Early Invasive Cervical Cancer: CT and MR Imaging in Preoperative ...

Note: This copy is for your personal non-commercial use only. To order presentation-ready copies for distribution to your colleagues or clients, contact us at www.rsna.org/rsnarights. ORIGINAL RESEARCH

Hedvig Hricak, MD, PhD Constantine Gatsonis, PhD Fergus V. Coakley, MD Bradley Snyder, MS Caroline Reinhold, MD Lawrence H. Schwartz, MD Paula J. Woodward, MD Harpreet K. Pannu, MD Marco Amendola, MD Donald G. Mitchell, MD

Purpose:

To retrospectively compare diagnostic performance and interobserver variability for computed tomography (CT) and magnetic resonance (MR) imaging in the pretreatment evaluation of early invasive cervical cancer, with surgical pathologic findings as the reference standard.

Materials and Methods:

This HIPAA-compliant study had institutional review board approval and informed consent for evaluation of preoperative CT (n ⫽ 146) and/or MR imaging (n ⫽ 152) studies in 156 women (median age, 43 years; range, 22– 81 years) from a previous prospective multicenter American College of Radiology Imaging Network and Gynecologic Oncology Group study of 172 women with biopsy-proved cervical cancer (clinical stage ⱖ IB). Four radiologists (experience, 7–15 years) interpreted the CT scans, and four radiologists (experience, 12–20 years) interpreted the MR studies retrospectively. Tumor visualization and detection of parametrial invasion were assessed with receiver operating characteristic curves (with P ⱕ .05 considered to indicate a significant difference). Descriptive statistics for staging and ␬ statistics for reader agreement were calculated. Surgical pathologic findings were the reference standard.

Results:

For CT and MR imaging, respectively, multirater ␬ values were 0.26 and 0.44 for staging, 0.16 and 0.32 for tumor visualization, and ⫺0.04 and 0.11 for detection of parametrial invasion; for advanced stage cancer (ⱖIIB), sensitivities were 0.14 – 0.38 and 0.40 – 0.57, positive predictive values (PPVs) were 0.38 –1.00 and 0.32– 0.39, specificities were 0.84 –1.00 and 0.77– 0.80, and negative predictive values (NPVs) were 0.81– 0.84 and 0.83– 0.87. MR imaging was significantly better than CT for tumor visualization (P ⬍ .001) and detection of parametrial invasion (P ⫽ .047).

Conclusion:

Reader agreement was higher for MR imaging than for CT but was low for both. MR imaging was significantly better than CT for tumor visualization and detection of parametrial invasion. The modalities were similar for staging, sharing low sensitivity and PPV but relatively high NPV and specificity.

1

From the Department of Radiology, Memorial SloanKettering Cancer Center, 1275 York Ave, New York, NY 10021 (H.H., L.H.S.); Center for Statistical Sciences, Brown University, Providence, RI (C.G., B.S.); Department of Radiology, University of California San Francisco, San Francisco, Calif (F.V.C.); Department of Diagnostic Radiology, McGill University Health Center, Montreal, Quebec, Canada and Synarc, San Francisco, Calif (C.R.); Department of Radiology, Armed Forces Institute of Pathology, Washington, DC (P.J.W.); Department of Radiology, Johns Hopkins University School of Medicine, Baltimore, Md (H.K.P.); Department of Radiology, University of Miami Medical School, Miami, Fla (M.A.); and Department of Radiology, Thomas Jefferson University, Philadelphia, Pa (D.G.M.). From the 2005 RSNA Annual Meeting. Received November 21, 2006; revision requested January 5, 2007; revision received April 24; final version accepted May 4. Supported by National Cancer Institute grants U01 CA079778 and U01 CA080098. Address correspondence to H.H. (e-mail: [email protected] ).

