic segmentation model of the segmentation module and the original images, and screens the data;

[0066] The modeling module is used to construct renal function index regression models, including a GFR regression model and an SRF regression model;

[0067] The verification module is used to verify the consistency and differences between the estimated values of the regression models and the measured values based on SPECT;

[0068] The application module uses the renal function index regression models to calculate GFR and SRF and distinguish kidneys in different health states.

[0069] Due to the adoption of this system and method, compared with the prior art, it has the following advantages:

[0070] 1. Establishing a model using non-contrast enhanced CT data avoids the radiation risk and nephrotoxicity risk brought by other data acquisition processes.

[0071] 2. An automatic segmentation model for kidney parenchyma and hydronephrosis regions based on UNETR is adopted. While using a transformer as the encoder, UNETR maintains the U-shaped structure of U-Net, has the advantage of extracting image features at different scales, and is suitable for segmenting kidney parenchyma and hydronephrosis regions of different sizes.

[0072] 3. Separate regression modeling for GFR and SRF can automatically estimate the glomerular filtration rate and split renal function of each kidney. For patients with renal atrophy or hydronephrosis, a rapid, convenient, and low-cost renal function assessment method is provided. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flowchart of a method for constructing an automatic renal function estimation model based on non-contrast-enhanced CT.

[0074] Figure 2 It is a correlation diagram between estimated values and SPECT measurement values. Among them: the dotted line indicates perfect agreement. (A) is rGFR, (B) is rcGFR, (C) is rcSRF, (D) is pSRF, (E) is hSRF, and (F) is rcphSRF.

[0075] Figure 3 It is a Bland Altman diagram of the estimated values of the regression model and the SPECT-based measurement values. Among them, the 95% limits of agreement (LoA) are between the dashed horizontal lines. The dotted line represents the mean difference. (A) is rGFR, (B) is rcGFR, (C) is rcSRF, (D) is pSRF, (E) is hSRF, and (F) is rcphSRF. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0076] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0077] As Figure 1 shown in the flowchart of the method of the present invention, the embodiments of the present invention provide an artificial intelligence (AI) technology based on non-contrast-enhanced CT (NCCT) for automatically estimating the glomerular filtration rate (GFR) and split renal function (SRF) of each kidney. The specific steps are as follows:

[0078] Step S1, data collection and preprocessing, specifically including the following steps:

[0079] Step S101, collect data of patients with renal atrophy or hydronephrosis who have undergone NCCT and SPECT imaging, and GFR and SRF are calculated by the SPECT workstation.

[0080] Step S102, preprocess the NCCT images, including cropping the range of voxel gray value I to the renal window (window width = 260 HU, window level = 60 HU) and normalizing it using the z-score method:

[0081] ,

[0082] where, is to calculate the mean gray value, is to calculate the standard deviation of gray value.

[0083] Step S2, automatic kidney segmentation, specifically including the following steps:

[0084] Step S201, use a radiologist to manually delineate the renal parenchyma and hydronephrosis regions as training samples for the deep learning-based automatic kidney segmentation model.

[0085] Step S202, train an automatic segmentation model for the renal parenchyma and hydronephrosis regions based on UNETR. While using a transformer as the encoder, UNETR maintains the U-shaped structure of U-Net, has the advantage of extracting image features at different scales, and is suitable for segmenting renal parenchyma and hydronephrosis regions of different sizes. UNETR is trained using Python 3.9 and PyTorch 1.9 in a workstation with NVIDIA GeForce RTX2080ti. The parameters are as follows: max epoch = 1000; the loss function is diceloss; the optimizer is adamw; the learning rate = 0.0001; the batch size = 4. Data augmentation is performed during model training, including random cropping, flipping, and rotation.

