The Application of CT Radiomics in the Diagnosis of Vein Wall Invasion in Patients With Renal Cell Carcinoma Combined With Tumor Thrombus

Abstract

Objective

The objective of this study was to explore the application of radiomics combined with machine learning to establish different models to assist in the diagnosis of venous wall invasion in patients with renal cell carcinoma and venous tumor thrombus and to evaluate the diagnostic efficacy.

Materials and Methods

We retrospectively reviewed the data of 169 patients in Peking University Third Hospital from March 2015 to January 21, who was diagnosed as renal mass with venous invasion. According to the intraoperative findings, 111 patients were classified to the venous wall invasion group and 58 cases in the non-invasion group. ITK-snap was used for tumor segmentation and PyRadiomics 3.0.1 package was used for feature extraction. A total of 1598 features could be extracted from each CT image. The patients were divided into training set and testing set by time. The elastic-net regression with 4-fold cross-validation was used as a dimension-reduction method. After feature selection, a support vector machines (SVM) model, a logistic regression (LR) model, and an extra trees (ET) model were established. Then the sensitivity, specificity, accuracy, and the area under the curve (AUC) were calculated to evaluate the diagnostic performance of each model on the testing set.

Results

Patients before September 2019 were divided into the training set, of which 88 patients were in the invasion group and 42 patients were in the non-invasion group. The others were in the testing set, of which 32 patients were in the invasion group and 16 patients were in the non-invasion group. A total of 34 radiomics features were obtained by the elastic-net regression. The SVM model had an AUC value of 0.641 (95% CI, 0.463-0.769), a sensitivity of 1.000, and a specificity of 0.062. The LR model had an AUC value of 0.769 (95% CI, 0.620-0.877), a sensitivity of 0.913, and a specificity of 0.312. The ET model had an AUC value of 0.853 (95% CI, 0.734-0.948), a sensitivity of 0.783, and a specificity of 0.812. Among the 3 models, the ET model had the best diagnostic effect, with a good balance of sensitivity and specificity. And the higher the tumor thrombus grade, the better the diagnostic efficacy of the ET model. In inferior vena cava tumor thrombus, the sensitivity, specificity, accuracy, and AUC of ET model can be improved to 0.889, 0.800, 0.857, 0.878 (95% CI, 0.745-1.000).

Conclusion

Machine learning combined with radiomics method can effectively identify whether venous wall was invaded by tumor thrombus and has high diagnostic efficacy with an AUC of 0.853 (95% CI, 0.734-0.948).

Using radiomics combined with machine learning, this article reports on models to assist in the diagnosis of venous wall invasion in patients with renal cell carcinoma and venous tumor thrombus, evaluating diagnostic efficacy.

Implications for Practice.

Deep invasive tumor thrombus, a pervasive condition affecting 40%-60% of patients, is particularly challenging to treat in patients with renal cell carcinoma and venous tumor thrombus. Results of this study showed that machine learning combined with the radiomics method could effectively identify whether venous wall was invaded by tumor thrombus and has high diagnostic efficacy with an area under the curve of 0.853 (95% CI, 0.734-0.948). The results suggest that the radiomics method can be used to assist in the diagnosis of tumor thrombus.

Introduction

Renal cell carcinoma (RCC) is a common malignancy of the urinary system, accounting for 2%–3% of adult malignant tumors and 1.8% of deaths.¹ RCC is a frequently occurring malignancy with over 270 000 new cases diagnosed and it is responsible for 110 000 deaths annually on a global basis.² Among them, 4%-10% of patients will develop venous tumor thrombus (VTT).³ Radical nephrectomy and thrombectomy can effectively improve the prognosis.⁴ A 5-year survival rate of more than 50% can be achieved when the tumor and thrombus are intactly removed. Otherwise, patients would face a dire 5-year survival of only 10% if the resection is incomplete.⁵ Deep invasive tumor thrombus (DITT), a pervasive condition affecting 40%-60% of patients, is particularly challenging to treat, requiring invasive surgical procedures that involve the removal of the vessel wall infiltrated by the tumor thrombus, or even segmental resection of the inferior vena cava (IVC) and reconstruction if the invasion exceeds 50% of the caval wall.^6,7 DITT is a crucial determinant of a patient’s prognosis and has been included in the 8th edition of the American Joint Committee on Cancer TNM system, which classifies it as stage T3c.^8,9 Therefore, preoperative diagnosis of DITT is important for surgical planning, prognosis assessment, and patient counseling.

