Jpn J Radiol (2017) 35:472–483

DOI 10.1007/s11604-017-0655-6

ORIGINAL ARTICLE

A predictive diagnostic model using multiparametric MRI

for differentiating uterine carcinosarcoma from carcinoma of the
uterine corpus
Yuki Kamishima1 · Mitsuru Takeuchi2 · Tatsuya Kawai1 · Takatsune Kawaguchi3 · Ken Yamaguchi4 ·
Naoki Takahashi5 · Masato Ito6 · Toshinao Arakawa7 · Akiko Yamamoto8 · Kazushi Suzuki1 · Masaki Ogawa1 ·
Moe Takeuchi9 · Yuta Shibamoto1

Received: 15 February 2017 / Accepted: 19 May 2017 / Published online: 5 June 2017
© Japan Radiological Society 2017

Abstract defect area volume percentage, and degree of enhancement.

Purpose To construct a diagnostic model for differentiat- Two predictive models—with and without contrast—were
ing carcinosarcoma from carcinoma of the uterus. constructed using multivariate logistic regression analysis.
Materials and methods Twenty-six patients with carcino- Fifteen other patients with carcinosarcomas and 30 patients
sarcomas and 26 with uterine corpus carcinomas consti- with carcinomas constituted a validation cohort. The sen-
tuted a derivation cohort. The following nine MRI features sitivity and specificity of each model for the validation
of the tumors were evaluated: inhomogeneity, predominant cohort were calculated.
signal intensity, presence of hyper- and hypointense areas, Results Inhomogeneity, predominant signal intensity on
conspicuity of tumor margin, cervical canal extension on T2WI, and presence of hyperintense areas on T1WI were
T2WI, presence of hyperintense areas on T1WI, contrast significant predictors in the unenhanced-MRI-based model.
Presence of hyperintensity on T1WI, contrast defect area
volume percentage, and degree of enhancement were sig-
* Mitsuru Takeuchi
nificant predictors in the enhanced-MRI-based model. The
m2rbimn@gmail.com
sensitivity/specificity of unenhanced MRI were 87/73 and
1
Department of Radiology, Nagoya City University 87/70% according to reviewer 1 and 2, respectively. The
Graduate School of Medical Sciences and Medical School, sensitivity/specificity of the enhanced-MRI-based model
1 Kawasumi Mizuho‑cho, Mizuho‑ku, Nagoya 467‑8601,
were 87/70% according to both reviewers.
Japan
2
Conclusions Our diagnostic models can differentiate carci-
Department of Radiology, Radiolonet Tokai, 3‑86‑2
nosarcoma from carcinoma of the uterus with high sensitiv-
Asaoka‑cho, Chikusa‑ku, Nagoya 464‑0811, Japan
3
ity and moderate specificity.
Department of Radiology, Kariya Toyota General Hospital,
5‑15 Sumiyoshi‑cho, Kariya 448‑8505, Japan
4
Keywords Magnetic resonance imaging · Uterus ·
Department of Radiology, Faculty of Medicine, Saga
Endometrial neoplasms · Carcinosarcoma · Carcinoma
University, 5‑1‑1 Nabeshima, Saga 849‑8501, Japan
5
Department of Radiology, Mayo Clinic, 200 First Street SW,
Rochester, MN 55905, USA
6
Introduction
Department of Radiology, Japanese Red Cross Nagoya Daini
Hospital, 2‑9, Myoken‑cho, Showa‑ku, Nagoya 466‑8650,
Japan Carcinosarcoma of the uterus (formerly known as a malig-
7 nant mixed Mullerian tumor) is defined as a tumor that has
Department of Radiology, Okazaki City Hospital,
3‑1 Gosyoai, Kouryuji‑cho, Okazaki 444‑8553, Japan a malignant epithelial component and a malignant mes-
8 enchymal component [1]. It has distinctive clinical and
Department of Radiology, Aichi Cancer Center Aichi
Hospital, 18 Kuriyado, Kake‑machi, Okazaki 444‑0011, pathologic features that warrant their separation from endo-
Japan metrial carcinoma. Carcinosarcoma has a high incidence
9
Department of Radiology, Nagoya City East Medical Center, of lymphatic spread, peritoneal seeding, and a higher rate
1‑2‑23 Wakamizu, Chikusa‑ku, Nagoya 464‑8547, Japan of pulmonary metastases, so it is associated with a poorer