[0086] Step S3, radiomics feature extraction, specifically including the following steps:

[0087] Step S301, extract 2260 radiomics features from the NCCT images of the renal parenchyma and hydronephrosis regions obtained from the automatic segmentation model in Step S2, including 28 shape features, 36 first-order statistical features, 150 texture features, and 2046 filtering features. Shape features, first-order statistical features, and texture features are extracted from the original image, and filtering features are extracted from the wavelet filtering, square filtering, logarithmic filtering, and gradient filtering images.

[0088] Step S302. To improve the reproducibility of radiomics, features with an ICC lower than 0.9 between manual and automatic segmentation were excluded. In addition, the pairwise feature SCC was calculated, and feature pairs with an SCC value > 0.9 were considered highly correlated. For each highly correlated feature pair, the SCC between one feature and all other features was calculated, and the feature with a larger mean SCC was considered redundant and excluded.

[0089] Step S4. Regression modeling of renal function indicators, specifically including the following steps:

[0090] Step S401. GFR regression modeling, specifically including the following steps:

[0091] Step S411. Use the ElasticNet model to select the most representative features from the screened features and determine the feature weights to estimate the unilateral kidney GFR, i.e., rGFR. The remaining features were input into ElasticNet, and the GFR calculated by the SPECT workstation for the patient was used as the dependent variable to generate the GFR (rGFR) based on radiomics features. ElasticNet can select the most representative features and determine the weights of the features to estimate the single kidney GFR. The hyperparameters of ElasticNet were determined by ten-fold cross-validation. Finally, the calculation formula for rGFR (mL / min×1.73m 2 ) is as follows:

[0092]

[0093] where: p_original_glszm_LargeAreaHighGrayLevelEmphasis is the large area high gray level emphasis value based on the gray level size zone matrix in the renal parenchyma region, p_original_glszm_SizeZoneNonUniformity is the regional size non-uniformity based on the gray level size zone matrix in the renal parenchyma region, p_gradient_ngtdm_Busyness is the busyness based on the gradient neighborhood gray level difference matrix in the renal parenchyma region, p_square_firstorder_TotalEnergy is the total energy in the renal parenchyma region, and h_square_ngtdm_Busyness is the busyness based on the gradient neighborhood gray level difference matrix in the hydronephrosis region.

[0094] Step S412: Combine rGFR with clinical features and use multivariable linear regression (MLR) to obtain the GFR that combines radiomics and clinical features, namely rcGFR (radiomics - clinical based GFR estimation, the GFR estimation value that combines radiomics features and clinical features). rcGFR (mL / min×1.73m 2 ) is calculated by the formula:

[0095] rcGFR = age × (-0.147) + rGFR × (1.224) - 1.303,

[0096] where rcGFR is the final GFR estimation result.

[0097] Step S402: SRF regression modeling, which specifically includes the following steps:

[0098] Step S421: Calculate the relative contribution ratio rcSRF (radiomics - clinical based SRF estimation, the SRF estimation value that combines radiomics features and clinical features) of the rcGFR of a single kidney to the overall rcGFR:

[0099] rcSRFl = rcGFRl / (rcGFRl + rcGFRr),

[0100] rcSRFr = 1 – rcSRFl,

[0101] where rcGFRl and rcGFRr are the rcGFR of the left and right kidneys respectively, calculated according to the method in Step S412; rcSRFl and rcSRFr are the rcSRF of the left and right kidneys respectively.

[0102] Step S422: Calculate the ratio RPV_P of the renal parenchymal volume RPV of a single kidney to the total renal parenchymal volume:

[0103] RPV_Pl = RPVl / (RPVl + RPVr),

[0104] RPV_Pr = 1 - RPV_Pl,

[0105] where RPVl and RPVr are the RPV of the left and right kidneys respectively, and RPV_Pl and RPV_Pr are the RPV_P of the left and right kidneys respectively.

[0106] Use linear regression to obtain the renal parenchyma - based SRF estimate pSRF:

[0107] pSRF = 0.991×RPV_P + 0.5%,

[0108] Among them, RPV_P is RPV_Pl or RPV_Pr, and the pSRF of the left kidney or the right kidney is calculated correspondingly.