䡲 GENITOURINARY IMAGING

Early Invasive Cervical Cancer: CT and MR Imaging in Preoperative Evaluation—ACRIN/GOG Comparative Study of Diagnostic Performance and Interobserver Variability1

娀 RSNA, 2007

姝 RSNA, 2007 Radiology: Volume 245: Number 2—November 2007

491

GENITOURINARY IMAGING: CT and MR Imaging in Early Invasive Cervical Cancer

A

multicenter study conducted jointly by the American College of Radiology Imaging Network (ACRIN) and the Gynecologic Oncology Group (GOG) (1) found that in patients with early invasive cervical cancer, magnetic resonance (MR) imaging was superior to computed tomography (CT) for tumor visualization but the modalities were equivalent for preoperative staging. Twenty-five institutions, including academic and community medical centers, participated in the study, and each MR imaging or CT study was interpreted prospectively by a radiologist at the institution where the patient was enrolled. Although the study generated valuable information about the accuracy of CT and MR imaging as used in clinical practice in a broad variety of settings (1), it could not provide an assessment of reader variability with respect to interpretation or accuracy. Thus, the purpose of our secondary study was to retrospectively compare diagnostic performance and interobserver variability for

Advances in Knowledge 䡲 MR imaging and CT perform similarly in the staging of cervical cancer, sharing low sensitivity and low positive predictive value but also high negative predictive value and specificity. 䡲 In patients with cervical cancer, MR imaging is better than CT for tumor visualization (average areas under the receiver operating characteristic curve [AUCs] were 0.77 and 0.58, respectively; P ⬍ .001) and depiction of parametrial invasion (average AUCs were 0.68 and 0.62, respectively; P ⫽ .047). 䡲 In the pretreatment evaluation of cervical cancer, reader agreement is higher for MR imaging than for CT but is relatively low for both modalities (for CT and MR imaging, respectively, multirater ␬ values were 0.26 and 0.44 for staging, 0.16 and 0.32 for tumor visualization, and ⫺0.04 and 0.11 for detection of parametrial invasion). 492

CT and MR imaging in the pretreatment evaluation of early invasive cervical cancer, with surgical pathologic findings as the reference standard.

Materials and Methods Patients and Reference Standard The ACRIN 6651/GOG 183 intergroup multicenter prospective study enrolled consecutive women with biopsy-confirmed and previously untreated cervical cancer (including invasive squamous cell carcinoma, adenocarcinoma, and adenosquamous carcinoma) of International Federation of Gynecology and Obstetrics (FIGO) clinical stage IB or greater who were scheduled for surgery on the basis of results of clinical assessment (1). Each of the 25 institutions that participated was required to submit a protocol-specific application to ACRIN and to be a GOG member with a proved record of 20 surgical cases of gynecologic cancer per year. Each site had a study team consisting of at least one gynecologic oncologist, one pathologist, and two radiologists. Patients were required to be willing to undergo both contrast material– enhanced CT and MR imaging before surgery. They were enrolled before surgical exploration and after they signed a study-specific informed consent form. All institutions had study-specific institutional review board approval, and copies of each institution’s approval notice and studyspecific consent form were filed at ACRIN headquarters before patient registration. The institutional review board approval and informed consent forms included permission to conduct this secondary retrospective study, which, like the primary study, was compliant with the Health Insurance Portability and Accountability Act. Patients unwilling or unable to undergo contrast-enhanced CT and MR imaging were excluded from the study. Also excluded were pregnant patients, patients who could not give informed medical consent, and patients who were not considered surgical candidates for reasons of comorbidity. The study commenced in March 2000 and was closed in

Hricak et al

November 2002 after 208 patients were enrolled. Of the 208 patients enrolled, 36 were excluded from the final data analysis because of incomplete data, including nine patients whose disease was deemed too extensive for surgery after imaging was performed. Imaging findings suggestive of metastatic nodal involvement were permitted to influence the decision to perform (or cancel) surgery. The analysis set for the primary, prospective ACRIN/GOG study consisted of 172 patients (1). For 29 of these patients, CT images were not submitted to the central ACRIN imaging archive and so could not be included in this secondary retrospective study. However, three eligible patients who could not be included in the primary study set (because prospective CT findings were missing or MR imaging was not performed) had CT images and surgical pathology data (the reference standard) available and were included in this secondary study, bringing the total number of patients in the CT data set to 146 (Fig 1). For 22 of the 172 patients included in the primary study, MR imaging studies were not submitted to the central ACRIN imaging archive and could not be included in this secondary study. However, two eligible patients who could not be included in the primary analysis set because prospective CT findings were missing had Published online 10.1148/radiol.2452061983 Radiology 2007; 245:491– 498 Abbreviations: ACRIN ⫽ American College of Radiology Imaging Network AUC ⫽ area under the ROC curve CI ⫽ confidence interval FIGO ⫽ International Federation of Gynecology and Obstetrics GOG ⫽ Gynecologic Oncology Group ROC ⫽ receiver operating characteristic Author contributions: Guarantors of integrity of entire study, H.H., C.G.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; literature research, C.G., C.R.; statistical analysis, C.G., B.S.; and manuscript editing, H.H., C.G., F.V.C., B.S., C.R., D.G.M. Authors stated no financial relationship to disclose. Radiology: Volume 245: Number 2—November 2007


