Ultra-early predictive assay for treatment failure using functional magnetic resonance imaging and clinical prognostic parameters in cervical cancer

Authors


Abstract

BACKGROUND:

The authors prospectively evaluated magnetic resonance imaging (MRI) parameters quantifying heterogeneous perfusion pattern and residual tumor volume early during treatment in cervical cancer, and compared their predictive power for primary tumor recurrence and cancer death with the standard clinical prognostic factors. A novel approach of augmenting the predictive power of clinical prognostic factors with MRI parameters was assessed.

METHODS:

Sixty-two cervical cancer patients underwent dynamic contrast-enhanced (DCE) MRI before and during early radiation/chemotherapy (2-2.5 weeks into treatment). Heterogeneous tumor perfusion was analyzed by signal intensity (SI) of each tumor voxel. Poorly perfused tumor regions were quantified as lower 10th percentile of SI (SI[10%]). DCE-MRI and 3-dimensional (3D) tumor volumetry MRI parameters were assessed as predictors of recurrence and cancer death (median follow-up, 4.1 years). Their discriminating capacity was compared with clinical prognostic factors (stage, lymph node status, histology) using sensitivity/specificity and Cox regression analysis.

RESULTS:

SI(10%) and 3D volume 2-2.5 weeks into therapy independently predicted disease recurrence (hazard ratio [HR], 2.6; 95% confidence interval [95% CI], 1.0-6.5 [P = .04] and HR, 1.9; 95% CI, 1.1-3.5 [P = .03], respectively) and death (HR, 1.9; 95% CI, 1.0-3.5 [P = .03] and HR, 1.9; 95% CI, 1.2-2.9 [P = .01], respectively), and were superior to clinical prognostic factors. The addition of MRI parameters to clinical prognostic factors increased sensitivity and specificity of clinical prognostic factors from 71% and 51%, respectively, to 100% and 71%, respectively, for predicting recurrence, and from 79% and 54%, respectively, to 93% and 60%, respectively, for predicting death.

CONCLUSIONS:

MRI parameters reflecting heterogeneous tumor perfusion and subtle tumor volume change early during radiation/chemotherapy are independent and better predictors of tumor recurrence and death than clinical prognostic factors. The combination of clinical prognostic factors and MRI parameters further improves early prediction of treatment failure and may enable a window of opportunity to alter treatment strategy. Cancer 2010. © 2010 American Cancer Society.

For advanced cervical cancer patients, the first-time radical treatment is the best and may be the only opportunity to achieve a complete cure. Unfortunately, favorable clinical prognostic factors do not always translate into the successful ultimate outcome of primary tumor control and disease-free survival. Advanced cervical cancer is a prime example of a disease in which the early prediction of ultimate treatment outcome is critical but has been challenging to achieve. Current gold standard clinical prognostic factors, stage,1 lymph node status2 and histology,3 are established criteria for the selection of first-line therapy, but have been challenged by the variability of outcomes within each prognostic category.4, 5 The efficacy of radiation and chemotherapy generally cannot be reliably assessed until months after completion of all therapy by estimating tumor bulk using clinical palpation or by cervical cytology.6 Once treatment failure occurs, further treatment options are severely limited, salvage therapy is rarely effective, and outcome is dismal.7, 8 Therefore, individualized and therapy-specific predictors of primary tumor control and survival that are available before or early during a specific therapy regimen are needed to provide a window of opportunity to modify treatment strategy up front and improve survival.

Ultra-early individualized outcome prediction that is available within weeks of therapy start critically depends on the ability to quantify tumor heterogeneity and heterogeneous responses in patients with the same pretreatment clinical prognostic factors. Reliable clinically practical and noninvasive methods to effectively evaluate the heterogeneous therapy response in solid tumors have been elusive in clinical practice. Perfusion, reflecting the effectiveness of delivery of chemotherapy,9, 10 and oxygenation for radiotherapy (RT),11, 12 can be highly variable within each solid tumor. Essential information regarding this heterogeneity of perfusion13 during treatment has not been incorporated into the treatment strategy despite its known profound impact on therapy outcome in cervical cancer.9, 10, 13-18 Furthermore, accuracy of tumor volume measurement, as a early response indicator during treatment, has also been limited because of the irregular tumor shapes and subtlety of early volume regression.19

