CT-derived Radiomics Predict the Growth Rate of Renal Tumors in von Hippel-Lindau Syndrome

Abstract

Purpose:

Current management guidelines for renal tumors in von Hippel-Lindau syndrome recommend active surveillance (scan every 6–12months) until renal tumors reach 3cm diameter, followed by surgery. This study aims to predict tumor growth pattern with volume doubling time by utilizing radiomic features, which can further assist in developing personalized surveillance plans leading to better patient outcomes.

Methods:

The study evaluated 78 renal tumors in 55 patients with pathologically-confirmed clear cell Renal Cell Carcinomas (ccRCCs) which were segmented and radiomics were extracted. Volumetric doubling time (VDT) classified the tumors into fast-growing (VDT <365 days) or slow-growing (VDT ≥365 days). Volumetric and diametric growth analyses were compared between the groups. Multiple logistic regression and random forest classifier were used to select best features and models based on their correlation and predictability of VDT.

Results:

Fifty-five patients (mean age 42.2 years ±12.2, 27 males) with a mean time difference of 3.8 ±2 years between baseline and pre-operative scan were studied. Twenty-five tumors were fast-growing (low VDT i.e. <365days), and 53 tumors were slow-growing (high VDT i.e. ≥365days). Median volumetric and diametric growth rate were 1.71cc/year and 0.31cm/year. Best feature with univariate analysis was wavelet-HLL_glcm_ldmn (AUC of 0.80, p<0.0001), and with random forest classifier was log-sigma-0–5mm-3D_glszm_ZonePercentage (AUC: 79). AUC of ROC curves made by multiple logistic regression was 0.74, and with random forest classifier was 0.73.

Conclusion:

Radiomic features correlated with VDT and were able to predict the growth pattern of renal tumors in patients with VHL syndrome.

MATERIALS AND METHODS:

The study evaluated 78 renal tumours in 55 patients with histopathologically-confirmed clear cell renal cell carcinomas (ccRCCs), which were segmented and radiomics were extracted. Volumetric doubling time (VDT) classified the tumours into fast-growing (VDT <365 days) or slow-growing (VDT ≥365 days). Volumetric and diametric growth analyses were compared between the groups. Multiple logistic regression and random forest classifiers were used to select the best features and models based on their correlation and predictability of VDT.

RESULTS:

Fifty-five patients (mean age 42.2±12.2 years, 27 men) with a mean time difference of 3.8±2 years between the baseline and preoperative scans were studied. Twenty-five tumours were fast-growing (low VDT, i.e., <365 days), and 53 tumours were slow-growing (high VDT, i.e., ≥365 days). The median volumetric and diametric growth rates were 1.71 cm³/year and 0.31 cm/year. The best feature using univariate analysis was wavelet-HLL_glcm_ldmn (area under the receiver operating characteristic [ROC] curve [AUC] of 0.80, p<0.0001), and with the random forest classifier, it was log-sigma-0–5-mm-3D_glszm_ZonePercentage (AUC: 79). The AUC of the ROC curves using multiple logistic regression was 0.74, and with the random forest classifier was 0.73.

CONCLUSION:

Radiomic features correlated with VDT and were able to predict the growth pattern of renal tumours in patients with VHL syndrome.

INTRODUCTION

Von Hippel–Lindau (VHL) syndrome is an inherited autosomal dominant disorder affecting the short arm of chromosome 3, predisposing individuals to developing benign and malignant lesions in the kidney and other organ systems ^1,2. Renal lesions in VHL syndrome vary from simple and complex cystic lesions to solid masses, which are clear cell renal cell carcinomas (ccRCCs) in most cases ³. Considering the low metastatic potential of small renal tumours, current management guidelines suggest active imaging surveillance using cross-sectional imaging until the maximum tumour diameter reaches 3 cm, followed by surgical resection of the tumour ⁴. Elective surveillance imaging is often too early or late for the dominant tumour to cross the 3 cm threshold, given the heterogeneity in growth patterns of the respective renal lesions in VHL syndrome. Personalising the frequency of surveillance can avoid tumour metastasis in rapidly growing tumours and reduce radiation exposure to the patient and mutual logistical burden in case of slow-growing tumours.