13
Jpn J Radiol (2017) 35:472–483 473

prognosis than carcinoma of the uterus [2, 3]. In contrast
to the treatment of uterine corpus carcinoma, primary treat-
ment of uterine carcinosarcoma includes total hysterectomy
and bilateral salpingo-oophorectomy with surgical staging,
peritoneal lavage for cytology, omental and peritoneal biop-
sies, and maximal tumor debulking for gross disease, even
if the lesion cannot be curatively resected [4]. The choice
of anticarcinogenic drugs is different for carcinosarcoma
and carcinoma, and hormonal treatment is not applied to
carcinosarcoma of the uterus [4, 5]. Clinically, carcino-
sarcoma often presents with symptoms similar to those of
endometrial adenocarcinoma, with vaginal bleeding and
abdominal distention as well as uterine enlargement. Pre-
operative diagnosis of uterine carcinosarcoma is suggested
by imaging and is made by endometrial sampling. How-
ever, carcinosarcoma is often diagnosed after hysterectomy
because the sensitivity of endometrial sampling for identi-
fying carcinosarcoma is insufficient (23.5–58.8%) [6, 7].
Several researchers have investigated imaging findings
for carcinosarcoma of the uterus [8–15], but only one of
them has addressed the diagnostic accuracy of imaging
diagnosis for differentiating carcinosarcoma from carci-
noma [14]. Consequently, no precise imaging diagnostic
method has been established to date. The purpose of the
study reported in the present paper was to construct a pre-
dictive diagnostic model for differentiating uterine carcino-
sarcoma and carcinoma using multiparametric MRI.
proven carcinoma of the uterine corpus before September
2013 in our hospital and had the same complete data sets
as the patients in the carcinosarcoma cohort. Patient age,
tumor size and tumor stage were not matched between
the carcinosarcoma and carcinoma groups. Thus, a total
of 52 patients were included in the derivation cohort in
order to construct a predictive diagnostic model. The
stages in the two groups are shown in Table 1.
Validation cohort
From January 2001 to September 2014, 15 patients with
pathologically proven carcinosarcoma of the uterus who
underwent preoperative MRI with the same complete
data sets as in the derivation cohort (11 consecutive cases
from Saga University Hospital, 4 consecutive cases from
Mayo Clinic) constituted the carcinosarcoma group.
Patients who were pathologically diagnosed with carci-
noma of the uterine corpus and underwent MRI with the
same complete data sets as in the carcinosarcoma group
constituted the carcinoma group. Tumor sizes of the car-
cinoma group were matched with those in the carcino-
sarcoma group in a 1:2 ratio. Thus, 30 patients (22 case
from Saga University Hospital and 8 cases from Mayo
Clinic) with carcinoma of the uterine corpus constituted
the carcinoma group used to evaluate the diagnostic per-
formance of the predictive diagnostic model.

MRI protocol
Materials and methods
MRI scanners, acquisition parameters, and contrast
This study was observational and retrospective, and was injection protocols were not uniform due to the nature
approved by the institutional review board of our hospital. of ­retrospective multi-institutional joint research. Axial
Patient informed consent was waived. or sagittal T1WI, sagittal T2WI, and axial or sagittal
contrast-enhanced T1WI with or without fat suppres-
Patient population sion in the venous phase were obtained for all derivation
and validation cohorts. The scan parameters are shown
Derivation cohort
Table 1  Tumor stages of carcinosarcoma and carcinoma of the
From April 2004 to September 2013, 26 patients (aged uterus
52–85 years; mean 68 years) with pathologically proven
Stage Derivation cohort Validation cohort
carcinosarcoma of the uterus who underwent preop-
erative MRI with axial or sagittal T1-weighted images Carcinosarcoma Carcinoma Carcinosarcoma Carcinoma
(T1WI), sagittal T2-weighted images (T2WI), and axial IA 4 10 6 7
or sagittal contrast-enhanced T1WI constituted the carci- IB 9 9 1 7
nosarcoma group. Of the 26 patients, 11, 8, 3, 2, 1, and II 2 3 1 3
1 were from Nagoya City University Hospital, Japanese IIIA 1 1 2 4
Red Cross Nagoya Daini Hospital, Kariya Toyota Gen- IIIB 0 1 0 0
eral Hospital, Toyokawa City Hospital, Okazaki City
IIIC1 4 0 3 1
Hospital and Nagoya City East Medical Center, respec-
IIIC2 4 1 0 5
tively. The carcinoma group consisted of the last 26 con-
IVA 1 0 0 0
secutive patients (aged 40–85 years; mean 62 years) to
IVB 1 1 2 3
undergo preoperative MRI imaging for pathologically

13
474 Jpn J Radiol (2017) 35:472–483

Table 2  Imaging parameters

Sequence Number of cases Repetition time (ms) Echo time (ms) Flip angle (°) Slice thickness (mm)

T2-weighted image Fast spin echo 97 1700–10,941 80–124 90–180 4–8

T1-weighted image Fast spin echo 87 417–743 8–18 90–180 4–8
3D gradient echo 7 4.4–5.0 1.9–2.3 15 2–3
2D gradient echo 3 133–225 4–5 60–80 3–5
Contrast-enhanced image Fast spin echo 81 390–904 8–18 90–180 4–8
3D gradient echo 11 3.7–5.1 1.5–2.3 6–15 2–3
2D gradient echo 5 105–225 3–5 60–80 3–5

in Table 2. All data were reviewed on a commercial pic-
ture archiving and communication system (EV Insite,
PSP Corporation, Tokyo, Japan).
Image interpretation
In the derivation cohort, 52 cases (26 carcinosarcomas and
26 carcinomas) were independently interpreted in random
order by two radiologists (T. Kawai and T. Kawaguchi),
both with 9 years of experience. The reviewers were blinded
to pathological diagnosis or any clinical information. Nine
MRI features of the tumors were visually evaluated using
an original grading system as follows: inhomogeneity on
T2WI (0: homogeneous, 1: slightly heterogeneous, 2: het-
erogeneous) [9, 11–13], predominant signal intensity (SI)
on T2WI (0: hypointense, 1: intermediate, 2: hyperintense)
[10–13], markedly hyperintense area in tumor on T2WI
(0: absent, 1: present) [15], markedly hypointense area in
tumor on T2WI (0: absent, 1: present), extension to the
uterine cervix with dilation of the cervical canal (0: absent,
1: present) [15, 16], conspicuity of tumor margin on T2WI
(0: well-defined, 1: intermediate, 2: ill-defined) [11], hyper-
intense areas in the tumors on T1WI (0: absent, 1: present)
[9, 11–13, 15], contrast defect area volume percentage in
tumor (0: 0–5%, 1: 6–25%, 2: 26–50%, 3: more than 50%)
[10, 11, 15], and degree of tumor enhancement (0: low, 1:
intermediate, 2: high) [9, 11–13, 15] (Table 3). All planes
in which the tumor was depicted were interpreted for each
MRI feature. An area with hyperintensity similar to the SI
of the urine was regarded as a markedly hyperintense area
on T2WI. An area with hypointensity similar to the SI of
the muscle was regarded as a markedly hypointense area
on T2WI. An area with a SI higher than that of the myo-
metrium was regarded as a hyperintense area on T1WI to

