ORIGINAL RESEARCH

Multiparametric MRI Radiomic Model for

Preoperative Predicting WHO/ISUP
Nuclear Grade of Clear Cell Renal Cell
Carcinoma
Qiong Li, MD,1,2# Yu-jia Liu, BS,3,4# Di Dong, PhD,3,4# Xu Bai, MS,2 Qing-bo Huang, MD,5
Ai-tao Guo, MD,6 Hui-yi Ye, MD,2 Jie Tian, PhD,4,7* and Hai-yi Wang, MD2*

Background: Nuclear grade is of importance for treatment selection and prognosis in patients with clear cell renal cell car-
cinoma (ccRCC).
Purpose: To develop and validate an MRI-based radiomic model for preoperative predicting WHO/ISUP nuclear grade in
ccRCC.
Study Type: Retrospective.
Population: In all, 379 patients with histologically confirmed ccRCC. Training cohort (n = 252) and validation cohort
(n = 127) were randomly assigned.
Field Strength/Sequence: Pretreatment 3.0T renal MRI. Imaging sequences were fat-suppressed T2WI, contrast-enhanced
T1WI, and diffusion weighted imaging.
Assessment: Three prediction models were developed using selected radiomic features, radiomic and clinicoradiologic
characteristics, and a model containing only clinicoradiologic characteristics. Receiver operating characteristic (ROC) curves
and area under the curve (AUC) were used to assess the predictive performance of these models in predicting high-grade
ccRCC.
Statistical Tests: The least absolute shrinkage and selection operator (LASSO) and minimum redundancy maximum rele-
vance (mRMR) method were used for the selection of radiomic features and clinicoradiologic characteristics, respectively.
Multivariable logistic regression analysis was used to develop the radiomic signature of radiomic features and
clinicoradiologic model of clinicoradiologic characteristics.
Results: The radiomic signature showed good performance in discriminating high-grade (grades 3 and 4) from low-
grade (grades 1 and 2) ccRCC, with sensitivity, specificity, and AUC of 77.3%, 80.0%, and 0.842, respectively, in the vali-
dation cohort. The radiomic model, combining radiomic signature and clinicoradiologic characteristics, displayed good
predictive ability for high-grade with sensitivity, specificity, and accuracy of 63.6%, 93.3%, and 88.2%, respectively, in
the validation cohort. The radiomic model showed a significantly better performance than the clinicoradiologic
model (P < 0.05).

View this article online at wileyonlinelibrary.com. DOI: 10.1002/jmri.27182

Received Mar 3, 2020, Accepted for publication Apr 17, 2020.

*Address reprint requests to: J.T., CAS Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, No. 95, Zhongguancun
East Road, Haidian District, Beijing 100191, China. E-mail: [email protected] or H.-y.W., Department of Radiology, First Medical Center, Chinese PLA General
Hospital, No. 28, Fuxing Road, Haidian District, Beijing 100853, China. E-mail: [email protected]
#
These authors contributed equally to this work.
Contract grant sponsor: National Natural Science Foundation of China; Contract grant numbers: 81971580, 91959130, 81971776, 81771924, 81930053; Contract
grant sponsor: National Key R&D Program of China; Contract grant numbers: 2017YFA0205200, 2017YFC1308700, 2017YFC1309100; Contract grant sponsor:
Beijing Natural Science Foundation; Contract grant number: L182061; Contract grant sponsor: Medical Big Data Research and Development Project of Chinese
PLA General Hospital; Contract grant number: 2018MBD-023; Contract grant sponsor: Youth Innovation Promotion Association CAS; Contract grant number:
2017175.

From the 1Department of Radiology, Tianjin Nankai Hospital (Tianjin Hospital of Integrated Traditional Chinese and Western Medicine), Tianjin, China;
2
Department of Radiology, First Medical Center, Chinese PLA General Hospital, Beijing, China; 3School of Artificial Intelligence, University of Chinese
Academy of Sciences, Beijing, China; 4CAS Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, Beijing, China;
5
Department of Urology, First Medical Center, Chinese PLA General Hospital, Beijing, China; 6Department of Pathology, First Medical Center, Chinese PLA
General Hospital, Beijing, China; and 7Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, School of Medicine, Beihang University,
Beijing, China
Additional supporting information may be found in the online version of this article

© 2020 International Society for Magnetic Resonance in Medicine 1

Journal of Magnetic Resonance Imaging

Data Conclusion: Multiparametric MRI-based radiomic model can predict WHO/ISUP grade in patients with ccRCC with
satisfying performance, and thus could help the physician to improve treatment decisions.
Level of Evidence: 3
Technical Efficacy Stage: 2
J. MAGN. RESON. IMAGING 2020.