[0109] Step S423, calculate the proportion RHV_P of the unilateral hydronephrosis volume RHV to the total hydronephrosis volume:

[0110] RHV_Pl = RHVl / (RHVl + RHVr),

[0111] RHV_Pr = 1 - RHV_Pl,

[0112] Among them, RHVl and RHVr are the RHV of the left and right kidneys respectively, and RHV_Pl and RHV_Pr are the RPV_P of the left and right kidneys respectively.

[0113] Use linear regression to obtain the SRF estimate hSRF based on hydronephrosis:

[0114] hSRF = (-0.223)×RHV_P + 61.1%,

[0115] Among them, RHV_P is RHV_Pl or RHV_Pr, and the hSRF of the left kidney or the right kidney is calculated correspondingly.

[0116] Step S424, use linear regression, take the SRF calculated by the SPECT workstation for the patient as the dependent variable, and obtain the final SRF estimate rcphSRF of the left kidney or the right kidney combining rcSRF, pSRF and hSRF:

[0117] rcphSRF = rcSRF×0.416 + pSRF×0.564 + hSRF×0.535 - 25.8%.

[0118] Step S5, verify the consistency and differences between the regression model estimated value and the value measured based on SPECT. Use the slope and intercept of the simplified long axis, the Lin's concordance coefficient (CCC), and the Bland-Altman plot with the limits of agreement (LoA).

[0119] As shown in Table 1, the characteristic distribution of the patient dataset used for model training and testing, where there are 128 cases in the training set and 117 cases in the test set. Except for RPV and single kidney GFR, there are no significant differences in other characteristics.

[0120] Table 1. Clinical characteristics of patients.

[0121] Clinical characteristics Training set Test set p Gender 0.853 Male 74 (57.8) 69 (59.0) Female 54 (42.2) 48 (41.0) Age (years) 52.76±13.70 55.79±16.39 0.070 BMI 23.90±3.29 23.86±3.23 RPV (ml) 152.83±51.58 132.70±60.43 <0.01 RHV (ml) 34.51±67.98 47.51±91.18 0.408 <![CDATA[Single kidney GFR (mL / min×1.73m 2 )]]> 40.36±14.43 36.27±19.41 0.002

[0122] Among them, for gender, the data are the number of patients, and the data in parentheses are percentages. Age, BMI, RPV, RHV, and single-kidney GFR are mean ± standard deviation. BMI = body mass index, RPV = renal parenchymal volume, RHV = hydronephrotic volume, GFR = glomerular filtration rate. p-values are from the Mann-Whitney U test (continuous parameters) and the χ2 test (categorical variables).

[0123] The correlation between the GFR estimated by the model and the GFR measured by SPECT is shown in Table 2 and Figure 2 as follows. Pearson correlation analysis showed that rGFR had a good correlation with the GFR measured by SPECT (r = 0.72, P < 0.001), and the paired-samples t-test showed no statistically significant difference between rGFR and the GFR measured by SPECT (P > 0.05). Meanwhile, the CCC of rGFR was relatively poor, at 0.60. Multiple linear regression showed that age and rGFR were independent predictors, and rcGFR was calculated as a linear combination of age and rGFR (Table 2). Compared with rGFR, rcGFR had a stronger correlation (r = 0.75 vs. 0.72), a smaller difference (0.34 vs. 1.39), and a higher CCC value (0.70 vs. 0.60). In addition, the slope of rcGFR decreased (1.46 vs. 1.85) and the intercept increased (-17.53 vs. -33.68), making the two closer to perfect agreement (slope = 1 and intercept = 0). These results indicate that rcGFR has better predictive ability. In addition, as Figure 3 confirmed by the Bland-Altman plot in, most of the differences between the estimated GFR and the GFR measured by SPECT were within the LoA, indicating good agreement.

