Immuno-oncologic profiling of renal masses using multiparametric MRI: a pilot study

Abstract

Purpose

Preoperative characterization of renal mass malignancy, subtype, and immuno-oncologic pathology could inform and improve tailored management decisions. We examine multiparametric MRI (mpMRI) for (1) sensitivity to immuno-oncologic markers of the tumor immune microenvironment and (2) classification of tumor malignancy, subtype, and grade.

Methods and materials

In a prospective, institutional review board-approved single-center study, 40 patients (13 female/27 males, 60.4±10.7 years) scheduled to undergo surgical management of solid renal masses underwent preoperative 1.5T MRI. This included T1, multi-b-value diffusion-weighted imaging (DWI as intravoxel incoherent motion (IVIM), and apparent diffusion coefficient (ADC)), R2*, arterial spin labeling (ASL), and dynamic contrast-enhanced (DCE-)MRI. Clear cell likelihood scores (ccLS) were assigned using clinical MR images. Tumor diagnoses were extracted from the surgical histopathology. Logistic regression models were built with leave-one-out cross-validation and bootstrapping. Interobserver measurements were obtained in a subset of 27 patients, and tests–retests were run for 2 patients. Tumors from 18 patients with clear cell renal cell carcinoma (ccRCC) underwent immunohistochemistry for CD3, CD4, CD8, CD68, PD-L1, NKp46, HIF-1α, and CD31. MR biomarkers of immunohistochemistry stains were identified with Pearson’s r correlation, and diagnostic OR for >5% or >15% cells stained positive, by cross-validated univariate logistic regression.

Results

Of the 40 solid renal masses (mean (range)=32 (8–68) mm), 22 were clear cell, 9 non-clear cell and 9 benign, with 10 Grade 1. IVIM f, D*, and fD* correlated with T cells (CD4, CD3, CD8), while R2* correlated positively with macrophage presence (CD68) and negatively with angiogenesis (CD31). DCE-MRI and ASL negatively correlated with CD68. ASL negatively correlated with CD8 T cells. IVIM D* returned a significant OR for CD68-positive stains (OR=55.0, p<0.001), while ASL renal blood flow returned a significant OR for CD8-positive stains (OR=24.0, p=0.03). mpMRI tumor volume and IVIM D heterogeneity returned the highest $AUC\left(95\%CI\right)=0.84\left(\mathrm{0.70,0.98}\right)$) for malignancy. ccLS had the highest $AUC\left(95\%CI\right)=0.75(0.56,0.92)$ for overall detection of ccRCC, and mpMRI distinguished ccRCC from non-ccRCC with increased IVIM D and ADC with $AUC\left(95\%CI\right)=0.95\left(\mathrm{0.86,1.0}\right))$.

Conclusion

In this pilot study, mpMRI was sensitive to immuno-oncologic biomarkers supporting preoperative MRI as a method of characterizing tumor immune microenvironment, malignancy, and subtype for informed and tailored treatment management.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Immune system biomarkers in the tumor microenvironment influence tumor progression and are therapeutic targets. Programmed death-ligand 1 (PD-L1) expression, CD8+cytotoxic T cell, CD4+helper T cells and macrophages (CD68+) infiltration, and angiogenesis (CD31+), have demonstrated prognostic promise in renal cell carcinoma. Preoperative characterization of both tumor and tumor immune microenvironment with MRI has potential to inform advanced treatment decisions and assess eligibility for clinical trials.

WHAT THIS STUDY ADDS

• Multiparametric MRI (mpMRI) sequences intravoxel incoherent motion, blood oxygen level-dependent R2*, arterial spin labeling, and dynamic contrast-enhanced-MRI demonstrated sensitivity to T cells (CD4, CD3, CD8), macrophages (CD68), and angiogenesis (CD31), but not to NKp46 (natural killer cells) or HIF-1α (hypoxia), while clinical clear cell likelihood score successfully separated clear cell renal cell carcinomas from other solid renal mass subtypes.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Pending large-scale clinical validation, preoperative MRI and mpMRI biomarkers can improve tumor classification and tumor microenvironment characterization for tailored management decisions.

Introduction

Solid renal masses (SRMs) are commonly detected incidentally as a small lesion that could be either benign or malignant. As clinical factors alone have shown poor discrimination separating malignant from benign tumors, many patients with SRMs undergo surgical intervention prior to histopathologic diagnosis.¹ However, up to 27% of tumors are found to be benign at pathology after surgical resection,² and up to 20% of patients who undergo nephrectomy develop chronic kidney disease within 3 years postoperatively.^{3 4} While preoperative biopsies can be performed to avoid resection of benign masses, they are associated with procedural risks, and have reported a non-diagnostic rate of 19% and negative predictive value of 63%.⁵ In addition to clinical pathology, immune system biomarkers of renal cell carcinoma (RCC) may help predict tumor progression and immunotherapy response.⁶ However, current methods of immuno-oncologic profiling require resection, and may not be an option for patients with advanced disease. Preoperative characterization of SRM malignancy, subtype, and tumor immune microenvironment with MRI could help reduce unnecessary invasive treatment, tailor treatment management, improve active surveillance, and preserve long-term kidney function.

Clear cell renal cell carcinoma (ccRCC) is the most common renal malignancy, and the most aggressive subtype, with higher rates of recurrence, metastasis, and mortality compared with papillary-RCC and chromophobe-RCC.⁷ Consequently, a scoring system for ccRCC diagnosis using clinical MRI has been proposed, called the clear cell likelihood score (ccLS).^{8 9} In a retrospective review of 434 patients, ccLS has demonstrated sensitivity of 86% and specificity of 58% for renal masses <4 cm that are confined to the kidney.¹⁰ Further, ccRCC is an immunologically distinct tumor with a high response rate to immunotherapies despite low tumor mutational burden.^{11 12} As such, while not included in the ccLS scoring system, immune biomarkers have demonstrated promise and may improve prediction of RCC response to advanced immunotherapy treatments pending large-scale clinical validation.^{6 13 14} High CD8+cytotoxic T-cell infiltration has shown varied results,¹⁵ and double positive CD8+CD4+,¹⁶ angiogenesis (CD31+),^{17 18} and macrophages (CD68+)19,21 expression have demonstrated promise for prediction of survival and treatment response. Preoperative characterization of both tumor and tumor immune microenvironment with MRI has potential to inform advanced treatment decisions and assess eligibility for clinical trials.

