Prognostic significance of the pretreatment CT scan on time to progression for patients with malignant gliomas

Transcription

1 J Neurosurg 52:62-67, 198 Prognostic significance of the pretreatment CT scan on time to progression for patients with malignant gliomas VICTOR A. LEVIN, M.D., WILLIAM F. HOFFMAN, M.D., DAVID C. HEILBRON, PH.D., AND DAVID NORMAN, M.D. Brain Tumor Research Center, Department of Neurological Surgery, and Department of Radiology, School of Medicine, and the Computer Center, University of California, San Francisco, California ~" Case histories of 61 patients receiving multimodality therapy for primary malignant brain tumors were reviewed for factors visible on the computerized tomography (CT) scan that correlated with the interval of time from diagnosis and pretherapy evaluation to the time of documented tumor progression. The initial pretreatment CT scan of each patient was reviewed. Midline shift, peritumor low density (edema), the greatest diameter of tumor enhancement, and the greatest diameter of the intratumor low-density area were measured prior to radiation therapy and chemotherapy. Using a Weibull survival probability model, time to tumor progression was most satisfactorily fitted using two covariates, the volume of enhancing tumor (for instance, contrast-enhancing tumor less intratumor low density), and the presence of a peritumor lowdensity area. Patients who exhibited a small amount of contrast-enhancing tumor with peritumor low density tended to have a longer time to progression than patients with large contrast-enhancing tumor volume and no peritumor low density. Midline shift was not as important a variable as the extent of tumor contrast enhancement or peritumor low density. KEY WORDS 9 malignant glioma 9 statistical modeling 9 computerized tomography 9 time to progression 9 contrast-enhancing tumor 9 peritumor low density U NTIL relatively recently, the optimum therapy for malignant primary brain tumors has been extensive tumor resection with postoperative radiation therapy. In current practice, chemotherapy with oncolytic agents, such as the nitrosoureas, has prolonged the time to progression of tumor growth. Although the computerized tomography (CT) scan correlates well with other criteria of response and deterioration due to chemotherapy and radiotherapy, the importance of specific CT parameters, such as peritumor low density, contrast-enchanced tumor, midline shift, and intratumor low density are not well documented in this group of patients. 12 This report presents data accumulated on 61 consecutive patients treated on the Brain Tumor Chemotherapy Service, Department of Neurological Surgery, School of Medicine, University of California, San Francisco (UCSF). All the patients were studied by sequential CT scans, radionuclide scans, and neurological examinations. Clinical Material and Methods The clinical course of 7 successive newly diagnosed patients with primary malignant supratentorial brain tumors entered on the UCSF Phase III protocol that included BCNU, hydroxyurea, and radiation therapy (BHR) were reviewed retrospectively. The protocol consisted of two arms, both of which received 1,3-bis(2-chloroethyl)- 1-nitrosourea (BCNU, 8 mg/sq m intravenously three times daily) and megavoltage radiation therapy (5 rads to the whole brain, 1 rads to the reduced field). Patients in one arm also received hydroxyurea (HU, 275 mg/sq m by mouth every 6 hours every other day) during irradiation. Throughout the study, the proportion of patients who received HU to those who did not remained relatively constant. All patients were randomized within 21 days of clinical diagnosis or operation. Histopathology on all patients undergoing operation was reviewed by the Neuropathology Department, UCSF. On entry into the study, all patients 62 J. Neurosurg. / Volume 52 / May, 198

2 Correlation of CT scan with prognosis had a baseline neurological examination, CT scan with and without contrast enhancement, and a radionuclide scan. The CT scans were made on the EMI head scanner, EMI body scanner, or General Electric body scanner; a matrix was used in all cases. Patients were irradiated at doses of 17 to 2 rads/day, for a total brain dose of 5 rads and a tumor dose of 6 rads. Patients returned for reevaluation 1 to 2 weeks after completing irradiation. At that and at subsequent evaluation periods, results of a clinical examination, radionuclide scan, Karnofsky rating, CT scan, and steroid dose were recorded. The clinical neurological examination, radionuclide scan, CT scan, and