[0124] The correlation between the SRF estimated by the model and the SRF measured by SPECT is shown in Table 2 and Figure 2As shown. rcSRF was calculated based on rcGFR and was significantly correlated with the SRF measured by SPECT (r = 0.83, p < 0.001), with a CCC of 0.78. Compared with rcSRF, pSRF showed similar correlations (r = 0.84 vs. 0.83) and CCCs (0.79 vs. 0.78), slopes (1.43 vs. 1.52), and intercepts (-0.26 vs. -0.22), but a larger mean difference (1.20 vs. 0.94). hSRF had the weakest correlation (r = 0.73). rcSRF, pSRF, and hSRF were determined as independent predictors of SRF by multiple linear regression and were subsequently combined to form rcphSRF. In the test set, rcphSRF had the strongest correlation (r = 0.92, p < 0.001), the smallest mean difference of 0.51, and the highest CCC of 0.88. Additionally, the slope (1.37) and intercept (-0.18) of rcphSRF were closest to perfect agreement. These results confirmed that rcphSRF had the best predictive ability. Figure 3 The Bland - Altman plot in Figure 3 further confirmed that more than 90% of the differences between the estimated SRF and the SRF measured by SPECT were within the LoA, indicating good agreement.

[0125] Table 2. Comparison of estimated values with SPECT - based measurements.

[0126] Estimated value r <![CDATA[p § > Difference <![CDATA[p * > CCC Slope Intercept rGFR Training set 0.67 <0.001 0.00±10.94 0.989 0.55(0.48,0.61) 1.93 -37.52 Test set 0.72 <0.001 1.39±13.93 0.085 0.60(0.54, 0.66) 1.85 -33.68 rcGFR Training set 0.68 <0.001 0.00±10.54 0.762 0.64(0.57, 0.70) 1.72 -18.70 Test set 0.75 <0.001 0.34±12.87 0.669 0.70(0.64, 0.75) 1.46 -17.53 rcSRF Training set 0.85 <0.001 1.03±8.50 0.011 0.77(0.73, 0.81) 1.51 -0.26 Test set 0.83 <0.001 0.94±14.23 0.314 0.78(0.74, 0.81) 1.52 -0.26 pSRF Training set 0.82 <0.001 2.13±7.81 <0.001 0.79(0.74, 0.83) 1.24 -0.12 Test set 0.84 <0.001 1.20±12.58 0.212 0.79(0.75, 0.83) 1.43 -0.22 hSRF Training set 0.68 <0.001 0.04±10.08 0.801 0.64(0.57, 0.70) 1.46 -2.28 Test set 0.73 <0.001 0.64±16.87 0.507 0.53(0.48, 0.59) 2.32 -0.66 rcphSRF Training set 0.90 <0.001 1.39±6.02 <0.001 0.89(0.86, 0.91) 1.12 -0.06 Test set 0.92 <0.001 0.51±9.63 0.406 0.88(0.86, 0.90) 1.37 -0.18

[0127] Among them, the differences are expressed as mean ± standard deviation, and the numerical unit of GFR is mL / min×1.73m 2 , and the numerical unit of SRF is %. r is the Pearson correlation coefficient. The slope and intercept were obtained from reduced major axis regression. CCC is the Lin's concordance coefficient, and the 95% confidence interval is in parentheses. GFR = glomerular filtration rate, SRF = split renal function, rGFR = estimated GFR based on radiomics features, rcGFR = estimated GFR combining radiomics features and clinical features. rcSRF was calculated as the relative contribution of single - kidney rcGFR to total rcGFR, pSRF was estimated based on the percentage of renal parenchymal volume (RPV), hSRF was estimated based on the percentage of hydronephrotic volume (RHV), and rcphSRF was calculated by multiple linear regression of rcSRF, pSRF, and hSRF. p § values were from Pearson correlation analysis, and p * values were obtained using paired - sample t - tests.