Multiparametric MRI (mpMRI) is a comprehensive in vivo imaging method that captures tumor biology, physiology, and microenvironment volumetrically.²² This includes T1 (marker of tissue composition), blood oxygen level-dependent (BOLD) R2* (marker of hypoxia), diffusion-weighted imaging (DWI, intravoxel incoherent motion (IVIM) for inflammation/microvascular perfusion and apparent diffusion coefficient (ADC) for tissue cellularity), arterial spin labeling (ASL; blood flow) and dynamic contrast-enhanced MRI (DCE-MRI; vascular density and permeability). To examine preoperative mpMRI characterization of solid renal masses and RCC microenvironment, we test (1) mpMRI for sensitivity to immuno-oncologic markers of the tumor immune microenvironment and (2) mpMRI and ccLS for classification of SRM malignancy, tumor subtype, and tumor grade.

Methods and materials

Patients

This prospective, institutional review board-approved (#Study-18–00374 approved as of 5 May 2020, Health Insurance Portability and Accountability Act compliant, cross-sectional pilot study consists of adult patients with SRMs scheduled to undergo partial nephrectomy from October 2018 to April 2023 at the Icahn School of Medicine at Mount Sinai. Informed consent was obtained, and patients underwent preoperative MRI followed by histopathology (Clinical Flow Chart: online supplemental figure 1). Exclusion criteria were age <18 years, advanced stage (T3-T4) or metastatic renal masses, non-resectable renal masses, contraindications to MRI, or pre-existing medical conditions including claustrophobic reaction or seizure. Cystic masses were excluded.

Clinical features

Clinical and demographic features were collected including age, sex, race, ethnicity, body mass index, and non-tumorous renal volume. An abdominal radiologist (SL, 14 years of experience) measured the renal volumes, excluding the tumor, and tumor volumes from T1-weighted (T1-WI) images using segmentation software (Vitrea, Vital Images, Minneapolis, Minnesota, USA).

Clear cell likelihood scores

A second abdominal radiologist (GA, 3 years of experience), assigned ccLS V.2²³ using the clinical MRI sequences (T2-WI, T1-WI in-and-opposed phase imaging, and T1-WI pre/post contrast imaging) blinded to research MRI. Each mass was assigned a value along the 5-point Likert score (ccLS=1, very unlikely; ccLS=2, unlikely; ccLS=3, intermediate likelihood; ccLS=4, likely; and ccLS=5, very likely) to convey the likelihood of being ccRCC.²³

MRI acquisition and post-processing

Preoperative MRI was performed with a 1.5T MRI (Aera, Siemens Healthineers, Erlangen, Germany) using an 18-channel flexible body array and 32-channel integrated spine array coils. Standard of care clinical sequences acquired included coronal and axial T2-weighted Half-Fourier Acquisition Single-shot Turbo spin Echo, T1 in-and-opposed phase, and pre-contrast/post-contrast T1-WI Volumetric Interpolated Breath-hold Examination (VIBE). mpMRI protocol included multi-b-value two-dimensional (2D) echo planar DWI, 2D blood oxygen level-dependent (BOLD) T2* with 12 echoes, pseudo-continuous ASL (pCASL) with 1,500 ms post-labeling delay, and DCE-MRI based on coronal T1-WI 3D spoiled gradient-recalled echo sequence VIBE. Detailed MRI acquisition parameters are summarized in table 1a, with an example mpMRI of ccRCC and oncocytoma (figure 1) provided for illustration. All sequences were post-processed voxel-wise, excluding DCE-MRI which was volumetric. All post-processing was performed in MATLAB 2024b with details on fitting methods and models, as well as MR parameters of interest, summarized in table 1b.

Table 1. MR sequence acquisition, post-processing methods, and parameters.

a) Sequence parameters	T1	BOLD/R2*	Multi-b-value DWI	ASL	DCE-MRI
Sequence	3D T1 SPGR	2D T2* 12 Echoes	2D EPI 9-bval DWI 3-dir (b=0, 10, 30, 50, 80, 120, 200, 400, 800), 2-averages for b<400, 4-averages for b=400,800	TGSE pCASL TI=3,000 ms duration=1,500 ms flip angle=28 spoiler=5,000$\mu$ s	DCE-MRI VIBE gadobutrol (Gadavist, Bayer, 0.1 mmol/kg dose) with 40 mL saline flush rate: 2 mL/s post-sequence delay: 25 s
TR (ms)	4.29	311	4,700	5,000	4.53
TE (ms)	1.14	2, 18, 14, 20, 26, 32, 38, 44, 50, 56, 68, 79	81	25.84	1.06
Flip angle (deg)	2, 10	35	90	180	12
Acceleration factor	1	2	2	1	2
Scan time (min:sec)	0:21	0:23	4:04	5:05	8:58
Voxel dim. (mm3)	1×1×3	0.9 ×0.9×10	2.1×2.1×6	3.5×3.5×3.5	1×1×2.5
FOV (mm)	380	360	400	280	370×440
Fat Suppr.	None	Fat Sat.	SPAIR	Fat Sat.	Fat Sat.