Preoperative diagnosis of DITT with MRI or CT imaging is fraught with difficulties. Traditional subjective imaging findings and semi-quantitative measurements of the IVC or renal vein diameters offer only moderate degrees of accuracy.⁷ The subjective findings obtained by radiologists suffer from low sensitivity, accuracy, and lack of interpreter consistency.¹⁰ Many researchers have tried to increase the accuracy of diagnosis by semi-quantitative models. For example, Psutkaet et al¹¹ reported that the maximum anterior–posterior diameter of the IVC ≥ 24.0 mm was a risk factor for veinous wall invasion. However, it can only achieve moderate accuracy and exhibit low specificity.

In 2012, Lambin et al¹² introduced the concept of radiomics, which involves extracting additional information such as texture, shape, intensity, and high-order features from medical images using advanced feature analysis. This technique provides a way to convert the features into collectible radiomics data that can aid in revealing the nature of lesions and help radiologists make the correct decision. However, there have been few related reports on whether radiomics can assist in determining DITT in patients with RCC and VTT. The purpose of this study was to improve the diagnosis of DITT events by using radiomic techniques.

Materials and Methods

Patients

We retrospectively analyzed the clinical data of patients of RCC with VTT who were admitted to the Department of Urology, Peking University Third Hospital from March 2015 to January 2021. The inclusion criteria were as follows: (1) Patients underwent radical nephrectomy and thrombectomy; (2) postoperative pathological diagnosis was renal cell carcinoma; (3) contrast-enhanced CT examination of the abdomen or kidneys was performed within 2 weeks before surgery. The exclusion criteria were as follows: (1) Patients with incomplete clinical data; (2) lack of documentation on whether the vein wall was invaded; (3) low image quality or obvious artifacts. A total of 169 patients were finally included, of which 111 were classified into the DITT group and the other 58 cases were classified into the non-invasive tumor thrombus (NITT) group. For renal vein tumor thrombus, the criterion for invading the venous wall is postoperative pathological results. For IVC tumor thrombus, the criterion of venous wall invasion is mainly based on intraoperative findings and postoperative pathological results.

CT Images Acquisition

Renal plain and enhancement CT scans were performed on all patients using a 64-row scanner (GE Discovery HD 750, GE Healthcare). Prior to scanning, patients were instructed to fast for 6-8 h. The scanning parameters included a tube voltage of 120 kV, automatic tube current adjustment technology, layer thickness of 5 mm, and matrix of 512 × 512. In addition, all patients underwent an enhanced CT scan, which involved the injection of 100 mL of iodine contrast agent through the elbow vein at a rate of 2-3 mL/s. The scanning sequences and their corresponding parameters consisted of a cortical phase, medullary phase, and excretion phase, each with a scanning duration of 12, 37, and 170 s after the injection of contrast agent, respectively.

Image Processing

CT images are acquired in DICOM format for image processing and analysis and then were loaded into the ITK-Snap (version 3.8.0) for segmentation. The delineation principles used involved a medial margin of 0-1 mm in the cortical phase, and layer-by-layer segmentation to obtain the volume of interest (VOI). For each patient, the VOI delineations were carried out for the tumor thrombus part (as shown in Fig. 1). The segmentation process was performed by 2 urologists, each with at least 5 years of clinical experience.

Figure 1.

Segmentation of the volume of interest (VOI). A: Image of the right renal cell carcinoma with tumor thrombus in cortical phase. B: Segmentation of VOI, the covered part is the tumor thrombus.