Table 3  Univariate analysis of patient age and MRI findings for the carcinosarcoma and carcinoma groups
Carcinosarcoma Carcinoma Kappa value Univariate analysis
(n = 26) (n = 26)

Age Mean 68 (52–85) Mean 61 (40–85) p = 0.039*

Tumor diameter Mean 77 (36–205) Mean 42 (14–97) p = 0.001*
ET/AP ratio Mean 0.68 Mean 0.52 p < 0.001*
(0.51–0.97) (0.12–0.90)
Score Number of cases Kappa value Univariate analysis
0 1 2 3 0 1 2 3

Inhomogeneity on T2WI 0 9 17 6 17 3 0.84 [0.70–0.97] p < 0.001*

Predominant signal intensity on T2WI 3 17 6 13 13 0.64 [0.44–0.84] p = 0.002*
Markedly hyperintense area on T2WI 11 15 20 6 0.84 [0.69–0.99] p = 0.023*
Markedly hypointense area on T2WI 7 19 18 8 0.77 [0.59–0.94] p = 0.005*
Extension to the uterine cervix with dilation of cervical canal 16 10 20 6 0.90 [0.77–1] p = 0.37
Conspicuity of tumor margin on T2WI 3 11 12 8 12 6 0.38 [0.18–0.59] p = 0.12
Hypreintense area on T1WI 10 16 23 3 0.75 [0.56–0.94] p < 0.001*
Contrast defect area 6 14 3 3 22 4 0 0 0.77 [0.62–0.92] p < 0.001*
Degree of tumor enhancement 1 9 16 9 16 1 0.58 [0.39–0.77] p < 0.001*

95% confidence interval for the data is shown in parentheses

* Significant difference between carcinosarcoma and carcinoma groups

13
Jpn J Radiol (2017) 35:472–483 475

Fig. 1a–c  71-year-old woman with carcinosarcoma of the uterus. a ▸

A sagittal T2-weighted image shows a large intrauterine mass with
exophytic growth (arrow). The mass shows predominantly heteroge-
neous hyperintense signal intensity. b An axial T1-weighted image
shows a hyperintense area (arrow) in the mass. c On a contrast-
enhanced T1-weighted image with fat suppression, there is an area
in the mass showing strong enhancement similar to the myometrium
(arrow). There is a contrast defect area (arrowhead) spanning 10% of
the tumor volume when all planes are evaluated. Thus, this case was
rated as 5 points in total by the simplified unenhanced-MRI-based
model and 4 points in total by the simplified enhanced-MRI-based
model, and was correctly diagnosed with carcinosarcoma

identify a subtle hemorrhage in the tumor. In the evalua-

tion of degree of tumor enhancement, only the delayed
enhanced phase was evaluated. Compared with the contrast
enhancement of the myometrium, similar or higher, slightly
lower, and obviously lower enhancement were considered
high, intermediate, and low enhancement, respectively.
Differences between the evaluations were resolved by con-
sensus. Another radiologist with 11 years of experience
(Mitsuru Takeuchi) measured the largest diameter of the
tumor, the largest anteroposterior dimensions of the uterus,
and the thickness of the endometrium including the tumor
using sagittal T2WI. The ratio of the thickness of the endo-
metrium to the largest anteroposterior dimensions of the
uterus (ET/AP ratio) was calculated for each case accord-
ing to Genever’s report [14].
In the validation cohort, 45 cases (15 carcinosarcomas
and 30 carcinomas) were interpreted in random order by
two reviewers (T. Kawai with 9 years of experience and Y.
K. with 5 years of experience) independently, without any
knowledge of pathological diagnosis or clinical informa-
tion. The reviewers focused on the MRI features identified
by the multivariate logistic regression model analysis in the
derivation cohort. The same visual grading scale system
used for the derivation cohort was employed to evaluate
the MRI features in the validation session. Y. K. measured
the largest anteroposterior dimensions of the uterus and the
thickness of the endometrium including the tumor using
sagittal T2WI. The ET/AP ratio was calculated for each
case. Y. K. also measured the largest tumor diameter.