b) Post-processing	T1	BOLD/R2*	Multi-b-value DWI	ASL	DCE-MRI
Equation model	$T1=-\frac{TR}{\ln \left(m\right)}$ from $\frac{S_n}{\sin \left(\theta \right)}=\ast$ $m\frac{S_n}{\tan \left(\theta \right)}+C$	$R{2}^{*}=\frac{1000}{T{2}^{*}}$ from ${\frac{S_n}{S_0}=e}^{-\frac{TE}{T{2}^{*}}}$	1)$A{e}^{-bADC}$ 2)${fe}^{-b{D}^{\mathrm{*}}}+\left(1-f\right)e^{-bD}$	$\begin{array}{l}\frac{S_{control}-S_{label}}{M_0}\\ \times \frac{6000\lambda e^{PLD/T{1}_{blood}}}{2\alpha T{1}_{blood}}\\ \times \frac{\left(1-e^{-\frac{TR}{T{1}_{tissue}}}\right)}{\left(1-e^{-\frac{PLD}{T{1}_{blood}}}\right)}\end{array}$	$\begin{array}{l}{C}_t\left(t\right)=v_p{C}_p\left(t\right)\\ +K_{trans}{\int }_0^t{C}_p\left(\tau \right)e^{-\frac{Ktrans}{v_e}\left(t-\tau \right)}d\tau \end{array}$
Fitting method	Slope from first order polynomial fit to the two VFA=2°,10° with a Vandermonde matrix	Non-linear least-squares estimation with trust region algorithm values normalized with starting value 50.	ADC as slope of the linear fit through logarithmic signal data of b=120–800 s/mm2 IVIM as Bayesian-log estimation with log D mean=−6.2±1 and log D* mean=−3.5±1 in mm2/s	Simplified single compartment model with $\begin{array}{l}\alpha =0.6,\lambda =0.9,\\ PLD=1500ms,\\ T{1}_{blood}=1480ms,\\ T{1}_{tissue}=1100ms,\\ TR=5000ms\end{array}$	Extended Tofts model, non-linear least-squares estimation with trust region algorithm, with a Gaussian weight to enhance the fitting’s weight on both the wash-in and wash-out phases.
Parameters (value)	T1 (tissue composition)	BOLD/R2* (blood oxygen lavel -dependent (BOLD)/R2* deoxyhemoglobin levels, and hypoxia)	ADC 2) IVIM diffusion (D), perfusion fraction (f), pseudo-diffusion coefficient (D*), and the product $f{D}^{*}$	RBF	vp (plasma volume fraction) ve (extravascular volume fraction) ktrans (volume transfer constant).

ADC, apparent diffusion coefficient; ADC, apparent diffusion coefficient; ASL, arterial spin labeling; BOLD, blood oxygen level-dependent; COR, Coronal; 2D, two-dimensional; 3D, three-dimensional; DCE, dynamic contrast-enhanced; DWI, diffusion weighted imaging; EPI, Echo Planar Imaging ; FOV, Field of View; IVIM, intravoxel incoherent motion; pCASL, pseudo-continuous ASL; RBF, renal blood flow; SPAIR, Spectral Adiabatic Inversion Recovery; SPGR, spoiled gradient-recalled echo sequence; TE, Echo Time; TGSE, Turbo Gradient Spin Echo; TI, Inversion Time; TR, Repetition Time; VFA, variable flip angles; VIBE, Volumetric Interpolated Breath-hold Examination.

Figure 1. (A–H) Example MRI images of a kidney with a 3.0 cm clear cell renal carcinoma mass. The sequences are (A) T1 (VFA=10), (B) BOLD/R2* (s⁻¹), (C) b0 DWI, (D) b800 DWI, (E) T2w, (F) ASL RBF (mL/min/100 g), (G) early phase DCE, and (H) delayed phase DCE. (I–P) Example MRI images of a kidney with a 3.2 cm oncocytoma. The sequences are (I) T1 (VFA=10), (J) BOLD/R2* (s⁻¹), (K) b0 DWI, (L) b800 DWI, (M) T2w, (N) ASL RBF (mL/min/100 g), (O) early phase DCE, and (P) delayed phase DCE. ASL, arterial spin labeling; BOLD, blood oxygen level-dependent; DCE, dynamic contrast-enhanced; DWI, diffusion-weighted imaging; RBF, renal blood flow; VFA, variable flip angle.

mpMRI features

Volumes of interest were placed encompassing the entire renal mass (SL, abdominal radiologist with 14 years of experience). A subset (n=27) of regions of interest (ROIs) was drawn (AMR, radiology resident with 1 year of experience) for interobserver reliability calculated with interclass correlation coefficient (ICC). Two patients had scans run two times for test–retest reliability of MR parameters of the kidney tissue calculated with coefficient of variation (CoV%). Volumes of interest for DCE-MRI analysis were drawn (XM, research assistant with 1 years of experience) under the guidance of SL. A subset (n=10) of ROIs was drawn (OB, MR physicist with 12 years of experience) for DCE-MRI interobserver agreement. To minimize contrast administrations, DCE-MRI test–retest reliability was not included. Voxel-wise mean, median, and SD were included as features for every mpMRI parameter excluding volumetric DCE.

Immuno-oncologic markers

Immuno-oncologic pathology was analyzed using chromogenic multiplex immunohistochemistry (IHC) in 18 patients (13 males/5 females, 59.4±12.4 years) with ccRCC whose resected tumor slides passed quality control. Analysis included CD3 (T cells), CD8 (cytotoxic T cells), CD4 (helper T cells), CD68 (macrophages), programmed death-ligand 1 (PD-L1) (immunotherapy target), NKp46 (natural killer immune cells), HIF-1$\alpha$ (hypoxia marker), and CD31 (angiogenesis marker). Example multiplex chromogenic analysis of these markers is provided in figure 2.

Figure 2. Example multiplex chromogenic analysis of immune-oncologic markers in clear cell renal carcinoma from figure 1A–H. The left column presents slide panels at 2.5 magnification. The right column shows 40× magnification of a selected region. Colored arrows that match the stain colors note the appearance of different markers. NK, natural killer; PD-L1, programmed death-ligand 1.