steroid dose were graded as unchanged (), unequivocally better (+2), or unequivocally worse (-2). Criteria for tumor progression, as described by Levin, et al.?,~ were -2 scores on any two of the following three tests: CT scan, radionuclide scan, or clinical examination on any evaluation conducted more than 3 months after irradiation was completed. This 3-month grace period after completion of radiation therapy allowed adequate time for resolution of postoperative artifacts, oncolytic effects of irradiation, and the delayed postirradiation reaction. ~ For the present purposes, we also assumed that resolution of the preceding three reactions had occurred within 3 months after the end of irradiation. If incorrect, these assumptions would lead to slight overestimates of time to progression for some patients. At each evaluation period, patients were re-treated with BCNU. Patients returned at 8-week intervals for similar evaluations and comparisons. Of the 7 patients admitted, 61 constitute the valid study group; four patients withdrew from treatment or were protocol violators, four patients died of unrelated medical problems, and one patient who was thought t~r have a necrotic glioblastoma at operation was found postmortem to have cerebral metastasis from a primary pancreatic carcinoma. Initial preirradiation CT scans, with and without contrast enhancement, were reviewed, and four measurements of the Dunn Photocopy were made with the appropriate correction for magnification. They were: 1) midline shift; 2) greatest diameter of peritumoral low-density area; 3) greatest diameter of contrast enhancement; and ) greatest diameter of intratumor low-density area within an enhanced tumor. Arbitrarily, contrast enhancement was used to define the periphery of each tumor; all tumors enhanced to some degree. In a few cases where the enhancing area was insufficient to define a tumor, imaginary lines were drawn in to produce a rounded or globular configuration to the area to be designated as tumor. Lowdensity areas outside of the tumor (peritumor low density) were presumed to represent edema, although we recognized that these regions included hypodense, nonenhancing tumor. Low density inside of the tumor was considered to indicate a poorly perfused region, which in many scans represented varying degrees of necrosis. Preliminary examination of the data using scatterplots and Kendall rank correlations between the tumor measurements and time to progression/censoring suggested that certain quantities derived from the original tumor measurements would be more useful predictors than the raw data. These transformations are summarized in the Results section. Two classes of survival time distribution models, the exponential and Weibull, were fitted by maximum likelihood. The Weibull was formulated in a manner given by Johnson and Kotz. 7 The general approach for handling censoring times followed that of Zippin and Armitage. 15 Model parameters were related to linear combinations of covariates as described by Cox and Snell 5 to force the parameter estimates to be positive; other approaches follow those used by Myers, et al. 9'1~ Tentative models involving all of the transformed predictors were refined with a "t-directed search" process. 8 Using t statistics, defined by an estimated coefficient value divided by its estimated standard deviation obtained from the estimated asymptotic covariance matrix, the predictor with smallest absolute t value less than 2., if any, was eliminated and the reduced model was fitted. For the final models obtained, the significance of each retained predictor also was tested by a "chi-square to remove" statistic, the difference in -2 (log likelihood) between the final model and a model further reduced by eliminating the given predictor. Such statistics have a limiting chisquare distribution with one degree of freedom, under standard assumptions. 3 Results The relationship of neurological examination, CT scan, and radionuclide scan evaluation scores to CT parameter measurements and time to tumor progression for each evaluation period is summarized in Table 1. At all evaluation periods, it appears that the best prognosis (no -2, or change of -2 to +2) is associated with a large peritumor low-density region and a small difference between the diameter of contrast enhancement and intratumor low density. Midline shift is less well correlated. Table 2 summarizes CT measurements for patients who continued beyond 26 weeks and beyond 52 weeks without progression, and for those who progressed earlier. As compared to the earlier progressing group, the group that continued beyond 26 weeks exhibits significantly lower means (p <.5) in contrast enhancement diameter, intratumor low-density diameter, and estimated volume of enhanced tumor (V3, defined below). The group that continued beyond 52 weeks, as compared to those who progressed earlier, exhibits a significantly higher mean in peritumor lowdensity diameter, as well as significantly lower means in diameter of contrast enhancement and estimated J. Neurosurg. / Volume 52 / May,