Feature Extraction and Selection

For feature extraction, the PyRadiomics 3.0.1 package was used. The radiomics features were extracted from the original CT images and derived CT images. The derived CT images were obtained from the original CT images processed using 17 different filters, including wavelet (consisting of LHH, HHH, LLH, HLH, LHL, HHL, LLL, and HLL), square, squareroot, logarithm, exponential, gradient, and local binary patterns (LBP; comprising of 2D, 3D-m1, 3D-m2, and 3D-k). The extracted radiomics features included 14 shape features, 18 first-order features, 24 gray-level co-occurrence matrices (GLCM) features, 16 gray-level run-length matrix (GLRLM) features, 16 gray-level size zone matrix (GLSZM) features, and 14 gray-level dependence matrix (GLDM) features. The shape features were only extracted from the original CT images. In total, 1598 (18 × 88 + 14) features were extracted from each patient prior to feature selection. Each of the extracted radiomics features was standardized using Z-score normalization by subtracting its mean and then dividing by its standard deviation (SD). Following this, the elastic-net regression was used for feature selection. A list of 50 values in a log-space from −3 to 3 was defined for the selection of the best alpha in the elastic-net regression with 4-fold cross-validation.

The Construction of Machine Learning Models

Following the feature selection process, a logistic regression (LR) model, a support vector machine (SVM) model, and an extra tree (ET) model were constructed using the training set and subsequently tested using the testing set. The construction of the models was based on the Scikit-Learn 1.0.1 package. A grid search method was used to select the best hyperparameters of the machine learning models in this study. For the LR model with L1 regularization, the best value of the inverse of regularization strength was 0.125, and the solver was saga. For the SVM model with linear kernel, the best values of the regularization parameter and kernel coefficient were 0.0625 and 0.0625, respectively. Furthermore, for the ET model, “entropy” was selected as the best criterion, and the best values of the max features, max depth, min samples split, and number of estimators were 4, 10, 4, and 50, respectively. The optimal cut-off was determined using the Youden index on the training set.

Statistical Analysis

Statistical analysis was performed using SPSS version 24.0 and Python version 3.8. The independent t-test and chi-square test were used to assess the distinctions between the training and testing sets. To evaluate the diagnostic performance of the different models, the receiver operating characteristic (ROC) curve was constructed. Additionally, metrics such as sensitivity (SEN), specificity (SPE), accuracy, and area under the ROC curve (AUC) were calculated. A P-value less than .05 with a 2-tailed test was considered statistically significant.

Results

Patients’ Characteristics

The present study included a total of 169 patients, aged between 15 and 83 years (mean 59.18 ± 11.07 years). Of these patients, 121 (71.6%) were male and 48 (28.4%) were female. Tumor location was predominantly in the right kidney (62.7%, n = 106) and clear renal cell carcinoma was the most common pathological type (85.8%, n = 145). Non-clear renal cell carcinoma accounted for 14.2% (n = 24) of cases. The Mayo classification system was used to assess venous tumor thrombus (VTT) with 20.1% (n = 34) classified as level 0, 18.3% (n = 31) as level I, 42.6% (n = 72) as level II, 9.5% (n = 16) as level III, and 9.5% (n = 16) as level IV. The patients were divided into 2 groups: 111 were classified as having direct invasion to the renal vein and inferior vena cava (DITT), while the remaining 58 were classified as having non-invasive tumor thrombus (NITT).

To improve the practicality and validation of the study, we divided the patients into training and testing sets based on their operation dates. Specifically, the training set consisted of 130 patients who underwent surgery between March 2015 and September 2019, with 88 cases in the DITT group and 42 in the NITT group. The testing set comprised 39 patients who underwent surgery between September 2019 and January 2021, including 23 cases in the DITT group and 16 in the NITT group. The ratio of training set to testing set was approximately 3:1. Clinical and pathological characteristics of the training and testing sets are presented in Table 1, with no significant differences observed between the 2 groups in terms of age, sex, body mass index (BMI), ASA score, Mayo classification, or pathological type (P > 0.05). The similar proportion of DITT in both groups (67.7% vs 59.0%, P = .340) indicated that the researchers’ assessments of venous wall invasion were consistent across different periods. These findings suggest that our method of dividing the training and testing sets was appropriate.

Table 1.

Comparison of clinical and pathologic features between training set and testing set.