Paraffin-embedded tissues obtained from Biorepository and Pathology Core (Icahn School of Medicine at Mount Sinai) were cut into 3 microns slides and IHC was performed on central tumor slices using VENTANA DISCOVERY ULTRA from Roche (Roche Diagnostics, Roche Diagnostics Corporation, Indiana, USA). Slides were stained in three panels: tumor immunogenicity/immune response (CD3, CD8, CD68, PD-L1); tumor microenvironment (CD4, CD31, HIF-1$\alpha$), and natural killer cells (NKp46 and cytokeratin, for marking tumor cells). For chromogenic multiplex IHC, stains were analyzed as percent cells stained positive. Each primary marker was followed by the correspondent secondary antibody (DISCOVERY OmniMap anti-Rb HRP (RUO) Catalog # 760–4311, DISCOVERY Anti-Rabbit HQ Catalog # 760–4815, DISCOVERY Anti-Mouse HQ Catalog # 760–4814) and the signal was developed with a different color respectively (figure 2): DISCOVERY ChromoMap DAB kit (RUO) Catalog # 760–159 (brown: PD-L1, HIF-1$\alpha$, NK-p46;), DISCOVERY Purple kit (RUO) Catalog # 760–229 (purple: cytokeratin, CD68, CD31) and DISCOVERY Teal HRP Kit Catalog # 760–247 (teal: CD8, CD4), DISCOVERY yellow kit (RUO) Catalog # 760–239 (yellow: CD3). After each sequential staining, slides went through a step of inhibition, heat denaturation and neutralization. Tissues were counterstained with hematoxylin to visualize the nuclei.

Whole tissue sections on the slide were converted into high-resolution digital data using a NanoZoomer S210 Digital slide scanner (Hamamatsu). The HALO image analysis platform was used for quantitative tissue analysis (Indica Labs). Multiplex IHC algorithm and color deconvolution were used to separate chromogenic stains together with nuclei segmentation classifier to set-up the system for quantitative analysis.

Endpoints

Clinicopathologic diagnoses were extracted from the surgical pathology report including malignancy and subtype (malignant: clear cell, papillary, clear cell papillary, chromophobe, or chromophobe with oncocytoma features benign: angiomyolipoma (AML), or oncocytoma), and WHO-ISUP Grade (1–4). These were grouped into the following diagnostic classifications: (1) ccRCC, (2) “non-cc RCC”: papillary RCC, chromophobe RCC, chromophobe with oncocytoma features, and clear cell papillary, and (3) “other SRMs”: non-ccRCC, AML, and oncocytoma. Due to the distribution of tumor grades, grades were divided into Grade 1 versus Grade 2, and low Grade (1–2) versus high Grade (3–4). IHC was reported as percent stain positivity for direct correlation. As there is no current standard cut-off for binarizing IHC in ccRCC, these percent positivities were then dichotomized based on the observed data into “high” positivity and “low” positivity. This meant dichotomy by >15% if the maximum percent positivity observed among the 18 samples exceeded 20%, and by >5% if it did not. These dichotomized values were considered immunophenotypes.

Statistical analysis

Significant features were determined by Mann-Whitney U test with p<0.05 and Benjamini-Hochberg multiple comparisons correction (false discovery rate <0.35). This was chosen to minimize false negatives, while allowing for exploration of multiparametric sequences. Univariate logistic regression models were built for each of the significant histogram parameters with leave-one-out cross-validation. Multivariate logistic regression models were generated with mpMRI parameters and with combined mpMRI and ccLS with feature selection per loop to eliminate data leakage. Mean area under the curve (AUC) and 95% CI was calculated via bootstrapping, and AUCs were compared via the DeLong test. The diagnostic ability of a model was determined by an AUC with a 95% CI entirely above the null hypothesis of 0.50 (ie, AUC p<0.05). Sensitivity (SN) and specificity (SP) were reported at the Youden’s J-statistic cut-off threshold for each parameter. Correlation of mpMRI parameters against IHC biomarker percent positivity was calculated with Pearson’s r, Spearman’s rank and linear regression was used to test corresponding monotonic correlation. Diagnostic OR for detection of dichotomized immunophenotype was calculated for significant mpMRI parameters from leave-one-out cross-validated univariate logistic regression models at the Youden’s J-statistic probability threshold. If a patient had an incomplete MRI acquisition or a stain that failed quality control, it was excluded from that subanalysis. All statistical analysis and machine learning was performed in Python V.3.11.4 (Anaconda, 2024).

Results

Patient cohort

A total of 40 patients with SRM are included in this study meeting the goals of a cross-sectional pilot observational study. The 40 SRMs (mean (range)=32 (8–68) mm) had the following clinicopathologic diagnoses: 22 ccRCC, 9 non-ccRCC, 9 benign lesions, with n=10 Grade 1. Clinical demographics are included in table 2. Work including a subset of these patients has been published in one previous publication on radiomics of clinical MRI,²⁴ and is included in a separate concurrent study on chronic kidney disease, neither of which overlap with the analyses presented here.

Table 2. Patient demographics and clinical characteristics.

Demographics and clinical features
Sex (M/F)	27/13
Race
White/Caucasian	30 (28 Not Hispanic)
Black/African American	9 (8 Not Hispanic)
Other	1 (1 Not Hispanic)
Age (years; mean±SD, range)	59.0±12 (range 33–86)
Weight (kg; mean±SD, range)	87.8±21.5 (range 46.3–137.0)
BMI (kg/m2, mean±SD, range)	29.3±5.9 (range 18.1–43.8)
Renal volume (mL; mean±SD, range)	376±95 (range 206–643)
Renal mass volume (mL; mean±SD, range)	34.5±48.1 (range 1.1–251.4)
Renal mass subtype
Clear cell	22
Papillary	4
Chromophobe	1
Chromophobe with/oncocytoma	1
Clear cell papillary	3
AML	4
Oncocytoma	5
Malignancy (malignant/benign)	31/9
Grade
1	10
2	17
3	2
4	2
Tumor size on histopathology (cm)	32.9±17.4
Baseline eGFR (CKD-EPI 2021 mL/min/1.73 m2; mean±SD, range)	70.4±19.4 mL/min/1.73 m2 (range 37.3–130.5 mL/min/1.73 m2)
eGFR<60 (mL/min/1.73 m2; mean±SD, range)	N=12; 48.6±7.15 mL/min/1.73 m2 (range 37.3–59.85 mL/min/1.73 m2)
eGFR$\ge$ 60 (mL/min/1.73 m2; mean±SD, range)	N=28; 80.0±14.85 mL/min/1.73 m2 (range 60.8–130.55 mL/min/1.73 m2)

AML, angiomyolipoma; CKD, chronic kidney disease; CKD-EPI, Chronic Kidney Disease Epidemiology Collaboration; eGFR, Estimated Glomerular Filtration Rate.