3 V. A. Levin, W. F. Hoffman, D. C. Heilbron and D. Norman TABLE 1 Relationship of evaluation scores, computerized tomography (CT) measurements, and median time to tumor progression* Evaluation Score CT Measurements (average (ram) 5: SE) Median Midline Peritumor Contrast Intratumor Time No. of Shift Low Density Enhancement Low Density (w~ks) Cases 1st evaluation no : or more : : to , no change : nd evaluation no : or more :.7 65:6 5: to : :8 38 5: 23 5: , no change 6.5 5:1. 6 5: : both 1st & 2nd evaluation --2, no change t to : rd evaluation no--2.-t or more : to : , no change 6.6 5: :8 39 5: *Evaluation scores: neurological exam, CT scan, and radionuclide scan (--2 = definitely worse; +2 = definitely better; --2 to +2 = initial deterioration reversed during subsequent courses of therapy; --2 no change = irreversible worsening and continued deterioration). volume of contrast-enhanced tumor. Because the peritumor low-density area was not evident in 27% of these cases ( value assumed), an indicator for the presence of peritumor low density was also examined (I2, defined below). The association of the presence of peritumor low density with later versus earlier tumor progression barely failed to be significant at 26 weeks (p =.51) and was significant (p =.39) at 52 weeks, by Fisher's exact test. To better evaluate the association of CT measurements with time to progression while properly accounting for the fact that eight patients had not suffered tumor progression (at times exceeding 6 weeks), analyses were undertaken of time to progression or censoring. The CT parameter measurements and time to progression were evaluated using an exponential model and a Weibull model. The models were best-fitted using a program developed by one of TABLE 2 Comparison of computerized tomography measurements to time to tumor progression (average + SE) Peritumor Intratumor Estimated Contrast Patient Midline Low- Enhancement Low- Volume of Shift Density Diameter Density Enhanced Group (mm) Diameter Diameter Tumor (mm) (mm) (mm) (cu cm) 8 patients continuing _> 26 wks (n = 6) 1 patients progressing at < 26 wks (n = 8) p, test of means.18 2 patients continuing _> 52 wks patients progressing at wks (n = 3) p, test of means.6 *Separate variances t-test, with adjusted degrees of freedom (df) as given _ " * (1 df) _ k * (5 dr) 6 J. Neurosurg. / Volume 52 / May, 198

4 Correlation of CT scan with prognosis us (D.C.H.); the Weibull proved to be the more appropriate model. Of the 61 patients, 58 had sufficient measurements for the analysis; the rest lacked values for all CT parameters. The following transformed variables were used: Z 1 = midline shift (cm) if measured;, if midline shift not measured ( of 58 cases) I2 = 1, if peritumor low density was measurable;, if absent V3 = estimated volume of enhanced tumor = r/6 (X38 - X3) V = estimated volume of intratumor low-density area = r/6 X s, where X3 = diameter of enhanced tumor (cm), and X = diameter of intratumor low-density area (cm). The tumor and intratumor low-density regions are considered to be approximated by spheres. The derived model for time to tumor progression states that the probability that time to tumor progression, T, will exceed any fixed time t is given by: P(T > t) = EXP(-[t/B]A), (1) where A = EXP (Oeo) and B = EXP (/3 +/32 I2 + r v3). Table 3 summarizes the parameters, estimated values, and significance of the parameters in this study. Figure 1 shows the estimated relationship of I2 and V3 to P(T > t), the probability of a tumor not progressing by time t (generically, the survival probability). Originally, the Weibull model was fitted with all four covariates (midline shift, tumor enhancement, intratumor low density, peritumor low-density in- TABLE 3 Parameters for Weibull model Estimated Chi-Square Param- Estimated S.D. of t-value to Remove P eter Value, Estimate* c~ / / / *S.D. = standard deviation. dicator) entering the B term of the model. Also considered were models with one covariate entering the A term, and the other three