Statistical analysis

Statistical analyses were performed using the following sta-

tistical software: JMP version 11.0.0 (SAS Institute; Cary,
NC, USA) and MedCalc version 14 (MedCalc Software
bvba, Ostend, Belgium).
In the derivation cohort, interobserver agreement was
calculated using kappa statistics. Kappa scores of 0.41–
0.60, 0.61–0.80, and >0.80 were regarded as indicating significance using Fisher’s exact test. Differences between
moderate, good, and excellent agreement, respectively continuous data were tested for significance using an inde-
[17]. Differences between categorical data were tested for pendent t test. A logistic regression model was constructed

13
476 Jpn J Radiol (2017) 35:472–483

Fig. 2a–c  59-year-old woman with endometrial carcinoma of the ▸

uterus. a A sagittal T2-weighted image shows a large intrauterine
mass (arrow) with invasion into the myometrium (arrowhead). The
mass shows homogeneous hypointensity. The ratio of the largest
anteroposterior dimensions of the uterus to the thickness of the endo-
metrium (ET/AP ratio) is 0.78 (35/45 mm). b There is no hyperin-
tense area in the mass on the axial T1-weighted image (arrow). c
On a contrast-enhanced T1-weighted image with fat suppression,
the tumor is enhanced much less than the myometrium, without any
contrast defect areas (arrow). This case was misdiagnosed as carci-
nosarcoma by Genever’s criteria, but was correctly diagnosed by both
the simplified unenhanced- and enhanced-MRI-based models, with a
total score of 0 assigned by both reviewers

to gauge the associations of age, longest tumor diameter,

ET/AP ratio, and seven unenhanced MRI features with
carcinosarcoma (unenhanced-MRI-based model). Another
logistic regression model was constructed to gauge the
associations of age, longest tumor diameter, ET/AP ratio,
and seven unenhanced and two enhanced MRI features
with carcinosarcoma (enhanced-MRI-based model). The
modeling was initially performed such that all independ-
ent variables were entered as covariates. Stepwise selec-
tion was then used until only variables with p values for
the likelihood ratio test of <0.05 were retained in the final
model. These two predictive diagnostic models were sim-
plified to facilitate the utilization of the models in clinical
situations. The area under the curve for the receiver opera-
tor characteristic (AUC), the sensitivity, the specificity, and
the accuracy of the simplified unenhanced- and enhanced-
MRI-based models for differentiating carcinosarcoma
and carcinoma were calculated. Cutoff values for total
scores in each predictive diagnostic model were estab-
lished by calculating the maximal Youden index (Youden
index = sensitivity + specificity − 1) [18]. The sensitivity,
specificity, and accuracy of Genever’s method for diagnos-
ing carcinosarcoma of the uterus were calculated. An ET/
AP of >0.63 was considered to be a positive finding of car-
cinosarcoma [14].
In the validation cohort, the sensitivity, specificity, and
AUC of each simplified unenhanced-MRI-based model and
simplified enhanced-MRI-based model for diagnosing car-
cinosarcoma were calculated for all 45 cases (15 carcino-
sarcomas and 30 carcinomas). The exact binominal test was
used to compare the sensitivities and specificities of the
two predictive diagnostic models and Genever’s method. A
p value of <0.05 was considered statistically significant for
a single comparison of the carcinosarcoma and carcinoma
groups. A p value of <0.016 (0.05/3) which was adjusted
by Bonferroni correction was taken to indicate significance
when comparing the two predictive diagnostic models and
Genever’s method.

13
 Jpn J Radiol (2017) 35:472–483
Table 4  Diagnostic performance of each predictive model constructed for the derivation cohort (n = 52)
Model Score Regression coefficient p value Sensitivity Specificity Accuracy AUC
0 1 2 3

Simplified unenhanced-MRI-based 88% (23/26) 88% (23/26) 88% (46/52) 0.93 [0.82–0.98]
model
Inhomogeneity on T2WI Homogeneous Intermediate Heterogeneous 2.4 0.004
Predominant signal intensity on T2WI Hypointense Intermediate Hyperintense 2.1 0.010
Hyperintense area on T1WI Absent Present 2.7 0.013
Simplified enhanced-MRI-based model 96%* (25/26) 81% (21/26) 88%† (46/52) 0.95 [0.85–0.99]
Hyperintense area on T1WI Absent Present 2.6 0.021
Contrast defect area <5% 5–25% 26–50% >50% 2.4 0.001
Degree of tumor enhancement Hypovascular Intermediate Hypervascular 2.7 0.006
Genever’s method 62%* (16/26) 73% (19/26) 67%† (35/52) NA

Each value in parentheses is the number of cases

Values in square brackets indicate the 95% confidence interval
AUC area under the receiver operating characteristic curve, NA not available
p < 0.016 was considered to indicate a statistically significant difference when comparing the three predictive diagnostic models in terms of diagnostic performance
* Significant difference in sensitivity between the simplified enhanced-MRI-based model and Genever’s method (p = 0.012)
†
Significant difference in accuracy between the simplified enhanced-MRI-based model and Genever’s method (p = 0.013)
13

477
478 Jpn J Radiol (2017) 35:472–483

Fig. 3a–c  85-year-old woman with carcinosarcoma of the uterus. a A ▸

sagittal T2-weighted image shows a large intrauterine mass with exo-
phytic growth (arrow). The mass shows predominantly heterogeneous
intermediate signal intensity. The ratio of the largest anteroposterior
dimensions of the uterus to the thickness of the endometrium (ET/
AP ratio) is 0.56 (37/66 mm). b An axial T1-weighted image shows
a hyperintense area (arrow) in the mass. c On a contrast-enhanced
T1-weighted image with fat suppression, there is an area in the mass
showing stronger enhancement than the myometrium (arrow). There
was a contrast defect area (arrowhead) spanning 10% of the tumor
volume when all planes were evaluated. Thus, this case was rated 4
points in total by the simplified unenhanced-MRI-based model and 4
points in total by the simplified enhanced-MRI-based model, and was
correctly diagnosed with carcinosarcoma