Characterization of tumor immune microenvironment with mpMRI

Pearson’s r is plotted as a heatmap for every combination of functional MRI parameters and pathology (figure 3). Those with both significant Pearson’s r and significant Spearman’s rank are plotted in figure 4 with linear regression and 95% CI.

Figure 3. Pearson’s r values for MRI sequence parameters against immuno-oncologic per cent positive cells. Sequences are color-coded (IVIM: green, ADC: purple, BOLD/R2*: dark blue, T1: red, DCE: light blue, clinical parameters: orange, ASL: pink). The color bar range shows the range from perfect positive correlation (red) to perfect negative correlation (blue). Significance is shown as *p$\le 0.05$, **p$\le 0.01$, and ***p$\le 0.001$. Those that also had significant monotonic Spearman’s correlation are marked with $\diamond$ for Spearman’s p$\le 0.05$. ADC, apparent diffusion coefficient; ASL, arterial spin labeling; BOLD, blood oxygen level-dependent; DCE, dynamic contrast-enhanced; HIF, hypoxia; IVIM, intravoxel incoherent motion; NK, natural killer; PD-L1, programmed death-ligand 1.

Figure 4. Linear regression and CIs for mpMRI parameters with both significant (p<0.05) Pearson’s r and Spearman’s rank correlation against IHC per cent positive cells. If more than one central tendency parameter was significant, only the strongest correlation is plotted to minimize redundancy. Note that 6 of 18 CD4 IHC stains did not pass quality control, and 6 stained cases did not have successfully acquired ASL reducing the datapoints for CD4 and ASL. Boxplots of the mpMRI parameters split by the dichotomized immunophenotypes are shown with median, IQR, whiskers, and individual data points as scatter plot on top (CD68+ as %CD68>15%, CD4+ as %CD4>5%, and CD8+ as %CD8>15%). ASL, arterial spin labeling; IHC, immunohistochemistry; IVIM, intravoxel incoherent motion; mpMRI, multiparametric MRI.

T-cell expression

%CD4 (helper T cells), %CD3 (T cells), and %CD8 (cytotoxic T cells): %CD4 correlated positively with IVIM perfusion fraction $f$ SD and perfusion $f{D}^{*}$ (r=0.64 to 0.65, p=0.03) and both demonstrated significant monotonic correlation (figure 4). %CD3 also correlated positively with IVIM f and fD* (r=0.50 to 0.53, p=0.02–0.03). %CD8 correlated positively with R2* (r=0.71 to 0.73, p=0.001), and R2* SD (r=0.80, p=0.0005) as well as IVIM D* SD (r=0.58, p=0.02) (figure 3). Meanwhile, %CD8 correlated negatively with ASL renal blood flow (RBF) (r=−0.59 to −0.56, p=0.04). R2* SD and ASL demonstrated significant monotonic correlation (figure 4).

Immune system biomarkers

%CD68 (macrophages), PD-L1 (immunotherapy target), and natural killer cells (NKp46): %CD68 correlated positively with IVIM D* and R2* SD (r=0.49, p=0.04; r=0.53, p=0.02) while correlating negatively with DCE-MRI Ktrans, ve, and va (r=−0.49 to -0.53, p=0.02–0.04) as well as ASL mean and SD (r=−0.63, p=0.03; r=−0.71, p=0.01) (figure 3). IVIM D*, R2*, and ASL also demonstrated monotonic correlation against %CD68 (figure 4). %PD-L1 correlated positively with IVIM D*, R2*, and T1 SD (r=0.048 to 0.89, p<0.001–0.04) although all stains in this cohort had PD-L1<0.1%, so correlations must be interpreted with caution. NKp46 did not correlate with any parameters in this study.

Angiogenesis and hypoxia

Angiogenesis (%CD31) and hypoxia (%HIF-1$\alpha$): %CD31 correlated negatively with R2* (r=−0.66 to -0.67, p=0.02–0.03) (figure 3). HIF-1$\alpha$ expression did not correlate with any parameters in this study.

Dichotomized immunophenotype

%CD68, %CD4, and %CD8 were dichotomized at >15%, >15%, and >5% stained positivity, respectively (figure 4). CD68-positive immunophenotype (n=12) was detected with IVIM mean $D^{*}$ (OR=55.0, p<0.001), IVIM $D^{*}$ SD (OR=15.0, p=0.03), R2* SD (OR=6.0, p=0.25), and ASL RBF (OR=4.0, p=0.73). CD4-positive immunophenotype (n=6) was detected with IVIM $f$ SD (OR=21.0, p=0.07), and IVIM $f{D}^{*}$ (OR=10.0, p=0.26). Finally, CD8-positive immunophenotype (n=7) was detected with R2* SD (OR=7.5, p=0.32) and ASL RBF (OR=24.0, p=0.03).

Characterization of solid renal masses with clinical and demographic features

Tumor volume was significantly higher in malignant masses (37±18 vs 18±7 mm, p=0.007) with AUC$\left(95\%CI\right)=0.81\left[\mathrm{0.66,0.96}\right]$) and SN=0.65 and SP=0.88 at the cut-off threshold=0.610. Tumor volume was not significant for distinguishing ccRCC from other SRMs, or from other RCCs. Total kidney volume and clinical demographic features were not significant for any diagnostic classification in this work.