C LEAR CELL RENAL CELL CARCINOMA (ccRCC) MRI on GE (Milwaukee, WI) 3.0T scanners, were included. The
is the most common subtype of renal cell carcinoma and initial query yielded a target population of 425 patients who were
accounts for ~90% of kidney tumors.1 The major risk factors considered eligible for inclusion in the study. Exclusive criteria
of renal cell carcinoma include smoking, obesity, and hyper- included: 1) patients who received previous treatment or experienced
postoperative recurrence of ccRCC; 2) MRI images with remarkable
tension.1,2 Tumor nuclear grade is one of the well-known
artifact (determined by three radiologists: Q.L., X.B., HY.Y.); 3)
prognostic factors of ccRCC and considered as an indepen-
WHO/ISUP nuclear grade was not available; 4) the diameter of the
dent predictor of cancer-specific survival.1,3 In 2016, the mass in an axial plane less than 1 cm; or 5) the time interval
WHO classification of tumors of the kidney, the WHO/ between MRI examination and following surgical resection was more
ISUP system, was recommended for ccRCC grading.4 Radical than 1 month.
operations are acceptable for high-grade ccRCC, while mini- The flowchart of the study population recruitment is shown
mally invasive techniques are more feasible management con- in Fig. 1. Finally, our study enrolled 379 patients (age range,
siderations for low-grade ccRCC, such as nephron-sparing 24–87 years old): 88 (23%) female and 291 (77%) male, with a
surgery, ablation, and even active surveillance.2 In order to median age of 54 years. In this study the tumor size ranged from
evaluate the malignant degree of tumor and make an optimal 1.23–14.87 cm. Of all tumors, 206 lesions were less than 4 cm,
treatment plan, preoperative pathological stratification is 119 lesions were between 4 cm and 7 cm, and 54 lesions were more
important. Percutaneous biopsy is of great help in confirming than 7 cm. The entire dataset was randomly assigned to a training
cohort (n = 252) and a validation cohort (n = 127) at a ratio of 2:1.
the grade of ccRCC; however, it is associated with a risk of
procedural complications, potential sampling errors, failure
to perform nuclear grading, and mismatch with pathology MRI Acquisition
outcome.5 MR sequences included axial fat-suppressed T2WI (weighted
Renal magnetic resonance imaging (MRI), a crucial tool images), diffusion-weighted imaging (DWI) (apparent diffusion
for preoperative diagnosis in ccRCC, facilitates noninvasive coefficient [ADC] calculation thereafter), and dynamic contrast-
enhanced scans. The details of the MRI equipment and MRI scan
assessment. Previous studies have found that some MRI fea-
protocol are shown in Supplementary A1 and Table S1.
tures contributed to preoperative grading of ccRCC, includ-
ing tumor size, hemorrhage, necrosis, enhancement degree,
and vein thrombosis.6–9 Histopathological Assessment of Nuclear Grade
Radiomics is an emerging and promising technique and WHO/ISUP nuclear grade was recorded from pathology reports of
histopathological examination in our institute. The specimens were
has been widely applied in the field of oncology. It converts
from the biopsy of one patient, radical nephrectomy of 127 patients,
medical images into high dimensional and quantitative image
and partial nephrectomy of 251 patients. A dedicated genitourinary
features by means of computer postprocessing technology. By
pathologist (A.T.G.) rechecked the WHO/ISUP nuclear grade of all
utilizing model-building algorithms, radiomic features can show the slides. These tumors were divided into low-grade (grades 1 and
associations with tumor histopathology and heterogeneity.10–14 2) and high-grade (grades 3 and 4).
Several studies have indicated that the increasing heterogeneity
on MRI reflected by radiomics was associated with the histo-
Clinicoradiologic Model Building
logical grade of cervical cancer and prostate cancer.15,16
The clinicoradiologic characteristics were collected from clinical risk
The purpose of this study was to develop and validate factors, tumor size, and subjective MRI scores. Clinical risk factors
an MRI-based radiomic model for preoperative prediction of included gender, age at diagnosis, body mass index (BMI), smoking
ccRCC WHO/ISUP nuclear grade. (former or current smoker) or not, and hypertension (previous or
current diagnosis of hypertension) or not. The tumor size was the
maximal diameter measured on the axial fat-suppressed T2WI image.
Materials and Methods The criteria for the subjective MRI scores is shown in Supplemen-
Patients tary A2 and Table S2.
Ethical approval was obtained from the Institutional Review Board Univariate analysis was performed to assess the potential asso-
of our institute for this retrospective study and the requirement for ciation between clinicoradiologic characteristics and nuclear grade.
informed consent was waived. The normality of all data was analyzed using the Shapiro–Wilk
Patients with pathologically proven ccRCC from January method. Differences in clinicoradiologic characteristics between low-
2016 to November 2017, with preoperative multiparametric renal and high-grade were assessed using the independent Student’s t-test