the B term; none of these significantly improved the fit (as indicated by the likelihood ratio test) relative to models with only the constant term in A. Models with high-density intratumor volume entering both the A and B terms were explored with similar results. Models that generalized the derived model by inclusion of quadratic and interaction effects did not significantly improve the fit (p >.1). Goodness of fit of the derived model cannot be rigorously tested with this type of model and data; however, two kinds of relevant assessment were made. First, survival probabilities corresponding to a Cox survival time model, with the hazard proportionality factor B A evaluated at the fitted coefficients for the Weibull model, were computed using an augmented version of the NONMEM program. 1 For this version FIG. 1. Probability of a tumor not progressing by time t (in weeks), computed using the Weibull model and Equation 1. Left." The probability when 12 = 1 (peritumor low density). Right." The probability when I2 = (no peritumor low density). The percentages associated with the curves are the percentile rank for V3 (enhanced tumor volume; 1% = 3 cc and 9% = 85 cc). The statistical significance of the curves is a function of the values given in Table 3. J. Neurosurg. / Volume 52 / May,

5 TABLE Observed versus expected survivors by quintiles of estimated survival probability V. A. Levin, W. F. Hoffman, D. C. Heilbron and D. Norman Survival No. of (wks) Quintile Observed Expected Cases I I I1 FIG. 2. Comparison of the estimated probability of survival (actually, failure of tumor to progress) based on the Cox model and the derived Weibull model. of the Cox model, the estimated hazard at time t for covariate values z is h(t,z) = B -A h(t,), (2) where h(t,) is the estimated Cox hazard. For the corresponding Weibull model, the estimated hazard is h(t,z) = B -A At ~A 1~. (3) TABLE 5 Observed versus expected survivors by 12 and V3 group* Survival Group Obser- Expec- No. of (wks) 12 V3 (cu cm) ved ted Cases > > > > > > = for no peritumor low density, 12 = 1 for peritumor low density. V3 = volume of tumor enhancement. As Fig. 2 indicates, these Cox model estimated survival probabilities agree fairly closely with the corresponding probabilities for the derived Weibull model. (Attempts to directly fit Cox models with this and related forms of hazard proportionality factor failed due to numerical difficulties.) In the second kind of assessment, estimates of the expected number of survivors computed from the derived Weibull model were compared with actual numbers of survivors, in various subsets of the sample. Table presents observed versus expected numbers of survivors at 26, 52, and 78 weeks, by quintiles of the estimated survival probability, and Table 5 by intervals of the covariates (I2, V3). Due to censoring, the number of patients evaluated at 78 weeks is reduced to 55. The deviates (Observed-Expected)/(Expected) v~ are all less than 1.9 in magnitude, suggesting a reasonable fit. Discussion It has become apparent with current treatment that the end point of therapeutic trials should not be survival time but the time to unequivocal progression of the tumor. Time to progression is frequently a long interval that would permit consideration of reoperation and initiation of alternative chemotherapeutic and/or radiotherapeutic regimens that could prolong life. Serial CT scanning has become an essential diagnostic tool for following patients under therapy. 8,1~ The scan-to-scan status of variable parameters related to the tumor can be important for maintaining the efficacy and determining the value of treatment. However, interpretation of low density versus contrast enhancement in CT scans is still incompletely understood? 2 It is reasonable to assume that the extent of contrast 66 J. Neurosurg. / Volume 52 / May, 198

6 Correlation of CT scan with prognosis enhancement (for instance, volume of enhancement, V3) reflects tumor cell burden. Although tumors are not always true spheres, the tumor cell burden can be estimated by subtracting the estimated tumor lowdensity volume from the total tumor volume. Patients whose tumor progressed at or after 26 weeks postirradiation had a mean cell burden of ~ while those whose tumor progressed in less than 26 weeks had a mean cell burden of ~ (assuming 1 cc = 19 cells). It is clear from Fig. 1 right that the estimated probability of patients continuing 1 year without tumor progression is fourfold greater (2% versus 1 I%) if the enhanced tumor volume (V3) is in the 1th