Characterization of solid renal masses with ccLS

ccLS was significantly higher in malignant tumors (p=0.017), but did not return a significant univariate model ($AUC\left(95\%CI\right)=0.70(0.45,\mathrm{}0.96)$; table 3a). Preoperative ccLS was significantly higher in ccRCC compared with other SRMs (p<0.001, $AUC\left(95\%CI\right)=0.75[0.56,0.92]$; table 3b). ccLS was also higher in ccRCC compared with non-ccRCC and returned a significant univariate model ($p=0.003,AUC\left(95\%CI\right)=0.75\left[\mathrm{0.52,0.98}\right]$; table 3c) but was not significant for WHO grade (table 3d).

Table 3. Renal mass classification. Shown here is a table of significant features determined by either a Mann-Whitney U test p<0.05 or an AUC with a 95%CI that is entirely above the null hypothesis (i.e. p<0.05).

Classification	Group 1 (n) mean±SD	Group 2 (n) Mean±SD	P value	AUC (95% CI)	Sens.	Spec.	Optimal cut-off
a) Malignant vs benign	Malignant (31)	Benign (9)
IVIM D SD (mm2/s)	345.7±133.25	263±61.1	0.061	0.72 (0.54 to 0.91)*	0.69	0.67	0.461
ADC SD (mm2/s)	421.4±115.2	339.8±66.1	0.035*	0.72 (0.53 to 0.90)*	0.83	0.56	0.391
cc likelihood score	4.26±0.84	2.78±1.55	0.017*	0.70 (0.45 to 0.96)	0.65	0.67	0.495
mpMRI model				0.84 (0.70 to 0.98)*	0.72	0.88	0.514

b) ccRCC vs other SRM	ccRCC (22)	Other SRM (18)
ADC SD (mm2/s)	435.4±124.5	360.8±73.4	0.023*	0.70 (0.53 to 0.87)*	0.65	0.71	0.514
ADC mean (mm2/s)	2001±401	1688±437	0.027*	0.69 (0.50 to 0.87)*	0.53	0.86	0.585
ADC median (mm2/s)	1980±435	1671±455	0.041*	0.63 (0.44 to 0.82)	0.65	0.67	0.482
cc likelihood score	4.59±0.58	3.11±1.29	<0.001*	0.75 (0.56 to 0.92)*	0.50	0.95	0.734
mpMRI model				0.70 (0.52 to 0.88)*	0.65	0.76	0.522
mpMRI+ccLS model				0.81 (0.59 to 1.00)*	0.63	0.95	0.668

c) ccRCC vs non-ccRCC	ccRCC (22)	Non-ccRCC (9)
IVIM D mean (mm2/s)	1855±373	1463±429	0.017*	0.77 (0.52 to 1.00)*	0.62	0.81	0.512
IVIM D median (mm2/s)	1861±396	1438±402	0.025*	0.75 (0.50 to 1.00)*	0.75	0.62	0.422
ADC mean (mm2/s)	2001±401	1493±209	0.003*	0.82 (0.61 to 1.00)*	0.62	0.90	0.686
ADC median (mm2/s)	1980±435	1495±349	0.007*	0.79 (0.56 to 1.00)*	0.62	0.81	0.623
cc likelihood score	4.59±0.58	3.44±0.83	0.003*	0.75 (0.52 to 0.98)*	0.56	0.64	0.508
mpMRI model				0.95 (0.86 to 1.00)*	0.88	0.95	0.683
mpMRI+ccLS model				0.93 (0.83 to 1.00)*	0.75	0.95	0.668

d) Grade 1 vs Grade 2	Grade 1 (10)	Grade 2 (17)
IVIM D mean (mm2/s)	1959±240	1654±491	0.027*	0.73 (0.52 to 0.95)*	0.67	0.90	0.556
IVIM D median (mm2/s)	1959±240	1637±494	0.031*	0.74 (0.53 to 0.94)*	0.60	0.90	0.573
ADC mean (mm2/s)	2143±321	1699±425	0.011*	0.77 (0.58 to 0.97)*	0.67	0.90	0.604
ADC median (mm2/s)	2143±356	1677±449	0.011*	0.78 (0.60 to 0.97)*	0.67	0.90	0.573
T1 mean (ms)	1709±530	1310±258	0.006*	0.81 (0.59 to 1.00)*	0.76	0.80	0.499
T1 median (ms)	1621±514	1238±290	0.008*	0.78 (0.56 to 0.99)*	0.82	0.70	0.449
mpMRI model				0.80 (0.59 to 1.00)*	0.80	0.80	0.538

e) Low grade vs high grade	Low Grade (27)	High Grade (4)
IVIM $f{D}^{\ast }$ mean (mm2/s)	74.9±40.8	110.0±47.5	0.164	0.73 (0.52 to 0.94)*	0.75	0.52	0.554
mpMRI model				0.79 (0.62 to 0.96)*	0.75	0.72	0.544

Each parameter is provided with the mean±SD of the two classifications, the Mann-Whitney U test p value, and individual cross-validated logistic regression model AUC with 95% CI, Sens., Spec., and Youden’s J-statistic optimal cut-off. A multiparametric model is generated for every classification using functional MRI, tumor volume, and ccLS with feature selection per loop to minimize data leakage. The mpMRI model was given all quantitative MR features to select from (eg, mean, median, SD from T1, ADC), while the mpMRI+ccLS was given the quantitative MRI features+ccLS to select from. All diffusion coefficients are in units of 10−6 mm2/s, while $f{D}^{\ast }$ is in 10−4 mm2/s. An asterisk (*) marks the significant U test p-values <0.05 and the significant AUC 95%CI ranges that are entirely above the null hypothesis 0.50.

ADC, apparent diffusion coefficient; AUC, area under the curve; cc, Clear Cell; ccLS, clear cell likelihood score; ccRCC, clear cell renal cell carcinoma; IVIM, intravoxel incoherent motion; mpMRI, multiparametric MRI; Sens, sensitivity; Spec, specificity; SRM, solid renal masse.