2
 Li et al.: Radiomic model for predicting ccRCC grade

FIGURE 1: Flowchart of study population recruitment.

or Mann–Whitney U-test for continuous variables, and Fisher’s MRI Radiomic Signature Building
exact test or χ2 test for categorical variables.
The minimum redundancy maximum relevance (mRMR) TUMOR SEGMENTATION AND RADIOMIC FEATURE
method17 was carried out for ranking the clinicoradiologic character- EXTRACTION. All MR images were exported from the picture
istics with mutual information. The clinicoradiologic model was archiving and communication systems (PACS) and then transferred
built using multivariable logistic regression on clinical characteristics to an independent workstation for manual segmentation using ITK-
in the training cohort. SNAP 3.6 software (www.itk-snap.org). For each tumor, an

FIGURE 2: Radiomic workflow in this study. (a) Tumor segmentation by radiologists. (b) Feature extraction from tumor region. (c)
Feature selection by ICC and LASSO method and clinicoradiologic characteristics by mRMR. (d) Model construction and validation.

3
Journal of Magnetic Resonance Imaging

abdominal radiologist (Q.L., with 10-year experience in abdominal

0.138
0.631
0.251
0.130
0.759
<0.05
MRI diagnosis) drew a free-hand 3D region of interest (ROI) on

P
each slice of the axial fat-suppressed T2WI, ADC map, and cor-

/
ticomedullary phase of contrast-enhanced T1WI (CE-T1WI). As
shown in Fig. 2, the ROI was drawn along the contour of the lesion,
and the top and bottom layers of the lesion were abandoned to avoid
the partial volume effect.
A total of 4224 radiomic features of the three sequences

25.09  4.21
high-grade

6.64  3.42
(1408 features per sequence) were extracted from the ROIs, included

18 M, 4 W
in four groups: 1) image intensity (first-order features); 2) shape and

59  11

12/22
10/22
size-based features; 3) textural features; and 4) wavelet features.18 All

22
features extraction methods were implemented using the

Validation cohort
Pyradiomics package (https://pyradiomics.readthedocs.io/en/latest/)
based on Python (v. 3.6.5, https://www.python.org/).

INTER- AND INTRAOBSERVER RADIOMIC FEATURE

26.22  3.68
81 M, 24 W
EXTRACTION REPRODUCIBILITY. Fifty patients were randomly

4.22  2.40
low-grade
selected for intraobserver and interobserver reproducibility test. The

55  12

39/105
44/105
radiologist (Q.L.) and another abdominal radiologist (X.B., with

105
5-year experience in abdominal MRI diagnosis) outlined the ROI
again 30 days after the initial segmentation. The intraclass correla-
tion coefficient (ICC) was used to assess the agreement of extracted
features by intraobserver and interobserver segmentations. An ICC
greater than 0.75 was considered in great agreement and highly
robust.19

<0.05
0.924
0.422
0.218
0.087
<0.05
P
/
RADIOMIC FEATURE SELECTION AND RADIOMIC
TABLE 1. Clinical Characteristics of Patients in the Training and Validation Cohorts

SIGNATURE BUILDING. Radiomic signature building was per-

formed in the training cohort. The feature selection and signature
building processes were separately performed on each sequence. The
steps were as follows: first, filtering out the features with ICCs ≤0.75
25.15  3.26
31 M, 10 W
high-grade

6.03  2.84
in the intraobserver or interobserver test; second, refining the most
59  11

useful prognostic features using the least absolute shrinkage and

14/41
21/41
selection operator (LASSO) method; finally, using logistic regression
41

of the selected features weighted by their coefficients to build a sig-

Training cohort

nature. After those steps, three signatures of three sequences (fat-

suppressed T2WI, CE-T1WI, and ADC) were separately built. Then
a radiomic signature was generated via the linear combination of the
three signatures.
161 M, 50 W
25.59  3.21

4.04  2.16
low-grade

BMI = body mass index; M = men; W = women.