[0128] Step S6. To evaluate the clinical application value of the model, using the cut-off values reported in previous studies, calculate the estimated values for performance indicators to distinguish kidneys in different health states: (1) impaired kidneys (single kidney GFR ≤ 30 mL / min×1.73m 2 ) or normal kidneys (single kidney GFR > 30 mL / min×1.73m 2 ); (2) non-functional kidneys (single kidney GFR < 10 mL / min×1.73m 2 ) or functional kidneys (single kidney GFR ≥ 10 mL / min×1.73m 2 ); (3) suitable for nephrectomy (SRF < 15%) or not suitable for nephrectomy (SRF ≥ 15%); (4) renal function decline (SRF < 45%) or no renal function decline (SRF ≥ 45%); (5) non-dominant kidney (SRF ≤ 60%) or dominant kidney (SRF > 60%). Calculate the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity for distinguishing kidneys in different health states.

[0129] The calculation results of the discrimination performance of rcGFR and rcphSRF for kidneys in different health states are shown in Table 3. The discrimination of impaired or normal, non-functional or functional kidneys is evaluated based on rcGFR, and other health states are evaluated based on rcphSRF. Generally speaking, rcGFR and rcphSRF show relatively high AUC, accuracy, and specificity.

[0130] Table 3. Performance of rcGFR and rcphSRF in the discrimination of kidneys in different health states.

[0131] Health status AUC(95% CI) Accuracy(95% CI) Sensitivity(95% CI) Specificity(95% CI) Impaired or normal Training set 0.840(0.779, 0.901) 85.2 [218 / 256](80.2, 89.3) 46.7 [28 / 60](33.7, 60.0) 96.9 [190 / 196](93.5, 98.9) Test set 0.862(0.813, 0.910) 81.2 [190 / 234](75.6, 86.0) 71.4 [70 / 98](61.4, 80.1) 88.2 [120 / 136](81.6, 93.1) Non-functional or functional Training set 0.977(0.956, 0.998) 98.4 [252 / 256](96.0, 99.6) 0 [0 / 4](0, 60.2) 100 [252 / 252](98.5, 100) Test set 0.911(0.856, 0.965) 91.0 [213 / 234](86.6, 94.4) 10.5 [2 / 19](1.3, 33.1) 98.1 [211 / 215](95.3, 99.5) Suitable for nephrectomy or non-nephrectomy Training set 1.000(1.000, 1.000) 100 [256 / 256](98.6, 100) 100 [2 / 2](15.8, 100) 100 [254 / 254](98.6, 100) Test set 0.959(0.929, 0.989) 97.0 [227 / 234](93.9, 98.8) 22.2 [2 / 9](2.8, 60.0) 100 [225 / 225](98.4, 100) Whether renal function is low Training set 0.938(0.910, 0.967) 86.7 [222 / 256](81.9, 90.6) 80.0 [60 / 75](69.2, 88.4) 89.5 [162 / 181](84.1, 93.6) Test set 0.980(0.966, 0.994) 92.3 [216 / 234](88.1, 95.4) 88.2 [90 / 102](80.4, 93.8) 95.5 [126 / 132](90.4, 98.3) Non-dominant or dominant kidney Training set 0.958(0.934, 0.982) 90.6 [222 / 256](86.4, 93.9) 96.6 [198 / 205](93.1, 98.6) 66.7 [34 / 51](52.1, 79.2) Test set 0.964(0.944, 0.983) 87.2 [204 / 234](82.2, 91.2) 93.2 [136 / 146](87.8, 96.7) 77.3 [68 / 88](67.1, 85.5) Mean value Training set 0.943 92.2 64.7 90.6 Test set 0.935 89.7 57.1 91.8

[0132] Wherein: the data are percentages, the numbers in square brackets are the number of kidneys, and the numbers in parentheses are the 95% confidence intervals. Kidneys in poorer health states (i.e., impaired, non-functional, undergoing nephrectomy, with low renal function, or non-dominant kidneys) are considered positive to calculate sensitivity and specificity. The performance indicators for discriminating impaired or normal, non-functional or functional kidneys are calculated from rcGFR, and other indicators are calculated from rcphSRF.