Characterization of solid renal masses with mpMRI

Malignant versus benign

Malignant tumors had higher IVIM D SD and ADC SD and significant univariate models (AUC=0.70–0.72). The central tendency values of ADC and IVIM D themselves were not significant (Mann-Whitney U test p=0.69–0.95). mpMRI demonstrated the highest AUC$\left(95\%CI\right)=0.84\left(\mathrm{0.70,0.98}\right)$ with SN=0.72 and SP=0.88 when tumor volume and IVIM D SD were combined (table 3a).

ccRCC versus other SRMs

ccRCCs had higher ADC SD than other SRMs and ADC SD alone returned the same AUC as mpMRI (AUC=0.70). Combining mpMRI and ccLS returned the highest $AUC\left(95\mathrm{\%}CI\right)=0.81\left(0.59,1.00\right)$ with mpMRI increasing the AUC from the ccLS-only model (AUC=0.75 up to 0.81, DeLong p=0.64) and improving sensitivity (SN=0.50 up to 0.63) (table 3b).

ccRCC versus non-ccRCC

ccRCC had higher IVIM D and ADC compared with non-ccRCC (AUC=0.75–0.82). The mpMRI model was a stronger predictor than the ccLS-only model $(AUC\left(95\%CI\right)=0.95\left[\mathrm{0.86,1.0}\right])$ and combining mpMRI and ccLS did not increase the mean AUC (table 3c).

Tumor grade

IVIM D, ADC, and T1 were higher in Grade 1 than Grade 2 lesions, and returned a significant mpMRI model $(AUC\left(95\%CI\right)=0.80[0.59,1.00]$). Separation of high and low-grade tumors was significant using elevated IVIM $f{D}^{*}$ alone ($AUC\left(95\%CI\right)=0.80[0.59,1.00]$) as well as an mpMRI model using elevated IVIM $f{D}^{*}$, elevated $f$, and increased $ADC$ heterogeneity ($AUC\left(95\%CI\right)=0.79[0.62,0.96]$) (table 3d–e).

Histopathologic classification with functional mpMRI

The functional MRI parameters of physiology such as IVIM perfusion (perfusion fraction f, pseudo-diffusion coefficient D*), and DCE-MRI (plasma volume fraction vp, extravascular volume fraction ve, and volume transfer constant Ktrans), ASL (renal blood flow, RBF), and BOLD/R2* (blood oxygen level dependent R2* for hypoxia) were not significant for renal mass histopathologic classification.

Interobserver correlation and test–retest variation

The ICC and CoV% for every MRI parameter are included in online supplemental table 1, although CoV% is limited as only two patients completed the test–retest scan. Interobserver agreement was excellent (ICC>0.75) for all non-contrast sequences, excluding T1 SD (ICC=0.23) and ADC SD (ICC=0.69). DCE-MRI returned moderate to excellent ICC (0.62–0.91). Central tendency of R2* and T1 returned the lowest CoV% (COV%=1.99–7.40) while ASL and DWI were higher (COV%=9.45–15.22). Central tendency of IVIM $f{D}^{*}$returned low CoV% and high ICC (COV%=2.07–2.26, ICC=0.95–0.99).

Discussion

Preoperative characterization of both tumor and tumor immune microenvironment with MRI has potential to inform advanced treatment decisions and assess eligibility for clinical trials. Functional MRI parameters that capture physiology demonstrated sensitivity to immunophenotypes while ccLS, ADC, and IVIM D characterized renal mass malignancy, subtype, and grade. Notably, functional MR parameters such as IVIM f and D*, blood oxygen level dependent (BOLD) R2*, ASL RBF, and DCE-MRI ve, vp, and Ktrans correlated with markers of T cells, macrophages, and angiogenesis while renal mass histopathology was best determined by ccLS, ADC, and T1. As grade, subtype, and immunophenotypes influence tumor growth and treatment response patterns,25,27 preoperative mpMRI may provide clinically relevant information regarding tumor histological subtype and tumor immune microenvironment, in vivo, prior to resection, allowing more precise patient-specific management.

Pending large-scale clinical validation, there are no current clinically accepted IHC immune biomarkers for RCC. However, non-invasive measures of immune microenvironment may be valuable for patient stratification for immunotherapy and antiangiogenic treatment as ccRCC is highly immune-infiltrated and has a high angiogenesis score.^{28 29} Preliminary results support immuno-oncologic biomarkers including CD4 (marker of immune-inflamed tumor associated with poor immunotherapy response),²¹ CD31 (marker of angiogenesis, suggesting longer disease-free survival),¹⁷ CD68 (marker of macrophages suggesting shorter progression-free survival),19,2130 and CD8 (marker of immune system activation suggesting tumor regression).³¹ Furthermore, recent evidence shows that change in tumor hypoxia and metabolism due to cellular proliferation may influence the dysregulated immune response in ccRCC and subsequently the immunotherapy response.^{28 32} Preoperative mpMRI of the entire tumor volume could detect hypoxia, perfusion, and cellular proliferation for complementary information on the tumor immune microenvironment.