	Training set (N = 130)	Testing set (N = 39)	P
Age, years	59.4 ± 10.5	58.5 ± 12.9	.670
BMI, kg/m2	24.1 ± 4.0	24.8 ± 3.9	.330
Tumor diameter, cm	8.9(7.0, 10.5)	9.0(6.2, 9.9)	.516
Hb, g/L	121(107, 138)	131(109, 146)	.124
Neu, 109/L	4.40(3.43, 5.39)	4.61(3.57, 5.99)	.473
Plt, 109/L	236(183, 299)	238(166, 339)	.905
ALP, U/L	83.5(68.0, 104.0)	90.0(78.0, 104.0)	.277
Alb, g/L	40.0(35.8, 43.0)	40.0(38.0, 43.1)	.337
SCr, μmol/L	91.0(79.0, 111.0)	88.0(71.0, 109.0)	.360
SCr after surgery, μmol/L	101.0(80.8, 122.0)	91.0(74.0, 113.0)	.129
Sex			.067
Male	98(75.4%)	23(59.0%)
Female	32(24.6%)	16(41.0%)
Side			.130
Left	44(33.8%)	19(48.7%)
Right	86(66.2%)	20(51.3%)
ASA score			.932
1	8(6.2%)	1(2.6%)
2	102(78.5%)	33(84.6%)
3	20(15.4%)	5(12.8%)
Venous wall invasion			.340
DITT	88(67.7%)	23(59.0%)
NITT	42(32.3%)	16(41.0%)
Mayo classification			.090
0	23(17,7%)	11(28.2%)
I	24(18.5%)	7(17.9%)
II	55(42.3%)	17(43.6%)
III	14(10.8%)	2(5.1%)
IV	14(10.8%)	2(5.1%)
IVC resection			.353
No	103(79.2%)	34(87.2%)
Yes	27(20.8%)	5(12.8%)
Pathology type			.805
Clear cell carcinoma	111(85.4%)	34(87.2%)
Non-clear cell carcinoma	19(14.6%)	5(12.8%)

Feature Selection

A total of 34 features were selected by applying the elastic-net regression with 4-fold cross-validation. The selected features and their coefficients are shown in Table 2 and Fig. 2. Among them, there were 14 features with coefficients above 0.01, including 8 positive correlation features (wavelet-LLL_glcm_Correlation, wavelet-HHL_glszm_SizeZoneNonUniformityNormalized, wavelet-HLH_glszm_LowGrayLevelZoneEmphasis, wavelet-LLL_glszm_LowGrayLevelZoneEmphasis, wavelet-HHH_firstorder_Skewness, lbp-3D-k_firstorder_TotalEnergy, wavelet-HHL_glrlm_RunVariance, wavelet-LHH_firstorder_Median) and 6 negative correlation features (original_shape_Flatness, logarithm_glcm_JointEnergy, wavelet-HHH_glrlm_ShortRunLowGrayLevelEmphasis, squareroot_firstorder_90Percentile, logarithm_glszm_LargeAreaLowGrayLevelEmphasis, squareroot_firstorder_RootMeanSquared).

Table 2.

The selected features and their coefficients.

N.	Features	Coefficients
1	wavelet-LLL_glcm_Correlation	0.10939104
2	wavelet-HHL_glszm_SizeZoneNonUniformityNormalized	0.029863465
3	wavelet-HLH_glszm_LowGrayLevelZoneEmphasis	0.023543015
4	wavelet-LLL_glszm_LowGrayLevelZoneEmphasis	0.021784179
5	wavelet-HHH_firstorder_Skewness	0.017581696
6	lbp-3D-k_firstorder_TotalEnergy	0.012058674
7	wavelet-HHL_glrlm_RunVariance	0.010167446
8	wavelet-LHH_firstorder_Median	0.010075244
9	wavelet-LHH_glcm_Correlation	0.009754829
10	wavelet-HHH_firstorder_Mean	0.006855402
11	original_glcm_Idn	0.004570159
12	exponential_glszm_SmallAreaEmphasis	0.004267094
13	wavelet-HHH_firstorder_Median	0.003243158
14	lbp-3D-m2_glrlm_LongRunHighGrayLevelEmphasis	0.001418523
15	lbp-3D-m2_glrlm_LongRunLowGrayLevelEmphasis	0.001371199
16	lbp-3D-m2_glrlm_LongRunEmphasis	0.001311341
17	lbp-2D_glrlm_LongRunEmphasis	0.00116546
18	lbp-3D-m1_glrlm_LongRunLowGrayLevelEmphasis	0.001089809
19	lbp-2D_glrlm_LongRunHighGrayLevelEmphasis	0.000925746
20	lbp-3D-m1_glrlm_LongRunHighGrayLevelEmphasis	0.000850654
21	lbp-2D_glrlm_LongRunLowGrayLevelEmphasis	0.000745071
22	lbp-3D-m1_glrlm_LongRunEmphasis	0.000704519
23	wavelet-LHL_glszm_SmallAreaLowGrayLevelEmphasis	−0.000949484
24	square_firstorder_Minimum	−0.002229518
25	wavelet-LLL_firstorder_10Percentile	−0.002802228
26	wavelet-LHH_glcm_InverseVariance	−0.004541284
27	original_shape_Elongation	−0.007431839
28	wavelet-LLL_glszm_SmallAreaEmphasis	−0.007594597
29	original_shape_Flatness	−0.013019546
30	logarithm_glcm_JointEnergy	−0.013754829
31	wavelet-HHH_glrlm_ShortRunLowGrayLevelEmphasis	−0.015341637
32	squareroot_firstorder_90Percentile	−0.021816315
33	logarithm_glszm_LargeAreaLowGrayLevelEmphasis	−0.025760622
34	squareroot_firstorder_RootMeanSquared	−0.038311654

