Sequential magnetic resonance imaging of cervical cancer
The predictive value of absolute tumor volume and regression ratio measured before, during, and after radiation therapy
Version of Record online: 13 JUL 2010
Copyright © 2010 American Cancer Society
Volume 116, Issue 21, pages 5093–5101, 1 November 2010
How to Cite
Wang, J. Z., Mayr, N. A., Zhang, D., Li, K., Grecula, J. C., Montebello, J. F., Lo, S. S. and Yuh, W. T. C. (2010), Sequential magnetic resonance imaging of cervical cancer. Cancer, 116: 5093–5101. doi: 10.1002/cncr.25260
- Issue online: 13 JUL 2010
- Version of Record online: 13 JUL 2010
- Manuscript Revised: 6 JAN 2010
- Manuscript Accepted: 6 JAN 2010
- Manuscript Received: 8 DEC 2009
- cervical cancer;
- magnetic resonance imaging;
- absolute tumor volume;
- tumor regression ratio;
- tumor local control;
- patient disease-specific survival
The objectives of this study were to investigate outcome prediction by measuring absolute tumor volume and regression ratios using serial magnetic resonance imaging (MRI) during radiation therapy (RT) for cervical cancer and to develop algorithms capable of identifying patients at risk of a poor therapeutic outcome.
Eighty patients with stage IB2 through IVA cervical cancer underwent 4 MRI scans: before RT (MRI1), during RT at 2 to 2.5 weeks (MRI2) at 4 to 5 weeks (MRI3), and 1 to 2 months after RT (MRI4). The median follow-up was 6.2 years (range, 0.2-9.4 years). Tumor volumes at MRI1, MRI2, MRI3, and MRI4 (V1, V2, V3, and V4, respectively) and tumor regression ratios (V2/V1, V3/V1, and V4/V1) were measured by 3-dimensional volumetry. Predictive metrics based on tumor volume/regression parameters were correlated with ultimate clinical outcomes, including tumor local recurrence (LR) and dying of disease (DOD). Predictive power was evaluated using the Mann-Whitney test, sensitivity/specificity analyses, and Kaplan-Meier analyses.
Both tumor volume and regression ratio were strongly correlated with LR (P = .06, P = 5 × 10−4, P = 1 × 10−6, and P = 2 × 10−8 for V1, V2, V3, and V4, respectively; and P = 7 × 10−5, P = 1 × 10−6, and P = 1 × 10−8 for V2/V1, V3/V1, and V4/V1, respectively) and DOD (P = .015, P = .004, P = .001, and P = 3 × 10−4 for V1, V2, V3, and V4, respectively; and P = .03, P = .009, and P = 3 × 10−4 for V2/V1, V3/V1, and V4/V1, respectively). Algorithms that combined tumor volumes and regression ratios improved predictive power (sensitivity, 61%-89%; specificity, 79%-100%). The strongest predictor, pre-RT volume and regression ratio at MRI3 (V1 > 40 cm3 and V3/V1 > 20%, respectively), achieved 89% sensitivity, 87% specificity, and 88% accuracy for LR and achieved 54% sensitivity, 83% specificity, and 73% accuracy for DOD.
The current results suggested that tumor volume/regression parameters obtained during primary therapy are useful in predicting LR and DOD. Both tumor volume and regression ratio provided important information as early outcome predictors that may guide early intervention for patients with cervical cancer who are at high risk of treatment failure. Cancer 2010. © 2010 American Cancer Society.
The ability of magnetic resonance imaging (MRI) to differentiate tumor from normal soft tissue and to delineate tumor extent with high precision has greatly improved the quantitative assessment of tumor volume in cervical cancer.1-4 Currently, pretreatment tumor volume, a well known prognostic factor in cervical cancer,5-8 can be measured much more accurately using 3-dimensional (3D) quantitative volumetry of the tumor with the improved lesion contrast and the multiplanar imaging ability of MRI. In addition, subtle volume changes during early therapy can be monitored closely as an early indicator of tumor response to an ongoing treatment. Irregular tumor shape and nonlinear treatment-related shrinkage make it difficult to identify such subtle changes using traditional methods, which measure only 1-dimensional or 2-dimensional maximal tumor dimensions. This challenge can be overcome by quantitating detailed temporal changes of tumor volume using 3D volumetry with serial MRI studies during the course of radiation therapy (RT) and chemotherapy (CT).