The extraction of computerised quantitative radiomic features from medical scans has emerged as a promising tool for quantifying the radiological characteristics of tumours ⁵. It allows for non-invasive, high-throughput analysis of large amounts of data and can potentially identify subtle changes in tumours that are not visible on standard imaging, thus empowering precision medicine. Radiomics encompasses shape-based features, first-order statistics (distribution of voxel intensities), texture features (grey-level matrix features) and wavelet features. The aim of this study was to investigate the use of radiomics derived from baseline abdominal computed tomography (CT) examinations of renal tumours in patients with VHL syndrome, as a tool for predicting the growth of these tumours. The predicted growth rate can ultimately assist in individualising the management plans and improving patient outcomes.

MATERIALS AND METHODS

This institutional review board-approved, single-centre study was a retrospective evaluation of a prospectively maintained database of patients with genetically detected VHL alterations who underwent partial or radical nephrectomies at the Clinical Center, National Institutes of Health (Institute) between January 2015 and June 2021. In the study, the majority of CT examinations were conducted using a 256-section CT system. A uniform dose of approximately 100 ml of iopamidol 61% solution was used as a contrast agent for all adult patients in the study. Tumours were labelled with a unique ID during the pre-surgical scans, and a direct histopathological correlation was established with the surgical histopathology report after resection of respective tumours. The study included only VHL patients who were confirmed to have ccRCCs via surgical histopathology. Each participant had undergone contrast-enhanced abdominal CT at a minimum of two different times, specifically at the beginning of the study and just before surgery. The initial scan was the one in which the tumour in question was first identified (as shown in Electronic Supplementary Material Figure S1). Any patients with poor-quality scans, such as those with motion artefacts or those who had <1-year interval between two scans, were not included in the study. Ultimately, 55 patients with a total of 78 ccRCCs that met these criteria were part of the study. A study with a similar subset of patients undergoing active surveillance with sequential MRI was undertaken by Anari et al. with a cohort overlap of 30 patients in the present study. The focus of their work was to demonstrate the utility of machine-learning models in predicting the growth rate in renal tumours using MRI-based radiomics ⁶. Another study by Farhadi et al. demonstrated the utility of diffusion-weighted MRI in identifying ccRCC with higher growth rates with a cohort overlap of 17 patients ⁷.

Image acquisition and segmentations

CT images were imported in Syngo.via, Siemens Healthineers software (Ehrlangen, Germany). Tumours were identified independently and segmentations were performed using the semi-automatic segmentation feature in the radiomics prototype (FRONTIER, Siemens Healthineers) by two postdoctoral research fellows blinded to the tumour growth pattern or time point. All sections showing the tumours were segmented on the venous phase (2 minutes after contrast medium injection as per the institute protocol) of both baseline and preoperative CT. The venous phase was chosen because it provides a more comprehensive view of the tumour’s overall tissue characteristics beyond just its vascularity. This 3D segmentation mask of the tumour provides a comprehensive and complete representation of the tumour in three dimensions. Segmentations were supervised by three radiologists (AAM, EBT, FH) experienced in the field of body imaging.

Radiomics feature extraction and application

Radiomics features of the segmented tumours were extracted using the PyRadiomics library, a widely used tool that is an in-built feature in the radiomics prototype ( Figure 1). A uniform voxel size was achieved by resampling with B-spline implemented by PyRadiomics. A total of 1,691 features were extracted, including 17 shape-based, and 1,202 texture features (324 first-order, 433 grey-level concurrence matrix [GLCM], 228 grey-level size zone [GLSZ], 253 grey-level difference matrix [GLDM], and 288 grey-level run length matrix [GLRLM]), all including the additional mathematical transformations such as squares, square roots, logarithmic, and exponentials of actual feature values.

Figure 1.

Workflow diagram of segmentation and radiomic features analyses.