COMBINED MODEL, MODEL VALIDATION, AND

53  11

94/211
78/211

NOMOGRAM CONSTRUCTION. Multivariable logistic regres-

211

sion analysis was used to develop a combined radiomic model using

radiomic signature and clinicoradiologic characteristics. Model per-
formance was validated in terms of the receiver operating characteris-
tic (ROC) curve and area under the curve (AUC). Sensitivity,
specificity, positive predictive value (PPV), negative predictive value
(NPV), and accuracy were also calculated. The Delong test was used
History of smoking

to explore whether the radiomic model performed better than the

clinicoradiologic model and radiomic signature. Additionally, the
Hypertension
Characteristic

nomogram of the best model was established for visible use. The cali-
Tumor size
Age (years)
Number

bration curves were applied to modify and reduce the bias of the
Gender

final model. Also, stratified analysis was applied on a different MR

BMI

scanner in the three models to figure out the difference between

scanners in predicting efficiency.

4
 Li et al.: Radiomic model for predicting ccRCC grade

Statistical Analysis
The statistical analyses were performed with R software (v. 3.5.3; R
Foundation for Statistical Computing, Vienna, Austria) and Python
software (v. 3.6.0). The tests were two-sided and P < 0.05 was con-
sidered statistically significant.

Results
Clinical Characteristics
The clinical characteristics of patients in the training and
validation cohorts are shown in Table 1. The training cohort
included 252 patients (192 male and 60 female; median
age, 54 years; age range, 27–86 years), and 127 patients
were included in the validation cohort (99 male and
28 female; median age, 55 years; age range, 24–87 years).
Except for age (P < 0.05) and tumor size (P < 0.05), there
was no significant difference in gender (P = 0.924), BMI
(P = 0.422), history of smoking (P = 0.218), and hyperten-
sion (P = 0.087) between the two groups in the training FIGURE 3: Intraclass correlation coefficient (ICC) analysis.
cohort. In the validation cohort, except for tumor size
(P < 0.05), no differences were found between the two factors of high-grade ccRCC, and thus they were used to
groups in terms of age (P = 0.138), gender (P = 0.631), build the clinicoradiologic model (Table 2).
BMI (P = 0.251), history of smoking (P = 0.130), and The clinicoradiologic model showed good performance of
hypertension (P = 0.759). ccRCC grading, which reached an AUC of 0.821 (95% confi-
dence interval [CI], 0.768–0.867) in the training cohort, and
Predictive Performance of the Clinicoradiologic AUC of 0.777 (95% CI, 0.695–0.846) in the validation cohort.
Model
There were significant differences between low-grade and high- Radiomic Feature Selection and Radiomic Model
grade ccRCC in the training cohort with age, tumor size, subjec- Building
tive MRI score of pseudocapsule, shape and margin, hemorrhage, According to the standard of the ICC >0.75 in the
hypervascularity, intratumoral neovascularity, peritumoral intraobserver and interobserver tests, 1368 radiomic features
neovascularity, cystic-solid, vein thrombosis, lymphadenopa- from T2WI images, 1332 features from CE-T1WI images,
thy, necrosis, and perinephric invasion (Table S3). and 1061 features from the ADC map were highly robust
The mRMR selected the top five clinicoradiologic char- and selected for further analysis (Fig. 3). LASSO and logistic
acteristics correlated with nuclear grade, including shape and regression revealed two features from T2WI to construct the
margin, vein thrombosis, lymphadenopathy, necrosis, and T2WI signature, three key features from CE-T1WI to build
perinephric invasion. In multivariable logistic regression anal- the CE-T1WI signature, and seven features from ADC
ysis, only shape and margin and necrosis were independent images to build the ADC signature. The AUCs of the T2WI,

TABLE 2. Multivariate Logistic Regression Analysis of the Clinicoradiologic Model

β OR 95% CI p
Intercept 4.253
Shape and margin −1.192 0.304 0.145–0.635 <0.05
Vein thrombosis −0.593 0.553 0.196–1.558 0.262
Lymphadenopathy −0.864 0.421 0.137–1.298 0.132
Necrosis −0.755 0.470 0.286–0.772 <0.05
Perinephric invasion −0.402 0.669 0.231–1.936 0.459

CI = confidence interval; OR = odds ratio.