MR images and surgical pathology data available. These two patients were included in this secondary study, bringing the total number of patients in the MR imaging set to 152 (Fig 1). The CT and MR imaging analysis sets were defined separately to maximize the number of patients in each set, but comparisons of the estimates of performance between the modalities (see below) were performed for the 142 patients common to the two sets. The median age of the 156 patients included in this secondary study (ie, patients in the CT analysis set, the MR imaging analysis set, or both) was 43 years (range, 22– 81 years).

CT and MR Imaging Techniques All MR imaging and CT examinations met or exceeded standards agreed to by the study investigators. Required standards for CT included spiral data acquisition at 5-mm collimation during suspended respiration after administration of 120 –150 mL of 60% iodinated contrast medium delivered by power injector at 2.0 –3.0 mL/sec, with scans extending from the diaphragm to the symphysis pubis. All patients received oral contrast material (1000 mL of diatrizoate sodium or equivalent) given in divided doses over the 60 –90 minutes before scanning, with the last dose given 10 –15 minutes before scanning. Rectal contrast material was administered at the discretion of each institution. Required standards for pelvic MR imaging included a field strength of 1.5 T, the use of phased-array surface coils, and the acquisition of rapid acquisition with relaxation enhancement T2-weighted transverse and sagittal images of the pelvis and spin-echo or gradient-echo T1-weighted transverse images extending from the symphysis pubis to above the renal hilum. For all sequences, the field of view was 20 –28 cm, the section thickness was 5 mm or less, the matrix was 256 ⫻ 192 or greater, and two or more signals were acquired. Image Interpretation CT and MR images for all patients enrolled were submitted to the central ACRIN imaging archive. After quality inRadiology: Volume 245: Number 2—November 2007

spection, the full set of available CT studies was sent to four radiologists (F.V.C., P.J.W., H.K.P., and M.A., whose experience in reading CT studies ranged from 7 to 15 years), and the full set of available MR imaging studies was sent to four radiologists (H.H., C.R., L.H.S., and D.G.M., whose experience in reading MR imaging studies ranged from 12 to 20 years). All of the CT and MR imaging studies were distributed in digital (ie, “soft copy”) form. The eight readers included six of the site investigators who performed the primary, on-site interpretations. The latter readers were not chosen on the basis of their levels of diagnostic performance in prospective interpretation, because the number of studies interpreted prospectively by each site investigator was very small—10 or fewer—and, therefore, individual readers’ diagnostic performance levels could not be determined. Instead, the readers were chosen on the basis of their experience in the practice of gynecologic oncoradiology, supplemented by their records of educational activities (presentations on the subject of gynecologic imaging at national or international meetings) and publication of articles on gynecologic

Hricak et al

imaging in peer-reviewed journals. Readers were aware that patients had been given a diagnosis of cervical cancer but were not informed of any other clinical, imaging, or histopathologic findings. CT readers independently evaluated all CT images to determine if tumor was visible and if parametrial invasion was present. Tumor visibility was rated on a five-point scale (1 ⫽ definitely not visible, 2 ⫽ probably not visible, 3 ⫽ equivocal or indeterminate, 4 ⫽ probably visible, and 5 ⫽ definitely visible). Readers assigned these scores by using their best expert judgment and the criteria established by the consensus of the group. In general, a tumor was considered not visible if the cervix was of normal uniform attenuation and without focal distortion, while a tumor was considered visible if a focal lowattenuation mass was seen within the cervix. Parametrial invasion was similarly rated on a five-point scale (1 ⫽ definitely not present, 2 ⫽ probably not present, 3 ⫽ equivocal or indeterminate, 4 ⫽ probably present, and 5 ⫽ definitely present). Again, readers assigned these scores by using their best expert judgment and the criteria established by the