percentile (3 cc) rather than the 9th percentile (85 cc). The presence of a peritumor low-density area increased the probabilities to 68% and 32%, respectively. The finding of significantly increased time to progression in patients with surrounding peritumoral low density as opposed to similar sized contrastenhancing lesions in other patients with absent low density was unexpected, and is difficult to explain. In this series of patients, there was no relationship between steroid usage during the first three evaluation periods and the presence or absence of peritumor low density; time to tumor progression was independent of dexamethasone dose for the first 18 weeks of therapy. It is possible that the favorable effect of low density around a tumor is related to vasogenic edema and breakdown in the blood-brain barrier at the tumor periphery, which makes the passage of immune reactive cells or proteins possible. Indirect support for this proposition comes from the observation that peritumor low density was more frequent in patients in the third and fourth decades of life who are, in general, more immunologically competent than older patients. While the correlation of peritumor low density and age with time to tumor progression may reflect a favorable immune reaction to the tumor, additional data are needed to confirm this possibility. It must be emphasized that the models derived here are tentative. The transformations of originally measured variables were at least partly chosen to improve the fit of models on this set of data; hence, the p values for their estimated coefficients may be somewhat biased toward lower values. The sample size is modest for determination of models of this degree of complexity, and the lack of significance of certain variables should not be taken as establishing their lack of effect. Estimates of volumes based on conventional CT scans can certainly be improved upon with the use of more sophisticated computer measurements of CT scan densities. Acknowledgments We are deeply indebted to the neurosurgeons in private practice and at the Kaiser-Permanente Medical Centers of the Greater Bay Area for referring patients to us. We thank Neil Buckley for editorial assistance and Stuart Beal, Ph.D., for assistance in the use of NONMEM Statistical Computer Package. References 1. Beal SL, Sheiner LB: NONMEM Users Guide -- Part 1. Technical Report. San Francisco: Division of Clinical Pharmacology, University of California, Boldrey E, Sheline G: Delayed transitory clinical manifestations after radiation treatment of intracranial tumors. Acta Radiol (Ther) 5:5-1, Bradley RA, Gart J J: The asymptotic properties of ML estimators when sampling from associated populations. Biometrica 9:25-21, Cox DR: Regression models and life tables. J R Stat Soc (Series B) 3:187-22, Cox DR, Snell E J: A general definition of residuals. J R Stat Soc (Series B) 3:28-275, 1968' 6. Daniel C, Woods F: Fitting Equations to Data. New York: John Wiley and Sons, Johnson NL, Kotz S: Continuous Univariate Distributions. New York: Houghton Mifflin, Levin VA, Crafts DC, Norman DM, et al: Criteria for evaluating patients undergoing chemotherapy for malignant brain tumors. J Neurosurg 7: , Mantel N, Myers M: Problems of convergence of maximum likelihood iterative procedures in multiparameter situations. J Am Stat Assoc 66:8--91, Marks JE, Gado M: Serial computed tomography of primary brain tumors following surgery, irradiation, and chemotherapy. Radiology 125: , Myers MH, Hankey BF, Mantel N: A logisticexponential model for use with response-time data involving regressor variables. Biometrics 29: , Norman D, Enzmann DR, Levin VA, et al: Computerized tomography in the evaluation of malignant glioma before and after therapy. Radiology 121:85-88, Pay NT, Carella R J, Lin JP, et al: The usefulness of computed tomography during and after radiation therapy in patients with brain tumors. Radiology 121:79-83, Wilson CB, Crafts DC, Levin VA: Brain tumors: criteria of response and definition of recurrence. Natl Cancer Inst Monogr 6:197-23, Zippin C, Armitage P: Use of concomitant variables and incomplete survival information in the estimation of an exponential survival parameter. Biometrics 22: , 1966 This work was supported in part by NIH Center Grant CA and an American Cancer Society Faculty Research Award FRA-155 to Dr. Levin. Address reprint requests to: Victor A. Levin, M.D., The Brain Tumor Research Center, Department of Neurological Surgery, University of California, San Francisco, California Neurosurg. / Volume 52 / May, i98 67