5
Journal of Magnetic Resonance Imaging

TABLE 3. Radiomic Features Selection From the MRI in the Training Cohort

Feature selected Group P-value

T2WI “exponential_gldm_LargeDependenceEmphasis” texture <0.05
“square_glrlm_RunPercentage” texture <0.05
CE-T1WI “original_shape_Sphericity” shape <0.05
“wavelet.HLL_glcm_Imc2” texture <0.05
“lbp.2D_glszm_SizeZoneNonUniformity” texture <0.05
ADC “original_firstorder_Mean” intensity <0.05
“original_firstorder_Median” intensity <0.05
“wavelet.HHL_glszm_LargeAreaLowGrayLevelEmphasis” texture <0.05
“exponential_glcm_Imc2” texture <0.05
“exponential_glrlm_ShortRunLowGrayLevelEmphasis” texture <0.05
“exponential_ngtdm_Busyness” texture <0.05
“square_firstorder_Median” intensity <0.05

ADC = apparent diffusion coefficient; CE-T1WI = contrast-enhanced T1WI; MRI = magnetic resonance imaging.
The description of the radiomic features mentioned above is displayed in the Supplementary A3.

CE-T1WI, and ADC signatures were 0.794 (95% CI, of high-grade ccRCC, which reached an AUC of 0.873 (95%
0.761–0.926), 0.804 (95% CI, 0.636–0.861), and 0.850 CI, 0.826–0.912) in the training cohort, and an AUC of 0.845
(95% CI, 0.639–0.864), respectively, in the training cohort. (95% CI, 0.770–0.903) in the validation cohort. The ROC
The three signatures were chosen for radiomic signature con- curves of the radiomic signature, clinicoradiologic model, and
struction (Table 3). radiomic model in the training and validation cohorts are shown
The AUC of the radiomic signature was 0.860 (95% in Fig. 4.
CI, 0.811–0.900) in the training cohort and 0.842 (95% CI, The performance of all models in the validation cohort
0.767–0.901) in the validation cohort. is shown in Table 4. The AUC of the radiomic model was
significantly better than that of the clinicoradiologic model
Construction of a Combined Radiomic Model and (Z = 2.960, P < 0.05). However, there was no significant dif-
Nomogram ference of AUC between the radiomic model and the radio-
The combined radiomic model of radiomic signature and mic signature (Z = 0.184, P = 0.854), or between the
clinicoradiologic characteristics showed a prediction performance radiomic signature and clinicoradiologic model (Z = 1.740,

FIGURE 4: Comparison of the AUC of radiomic signature, clinicoradiologic model, and radiomic model in the training and validation
cohorts. (a) ROC curves of three models in training cohort. (b) ROC curves of three models in the validation cohort.

6
 Li et al.: Radiomic model for predicting ccRCC grade

TABLE 4. Predictive Performance of the Three Models for Preoperative ccRCC Grading in the Validation Cohort

Model AUC [95% CI] Sensitivity (%) Specificity (%) PPV (%) NPV (%) Accuracy (%)
Clinicoradiologic 0.777 [0.695–0.846] 68.2 (15/22) 84.8 (89/105) 48.4 (15/31) 92.7 (89/96) 81.9 (104/127)
model
Radiomic 0.842 [0.767–0.901] 77.3 (17/22) 80.0 (84/105) 44.7 (17/38) 94.4 (84/89) 79.5 (101/127)
signature
Radiomic model 0.845 [0.770–0.903] 63.6 (14/22) 93.3 (98/105) 66.7 (14/21) 92.5 (98/106) 88.2 (112/127)

AUC = area under the curve; ccRCC = clear cell renal cell carcinoma; CI = confidence interval; NPV = negative predictive value;
PPV = positive predictive value.

P = 0.082). Compared with the other two models, the radio- Discussion
mic model yielded the highest specificity and accuracy. In this retrospective study we developed and validated a mul-
The nomogram of radiomic model is illustrated in tiparametric MRI-based radiomic model for noninvasively
Fig. 5a, which could facilitate nuclear grading in clinical prac- predicting high-grade ccRCC.
tice (Fig. 6). Figure 5b,c demonstrates the calibration curves With recent advances in imaging technique and post-
of the radiomic model, which indicated that there was a good processing analysis, MRI can serve as a diagnostic, therapeu-
calibration of the prediction result of the radiomic model and tic, and prognostic biomarker for renal cell carcinoma.20
the real result. The stratified analysis demonstrated that there However, the MRI findings of low-grade and high-grade
was no obvious difference in predicting high-grade ccRCC ccRCC overlap in some cases. Moreover, previous studies are
between the two scanners (Supplementary A4 and Fig. S1). not completely consistent on which MRI features are valuable

FIGURE 5: Radiomic nomogram and its calibration curves. (a) Nomogram of radiomic model for prediction of high-grade in patients
with ccRCC. (b) Calibration curve of the nomogram in the training cohort. (c) Calibration curve of the nomogram in the validation
cohort.