Figure 1

Figure 1:

Patient flowchart. 493


consensus of the group. In general, parametrial invasion was considered absent when no tumor was visible and the cervix had a smooth regular outline with the surrounding parametrium, while parametrial invasion was considered present if the cervix demonstrated a focal irregular interface with the surrounding parametrium (particularly if there was a visible mass that extended through the full thickness of the cervix) or if gross tumor extension into the parametrium was evident. MR imaging readers independently evaluated all MR images to determine if tumor was visible and if parametrial invasion was present. Tumor visibility was rated on a five-point scale (1 ⫽ definitely not visible, 2 ⫽ probably not visible, 3 ⫽ equivocal or indeterminate, 4 ⫽ probably visible, and 5 ⫽ definitely visible). Tumor was considered to be visible if a mass with higher signal intensity than that of the adjacent stroma was found. Parametrial invasion was similarly rated on a five-point scale (1 ⫽ definitely not present, 2 ⫽ probably not present, 3 ⫽ equivocal or indeterminate, 4 ⫽ probably present, and 5 ⫽ definitely present). Readers assigned these scores by using their best expert judgment and the criteria established by the consensus of the group. In general, parametrial invasion was considered absent when on T2-weighted images no tumor was visible, cervical stroma demonstrated uniform low signal intensity,

Hricak et al

and the cervix had a smooth regular outline with the surrounding parametrium. Parametrial invasion was considered present if full-depth stromal invasion was seen along with one or more of the following findings: tumor tissue in the parametrium, vascular obliteration, focal bulge or asymmetry of the cervical border, spiculation of the cervical margin, and thickening of the uterosacral ligaments. All readers were asked to assign the single most likely overall tumor stage by using both the FIGO and the TNM classification systems (2,3).

Statistical Analysis Data were collected, managed, and analyzed by the Biostatistics and Data Management Center of ACRIN. The data analysis examined the performance of CT and MR imaging readers in assessing overall stage, cervical involvement (tumor visualization), and parametrial invasion. Summaries of performance (area under the receiver operating characteristic [ROC] curve [AUC] or sensitivity, specificity, and positive and negative predictive values) were first derived separately for each reader, and the averages of these metrics were then computed to make comparisons. The accuracy of the determination of overall stage was assessed by dichotomizing stage determinations and estimating sensitivity, specificity, positive predictive value, and negative predictive value for

Table 1 Reader Agreement in Retrospective Interpretation of CT and MR Imaging Studies Parameter Tumor visualization Invasion of right parametrium Invasion of left parametrium Overall parametrial invasion‡ Staging§

CT

Multirater ␬ Value* MR Imaging

0.16 (0.12 to 0.29) ⫺0.04 (⫺0.02 to 0.13) ⫺0.05 (⫺0.01 to 0.11) ⫺0.04 (⫺0.02 to 0.13) 0.26 (0.23 to 0.34)

0.32 (0.22 to 0.41) 0.10 (0.06 to 0.27) 0.12 (0.05 to 0.29) 0.11 (0.05 to 0.29) 0.44 (0.34 to 0.56)

CT ⬍.001 .961 .981 ... ⬍.001

P Value† MR Imaging ⬍.001 ⬍.001 ⬍.001 ... ⬍.001

* Data in parentheses are ranges of pairwise ␬ values. A pairwise ␬ value of greater than 0.00 but less than 0.40 was considered to represent poor agreement; a value of 0.40 – 0.75, fair to good agreement; and a value greater than 0.75, excellent agreement (8). †

For testing whether multirater ␬ values were significantly greater than zero.

Average of multirater ␬ values in left and right parametrium. (Data in parentheses are ranges of pairwise values over both left and right parametrium.)