Radiomics and statistical analysis

Volumetric and diametric sizes were used from the extracted features for growth analysis of tumours between the baseline and preoperative time point. Using the volume and the time difference between the baseline (t₀) and preoperative time point (t₁), VDT was calculated with the widely used formula⁷:

$$\text{VDT}=(t_1-t_0)\text{log}\left(2\right)/\text{log}({\text{V x t}}_1/{\text{V x t}}_0)$$

All tumours were divided into two groups based on the VDT: fast-growing tumours with low VDT of <365 days ( Figure 2), and slow-growing tumours with high VDT of ≥365 days ( Figure 3). Paired t-test was used to compare baseline and preoperative clinical characteristics and Mann–Whitney U-test was performed to compare unequal sample sizes of low and high VDT parameters. Volumetric and diametric growth analysis were compared between both groups. Univariate analysis was performed to select the top 10 radiomic features with best correlation coefficient. Features with a p-value of <0.3 (n=801) were considered for further analysis with multiple logistic regression and random forest classifier. A lenient p-value threshold of 0.3 was selected for this exploratory analysis. This approach, deviating from standard practices, aims to ensure that potentially important features are not omitted due to overly stringent threshold criteria. After dividing the dataset into training and testing set (80:20), the training set was resampled with synthetic minority oversampling technique (SMOTE). Logistic regression and random forest models were trained on train set and run on test set with 10-fold cross-validation. Receiver operating curves (ROC) were extracted and model performance was reported in terms of area under the curve (AUC), F1 scores, and accuracy.

Figure 2.

Axial and coronal sections in venous phase of baseline (a,c) segmented with yellow and preoperative scans (b,d) segmented with red show a fast-growing tumour within a time difference of 1.5 years between baseline and preoperative scan and VDT of 162 days (0.4 years).

Figure 3.

Axial and coronal sections in the venous phase of baseline (a,c) segmented with yellow and preoperative scans (b,d) segmented with red show a slow growing tumour within a time difference of 5.2 years between baseline and preoperative scan and VDT of 5,546 days (15 years).

RESULTS

Demographic characteristics

A total of 55 patients (27 men, 28 women) fulfilled the inclusion criteria, and 78 histopathologically confirmed ccRCC were studied. Median age of patients at baseline scan was 42 years. The median time difference between the two time points was 3.7 (3.2–3.9) years (Table 1). Median volume of tumours at baseline was 1.5 cm³ (0.8–2.2; high VDT tumours: 2.17 cm³ [0.55–5.17], low VDT tumours: 0.48 cm³ [0.19–1.70]; p<0.001) whereas at the preoperative scan was 8.2 cm³ (6.4–11.6; high VDT tumours: 8.13 cm³ [4.2–15.26], low VDT tumours: 10.11 cm³ [4.26–20.45]; p=0.40). Fifty-three tumours fell in the slow-growing range (high VDT), with a median doubling time of 721 days (644–946) whereas 25 tumours were fast-growing with a median doubling time of 243 days (213–269). The median overall volumetric growth rate was 1.71 cm³/year (1.33–2.74; Electronic Supplementary Material Fig. S2), and median overall diametric growth was 0.31 cm/year (0.25–0.41). The median volumetric growth rate in high and low VDT were 1.38 cm³/year (0.84–1.92) and 3.54 cm³/year (2.37–6.43), respectively (p<0.001). The median diametric growth rate in high and low VDT were 0.24 cm/year (0.18–0.30) and 0.71 cm/year (0.50–0.82), respectively (p<0.001).

Table 1.

Demographic and growth characteristics in patients with VHL syndrome included in this study.