Figure 2.

The selected features and their coefficients. The characteristics represented by numbers are the same as Table 2.

Predictive Performance of Models

The diagnostic performance of the 3 predictive models for distinguishing DITT from NITT in the testing set is summarized in Table 3. Figure 3 shows the ROC curves of the 3 models, respectively. The LR model exhibited an AUC value of 0.769 (95% CI, 0.620-0.877), with a sensitivity of 0.913 and a specificity of 0.312. The SVM model demonstrated an AUC value of 0.641 (95% CI, 0.463-0.769), with a sensitivity of 1.000 and a specificity of 0.062. Finally, the ET model demonstrated the best diagnostic performance, with an AUC value of 0.853 (95% CI, 0.734-0.948), a sensitivity of 0.783, and a specificity of 0.812. The ET model demonstrated a superior balance of sensitivity and specificity compared to the other 2 models.

Table 3.

The diagnostic effect of the prediction models in the testing set.

Model	Cutoff	AUC	95% CI of AUC	Accuracy	Sensitivity	Specificity
LR	0.544	0.769	0.620-0.877	0.667	0.913	0.312
SVM	0.522	0.641	0.463-0.769	0.615	1.000	0.062
ET	0.702	0.853	0.734-0.948	0.795	0.783	0.812

Abbreviations: AUC, area under the curve; CI, confidence interval; ET, extra trees; LR, logistic regression; SVM, support vector machine..

Figure 3.

Receiver operating characteristic (ROC) curve analysis of 3 radiomics models for differentiating deep invasive tumor thrombus from non-invasive tumor thrombus in the testing set. (a) The ROC curve of LR model, (b) The ROC curve of SVM model, (c) The ROC curve of ET model.

Subgroup Analysis

Subgroup analysis of the ET model was carried out to further evaluate its diagnostic performance (Table 4). The diagnostic effect of the ET model was consistent across different Fuhrman grades and pathological types. Regarding Mayo classification, the higher the level of VTT, the better the diagnostic performance of the model, with 100% accuracy observed in patients with level III and IV VTT. In patients with level I and II VTT, the ET model demonstrated an AUC value of 0.857 (95% CI, 0.696-1.000), a sensitivity of 0.857, and a specificity of 0.800, which was slightly higher than that observed in all patients. However, the diagnostic performance for renal vein tumor thrombus was poor, with a sensitivity of 0.400. This finding may be attributed to the structural differences between the renal vein and the inferior vena cava. In terms of different sides, the diagnostic performance of the ET model was better for the right side (sensitivity 0.917; specificity 0.875; AUC 0.927, 95% CI, 0.810-1.000) than the left side (sensitivity 0.636; specificity 0.750; AUC = 0.773, 95% CI, 0.554-0.992), possibly due to the higher proportion of patients with pure renal vein tumor thrombus in the left renal cell carcinoma.

Table 4.

The diagnostic effect of the ET model in different subgroups in the testing set.