Early results evaluating this new 3D imaging-based information suggest that therapy-induced tumor volume changes can be predictive of ultimate local tumor control and patient survival.5-8 Various tumor parameters and time points for MRI measurements have been suggested for outcome prediction, including before therapy, during therapy, and post-therapy follow-up.9-13 However, applicable practical paradigm and parameter thresholds for clinical implementation based on such a large and varied quantity of imaging data, including imaging timing, has not been well defined or validated with a long-term clinical study. The objectives of the current study were to investigate the efficacy of outcome prediction using serial MRI examinations to measure both tumor volumes and regression ratios before, during, and after RT/CT and to develop algorithms to identify cervical cancer patients at risk of treatment failure.
MATERIALS AND METHODS
Patient Population and Treatment
Eighty patients with cervical cancer (stages IB2-IVA), who received standard RT/CT, underwent serial MRI on an Institutional Review Board-approved imaging study. Patient ranged in age from 25 years to 85 years (median age, 51 years). RT consisted of standard external-beam RT (EBRT) and low-dose-rate (LDR) brachytherapy. The dose prescription was between 45 grays (Gy) and 50 Gy of EBRT delivered in daily fractions of 1.8 Gy to 2 Gy and 40 Gy of LDR brachytherapy delivered in 1–2 fractions of 20 Gy. The median follow-up was 6.2 years (range, 0.2-9.4 years).
The imaging protocol included serial MRI studies at 4 well defined time points: 1) MRI1, at the beginning of RT; 2) MRI2, early during RT (2-2.5 weeks with an RT dose of 20-25 Gy); 3) MRI3, midway during RT (4-5 weeks with 45-50 Gy); and 4) MRI4, at the follow-up visit (1-2 months after the completion of all therapy). The MRI examinations were conducted using a standard body coil with a 1.5-Tesla superconductive scanner, including Signa (General Electric Medical Systems, Milwaukee, Wis) and Siemens Vision (Siemens Medical, Inc., Erlangen, Germany). The change in platforms was necessitated by a change in imaging systems at our institutions. Imaging included sagittal, 5-mm (4-mm thickness with a 1-mm gap), conventional fast spin echo, T2-weighted images (effective echo time [TEeff] = 104 msec; repetition time [TR] = 4000 msec; echo train length [ETL] = 10 msec; number of excitations [NEX] = 2); and axial, 7-mm (5-mm thickness with a 2-mm gap), T2-weighted and T1-weighted images (echo time [TE] = 16 msec, TR = 600 msec, NEX = 2). No MRI was performed with dwelling brachytherapy applicators.
Only patients who completed all 4 MRI studies were included in this study. The therapy was standard RT with CT for patients with cervical cancer and was not influenced by the imaging findings.
Tumor Volume and Temporal Change
The imaging datasets were evaluated by 3 reviewers, including 2 MRI radiologists (with 14 years and 10 years of experience) and 1 radiation oncologist (with 10 years of experience in MRI-based RT planning). The tumor was defined as an abnormal area with intermediate-to-high signal intensity on T2-weighted images with respect to the surrounding cervical stroma and uterus and lower than the fluid signal in the urinary bladder.1, 12, 14 Discrepancies in tumor delineation among the reviewers were resolved by consensus. For 3D volumetry, the region-of-interest (ROI) was delineated on each imaging slice on the sagittal T2-weighted image, and the 3D ROI-based volumes were then calculated by the summation of all tumor areas and multiplication by the slice thickness. More recently, we also calculated tumor volumes separately using pixel summation and produced consistent results. Our developed code for image data analysis was based on the MATLAB platform (MATLAB R2006a; MathWorks Inc., Natick, Mass).