7
Journal of Magnetic Resonance Imaging

FIGURE 6: Two cases with ccRCC. a–d: Preoperative MRI in a 73-year-old man with ccRCC of WHO/ISUP grade 2 in the right kidney.
Considering the radiomic signature of 0.049, shape and margin of 1, and necrosis of 0, the risk of high-grade ccRCC was smaller
than 0.1 according to the nomogram. (a) Axial fat-suppressed T2WI showed a round-shaped, well-defined, solid mass. (b) The lesion
showed similar signal compared with renal parenchyma on the ADC map. (c) On corticomedullary phase of contrast-enhanced T1WI,
the lesion showed significant enhancement. (d) The tumor cells were relatively small, with clear cytoplasm, small to moderate
nucleus with round shape, small nucleolus (hematoxylin–eosin staining, 40× magnification). e–h: Preoperative MRI in a 52-year-old
man with ccRCC of WHO/ISUP grade 3 in the left kidney. Considering the radiomic signature of 0.586, shape and margin of 1, and
necrosis of 2, the risk of high-grade ccRCC was 0.6 according to the nomogram. (e) Axial fat-suppressed T2WI showed a round-
shaped, well-defined, solid mass with central necrosis. (f) The lesion showed lower signal than renal parenchyma on the ADC map.
(g) Compared with the renal cortex, the lesion showed slight enhancement with intratumoral vessel on corticomedullary phase of
contrast-enhanced T1WI. (h) The tumor cells were large, with clear or eosinophilic cytoplasm and obvious nucleolus (hematoxylin–
eosin staining, 10× magnification).

for nuclear grading.6,8 Therefore, it is challenging to use mor- regular, and a round shape was often observed. The high-
phological evaluation of MRI for nuclear grading in ccRCC. grade ccRCC lesions tend to be aggressive, with an irregular
Fortunately, radiomics can reveal the subtle differences of the shape and ill-defined margin, with the presence of necrosis
intensity distribution in medical images, which cannot be easily and perinephric fat invasion or distant metastasis. Necrosis
recognized by human eyes, with the pattern of signal distribu- has been demonstrated to be an independent predictor of
tion reflecting the heterogeneity of tumors.21 Previous studies high-grade ccRCC in previous studies.6,8 Coy et al reported
have investigated the potential of computed tomography (CT)- that high-grade ccRCC lesions showed ill-defined tumor mar-
based radiomic to predict Fuhrman nuclear grade of gins.27 Wei et al showed that the tumor shape differed signifi-
ccRCC.22,23 However, there are some limitations to using cantly between low-grade and high-grade ccRCC.28 Two
CT. First, the Fuhrman grading system has poor interobserver MRI characteristics (shape and margin, necrosis) extracted in
reproducibility.24 Second, in contrast to CT, MRI provides this study were consistent with previous studies,8,27,28
multiple forms of soft-tissue contrast, as well as functional reflecting the features of invasive growth and uneven distribu-
parameters and permits a comprehensive evaluation of tion of blood supply in high-grade ccRCC.
ccRCC.25 Recently, a study using MRI and radiomics to pre- By revealing the heterogeneity of tumors, the radiomic
dict nuclear grade in ccRCC patients demonstrated that the pre- signature has been more and more used to predict the degree
dictive accuracy of their radiomic signature (74%) was lower of malignancy. In our study, the radiomic signature from MRI
than ours (88.2%).26 We think that the ADC data and wavelet showed excellent performance in predicting high-grade ccRCC.
analysis used in our method improve predictive accuracy. The 12 radiomic features in the final radiomic signature
In this study the clinicoradiologic model performed well included two features from T2WI, three features from CE-
in predicting high-grade ccRCC. In multivariate analysis of T1WI, and seven features from the ADC map. Most of these
clinical risk factors and subjective MRI scores, only shape and radiomic features were obtained from exponential or wavelet fil-
margin and necrosis were independent predictors of high- tered images, which were high-dimensional features and could
grade ccRCC. The margin of low-grade ccRCC lesions was not be easily deciphered by humans. The results suggest that