Results

Derivation cohort

Detailed results from the qualitative and quantitative evalu-

ations of the derivation cohort are shown in Table 3. The
interobserver agreement regarding the conspicuity of the
tumor margin on T2WI was low. The interobserver agree-
ment on the degree of tumor enhancement was moderate.
The interobserver agreement for the other MRI features
was either good or excellent. Univariate analysis demon-
strated that carcinosarcomas were more heterogeneous on
T2WI (p < 0.001), and the predominant SIs of carcinosar-
comas were higher than those of carcinomas (p = 0.002)
(Figs. 1, 2). Markedly hyperintense areas on T2WI, mark-
edly hypointense areas on T2WI, and hyperintense areas
on T1WI were more frequently seen in carcinosarcomas
(p = 0.023, 0.005, and <0.001, respectively) (Figs. 1, 2).
The intratumoral contrast defect area volume percent-
age was larger in carcinosarcomas than in carcinomas
(p < 0.001). Carcinosarcoma was enhanced more strongly
than carcinoma (p < 0.001) (Figs. 1, 2). The largest diam-
eter and ET/AP ratio of carcinosarcomas were larger than
those of carcinomas (p = 0.001 and p < 0.001, respec-
tively). The mean age of patients with carcinosarcomas was
higher than that of carcinoma patients (p < 0.039). There
was no significant difference in the presence of extension
to the uterine cervix with dilation of the cervical canal and
conspicuity of the tumor margin on T2WI between carcino-
sarcoma and carcinoma (p = 0.37 and 0.12, respectively).
In multivariate logistic regression analysis, inhomogene-
ity on T2WI, predominant SIs on T2WI, and presence of
hyperintense areas in tumors on T1WI were significant pre-
dictors of carcinosarcoma in the unenhanced-MRI-based
model (Table 4). When contrast-enhanced MRI features

13
Jpn J Radiol (2017) 35:472–483
were included, presence of hyperintense areas in tumors
on T1WI, larger intratumoral contrast defect area volume
percentage, and stronger tumor enhancement were signifi-
cant predictors of carcinosarcoma (Table 4). In the multi-
variate logistic regression analysis, there were only small
differences among the regression coefficients of the vari-
ables in both unenhanced- and enhanced-MRI-based mod-
els; thus, the contributions of the variables were considered
to be the same in each of the simplified unenhanced- and
simplified enhanced-MRI-based models. Therefore, the
total predictive score was calculated by adding the scores
for each variable, and this total score ranged from 0 to 5
in the simplified unenhanced-MRI-based model and from
0 to 6 in the simplified enhanced-MRI-based model. A total
score of ≥3 was considered to indicate carcinosarcoma in
the unenhanced-MRI-based model. A total score of ≥2 was
considered to indicate carcinosarcoma in the enhanced-
MRI-based model.
Diagnostic performance of the simplified predictive
diagnostic models
The sensitivity, specificity, and accuracy of the simplified
unenhanced-MRI-based model for diagnosing carcinosar-
coma were 88% (23/26), 88% (23/26), and 88% (46/52),
respectively. The AUC for the unenhanced-MRI-based
model was 0.93 [95% confidence interval (CI), 0.82–0.98]
(Table 4). The sensitivity, specificity, and accuracy of the
simplified enhanced-MRI-based model for diagnosing car-
cinosarcoma were 96% (25/26), 81% (21/26), and 88%
(46/52), respectively. The AUC for the enhanced-MRI-
based model was 0.95 (95% CI, 0.85–0.99) (Table 4). The
sensitivity, specificity, and accuracy of Genever’s method
were 62, 73, and 67%, respectively (Table 4). There was no
significant difference in sensitivity (p = 0.065), specificity
Table 5  Diagnostic performance of the predictive models in the vali-
dation cohort (n = 45)
All cases
Sensitivity Specificity AUC
Simplified unenhanced-MRI-based model
Reviewer 1 87% (13/15) 73% (22/30) 0.84 [0.70–0.98]
Reviewer 2 87% (13/15) 70% (21/30) 0.84 [0.68–0.99]
Simplified enhanced-MRI-based model
Reviewer 1 87% (13/15) 70% (21/30) 0.77 [0.62–0.93]
Reviewer 2 87% (13/15) 70% (21/30) 0.77 [0.62–0.93]
Genever’s method 80% (12/15) 60% (9/15) NA
Each value in parentheses is the number of cases
Values in square brackets indicate the 95% confidence interval
AUC area under the receiver operating characteristic curve, NA not
available
479
(p = 0.29), and accuracy (p = 0.019) between the unen-
hanced-MRI-based model and Genever’s method. The sen-
sitivity and accuracy of the enhanced-MRI-based model
were higher than those of Genever’s method (p = 0.012
and 0.013, respectively), but there was no difference in
specificity between them (p = 0.69) (Fig. 3). There was no
difference in sensitivity (p = 0.50), specificity (p = 0.73),
and accuracy (p = 1.0) between the unenhanced-MRI-
based and enhanced-MRI-based models.
Validation cohort
The two reviewers interpreted 45 cases by focusing on five
MRI features identified as independent predictors of car-
cinosarcoma via multivariate logistic regression analyses,
i.e., inhomogeneity of tumors on T2WI, predominant SIs of
tumors on T2WI, presence of hyperintense areas on T1WI,
intratumoral contrast defect area volume percentage, and
degree of tumor enhancement. Detailed results are shown
in Table 5. Interobserver agreement regarding predominant
SIs on T2WI was moderate, while interobserver agreement
for the other predictors was excellent.
Diagnostic performance of the predictive diagnostic
models
With the simplified unenhanced-MRI-based model, the
sensitivity, specificity, and accuracy for diagnosing car-
cinosarcoma were, respectively, 87% (13/15) and 87%
(13/15), 73% (22/30) and 70% (21/30), and 78% (35/45)
and 76% (34/45), according to reviewers 1 and 2, respec-
tively (Table 5). With the simplified enhanced-MRI-based
model, the sensitivity, specificity, and accuracy for diag-
nosing carcinosarcoma were 87% (13/15), 70% (21/30),
and 76% (34/45), respectively, according to both review-
ers (Table 5; Figs. 4, 5). With Genever’s method, the
sensitivity, specificity, and accuracy for diagnosing car-
cinosarcoma were 80% (12/15), 60% (18/30), and 67%
(30/45), respectively (Table 5). There was no significant
difference in sensitivity, specificity, and accuracy among
the three methods.
Discussion
Our multivariate logistic regression analyses suggested
that inhomogeneity on T2WI, predominant SIs on T2WI,
and the presence of hyperintense areas in tumors on
T1WI were significant predictors of carcinosarcoma in
the unenhanced-MRI-based model, and the presence
of hyperintense areas in tumors on T1WI, intratumoral
contrast defect area volume percentage, and degree
of tumor enhancement were significant predictors of
13
480 Jpn J Radiol (2017) 35:472–483