Tumor volumes (V1, V2, V3, and V4) measured by MRI 3D volumetry at the MRI1, MRI2, MRI3, and MRI4 time points, respectively, were used to compute the regression ratios (V2/V1, V3/V1, and V4/V1) for each patient. Both tumor volume data and regression ratio data were correlated with patient outcome endpoints, local recurrence (LR) of the tumor and death of disease (DOD), based on clinical follow-up. LR in the pelvis was defined as tumor regrowth or persistence of tumor in the cervix after the completion of treatment. Cases other than LR were considered as local control (LC) of tumor. For DOD, deaths from cervical cancer or cancer complications were scored as events, and deaths from intercurrent disease were censored. Patients other than DOD were considered as disease-specific survival (DSS) analysis. The correlation between the volume and ratio parameters and the outcomes of LR and DOD were analyzed using the Mann-Whitney rank-sum test.15
Receiver operating characteristics (ROC) analysis was performed to define the best cutoff for each volume and regression ratio. The resulting parameter cutoffs to differentiate poor from favorable outcome groups were then subjected to sensitivity/specificity analysis, Kaplan-Meier actuarial life table analysis, and log-rank tests so that improved prediction algorithms could be established by combining the predictive powers of tumor volumes and regression ratios at various time points. All statistical analyses were performed using the SPSS software package (SPSS version 16; SPSS Inc., Chicago, Ill).
Tumor Volume and Regression Ratio
The means, ranges, and standard deviations of the serial V1, V2, V3, and V4 MRI tumor volumes are summarized in Table 1 for the 2 outcome groups, LR and DOD, that included all patients. On the basis of the measurements of the 80 patients who were included in this study, the mean initial tumor volume (V1) was 80.5 cm3 and varied widely from 3.4 cm3 to 700 cm3. During therapy, the tumor volume regressed in an approximately exponential fashion with mean regression ratios of 56%, 18%, and 5% for V2/V1, V3/V1, and V4/V1 measured at MRI2, MRI3, and MRI4, respectively. Overall, the mean tumor volumes and standard deviations were much smaller in the LC and DSS groups.
|Therapy Outcome||No. of Patients||Tumor Volumea: Mean±SD (Range), cm3|
|Yes||62||66±55 (3.4-342)||36±39 (0.4-248)||8.4±11.5 (0-48)||0.59±1.9 (0-13)|
|No||18||130±162 (41-700)||96±119 (20-493)||47±52 (1.9-192)||26±36 (0-121)|
|Yes||52||64±57 (7.3-342)||36±40 (1.3-248)||9±13 (0-48)||1±3 (0-18)|
|No||28||111±133 (3.4-700)||76±100 (0.4-493)||31±46 (0-192)||16±32 (0-121)|
|All patients||80||81±93 (3.4-700)||50±70 (0.4-493)||17±31 (0-192)||6.3±20 (0-121)|
Prediction Power of Individual MRI Studies
The overall tumor LC rate for all 80 patients was 78%, and the DSS rate was 65%. The prediction power of parameters from all 4 MRI studies for LC and DSS are presented in Table 2. All tumor volumes (V2, V3, and V4) except for V1 and all regression ratios (V2/V1, V3/V1, and V4/V1) were correlated significantly with LC (P values ranged from 5 × 10−4 to 1 × 10−8) and DSS (P values ranged from .015 to 3 × 10−4) (Table 2). Based on P values, V1 was the least significant for predicting therapy outcome compared with the remaining MRI predictors.
|Tumor Volume||Regression Ratio|
|All 80 patients|
|Local control of tumor||.06||5×10−4||1×10−6||2×10−8||7×10−5||1×10−6||1×10−8|
|Patients stratified by tumor stage|
|Stage I/II, n=42|
|Local control of tumor||.16||.07||.02||.002||.29||.08||.003|
|Stage III/IVA, n=38|
|Local control of tumor||.78||.06||8×10−5||2×10−5||9×10−4||1×10−5||8×10−6|
|Patients stratified by LN status|
|LN uninvolved, n=59|
|Local control of tumor||.17||.008||8×10−5||2×10−5||.0018||1×10−4||1×10−5|
|LN involved, n=21|
|Local control of tumor||.16||.02||.004||.0004||.013||.004||.0004|
Because of large variations in tumor stage (ranging, from IB2 to IVA) and the initial tumor size (V1 ranging from 3.4 cm3 to 700 cm3), we also stratified the patient population into 2 groups based on tumor stage: 42 patients had stage I and II disease (Group 1), and 38 patients had stage III and IVA disease (Group 2). The mean initial tumor volume (V1) was 53 cm3 (range, 3.4-259 cm3) in Group 1 and 111 cm3 (range, 24-700 cm3) in Group 2, for a significant difference in V1 between the 2 groups (P = .008). The prediction values of the volume parameters for the 2 groups are presented in Table 2. The trend in P values for each stage group was overall consistent with that of the entire patient group, although the prediction power was not as significant as that of the entire patient population, likely because of the smaller patient numbers in each group.