‡

§

494

For staging of tumors as IIA or lower versus IIB or higher.

each reader. The main focus was on stages dichotomized as IIA or lower (“negative”) versus IIB or higher (“positive”). Additional thresholds were also evaluated. A mixed-model approach was used in the derivation of confidence intervals for averages and the comparison of average values of sensitivity and specificity (4–6). The approach accounts for the correlation caused by the fact that all readers of a particular modality interpreted studies from the same patients. Agreement among readers was assessed with unweighted ␬ statistics for pairs and with multirater ␬ statistics for the full set of readers (7). The ordinal rating scale was not dichotomized before calculation of the unweighted ␬ values. ␬ Values were assessed as follows: 0.00 ⱕ ␬ ⬍ 0.40 indicated poor agreement; 0.40 ⱕ ␬ ⱕ 0.75, fair to good agreement; and ␬ ⬎ 0.75, excellent agreement (8). The diagnostic performance of CT and MR imaging readers for tumor visualization and parametrial involvement was assessed by using ROC analysis of the scores for the degree of suspicion elicited from each reader. A nonparametric approach was followed for estimating and comparing AUC values while accounting for the correlation due to multiple readings of the same studies (9). In the case of the analysis of detection of parametrial involvement, separate degree-ofsuspicion data were elicited for the left and the right parametrium. The analysis accounted for clustering within the patient (10). For all inferences in which P values were reported, we used .05 as the critical value for declaring statistical significance.

Results Patient Characteristics One hundred eighteen (76%) of the 156 patients included in our study had surgicopathologic findings consistent with a FIGO stage in the range of IA to IIA, and 32 (21%) had surgicopathologic findings consistent with a FIGO stage of IIB or higher. Reader Agreement Generally, reader agreement was greater for MR imaging than for CT (Table 1). Radiology: Volume 245: Number 2—November 2007


For tumor visualization, ␬ values ranged from 0.12 to 0.29 for CT (multirater ␬ ⫽ 0.16) and from 0.22 to 0.41 for MR imaging (multirater ␬ ⫽ 0.32). For parametrial invasion, pairwise ␬ values ranged from ⫺0.02 to 0.13 for CT (multirater ␬ ⫽ ⫺0.04 averaged over left and right sides) and from 0.05 to 0.29 for MR imaging (multirater ␬ ⫽ 0.11 averaged over left and right sides). For staging, pairwise ␬ values ranged from 0.23 to 0.34 for CT (multirater ␬ ⫽ 0.26) and from 0.34 to 0.56 for MR imaging (multirater ␬ ⫽ 0.44). For tumor visualization, Spearman rank correlations ranged from 0.39 to 0.52 among CT readers, from 0.41 to 0.71 among MR imaging readers, and from 0.15 to 0.39 among pairs of CT and MR imaging readers. For parametrial involvement, the Spearman rank correlations ranged from 0.07 to 0.26 among CT readers, from 0.26 to 0.59 among MR imaging readers, and from 0.05 to 0.30 for pairs of CT and MR imaging readers.

Tumor Visualization Tumor visualization was significantly better with MR imaging than with CT (Table 2). For MR imaging readers, the AUC ranged from 0.67 to 0.86 (average ⫽ 0.77; 95% CI: 0.68, 0.87), while for CT readers, it ranged from 0.52 to 0.63 (average ⫽ 0.58; 95% CI: 0.46, 0.69) (Fig 2). The difference in average AUC between MR imaging and CT was 0.20 (95% CI: 0.12, 0.27), with a P value of less than .001. Evaluation of Parametrium MR imaging performed significantly better than CT for the detection of parametrial invasion (Table 2). For MR imaging readers, the AUC estimates (treating the left and right sides of the parametrium as clusters within the same subject) ranged from 0.64 to 0.75 (average ⫽ 0.68; 95% CI: 0.57, 0.80); for CT readers, the AUC estimates ranged from 0.54 to 0.68 (average ⫽ 0.62; 95% CI: 0.47, 0.76) (Fig 3). The difference in average AUC between MR imaging and CT readers was 0.07 (95% CI: 0.001, 0.15; P ⫽ .047). Table 3 provides data on understaging and overstaging by CT and MR imaging readers Radiology: Volume 245: Number 2—November 2007

Hricak et al

Table 2 AUC Values for Retrospective Interpretation of CT and MR Imaging Studies Parameter Tumor visualization Parametrial invasion‡

CT

Mean AUC* MR Imaging

0.58 (0.52–0.63) 0.62 (0.54–0.68)