In this current work, IVIM perfusion fraction $f$ and $f{D}^{*}$ showed positive correlation with %CD4 (helper T cells) and %CD3 (T cells) potentially due to dilated capillaries allowing tumor infiltration by T cells. In addition, CD4 can promote TGF$\beta$1 expression leading to increased cell proliferation.³³ As high cell proliferation requires angiogenesis and increased blood flow with high cellular metabolism and oxygen use, IVIM may detect increased, and more heterogeneous, perfusion with elevated $f$ and $f{D}^{*}$.²⁸ As such, elevated and heterogeneous perfusion detected by IVIM may signal higher %CD4 and %CD3, which indicate worse prognosis³⁴ and more advanced stage of disease.³² IVIM D*, R2*, and ASL heterogeneity were increased with increased macrophage expression (%CD68), a possible sign of heterogeneous tumor immune microenvironment,³⁵ imaged as spatially variable perfusion and hypoxia.^{28 36} This is different from angiogenesis, which may demonstrate elevated perfusion as opposed to just heterogeneity, and is another potential prognostic marker.³⁵ Hypoxia was also detected by DCE-MRI and ASL, which both negatively correlated with %CD68, possibly detecting damaged vasculature and hypoxia from the cellular proliferation exhausting the immune system.²⁸ R2* positively correlated with %CD8, detecting hypoxia resulting from T-cell proliferation and agreeing with negatively correlated ASL from reduced oxygenated blood supply.³⁷ R2*, which is elevated in hypoxic conditions, showed negative correlation with %CD31 (angiogenesis marker), with increased hypoxia consistent with reduced tumor angiogenesis.³⁸ The immunotherapy target PD-L1³⁹ correlated with IVIM D* SD, R2* SD, and T1 SD; higher tumor aggressiveness has also been found to correlate with heterogeneous⁴⁰ and increased PD-L1 expression.⁴¹ However, all stains in this cohort had PD-L1<0.1%, so PD-L1 correlations must be interpreted with caution.

Work on immuno-oncologic profiling of RCC has found that double positive CD8+CD4+ demonstrates prognostic promise,¹⁶ and macrophages (CD68+) predict poor survival outcomes of metastatic RCC.19,21 Functional mpMRI, including multi-b-value diffusion and arterial spin labeling perfusion,³⁶ demonstrates potential for non-invasive immuno-oncologic profiling, specifically %CD68, %CD4, and %CD8. Pre-resection IVIM D* predicted CD68-positive stains, while ASL RBF predicted CD8-positive stains. IVIM also returned a high OR for CD4-positive stains, but the small sample size may have limited the significance (p=0.07) and limited the ability to examine double positives. These findings support further research on mpMRI immunotherapy biomarkers to detect the physiologic differences of immunophenotypes and extract relevant information for treatment management teams.

Diffusion and clinical MRI demonstrated use for tumor subtype and malignancy of SRMs. Increased DWI heterogeneity was associated with malignancy,⁴² as is common due to heterogeneous genotype, varying molecular characteristics, irregular tumor angiogenesis, and vascular proliferation.^{28 36} DWI also returned higher IVIM and ADC diffusion coefficients in ccRCC compared with non-ccRCC and other SRMs. ccRCC is often hypervascularized and heterogeneous, with high lipid and glycogen-rich cytoplasm, all of which would support greater diffusion.^{2 43} In comparison, papillary RCCs have more densely packed papillary architecture with scant to moderate cytoplasm⁴⁴ and chromophobe RCCs are commonly solid, compact, with focal scarring and calcification.⁴⁵ This suggests higher diffusion coefficients in ccRCC reflect its histopathological features that favor the random motion of free water protons. For RCCs, ADC and IVIM D were higher in Grade 1 tumors than Grade 2 tumors, supporting a previous report of ADC negatively correlating against ccRCC grade.⁴⁶ For comparison between high-grade and low-grade tumors, as IVIM $f{D}^{*}$ and $f$ are markers of blood flow and blood volume, the elevated values in high-grade tumors may represent the increased angiogenesis and vascularization of higher grade RCCs.⁴⁶ Contrary to a previous study with 1.5T MRI, T1 was higher for low grade RCCs, although this difference may be due to the previous study using a modified look-locker inversion ecovery (MOLLI) sequence optimized for cardiac T1 mapping.⁴⁷ This work supports mpMRI containing information on immunophenotype, grade, malignancy, and subtype that may be complementary to ccLS and future clinical score models.

Interobserver reliability was high for R2* and IVIM parameters, and acceptable for ASL and T1 mapping. All test–retest covariance was acceptable and R2* and T1 demonstrated the lowest best test–retest covariance. This may be due to their relatively short scan times; the longer scans for multi-b-value DWI and ASL increase motion and respiratory artifacts, which vary between test–retest runs and influence post-processing. Improved denoising, motion correction, and higher signal-to-noise ratio at 3T MRI may further improve test–retest agreement.

There are several limitations to this study. This was a pilot study, meaning a larger sample size is needed to validate conclusions and test in clinically relevant applications. Not all MRI sequences could be acquired in each patient, which may have reduced the significance, specifically for ASL. Analysis of mpMRI parameter stability would also benefit from higher sample size of patients with test–retest scans in future studies. 3T MRI may improve signal-to-noise ratio for stronger correlation and predictive modeling of functional MRI. Further, highly immune-infiltrated tumors may have a different prognosis and response to therapy than angiogenic tumors²⁹ despite mpMRI detecting both as having faster diffusion; mpMRI for prediction of efficacy of immunotherapies and antiangiogenic drugs may benefit from advanced diffusion and perfusion MRI for improved characterization of tumor immunophenotypes and angiogenesis.2848,50 Immuno-oncologic profiling was restricted to the subset of patients with ccRCC, the most common RCC that undergoes immunotherapy, preventing analysis of biomarkers of varying subtypes; mpMRI compared with IHC in non-ccRCC needs further study. While stains were performed on central slices of resected tumors as a representative sample of the tumor, future work may benefit from placing fiducial markers (surgeon’s stitches) on resected tumors for more direct comparison. Due to the preliminary nature of the study, patient outcome is not addressed in this work.

Functional mpMRI demonstrated sensitivity to immuno-oncologic profiles, specifically correlating with T cells and macrophage presence, and angiogenesis, while mpMRI and ccLS characterized tumor malignancy, subtype, and grade. This work supports mpMRI as an imaging method for preoperative characterization of RCC tumor and tumor microenvironment. mpMRI may be of use in future large multicenter studies of localized and advanced renal masses to examine how mpMRI, ccLS, and immuno-oncologic profiling may improve patient stratification, reduce unnecessary invasive treatment, preserve long-term patient outcome, improve active surveillance, and inform patient-specific tailored treatment decisions.

Data availability statement

Data are available upon reasonable request.