Group	AUC	95% CI of AUC	Accuracy	Sensitivity	Specificity
Clear RCC	0.846	0.711-0.982	0.794	0.800	0.786
Renal vein tumor thrombus	0.767	0.462-1.000	0.636	0.400	0.833
IVC tumor thrombus	0.878	0.745-1.000	0.857	0.889	0.800
Mayo I-II	0.857	0.696-1.000	0.833	0.857	0.800
Mayo III-IV	1.000	1.000-1.000	1.000	1.000	1.000
Left tumor	0.773	0.554-0.992	0.684	0.636	0.750
Right tumor	0.927	0.810-1.000	0.900	0.917	0.875
Fuhrman grades 1-2	0.905	0.717-1.000	0.769	0.857	0.667
Fuhrman grades 3-4	0.852	0.689-1.000	0.792	0.750	0.875

Abbreviations: IVC, inferior vena cava; RCC, renal cell carcinoma;.

Discussion

Preoperative prediction of the invasion of venous wall by tumor thrombus is crucial for optimal surgical planning and rational evaluation of patient prognosis. In the current study, machine learning-based diagnostic models have been developed using radiomics features extracted from arterial phase images of preoperative CT. Among the 3 models we constructed, the extra trees (ET) model exhibited the best diagnostic performance, achieving a sensitivity of 0.783, specificity of 0.812, accuracy of 0.795, and AUC of 0.853 in the testing set. Notably, a diagnostic accuracy of 0.857 was achieved for inferior vena cava tumor thrombus.

The accuracy of preoperative imaging in diagnosing venous wall invasion by tumor thrombus remains a controversial topic. Although some studies have shown that CT or MRI can detect venous wall invasion, the accuracy has been reported to be unsatisfactory.^10,13 In contrast to the low diagnostic accuracy of preoperative CT and MRI reported by Guo et al¹⁰ for detecting deep inferior epigastric perforator flap, Adams et al¹⁴ found that complete occlusion of the IVC or disruption of the vessel wall was a reliable sign of tumor thrombus invasion, with high sensitivity and specificity of 92.3% and 86.4%, respectively. However, the accuracy reported by Adams was higher than that reported in previous studies, and the reason for this discrepancy was attributed to the extensive experience of the experts. Alayed et al⁷ observed a high false-negative rate for the physician’s subjective impression in diagnosing tumor thrombus invasion based on MRI data, indicating the limited accuracy and repeatability of imaging examinations, which are highly dependent on the expertise of experts and the consistency of examination equipment.

Several studies have attempted to improve the ability to diagnose tumor thrombus invasion by constructing semi-quantitative models. For instance, renal vein and IVC diameter have been used to predict tumor thrombus invasion, with an AUC of 0.78-0.86.⁶ Psutka et al¹¹ identified several risk factors for tumor thrombus invasion, such as right RCC, maximum anterior-posterior diameter of the IVC at the renal vein ostium ≥ 24.0 mm, and complete occlusion of the IVC, with an accuracy of the model reaching 0.81. However, Overholser et al reported that these factors did not predict vascular reconstruction.¹⁵ Different from 24 mm, some studies reported that the diameter of tumor thrombus >17.0 mm was an independent risk factor for DITT, but has no diagnostic effect.¹⁶ Liu et al¹⁷ developed a semi-quantitative model based on residual IVC blood flow, maximum coronal IVC diameter, and the presence of bland thrombus, achieving a diagnostic accuracy of 80.8%. Nevertheless, these models were highly subjective, and reported large differences among studies, generally lacking validation.

In previous studies, radiomics has demonstrated good results in renal mass differentiation, nuclear grade prediction, gene expression-based molecular signatures, and patient outcome prediction for RCC.¹⁸ However, radiomics research on the diagnosis of tumor thrombus invasion has been limited. In this study, radiomics was applied to the diagnosis of venous wall invasion and achieved promising diagnostic outcomes. We extracted a large number of radiomics features (1598 features) through a comprehensive open-source platform called PyRadiomics, which is common in radiomics studies,¹⁹ and they can reflect tumor characteristics in different ways. A large number of radiomics features could represent abundant characteristics of a lesion. But not all extracted radiomics features are used to build the models. After feature filtering to reduce the dimensionality of our feature space, there are only 34 selected features are used to build the models.