Fig. 4a–c  65-year-old woman with a small-sized carcinosarcoma ▸

of the uterus. a A sagittal T2-weighted image shows a slightly het-
erogeneous, predominantly hyperintense mass (arrow). The ratio
of the largest anteroposterior dimensions of the uterus to the thick-
ness of the endometrium (ET/AP ratio) is 0.40 (12/30 mm). b There
is no hyperintense area on the axial T1-weighted image in the mass
(arrow). c On a contrast-enhanced T1-weighted image with fat sup-
pression, part of the mass enhances more strongly than the myome-
trium (arrow). The contrast defect volume percentage (arrowhead)
was less than 5% when all planes were evaluated. This case was
misdiagnosed as carcinoma by Genever’s criteria, but was correctly
diagnosed by both simplified unenhanced- and enhanced-MRI-based
models with total scores of 3 and 2, respectively

carcinosarcoma in the enhanced-MRI-based model. The

presence of hyperintense areas on T1WI representing
intratumoral hemorrhage has been reported to be one of
the characteristic MRI features of carcinosarcomas [9,
11–13, 15]. The prevalence of this feature in this study
(62%) was higher than in previous reports (12–50%) [9,
11, 13, 15]. This discrepancy may be due to the defini-
tion of hyperintensity on T1WI: namely, we regarded a
SI higher than that of the myometrium as hyperintense,
whereas the other authors did not provide an explicit defi-
nition of hyperintensity on T1WI.
Heterogeneous SI on T2WI has also been reported to
be a characteristic MRI feature of carcinosarcoma of the
uterus [9, 11–13]. Carcinosarcomas have both epithelial
and stromal tumor tissue, and various stromal tissues
coexist in the sarcomatous component. Hemorrhage and
necrosis are common, which might contribute to the het-
erogeneous SI of carcinosarcoma on T2WI. Although a
number of carcinomas showed heterogeneous SI, there
were no homogeneous carcinosarcomas in the derivation
cohort; therefore, homogeneous SI on T2WI is a reli-
able indicator of carcinoma when attempting to differ-
entiate carcinoma and carcinosarcoma. Our observation
of a predominant hyperintensity of carcinosarcoma on
T2WI was in good agreement with previous reports [11,
13]. Only 3 out of 26 (12%) carcinosarcomas showed
predominant hypointensity, whereas 13 out of 26 (50%)
carcinomas showed predominant hypointensity in the
derivation cohort. This suggests that a predominantly
hypointense mass on T2WI is unlikely to be carcinosar-
coma. Furthermore, predominant hyperintensity was a
reliable indicator of carcinosarcoma because none of the
carcinomas showed predominant hyperintensity in the
derivation cohort. The presence of hemorrhage, necrosis,
cystic degeneration, or cartilage tissue in the chondro-
sarcomatous component in carcinosarcomas might con-
tribute to the predominant hyperintensity of the tumor
[1, 19]. However, a radiological pathological correlation
is warranted to confirm the mechanism for the predomi- more frequently seen in carcinosarcomas than in carci-
nant hyperintensity of carcinosarcoma on T2WI. Mark- nomas in the univariate analysis. The reason for this is
edly hyperintense and hypointense areas on T2WI were not known. Although we did not correlate these findings