0.77 (0.67–0.86) 0.68 (0.64–0.75)

Difference in AUC between Studies†

P Value

0.20 (0.12, 0.27) 0.07 (0.001, 0.15)

⬍.001 .047

* Data in parentheses are ranges over the readers. † Comparisons between MR imaging and CT were performed for the patients common to the MR imaging and CT analysis sets. Data in parentheses are 95% confidence intervals (CIs). ‡

The left and right sides of the parametrium were treated as clusters within the same subject.

with regard to the distinction between stage IIA or lower disease and stage greater than IIA disease (ie, disease with parametrial invasion). Of the 146 cases read by the CT readers, 30 (20.5%) had a pathologic stage greater than IIA, 110 (75.3%) had a pathologic stage of IIA or lower, and six (4.1%) had an unknown pathologic stage. Of the 152 cases read by the MR imaging readers, 30 (19.7%) had a pathologic stage greater than IIA, 117 (77.0%) had a pathologic stage of IIA or lower, and five (3.3%) had an unknown pathologic stage.

Assessment of Overall Stage Both modalities had low sensitivity and relatively high specificity in identifying advanced stage cervical cancer (ie, stage IIB or higher) (Table 4). For CT readers, the average sensitivity for advanced stage cancer was 0.28 (95% CI: 0.03, 0.52), and the average specificity was 0.90 (95% CI: 0.79, 1.00). MR imaging readers had higher average sensitivity (0.47; 95% CI: 0.23, 0.71) but lower average specificity (0.79; 95% CI: 0.71, 0.86). The difference between the average sensitivities of CT and MR imaging readers was ⫺0.19 (95% CI: ⫺0.43, 0.05; P ⫽ .104). The difference between the average specificities was 0.12 (95% CI: ⫺0.03, 0.26; P ⫽ .099). Achieved pairs of sensitivities and specificities were examined at other stage thresholds besides IIB or greater (ie, greater than IA, greater than IB1, greater than IB2, and greater than IIB). At each threshold, CT readers tended to have lower sensitivities and higher specificities than MR imaging readers. Positive predictive values were low

for both CT (average ⫽ 0.55; 95% CI: 0.41, 0.69) and MR imaging (average ⫽ 0.36; 95% CI: 0.24, 0.48). A single reader accounted for the high average positive predictive value for CT, but this reader also had the lowest sensitivity. The negative predictive values were similar for readers of CT (average ⫽ 0.83; 95% CI: 0.77, 0.88) and MR imaging (average ⫽ 0.85; 95% CI: 0.8, 0.91) studies.

Discussion Our study assessed the two main dimensions of variations among radiologists: degree of agreement in making specific diagnostic determinations and variations in diagnostic performance. We found substantial variability among both MR imaging and CT readers in making specific diagnostic determinations. ␬ Values indicating levels of reader agreement in tumor visualization, detection of parametrial invasion, and staging were higher for MR imaging than for CT but were relatively low for both modalities. Estimates of diagnostic performance per reader also varied substantially for both modalities. These findings suggest that MR imaging and CT are inherently imperfect for the evaluation of cervical cancer and that further technologic advances are required to improve the imaging assessment of cervical cancer. On average, MR imaging readers performed similarly to CT readers in overall staging but significantly better than CT readers in tumor visualization and detection of parametrial invasion. CT and MR imaging each had a relatively high negative predictive value for stage IIB cervical cancer (ie, disease 495


Hricak et al

Figure 2

Figure 2:

Empirical ROC curves for tumor visualization (cervical involvement) at (a) CT and (b) MR imaging stratified by reader (eg, CT1 ⫽ CT reader 1).