The most important predictive feature is wavelet-LLL_glcm_Correlation. The gray-level co-occurrence matrix (GLCM) characterizes the texture of an image by calculating how often pairs of pixels with specific values and in a specified spatial relationship occur in an image. The glcm_Correlation represents the linear dependence of gray-levels in GLCM. The correlation between pixels means that there is a predictable and linear relationship between the 2 neighboring pixels within the window, as expressed by the regression equation. In previous studies, GLCM correlation can be used to predict the presence of adherent perinephric fat (APF) in renal tumor surgery.²⁰ The glcm_correlation can also identify undifferentiated soft-tissue sarcomas at high risk of relapse and death, thus helping to optimize the perioperative treatment of patients.²¹ It is also of great significance in predicting the risk of metastatic progression in high-risk prostate cancer.²² There are also 3 positive predictive factors with correlation coefficients above 0.02, all of which are texture features, including wavelet-HHL_glszm_SizeZoneNonUniformityNormalized, wavelet-HLH_glszm_LowGrayLevelZoneEmphasis, wavelet-LLL_glszm_LowGrayLevelZoneEmphasis. A gray level size zone (GLSZM) quantifies gray level zones in an image. A gray level zone is defined as the number of connected voxels that share the same gray level intensity. These 3 features represent the enhancement uniformity of tumor thrombus, which is related to the microscopic changes in local metabolic activity and is a measure of regional tumor heterogeneity. The glszm_SizeZoneNonUniformityNormalized feature was reported to be selected for identifying the tumor grading of tongue squamous cell carcinomas.²³ And glszm_LowGrayLevelZoneEmphasis was correlated to pathologic response in patients with esophageal cancer treated with induction chemotherapy followed by neoadjuvant chemo-radiotherapy.²⁴ For negative correlation factors, there are 2 important first-order features: squareroot_firstorder_90Percentile and squareroot_firstorder_RootMeanSquared. Their decrease would be attributed to a relatively lesser degree of enhancement of DITT than that of NITT. These 2 first-order features can also help differentiate renal epithelioid angiomyolipoma from clear cell carcinoma.²⁵ The negative predictive factor with a correlation coefficient above 0.02 also includes a logarithm_glszm_LargeAreaLowGrayLevelEmphasis, which was believed to be related to the immune-related gene expression status.²⁶

The ET model exhibited the best diagnostic performance with a sensitivity of 0.783, specificity of 0.812, accuracy of 0.795, and AUC of 0.853. Some subgroup analyses have been done in our experiments, as shown in Table 4. The performance of the ET model on the group of renal vein tumor thrombus is poor. This may be due to the structural difference between the renal vein tumor thrombus and the IVC tumor thrombus. The shape of IVC is different from that of the renal vein. On the CT transverse plane, the IVC is nearly circular, while the renal vein is nearly rectangular, which is equivalent to the sagittal plane. Therefore, there may be great differences in segmentation and feature extraction. In addition, the CT images used in this study have a scanning layer thickness of 5 mm, which is too large for the renal vein, which is also one of the reasons why the model’s performance in renal vein tumor thrombus is poor. For IVC tumor thrombus, 4 samples do not perform well with the ET model, including 3 patients with sarcomatoid differentiation, necrosis, or bland thrombus. These rare factors may be the cause of model failure.

Despite these findings, this study has several limitations, such as a relatively small sample size and limited external validation, retrospective study design, no attempt at any neural network methods, and lack of distinction between the renal vein and IVC tumor thrombus. Future studies could focus on radiomics analysis and model construction for IVC tumor thrombus. Furthermore, the inclusion of a large number of features in the model poses potential risk of overfitting.

Funding

This study was supported by the National Nature Science Foundation of China (NO. 81771842).

Ethical Approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the Peking University Third Hospital ethics committee. This study has been approved by the Ethics Committee of Peking University Third Hospital (the number of the ethics approval: No 202214701).

Conflict of Interest

The authors indicated no financial relationships.

Author Contributions

Conception/design: X.Z., Y.Y., W.X. Provision of study material or patients: S.Z., C.L., L.M. Collection and/or assembly of data: X.Z., Y.Y., S.Z., C.L., L.M. Data analysis and interpretation: W.X., L.Z., J.L. Manuscript writing: X.Z., Y.Y., W.X., J.L., C.L., L.M. Final approval of manuscript: All authors.

Data Availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.