13
Jpn J Radiol (2017) 35:472–483 481

Fig. 5a–c  74-year-old woman with a large-sized endometrial carci- ▸

noma of the uterus. a A sagittal T2-weighted image shows a poorly
demarcated invasive mass (arrow). The signal intensity of the mass
is heterogeneous. Reviewers 1 and 2 rated the predominant signal
intensity as intermediate and hypointense, respectively. The ratio of
the largest anteroposterior dimensions of the uterus to the thickness
of the endometrium (ET/AP ratio) is 0.92 (45/49 mm). b An axial
T1-weighted image shows a hyperintense area (arrow) in the mass.
c On a contrast-enhanced T1-weighted image with fat suppression,
the contrast defect volume percentage (arrowhead) was more than
50% when all planes were evaluated. Reviewers 1 and 2 regarded the
tumor enhancement (arrow) as intermediate and strong, respectively.
This case was misdiagnosed as carcinosarcoma when both unen-
hanced- and enhanced-MRI-based models were used by both review-
ers and based on Genever’s criteria

with pathological findings, the presence of necrosis,

mucinous tissue, or chondrosarcomatous components
in carcinosarcomas might contribute to the markedly
hyperintense areas on T2WI, and the presence of a clot
or malignant skeletal muscle tissue in the sarcomatous
part might be associated with markedly hypointense
areas on T2WI [1, 19].
Extension to the uterine cervix with dilation of the cer-
vical canal has been reported to be a characteristic imag-
ing appearance for carcinosarcomas [15], but in this radio-
logical evaluation there was no difference in its frequency
between carcinosarcoma and carcinoma [1]. Regarding the
conspicuity of the tumor margin on T2WI, we expected
that carcinosarcomas would show a clearer border between
the tumor and myometrium because a carcinosarcoma is
typically a well-demarcated exophytic mass [11]; however,
there was no significant difference in this respect between
carcinosarcomas and carcinomas. Large carcinosarcomas
with both exophytic and endophytic growth may show a
poorly demarcated margin between the tumor and myome-
trium. Carcinosarcomas were enhanced equally as much
or more strongly than the myometrium, and were more
strongly enhanced than carcinomas, in good agreement
with previous reports [9, 11–13, 15]. As only 1 of 26 (4%)
carcinomas in the derivation cohort and 4 of 30 (13%) car-
cinomas in the validation cohort presented enhancement
equal to or stronger than that of the myometrium, strong
enhancement is suggestive of carcinosarcoma. Necrosis is a
common radiological and pathological finding for carcino-
sarcomas [1, 11, 15]. We regarded the contrast defect area
as representing tumor necrosis, and hypothesized that a
larger contrast defect area would be more likely to be seen
in more aggressive tumors such as carcinosarcomas. To our
best knowledge, there is no report of an investigation of the