Similarly, we examined the prediction power of tumor volumes and regression ratios for patient subgroups stratified by lymph node status (Table 2). The P values for outcome prediction were higher for the patient subgroup with uninvolved lymph nodes than for the patient subgroup with involved lymph nodes.
The ROC method was used to determine the optimal cutoff points for each parameter for the best outcome prediction. The ROC results are shown in Figure 1, and the derived optimal thresholds are presented in Table 3. Parameter values for tumor volumes and regression ratios greater than the cutoff points were associated with a poor outcome. The values of the optimal cutoff points were fairly close for the 2 outcome endpoints LR and DOD (Table 3). Therefore, to streamline and simplify the predictors, only 1 set of cutoffs from LR prediction was used for further analyses of both LR prediction and DOD prediction and to develop the predictive algorithms.
|Endpoint||Tumor Volume, cm3||Regression Ratio, %|
|Local tumor recurrence||40||20||11||3||75||20||10|
|Dead of disease||43||34||9||2.5||65||15||7|
Prediction Algorithms and Their Sensitivity and Specificity
On the basis of the cutoff points derived from ROC analyses, algorithms with the combined parameters of tumor volume and regression ratio at various time points were explored to identify patients at risk for LR and DOD. Judging by the sensitivity/specificity analyses, the following algorithms had potential value for clinical application: 1) the regression ratio at MRI4 (V4/V1 > 10%), 2) the regression ratio at MRI3 (V3/V1 > 20%), 3) the initial tumor volume and the regression ratio at MRI3 (V1 > 40 cm3 and V3/V1 > 20%, respectively), and 4) the initial tumor volume and the regression ratio at MRI2 (V1 > 40 cm3 and V2/V1 > 75%, respectively). The sensitivity and specificity, confidence intervals, and accuracy of these 4 algorithms are summarized in Table 4. The efficacy of these algorithms was analyzed statistically along with the timing for outcome prediction.
|Prediction Algorithma||Sensitivity (95% CI), %||Specificity (95% CI), %||Accuracy Rate, %|
|Local tumor recurrence|
|1. (V4/V1 > 10%)||67 (41-87)||100 (94-100)||93|
|2. (V3/V1 > 20%)||89 (65-99)||79 (66-88)||81|
|3. (V1 > 40 cm3) and (V3/V1 > 20%)||89 (65-99)||87 (76-94)||88|
|4. (V1 > 40 cm3) and (V2/V1 > 75%)||61 (36-83)||94 (84-98)||86|
|Dead of disease|
|1. (V4/V1 > 10%)||36 (19-56)||96 (87-100)||75|
|2. (V3/V1 > 20%)||54 (34-72)||73 (59-84)||66|
|3. (V1 > 40 cm3) and (V3/V1 > 20%)||54 (34-72)||83 (70-92)||73|
|4. (V1 > 40 cm3) and (V2/V1 > 75%)||36 (19-56)||90 (79-97)||71|
Algorithm 3 (V1 > 40 cm3 and V3/V1 > 20%), which was based on the initial tumor volume V1 and the volume regression ratio at MRI3 (4-5 weeks of RT), had high sensitivity (89%), high specificity (87%), and an accuracy of 88% for LR; whereas, for DOD, sensitivity was 54%, specificity was 83%, and accuracy was 73%. Algorithm 4 (V1 > 40 cm3 and V2/V1 > 75%), which was based on an earlier measurement of V2 at MRI2 (2-3 weeks after the start of RT), had sensitivity of 61%, specificity of 94%, and accuracy of 86% for LR and had sensitivity of 36%, specificity of 90%, and accuracy of 71% for DOD. Kaplan-Meier survival analyses for the 2 algorithms for LC and DSS are presented in Figure 2. The P values calculated from log-rank tests indicate that the 2 algorithms (P ≤ .001) were highly significant during RT for identifying patients at risk for ultimate LR and DOD. For the best prediction with Algorithm 3 (V1 > 40 cm3 and V3/V1 > 20%), the 5-year LC rates for the 2 groups differed by 63% (96% vs 33%; P < .001), and the 5-year DSS rates differed by 37% (75% vs 38%; P < .001). Similarly, for prediction with Algorithm 4 (V1 > 40 cm3 and V2/V1 > 75%), the 5-year LC rates for the 2 identified groups was 89% versus 27% (P < .001), and the 5-year DSS rate was 70% versus 33%, respectively (P = .001) (Fig. 2).