[0133] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing an automatic renal function estimation model based on non-contrast CT, characterized in that, Including the following steps: Step S1: Collect the image data of patients with renal atrophy or hydronephrosis, and preprocess the images. Step S2: Use the renal parenchyma and hydronephrosis regions manually delineated by radiologists to train an automatic segmentation model for renal parenchyma and hydronephrosis regions based on UNETR. Step S3: Extract radiomics features from the images of renal parenchyma and hydronephrosis regions obtained from the automatic segmentation model in Step S2 and the original images, and screen the data. Step S4: Construct renal function index regression models, including the GFR regression model and the SRF regression model. Step S5: Verify the consistency and differences between the estimated values of the regression models and the measured values based on SPECT. Step S6: Use the renal function index regression models to calculate GFR and SRF, and distinguish kidneys in different health states. Step S1 includes the following steps: Step S101: Collect the data of patients with renal atrophy or hydronephrosis who have undergone NCCT and SPECT imaging. The gold standard G of GFR and SRF of the patients calculated by the SPECT workstation is used as the dependent variable for regression modeling. Step S102: Preprocess the NCCT images, including cropping the range of voxel gray value I to the renal window and normalizing it using the z-score method. The formula is as follows: where mean(I) is to calculate the gray mean, std(I) is to calculate the gray standard deviation, and I′ is the normalized voxel gray value. Step S2 includes the following steps: Step S201: Use the renal parenchyma and hydronephrosis regions manually delineated by radiologists as the training samples for the deep learning-based renal automatic segmentation model. Step S202: Train an automatic segmentation model for renal parenchyma and hydronephrosis regions based on UNETR. Step S3 includes the following steps: Step S301: Extract radiomics features from the NCCT images and original images of renal parenchyma and hydronephrosis regions obtained from the automatic segmentation model in Step S2, including shape features, first-order statistical features, texture features, and filtering features. Step S302: Screen the data with reference to ICC and SCC. The screening of data with reference to ICC is to exclude features with an ICC lower than 0.9 between manual and automatic segmentations to improve the reproducibility of radiomics. The screening of data with reference to SCC is to calculate the SCC between one feature and all other features in each pair of highly correlated features, and features with SCC > 0.9 are excluded. The construction method of the GFR regression model includes the following steps: Step S411: Use the ElasticNet model to select the most representative features from the filtered features and determine the feature weights to estimate rGFR. The remaining features are input into ElasticNet to obtain the calculation formula for rGFR (mL / min×1.73m 2 ): rGFR = p_original_glszm_LargeAreaHighGrayLeve1Emphasis × (3.592 × 10 -7 ) + p_original_glszm_SizeZoneNonUniformity × (2.147 × 10 -5 ) + p_gradient_ngtdm_Busyness × (2.880) + p_square_firstorder_TotalEnergy × (1.353 × 10 -8 ) + h_square_ngtdm_Busyness × (-0.1498) + 19.585; Among them, p_original_glszm_LargeAreaHighGrayLevelEmphasis is the large-area high-gray-level emphasis value based on the gray-scale size zone matrix in the renal parenchyma region, p_original_glszm_SizeZoneNonUniformity is the regional size non-uniformity based on the gray-scale size zone matrix in the renal parenchyma region, p_gradient_ngtdm_Busyness is the busyness based on the gradient neighborhood gray-scale difference matrix in the renal parenchyma region, p_square_firstorder_TotalEnergy is the total energy in the renal parenchyma region, and h_square_ngtdm_Busyness is the busyness based on the gradient neighborhood gray-scale difference matrix in the hydronephrosis region; Step S412: Combine rGFR with