Clinical variable	Overall median (IQR); proportion	Overall mean±SD	High versus low VDT median (IQR), Mean; Proportion		p-Value
			Low VDT (n=25)	High VDT (n=53)
Age (years)	42 (30.3–51.8)	42.2±12.3	38 (29–48), 40	38 (29–50), 41.2	0.9
Male:female	27:28:00	-	15:09	20:23	0.1
Time interval (T1–T2) (years)	3.7 (2.28–5.17)	3.8±2	3.3 (1.5–3.9), 2.9	3.8 (2.7–5.9), 4.2	0.01
Baseline volume (cm3)	1.5 (0.4–3.3)	4.3±13.9	0.48 (0.19–1.70), 1.04	2.17 (0.55–5.17), 6.12	<0.001
Preoperative volume (cm3)	8.2 (4.2–20.2)	14.5±18.6	10.11 (4.26-20.45), 15.1	8.18 (4–15.26), 14.51	0.4
Volumetric Growth rate (cm3/year)	1.71 (0.8–4.2)	3.0±3.3	3.54 (2–7.6), 4.96	1.37 (0.53–2.75), 2.1	<0.001
Diametric Growth rate (cm/year)	0.31 (0.17–0.51)	0.4±0.33	0.71 (0.47–0.88), 0.73	0.24 (0.15–0.35), 0.25	<0.001
Volume doubling time (days)	538 (278–539)	897±1517	243 (171–278), 97	721 (484–1212), 1303	-

Feature selection and statistical analysis

A cluster map of 10 best features with their z-scores was extracted (Electronic Supplementary Material Fig. S3). The best feature selected from univariate analysis according to significant correlation with doubling time was wavelet-HLL_glcm_ldmn (AUC of 0.80, p<0.001). Two out of 10 features were first order (3d shape and wavelet filter), and the rest were higher-order features. The most significantly associated feature by multiple logistic regression was wavelet-LLH_glcm_ClusterShade (AUC: 0.86) and by random forest classifier was log-sigma0–5-mm-3D_glszm_ZonePercentage (AUC: 0.79). The mean AUC, F1-score and accuracy for the multiple logistic regression analysis were 0.74, 0.74, and 0.69 respectively. The mean AC, F1-score, and accuracy for the random forest classifier was 0.73, 0.76, and 0.69, respectively (Figure 4). Patient gender and age were not significantly associated with doubling time, separately or in combination with the radiomics features.

Figure 4.

ROC curve of multiple logistic regression and random forest classifier using the selected features with AUC of 0.74 and 0.73, respectively, in predicting the doubling time of renal tumours.

DISCUSSION

The present study demonstrated that radiomic features from baseline CT can predict the growth rate of ccRCC in patients with VHL. Wavelet-LLH_glcm was the best feature predictive of tumour doubling time as selected by multiple logistic analysis (AUC of 0.86), whereas log-sigma-0–5-mm-3D_glszm_ZonePercentage performed best in the random forest analysis (AUC: 0.79). Previous studies on the utility of radiomics in kidney cancers have investigated its role in differentiating benign from malignant, classifying RCC subtypes, low-grade from high-grade tumours, and treatment response ^8–12. The present study reports the utility of CT-derived radiomics in predicting the growth rate of renal tumours in VHL syndrome for the first time.

ccRCCs in patients with VHL can be bilateral, multifocal, and recurrent, thereby requiring lifetime screening for monitoring tumour growth until they are resected at 3 cm ^4,13–15. This imposes a complex clinical challenge of predicting the timing of resection and tailoring the frequency of imaging surveillance, which necessitates research on the potential utility of artificial intelligence and machine learning to address this clinical problem.

Anari et al. used radiomics obtained from MRI to predict growth rate in patients with VHL and found a mean AUC of 0.795 and the best feature was original_shape_major_axis_length, compared to the present study using CT, which demonstrates that higher-order features are more predictive of VDT⁶.The present findings demonstrate that two features were first order and the remaining eight were higher order, implying that specific parameters, unidentifiable by the human eye, are largely responsible for predicting the growth rate, especially when tumours are small. This is synonymous with the findings seen by Anari et al. on MRI, where the majority of features were of higher order ⁶, and a model utilising higher-order features alone was comparable in accuracy to one incorporating all features, suggesting that higher-order features alone are sufficient for prediction of growth kinetics ⁶.