Advanced cervical cancer continues to be a disease with significant mortality.16, 17 Once the tumors recur after the completion of therapy, options for salvage therapy are poor, and the outcome almost uniformly is fatal.18 Therefore, the prediction of therapy outcome at the earliest possible time point before or during the initial treatment course is of eminent importance to optimize therapy regimens and to increase the chance of a cure for individual patients. If a poor response to an ongoing standard therapy is predicted sufficiently early then a therapeutic window may be provided to modify the therapy strategy for better outcome.
Tumor size has long been established as an important prognostic factor in cervical cancer.5-8 Numerous studies using tumor size estimated by clinical palpation have reported a correlation with clinical outcome.5-8 The use of 3D tumor volume measurements through quantitative imaging analysis with MRI has greatly improved the accuracy of tumor volume assessment1-4 compared with clinical palpation for the estimation of tumor size. This greater precision of tumor delineation enables not only more accurate pretherapy measurement but also serial measurements of tumor volume during and after the course of therapy.1, 2 Tumor regression can be quantified better, and early results have demonstrated that tumor regression correlates with outcome. Most such studies of tumor volume and outcome in cervical cancer have assessed tumors before therapy or after the completion of therapy,9, 10, 13 and only a few have obtained intratreatment measurements.11, 12 With the large individual variations in initial tumor volumes and tumor regression rates from the reported data, the best parameter threshold values and optimal timing for tumor volume measurement and volume regression rate have not been well established for practical clinical translation.
With its large patient number and longest follow-up of studies of this kind to date, and prospective serial imaging schedule, our study is well suited to identify promising MRI tumor volumetric parameters that are associated with radioresponsiveness, tumor control, and survival. Serial imaging pretherapy, at Week 2, at Week 5, and post-therapy has allowed us to evaluate and compare the sensitivity and specificity of the tumor regression metrics along with the absolute tumor volumes from the 4 serial measurements.
Our results suggest that a combination of tumor volume-related parameters, incorporating both the absolute tumor volume and the tumor regression ratio, is necessary to increase the accuracy of outcome prediction. Although the regression ratio V4/V1 at the post-RT time point (MRI4) had an accuracy of 93%, this accuracy was based largely on high specificity for tumor control rate in our patient population (Table 4). The combination of initial tumor volume information (V1 > 40 cm3) and the earlier intratreatment regression parameter (V3/V1 > 20%; obtained at 40-50 Gy at 4-5 weeks of RT) improved the sensitivity of tumor recurrence prediction from 67% to 89% (Table 4). This finding suggests that the regression rate alone may not fully assess the impact of absolute tumor cell numbers during the dynamic events of cytotoxicity-induced tumor reduction. For tumors with a small initial volume (V1 < 40 cm3), standard therapy may be adequate to eradicate all tumor cells regardless the tumor regression rate, which may be influenced by either radiosensitivity or tumor microenvironmental factors, eg tumor perfusion. The algorithm with combined parameters (V1 > 40 cm3 and V3/V1 > 20%) identified patients with both large initial tumor volumes (V1 > 40 cm3) and radioresistant tumor cells (V3/V1 > 20%, indicating a poor response at 40-50 Gy) who were likely to ultimately fail standard treatment, as indicated by the poor LC rate of only 33% in the Kaplan-Meier analysis (P < .001; log-rank test) (Fig. 2a). Similarly, for DSS, the addition of initial absolute volume V1 > 40 cm3 improved the sensitivity of the post-therapy regression rate (V4/V1) from 36% to 54% with Algorithm 3 (V1 > 40 cm3 and V3/V1 > 20%) and resulted in a strong outcome correlation (P < .001; log-rank test) (Fig. 2c). In these patients, the pre-RT parameter V1 in combination with the mid-RT parameters (V3/V1 > 20%) can be used to guide timely therapy intervention (eg, RT dose escalation or more intense, novel chemotherapy approaches) that may be necessary to overcome the compounded challenge of large tumor burden and low radiosensitivity.