clinical features and use multivariable linear regression to obtain rcGFR. The calculation formula for rcGFR (mL / min×1.73m 2 ) is as follows: rcGFR = age × (-0.147) + rGFR × (1.224) - 1.303, where rcGFR is the estimated result of GFR; The method for constructing the SRF regression model includes the following steps: Step S421, calculate the relative contribution ratio rcSRF of the rcGFR of a single kidney to the overall rcGFR, expressed by the following formula: rcSRFl = rcGFRl / (rcGFRl + rcGFRr), rcSRFr = 1 – rcSRFl, where rcGFRl and rcGFRr are the rcGFR of the left and right kidneys respectively, and rcSRFl and rcSRFr are the rcSRF of the left and right kidneys respectively; Step S422, calculate the proportion RPV_P of the single-sided renal parenchymal volume RPV to the total renal parenchymal volume, and the calculation formula is: RPV_Pl = RPVl / (RPVl + RPVr), RPV_Pr = 1 - RPV_Pl, where RPVl and RPVr are the RPV of the left and right kidneys respectively, and RPV_Pl and RPV_Pr are the RPV_P of the left and right kidneys respectively; Use linear regression to obtain the SRF estimated value pSRF based on the renal parenchyma, and the calculation formula is: pSRF = 0.991 × RPV_P + 0.5%, where RPV_P is RPV_Pl or RPV_Pr, and the corresponding pSRF of the left or right kidney is calculated; Step S423, calculate the proportion RHV_P of the single-sided hydronephrosis volume RHV to the total hydronephrosis volume, and the calculation formula is: RHV_Pl = RHVl / (RHVl + RHVr), RHV_Pr = 1 - RHV_Pl, where RHVl and RHVr are the RHV of the left and right kidneys respectively, and RHV_Pl and RHV_Pr are the RPV_P of the left and right kidneys respectively; Use linear regression to obtain the SRF estimated value hSRF based on the hydronephrosis, and the calculation formula is: hSRF = (-0.223) × RHV_P + 61.1%, where RHV_P is RHV_Pl or RHV_Pr, and the corresponding hSRF of the left or right kidney is calculated; Step S424, use linear regression to obtain the final SRF estimate rcphSRF of the left or right kidney combining rcSRF, pSRF, and hSRF. The calculation formula is: rcphSRF = rcSRF × 0.416 + pSRF × 0.564 + hSRF × 0.535 - 25.8%; In step S6, the kidneys in different health states are distinguished according to the following rules: The damaged kidney is a single kidney with GFR ≤ 30 mL / min × 1.73 m 2 , and the normal kidney is a single kidney with GFR > 30 mL / min × 1.73 m 2 ; A non-functional kidney is defined as having a single kidney GFR < 10 mL / min × 1.73 m² 2 , and a functional kidney is defined as having a single kidney GFR ≥ 10 mL / min × 1.73 m² 2 ; Suitable for nephrectomy is SRF < 15%, and not suitable for nephrectomy is SRF ≥ 15%; Renal function decline is SRF < 45%, and no decline is SRF ≥ 45%; The non-dominant kidney is SRF ≤ 60%, and the dominant kidney is SRF > 60%.

2. A system for constructing an automatic renal function estimation model based on non-contrast CT, characterized in that, It includes a collection module, a segmentation module, an extraction module, a modeling module, a verification module, and an application module, where: The collection module is used to collect the image data of patients with renal atrophy or hydronephrosis and preprocess the images; The segmentation module uses the renal parenchyma and hydronephrosis regions manually delineated by radiologists to train an automatic segmentation model for the renal parenchyma and hydronephrosis regions based on UNETR; The extraction module extracts radiomics features from the images of the renal parenchyma and hydronephrosis regions obtained from the automatic segmentation model of the segmentation module and the original images, and screens the data; The modeling module is used to construct renal function index regression models, including the GFR regression model and the SRF regression model; The verification module is used to verify the consistency and differences between the estimated values of the regression model and the measured values based on SPECT; The application module uses the renal function index regression model to calculate GFR and SRF and distinguish the kidneys in different health states.

Images