The present study noted an overall mean volumetric tumour growth rate of 3 (2.3–3.8) cm³/year and a VDT of 897 (553–1242) days, whereas Farhadi et al. worked on diffusion-weighted MRI to report a mean growth rate of 4.5±3.2 cm³/year and a VDT of 614.2 ± 432.3 days in these patients ⁷. Among the comparison of slow- and fast-growing tumours, the mean volumetric growth rate for slow-growing tumours (high VDT) was found to be 2.1 cm³/year, as compared to Anari at el., which was 3.79 cm³/year, respectively. For fast-growing tumours (low VDT), mean volumetric growth rate in the present study was 4.9 cm³/year, and that of Anari at el. was 13.96 cm³/year ⁶. These differences in mean growth rates could be a result of the exponentially increasing size-dependent growth rate of clear cell carcinomas during the tumour’s evolution. The present study incorporated a baseline time point where tumours were small, whereas Anari et al. used pre-surgical stages of tumour for growth rate and radiomics evaluation (mean baseline volume: 1 versus 3.7 cm³ in low VDT and 6.1 versus 8.6 cm³ in high VDT tumours, respectively. Exponential growth rates of tumours in VHL syndrome has been reported previously by Schuhmacher et al. ¹⁶, who stated that tumours usually grow at a faster rate as they evolve. This concept is compatible with the Gompertzian model of tumour growth, which states that growth is exponential in early stages of tumour and then plateaus at later stages ¹⁷.

Past studies have used a unidimensional approach via maximal cross-sectional diameter for evaluating the growth kinetics of ccRCCs in VHL patients ^18–21. Zhang et al., Ball et al., Jilg et al., and Li et al. reported the median tumour growth rate in VHL as 0.445, 0.37, 0.41, and 0.43 cm/year, respectively, which in spite of being slightly higher than the present study (0.31 cm/year) fall in the range reported (0.17–0.51 cm/year) ^18–21. Other studies have investigated imaging-based characteristics and clinical/genetic information to influence the growth kinetics of ccRCCs, which may not take into account the inherent complexity of the tumours as that can be detected by the radiomics ^18,20,22,23.

The study has some limitations. There was an unpremeditated statistically significant difference in baseline volumes of low versus high VDT tumours (p<0.001), which might have been due to unequal sample size of these two groups. The study primarily focuses on patients with VHL syndrome having ccRCC, potentially introducing selection bias. The study sample is small owing to the rarity of VHL syndrome, which may restrict its utility as sporadic clear cell carcinomas form the bulk of kidney cancers; however, previous studies have found that certain mutations of VHL are also seen in sporadic cancers, possibly suggesting overlap in cancer pathophysiology. Therefore, the insights gained in this study may also be applicable to sporadic cases of ccRCC, given their common histological characteristics and further research can help validate the findings seen in this study in the cohort of sporadic clear cell carcinomas. Additionally, validation of the study’s findings on external data under different scanners and protocols can further establish its reliability and clinical use. The study’s methodology of using semi-automated segmentation from different readers could have introduced some inter-reader variability, thereby possibly influencing the results of the study; however, this effect was minimised by review and modifications of all segmented images from experienced body radiologists.

In conclusion, this study demonstrates that radiomics features derived from contrast-enhanced CT have a high degree of accuracy in predicting the growth rate of renal cancer in VHL patients. This is an important advancement in cancer diagnosis and treatment, as it allows for earlier and more precise identification of patients at risk of disease progression. The results of this study highlight the importance of incorporating these advanced imaging techniques into routine clinical practice and underscore the potential for further research to improve patient care. Future directions could focus on making an intrinsically trained and consistently reproducible radiomics model that predicts the doubling time of small renal tumours, allowing surveillance to be to tailored accordingly.

Highlights:

1. Median renal tumor doubling time in VHL syndrome was 538 (278–539).

2. Median diametric and volumetric growth rates were 0.31 (0.17–0.51) and 1.71 (0.8–4.2).

3. Radiomic features fairly predicted tumor growth patterns in VHL syndrome patients.

4. Models employed for feature selection: multiple logistic regression and random forest.

5. Potential of radiomics to improve outcomes by personalized surveillance strategies.

Abbreviations:

VHL

von Hippel-Lindau

ccRCC

clear cell Renal cell carcinoma

VDT

Volume doubling time

ROC

Receiver operating characteristics

AUC

Area under curve

GLCM

Gray Level Concurrence Matrix

GLSZ

Grayl Level Size Zone

GLDM

Gray Level Difference Matrix

GLRLM

Gray Level Run Length Matrix