Evaluation of the timing of volume measurement throughout the course of therapy indicated that sensitivity and specificity generally increased at the later time points. For LR, although ROC analysis indicated poor results for the pretherapy V1 alone, and ROC curves increasingly improved with volumes that were obtained later in the course of therapy (Fig. 1a). V2 at 20 to 25 Gy (Week 2 of therapy) did not perform as well as V3 at 40 to 50 Gy (Week 4–5 of therapy) or V4 (1-2 months post-therapy) (Fig. 1a). The mean V2 value (50 cm3) remained large and ranged widely (Table 1) early in the RT course at 20 to 25 Gy, whereas much a greater volume decrease was evident later (V3 and V4). The observation that later measurements in the RT course improved the predictive power of absolute volume may be explained by a lag time between the actual tumor cell kill and measurable tumor shrinkage on anatomic imaging. Although 99% of tumor cells typically are killed during the first 2 weeks of RT, morphologic tumor volume reduction is not measurable until cell clearance has occurred. The intratreatment MRI scans reflect not only the RT-induced cell killing but also the effectiveness of dead cell clearance, which, to a certain extent, is related to the tumor microenvironment and fluid circulation.19, 20 Later in the course of RT, after cell clearance from earlier cell killing has occurred, the effect of cell killing becomes more measurably evident. This may explain why the later V3 and V4 values correlated better overall with outcome than the earlier V2 value (Table 2).
We expect that the post-RT volume V4 is most likely dominated by the repopulation of residual tumor, as demonstrated in a separate modeling study.19, 20 The information from both tumor regression because of radiation cell killing and tumor repopulation may explain the high sensitivity of V4 as a surrogate endpoint for LC. However the value of V4 for clinical management is offset by the late time point of post-therapy assessment, when V4 information is available.
Although these parameters also are useful for the prediction of survival, their sensitivity and specificity for DOD are less optimal than those for LR (Tables 2, 4; Figs. 1, 2). This may be related to the finding that DOD probably is also determined by other biologic or tumor microenvironmental factors, which may not be determined by local tumor bulk alone. Although sensitivities were lower, the algorithm that combined the pre-RT volume (V1 > 40 cm3) and the volume regression later during RT (V3) was effective in differentiating patients with DSS from patients in the DOD group (Table 4, Fig. 2).
When the patients were stratified according to their lymph node status (Table 2), the prediction power of tumor volume and regression ratio was greater for the subgroup with uninvolved lymph nodes than for the subgroup with involved lymph nodes. This may be interpreted as the tumor volume and its regression providing a local measure of tumor response; therefore, these parameters provide better prediction power for LC and DSS among patients with uninvolved lymph node status. The other reason may be the disparity in involved numbers in the 2 groups (59 patients with uninvolved lymph nodes vs 21 patients with involved lymph nodes).
On the basis of the current results, we would take the following approaches for future studies to validate these prognostic factors: 1) analyze larger database to validate the current findings, 1) refine and optimize the algorithms developed in the current study, and 3) combine with functional MRI21-26 to further improve the overall prediction power for clinical outcome. The overall imaging metric–outcome correlation results from our study confirm the value of serial MRI imaging in the assessment of tumor volume and the monitoring of tumor regression/response to ongoing treatment. Our parameters provide a basis for the development of algorithms to refine the efficacy of response assessment and to optimize outcome prediction. Such parameters have demonstrated promise in classifying radioresponsiveness and predicting tumor control and survival in patients with cervical cancer.
In conclusion, high-precision, 3D MRI-based serial tumor measurements provide important information about subtle morphologic changes as the early therapy response to ongoing treatment in individual patients. Our results with long-term follow-up data demonstrated that the combined pretherapy volumes and midtherapy volume regression ratios were highly accurate in predicting local failure and cancer death. Outcome predictions can be made as early as 2 weeks, and the best prediction was at 4–5 weeks into the treatment course. Therefore, the algorithms that we developed in the current study enable the early identification of patients who are at risk of treatment failure and may benefit from more aggressive interventions.
CONFLICT OF INTEREST DISCLOSURES
This work was supported in part by the National Institutes of Health (R01 CA71906).
- 15The Mann-Whitney rank-sum test. In: JeffersJD, EnglisMR, eds. Primer of Biostatistics. 3rd ed. New York: McGraw-Hill, Inc.; 1992: 262-334..