Figure 3

Figure 3: Empirical ROC curves for parametrial invasion stratified by modality and reader (eg, CT1 ⫽ CT reader 1). Top left: Curves for left parametrial invasion as assessed by CT readers. Bottom left: Curves for left parametrial invasion as assessed by MR imaging readers. Top right: Curves for right parametrial invasion as assessed by CT readers. Bottom right: Curves for right parametrial invasion as assessed by MR imaging readers. 496

Radiology: Volume 245: Number 2—November 2007


with parametrial invasion), indicating that both modalities are helpful in identifying patients who may be candidates for surgery. Furthermore, because MR imaging is excellent (and superior to CT) for tumor visualization and the assessment of cervical tumor size (as was shown in a separate analysis of our multicenter clinical trial data [11]) in patients with stage IB cancer, it is also valuable for choosing between surgery and concurrent chemotherapy and radiation therapy (11–13). A limitation of our study was that, by design, it included only women with early invasive cervical cancer who were considered candidates for surgery at initial clinical examination. The low positive predictive values of CT and MR imaging that we found may reflect this choice of patient cohort. Similarly, estimates of sensitivity, specificity, and ROC curves in such a cohort do not necessarily generalize to a broader population of women with cervical cancer. The inclusion criteria of the protocol resulted in the exclusion of most women with clinical findings that were highly suggestive or clearly indicative of parametrial invasion, although these may be the patients in whom imaging is very accurate. For medical and ethical reasons, it would have been impossible to demand that such patients undergo surgery. Another possible limitation was that our study was based on imaging studies that were obtained from a variety of sources and are now up to 6 years old. Both CT and MR imaging technology have evolved in the interim. Although no major advances have been made in MR imaging sequences for imaging of the uterus and cervix, one interval development worth noting is the use of dynamic imaging of cervical cancer. However, there is no consensus regarding the value of this technique, because published study results have not demonstrated it to be markedly better than conventional T1- and T2-weighted MR imaging (14,15). Although the use of contrast material has gained acceptance in the imaging of endometrial cancer, it has not been universally accepted in the Radiology: Volume 245: Number 2—November 2007

Hricak et al

Table 3 Understaging and Overstaging by CT and MR Imaging Readers Modality and Reader

Pathologic Stage

No. of Cases not Staged

No. of Cases Understaged

No. of Cases Correctly Staged

No. of Cases Overstaged

Total

ⱕIIA ⬎IIA ⱕIIA ⬎IIA ⱕIIA ⬎IIA ⱕIIA ⬎IIA

0 1 (3) 0 1 (3) 1 (0.9) 1 (3) 0 1 (3)

... 25 (83) ... 21 (70) ... 18 (60) ... 20 (67)

109 (100) 4 (13) 98 (89.1) 8 (27) 91 (82.7) 11 (37) 97 (88.2) 9 (30)

0 ... 12 (10.9) ... 18 (16.4) ... 13 (11.8) ...

109* 30 110 30 110 30 110 30

ⱕIIA ⬎IIA ⱕIIA ⬎IIA ⱕIIA ⬎IIA ⱕIIA ⬎IIA

0 0 2 (1.7) 0 0 0 4 (3.4) 0

... 13 (43) ... 16 (53) ... 17 (57) ... 18 (60)

90 (76.9) 17 (57) 92 (78.6) 14 (47) 93 (79.5) 13 (43) 88 (75.2) 12 (40)

27 (23.1) ... 23 (19.7) ... 24 (20.5) ... 25 (21.4) ...

117 30 117 30 117 30 117 30

CT 1 2 3 4 MR imaging 1 2 3 4

Note.—Data in parentheses are percentages. Data in participants with disease of unknown pathologic stage have been excluded from this table. * CT reader 1 did not read one case.

Table 4 Detection of Advanced Stage (>IIB) Cancer by Retrospective Readers of CT and MR Imaging Studies Parameter Mean sensitivity Mean specificity Mean positive predictive value Mean negative predictive value

CT*

MR Imaging*

P Value

0.28 (0.14–0.38) 0.90 (0.84–1.00) 0.55 (0.38–1.00) 0.83 (0.81–0.84)

0.47 (0.40–0.57) 0.79 (0.77–0.80) 0.36 (0.32–0.39) 0.85 (0.83–0.87)

.104 .099 .001 .305

* Data in parentheses are ranges over the readers.

imaging of cervical cancer. No new evidence-based imaging guidelines differing from those used in our study have been published for either CT or MR imaging since our study was closed to enrollment, and we believe that the CT and MR imaging techniques used were valid. Results of prior single-institution studies have shown better performance for both MR imaging and CT and have also shown MR imaging to be more accurate than CT in cervical cancer staging (16,17). The design of our study,

including the specific patient cohort examined and the participation of multiple institutions, may have contributed to the differences between our results and those of previous studies. In summary, although agreement was higher among MR imaging readers than among CT readers, the level of agreement was low for both modalities in the pretreatment assessment of early invasive cervical cancer. MR imaging was significantly better than CT for tumor visualization and detection of parametrial involvement. The two mo497


dalities were similar for overall staging, sharing low sensitivity and positive predictive value but also relatively high negative predictive value and specificity. Acknowledgment: The authors thank Muellner, BA, for editing the manuscript.