13
482
contrast defect area volume percentage of carcinosarcoma
by radiological evaluation. Our results demonstrated that
the contrast defect area volume percentage was larger in
carcinosarcomas than in carcinomas, and this feature was
one of the most significant predictors of carcinosarcoma.
To our best knowledge, only Genever et al. [14] have
investigated the diagnostic capacity of imaging to dis-
criminate carcinosarcoma and carcinoma of the uterus.
They attempted to differentiate these two using the ET/
AP ratio, but the performance of this method was found
to be limited. In the present study, the ET/AP ratio was
not a predictor of carcinosarcoma when using multivari-
ate analysis, but it was when using univariate analysis.
The ET/AP ratio method is based on the hypothesis that
a carcinosarcoma is likely to show exophytic growth
with expansion of the endometrial cavity. However, some
aggressive carcinosarcomas actually show endophytic
growth without expanding the cavity and some carcino-
mas show exophytic growth; this results in significant
overlap of the ET/AP ratio between the carcinosarcoma
and carcinoma groups. In contrast, our simplified pre-
dictive diagnostic models demonstrated high sensitivity
and specificity in the derivation cohort, and the sensitiv-
ity and accuracy of the simplified enhanced-MRI-based
model were superior to those of the ET/AP ratio method.
The application of a combination of three significant MRI
features that are useful for discriminating carcinosarcoma
and carcinoma may have contributed to this high diag-
nostic performance. However, the specificities of both the
simplified unenhanced- and enhanced-MRI-based models
for diagnosing carcinosarcoma in the tumor-size-matched
validation cohort were moderate (73% by reviewer 1,
70% by reviewer 2) when all cases were analyzed. This
indicates that large carcinomas of the uterus can show
similar features to carcinosarcomas (such as hyperintense
areas in tumors on T1WI, heterogeneous SI on T2WI,
and large contrast defect areas), resulting in a substantial
number of false-positive cases.
Our study had several limitations. MRI scanners,
devices, and imaging parameters were not uniform due
to the nature of the multi-institutional retrospective
study. We could not evaluate the usefulness of diffusion-
weighted MRI, perfusion MR imaging, or MR spectros-
copy because some cases lacked such sequences. As a
relatively high apparent diffusion coefficient and a low
choline concentration are reported to be characteristic
findings of carcinosarcomas [10, 15, 20], further stud-
ies are required to assess the ability of multiparametric
MRI—including diffusion-weighted MRI, perfusion
MRI, and MR spectroscopy—to discriminate carcinosar-
comas from carcinomas. The second limitation was that
we did not evaluate the presence of lymphadenopathy or
peritoneal dissemination, as the scan range varied among
13
Jpn J Radiol (2017) 35:472–483
the cases. The presence of such lesions may contribute
to a diagnosis of carcinosarcoma because uterine carci-
nosarcoma has a higher incidence of lymphatic spread
and peritoneal seeding than carcinoma of the uterus [21].
Thirdly, we did not compare the MRI findings with the
pathological findings. The imaging features of carcino-
sarcomas identified in this study should be supported by
radiological pathological correlation in future studies.
In conclusion, our MRI-based predictive models can
differentiate carcinosarcoma from carcinoma of the
uterus with high sensitivity and moderate specificity.
Acknowledgments We thank Dr. Kohei Sasaguri for his cooperation
during the collection of MRI data on carcinosarcomas and carcinoma
of the uterus.
Compliance with ethical standards
Grant support No grant support for this study.
Conflict of interest The authors declare that they have no conflict of
interest.
References
1. D’Angelo E, Prat J. Pathology of mixed Mullerian tumours.
Best Pract Res Clin Obstet Gynaecol. 2011;25:705–18.
2. Barwick KW, LiVolsi VA. Malignant mixed mullerian tumors of the
uterus. A clinicopathologic assessment of 34 cases. Am J Surg
Pathol. 1979;3:125–35.
3. Gadducci A, Sartori E, Landoni F, Zola P, Maggino T, Cosio S, et al.
The prognostic relevance of histological type in uterine sarcomas:
a Cooperation Task Force (CTF) multivariate analysis of 249 cases.
Eur J Gynaecol Oncol. 2002;23:295–9.
4. Koh WJ, Greer BE, Abu-Rustum NR, Apte SM, Campos SM, Chan
J, et al. Uterine neoplasms, version 1.2014. J Natl Compr Cancer
Netw. 2014;12:248–80.
5. Homesley HD, Filiaci V, Markman M, Bitterman P, Eaton L, Kilgore
LC, et al. Phase III trial of ifosfamide with or without paclitaxel in
advanced uterine carcinosarcoma: a gynecologic oncology group
study. J Clin Oncol. 2007;25:526–31.
6. Bansal N, Herzog TJ, Burke W, Cohen CJ, Wright JD. The utility of
preoperative endometrial sampling for the detection of uterine sar-
comas. Gynecol Oncol. 2008;110:43–8.
7. Sagae S, Yamashita K, Ishioka S, Nishioka Y, Terasawa K, Mori
M, et al. Preoperative diagnosis and treatment results in 106
patients with uterine sarcoma in Hokkaido, Japan. Oncology.
2004;67:33–9.
8. Yamashita Y, Takahashi M, Miyazaki K, Okamura H. Contrast-
enhanced MR imaging of malignant mixed mullerian tumor of the
uterus. AJR Am J Roentgenol. 1993;160:1150–1.
9. Ohguri T, Aoki T, Watanabe H, Nakamura K, Nakata H, Mat-
suura Y, et al. MRI findings including gadolinium-enhanced
dynamic studies of malignant, mixed mesodermal tumors of the
uterus: differentiation from endometrial carcinomas. Eur Radiol.
2002;12:2737–42.
10. Kato H, Kanematsu M, Furui T, Imai A, Hirose Y, Kondo H, et al.
Carcinosarcoma of the uterus: radiologic-pathologic correlations
with magnetic resonance imaging including diffusion-weighted
imaging. Magn Reson Imaging. 2008;26:1446–50.
Jpn J Radiol (2017) 35:472–483
11. Tanaka YO, Tsunoda H, Minami R, Yoshikawa H, Minami M. Car-
cinosarcoma of the uterus: MR findings. J Magn Reson Imaging.
2008;28:434–9.
12. Teo SY, Babagbemi KT, Peters HE, Mortele KJ. Primary malig-
nant mixed mullerian tumor of the uterus: findings on sonogra-
phy, CT, and gadolinium-enhanced MRI. AJR Am J Roentgenol.
2008;191:278–83.
13. Bharwani N, Newland A, Tunariu N, Babar S, Sahdev A, Rockall
AG, et al. MRI appearances of uterine malignant mixed mullerian
tumors. AJR Am J Roentgenol. 2010;195:1268–75.
14. Genever AV, Abdi S. Can MRI predict the diagnosis of endometrial
carcinosarcoma? Clin Radiol. 2011;66:621–4.
15. Takeuchi M, Matsuzaki K, Harada M. Carcinosarcoma of the
uterus: MRI findings including diffusion-weighted imaging and
MR spectroscopy. Acta Radiol. 2016;57:1277–84.
16. D’Angelo E, Prat J. Uterine sarcomas: a review. Gynecol Oncol.
2010;116:131–9.
483
17. Landis JR, Koch GG. The measurement of observer agreement for
categorical data. Biometrics. 1977;33:159–74.
18. Youden WJ. Index for rating diagnostic tests. Cancer.
1950;3:32–5.
19. Murphey MD, Walker EA, Wilson AJ, Kransdorf MJ, Temple HT,
Gannon FH. From the archives of the AFIP: imaging of primary
chondrosarcoma: radiologic-pathologic correlation. Radiographics.
2003;23:1245–78.
20. Takahashi M, Kozawa E, Tanisaka M, Hasegawa K, Yasuda M,
Sakai F. Utility of histogram analysis of apparent diffusion coef-
ficient maps obtained using 3.0 T MRI for distinguishing uterine
carcinosarcoma from endometrial carcinoma. J Magn Reson Imag-
ing. 2016;43:1301–7.
21. Amant F, Cadron I, Fuso L, Berteloot P, de Jonge E, Jacomen G,
et al. Endometrial carcinosarcomas have a different prognosis and
pattern of spread compared to high-risk epithelial endometrial can-
cer. Gynecol Oncol. 2005;98:274–80.
13