8
 Li et al.: Radiomic model for predicting ccRCC grade

combining multiple MRI sequences allowed the collection of organs—Part a: Renal, penile, and testicular tumours. Eur Urol 2016;70:
93-105.
more valuable information of the tumor than a single sequence.
5. Leveridge MJ, Finelli A, Kachura JR, et al. Outcomes of small renal
T2WI could provide morphological features of ccRCC, includ- mass needle core biopsy, nondiagnostic percutaneous biopsy, and the
ing pseudocapsule, cystic components, and necrosis, which are role of repeat biopsy. Eur Urol 2011;60:578-584.
often seen in high-grade ccRCC.6 ccRCC commonly exhibits 6. Pedrosa I, Chou MT, Ngo L, et al. MR classification of renal masses with
hypervascularity, and CE-T1WI can show the enhancement pathologic correlation. Eur Radiol 2008;18:365-375.

degree and neovascularization within the tumor.29 DWI could 7. Parada Villavicencio C, Mc Carthy RJ, Miller FH. Can diffusion-weighted
magnetic resonance imaging of clear cell renal carcinoma predict low
provide excellent tissue contrast based on the molecular diffu- from high nuclear grade tumors. Abdom Radiol 2017;42:1241-1249.
sion of water within tumors, and the reported ADC has been 8. Rosenkrantz AB, Niver BE, Fitzgerald EF, Babb JS, Chandarana H,
associated with the degree of malignancy in tumors.8 Melamed J. Utility of the apparent diffusion coefficient for dis-
tinguishing clear cell renal cell carcinoma of low and high nuclear
The radiomic model combined with radiomic signature
grade. AJR Am J Roentgenol 2010;195:W344-W351.
and clinicoradiologic characteristics significantly improved the
9. Cornelis F, Tricaud E, Lasserre AS, et al. Multiparametric magnetic reso-
predictive performance in contrast to the clinicoradiologic nance imaging for the differentiation of low and high grade clear cell
model. A meta-analysis of the percutaneous biopsy revealed renal carcinoma. Eur Radiol 2015;25:24-31.