Ada

References 1. Hricak H, Gatsonis C, Chi DS, et al. Role of imaging in pretreatment evaluation of early invasive cervical cancer: results of the intergroup study American College of Radiology Imaging Network 6651-Gynecologic Oncology Group 183. J Clin Oncol 2005;23:9329 – 9337. 2. Green FL, Page DL, Fleming ID, et al, eds. AJCC cancer staging manual. New York, NY: Springer-Verlag, 2002. 3. Pecorelli S, Benedet JL, Creasman WT, Shepherd JH. FIGO staging of gynecologic cancer. 1994 –1997 FIGO Committee on Gynecologic Oncology. International Federation of Gynecology and Obstetrics. Int J Gynaecol Obstet 1999;64:5–10. 4. Hillis SL, Obuchowski NA, Schartz KM, Berbaum KS. A comparison of the DorfmanBerbaum-Metz and Obuchowski-Rockette methods for receiver operating characteristic (ROC) data. Stat Med 2005;24:1579 – 1607.

498

Hricak et al

5. Obuchowski NA. Multireader, multimodality receiver operating characteristic curve studies: hypothesis testing and sample size estimation using an analysis of variance approach with dependent observations. Acad Radiol 1995;2(suppl 1):S22–S29.

rent weekly cisplatin and radiotherapy for cervical carcinoma with assessment of acute and late side-effects. Clin Oncol (R Coll Radiol) 2006;18(1):38 – 45. [Published correction appears in Clin Oncol (R Coll Radiol) 2006;18(4):368.]

6. Zhou X-H, Obuchowski N, McClish D. Statistical methods in diagnostic medicine. New York, NY: Wiley, 2002.

13. Narayan K, McKenzie A, Fisher R, Susil B, Jobling T, Bernshaw D. Estimation of tumor volume in cervical cancer by magnetic resonance imaging. Am J Clin Oncol 2003;26: e163– e168.

7. Fleiss J, Levin B, Paik MC. Statistical methods for rates and proportions, 3rd ed. Hoboken, NJ: Wiley-Interscience, 2003. 8. Fleiss JL. Statistical methods for rates and proportions. New York, NY: Wiley, 1981. 9. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated ROC curves: a non parametric approach. Biometrics 1988;44:837– 845. 10. Obuchowski NA. Nonparametric analysis of clustered ROC data. Biometrics 1997;53: 567–578. 11. Mitchell DG, Snyder B, Coakley F, et al. Early invasive cervical cancer: tumor delineation by magnetic resonance imaging, computed tomography, and clinical examination, verified by pathologic results, in the ACRIN 6651/GOG 183 intergroup study. J Clin Oncol 2006;24:5687–5694. 12. King M, McConkey C, Latief TN, Hartley A, Fernando I. Improved survival after concur-

14. Iwata S, Joja I, Okuno K, et al. Cervical carcinoma with full-thickness stromal invasion: efficacy of dynamic MR imaging in the assessment of parametrial involvement. Radiat Med 2002;20:247–255. 15. Yamashita Y, Baba T, Baba Y, et al. Dynamic contrast-enhanced MR imaging of uterine cervical cancer: pharmacokinetic analysis with histopathologic correlation and its importance in predicting the outcome of radiation therapy. Radiology 2000;216:803– 809. 16. Kim SH, Choi BI, Han JK, et al. Preoperative staging of uterine cervical carcinoma: comparison of CT and MR imaging in 99 patients. J Comput Assist Tomogr 1993;17:633– 640. 17. Subak LL, Hricak H, Powell CB, Azizi L, Stern JL. Cervical carcinoma: computed tomography and magnetic resonance imaging for preoperative staging. Obstet Gynecol 1995;86:43–50.

Radiology: Volume 245: Number 2—November 2007