that the median concordance rate between grading on biopsy 10. Dong D, Tang L, Li ZY, et al. Development and validation of an individ-
ualized nomogram to identify occult peritoneal metastasis in patients
and the surgical specimen was 87% using the two-tier (high with advanced gastric cancer. Ann Oncol 2019;30:431-438.
vs. low) grading system,30 while our radiomic model achieved
11. Peng H, Dong D, Fang MJ, et al. Prognostic value of deep learning
a similar diagnostic accuracy of 88.2% with noninvasive pre- PET/CT-based radiomics: Potential role for future individual induction
operative MRI. Therefore, the radiomic model may eventu- chemotherapy in advanced nasopharyngeal carcinoma. Clin Cancer
Res 2019;25:4271-4279.
ally be a supplement to percutaneous biopsy.
12. Song J, Shi J, Dong D, et al. A new approach to predict progression-
free survival in stage IV EGFR-mutant NSCLC patients with EGFR-TKI
Limitations therapy. Clin Cancer Res 2018;24:3583-3592.
The present study had several limitations. First, our model 13. Zhang B, Tian J, Dong D, et al. Radiomics features of multiparametric
was trained and validated using retrospective data obtained MRI as novel prognostic factors in advanced nasopharyngeal carci-
noma. Clin Cancer Res 2017;23:4259-4269.
from a single institution, which could yield an unintentional
14. Dong D, Zhang F, Zhong LZ, et al. Development and validation of a
selection bias. A large-scale prospective and multicenter study novel MR imaging predictor of response to induction chemotherapy in
is needed to validate our results. Second, in this study the locoregionally advanced nasopharyngeal cancer: A randomized con-
trolled trial substudy (NCT01245959). BMC Med 2019;17:190.
diameter of lesions in many cases was more than 4 cm.
Future studies are needed to further verify the value of the 15. Liu Y, Zhang Y, Cheng R, et al. Radiomics analysis of apparent diffusion
coefficient in cervical cancer: A preliminary study on histological grade
radiomic model for small ccRCC, which is of great impor- evaluation. J Magn Reson Imaging 2019;49:280-290.
tance to urologists. For a radiomics study, our numbers were 16. Niu XK, Chen ZF, Chen L, Li J, Peng T, Li X. Clinical application of
on the small side. biparametric MRI texture analysis for detection and evaluation of high-
grade prostate cancer in zone-specific regions. AJR Am J Roentgenol
2018;210:549-556.
Conclusion 17. Peng H, Long F, Ding C. Feature selection based on mutual informa-
The radiomic signature from multiparameter MRI can distin- tion: Criteria of max-dependency, max-relevance, and min-redundancy.
IEEE Trans Pattern Anal Mach Intell 2005;27:1226-1238.
guish high-grade ccRCC from low-grade ccRCC with good
18. Vallieres M, Zwanenburg A, Badic B, Cheze Le Rest C, Visvikis D,
performance. The radiomic model combining a radiomic sig- Hatt M. Responsible radiomics research for faster clinical translation.
nature and clinicoradiologic characteristics could improve the J Nucl Med 2018;59:189-193.
predictive performance of high-grade ccRCC, and thus can 19. Ma Z, Fang M, Huang Y, et al. CT-based radiomics signature for differ-
entiating Borrmann type IV gastric cancer from primary gastric lym-
provide physicians noninvasively preoperative characterization
phoma. Eur J Radiol 2017;91:142-147.
of ccRCC.
20. Wu Y, Kwon YS, Labib M, Foran DJ, Singer EA. Magnetic resonance
imaging as a biomarker for renal cell carcinoma. Dis Markers 2015;
2015:648495.
References
21. Liu Z, Wang S, Dong D, et al. The applications of radiomics in precision
1. Motzer RJ, Jonasch E, Agarwal N, et al. Kidney cancer, version 2.2017,
diagnosis and treatment of oncology: Opportunities and challenges.
NCCN clinical practice guidelines in oncology. J Natl Compr Canc
Theranostics 2019;9:1303-1322.
Netw 2017;15:804-834.
22. Shu J, Tang Y, Cui J, et al. Clear cell renal cell carcinoma: CT-based
2. Escudier B, Porta C, Schmidinger M, et al. Renal cell carcinoma: ESMO radiomics features for the prediction of Fuhrman grade. Eur J Radiol
clinical practice guidelines for diagnosis, treatment and follow-up. Ann 2018;109:8-12.
Oncol 2016;27:v58-v68.
23. Bektas CT, Kocak B, Yardimci AH, et al. Clear cell renal cell carcinoma:
3. Dagher J, Delahunt B, Rioux-Leclercq N, et al. Clear cell renal cell carci- Machine learning-based quantitative computed tomography texture
noma: Validation of World Health Organization/International Society of analysis for prediction of Fuhrman nuclear grade. Eur Radiol 2019;29:
Urological Pathology grading. Histopathology 2017;71:918-925. 1153-1163.

4. Moch H, Cubilla AL, Humphrey PA, Reuter VE, Ulbright TM. The 2016 24. Delahunt B, Eble JN, Egevad L, Samaratunga H. Grading of renal cell
WHO classification of tumours of the urinary system and male genital carcinoma. Histopathology 2019;74:4-17.

9
Journal of Magnetic Resonance Imaging

25. Sasaguri K, Takahashi N. CT and MR imaging for solid renal mass char- 28. Wei J, Zhao J, Zhang X, et al. Analysis of dual energy spectral CT and
acterization. Eur J Radiol 2018;99:40-54. pathological grading of clear cell renal cell carcinoma (ccRCC). PLoS
One 2018;13:e0195699.
26. Cui E, Li Z, Ma C, et al. Predicting the ISUP grade of clear cell renal cell
carcinoma with– multiparametric MR and multiphase CT radiomics. Eur 29. Rosenkrantz AB, Wehrli NE, Melamed J, Taneja SS, Shaikh MB. Renal
Radiol 2020;30:2912-2921. masses measuring under 2 cm: Pathologic outcomes and associations
with MRI features. Eur J Radiol 2014;83:1311-1316.
27. Coy H, Young JR, Douek ML, et al. Association of qualitative and quan-
titative imaging features on multiphasic multidetector CT with tumor 30. Marconi L, Dabestani S, Lam TB, et al. Systematic review and meta-
grade in clear cell renal cell carcinoma. Abdom Radiol 2019;44: analysis of diagnostic accuracy of percutaneous renal tumour biopsy.
180-189. Eur Urol 2016;69:660-673.

10