Advances in morphologic and functional imaging provide an opportunity to quantify each tumors' heterogeneity with dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI),20-22 and its heterogeneous early response to a specific ongoing treatment with improved 3-dimensional (3D) MRI volume measurement.19, 23-25 In a prior exploratory study we have identified MRI parameters of low tumor perfusion during early therapy and high 3D tumor volume as predictors of ultimate treatment failure.26

The purpose of this study was to 1) validate and refine DCE and 3D volumetry MRI parameters for the prediction of ultimate treatment outcome (tumor recurrence and death) of the ongoing cytotoxic treatment in cervical cancer, 2) compare the discriminatory value of the MRI parameters with that of the clinical prognostic factors, and 3) develop a novel predictive paradigm by synergizing the prognostic power of the pretreatment clinical prognostic factors and those of the MRI parameters obtained during early therapy.

MATERIALS AND METHODS

Patient Characteristics

Sixty-two patients with biopsy-proven advanced cervical cancer (excluding small cell/neuroendocrine histology), who were treated with RT and concurrent cisplatin-based chemotherapy, were studied prospectively on an institutional review board-approved imaging protocol. Pretreatment evaluations included routine workup following International Federation of Gynecology and Obstetrics (FIGO) guidelines.27 Therapy was comprised of standard RT in combination with concurrent weekly cisplatin chemotherapy (40-50 mg/m2). RT included external beam radiation (dose to the pelvis, 31-66 grays [Gy]; mean, 46.2 Gy), combined with brachytherapy in all but 2 patients. Fifty patients received low‒dose rate brachytherapy (dose, 17-62 Gy; mean, 42 Gy), and 12 received high‒dose rate brachytherapy (30 Gy in 5 fractions). An average of 4.7 (range, 3-6) courses of cisplatin were given with RT. In 3 patients, the cisplatin was combined with 5-fluorouracil, and in 3 with paclitaxel.

MRI

MRI was obtained pretherapy and early during RT, after a dose of 20 to 25 Gy in 2 to 2.5 weeks and 2 cycles of chemotherapy. MRIs were performed with a 1.5-Tesla scanner and included sagittal T2-weighted imaging for 3D tumor volumetry measurement (Fig. 1) described in detail earlier.24, 28 DCE MRI was acquired for 158 to 168 seconds using a T1-weighted 3D gradient echo multislice sequence (repetition time/echo time, 12 ms/5 ms; flip, 25 degrees; matrix, 138 × 256; partition, 12; number of excitations, 1; slab, 8.0 cm) after 3 precontrast baseline acquisitions, and followed by a bolus of MRI contrast agent (0.1 mmol/kg) at 5 mL/second. To cover the entire tumor volume, the temporal resolution was reduced to 26 to 28 seconds, remaining appropriate for plateau phase signal intensity (SI) assessment for each tumor voxel. The patients' therapy was not influenced by the MRI findings.

Figure 1.

(A) High dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) at 2.5 weeks of chemotherapy/radiotherapy (RT) is shown. In this 52-year-old woman with stage IIB squamous cell carcinoma of the cervix, MRI was obtained 2.5 weeks after the initiation of RT at a dose of 21.6 grays (Gy) in 12 fractions and 2 cycles of weekly chemotherapy with cisplatin. The volumetric MRI demonstrates the cervical tumor region (indicated by arrows). The DCE MRI shows intense heterogeneous enhancement (indicated by arrows). The voxel histogram distribution of DCE demonstrates high and heterogeneous signal intensity (SI) of the tumor and a high lower 10th percentile of SI (SI[10%]) of 3.13. The patient was alive and well without primary tumor recurrence or metastasis at the time of last follow-up, 5 years and 2 months after the completion of therapy. (B) Low DCE at 2.5 weeks of chemotherapy/RT is shown. In this 42-year-old woman with stage IIB adenocarcinoma of the cervix, MRI was obtained 2.5 weeks after the initiation of RT at a dose of 21.6 Gy in 12 fractions and 2 cycles of weekly chemotherapy with cisplatin. The volumetric MRI demonstrates the cervical tumor region (indicated by arrows). The DCE MRI shows very low and heterogeneous enhancement (indicated by arrows). Voxel histogram distribution of DCE demonstrates overall much lower SI than in the patient in shown in Panel A. The SI(10%) was 1.70. This patient had a recurrence of the primary tumor 1 month after the completion of therapy. Pelvic exenteration was attempted, but the tumor was unresectable and the patient died of cervical cancer 12 months later. U indicates uterus; B, bladder.

DCE Tumor Heterogeneity Analysis and 3D Volumetry Analysis

The DCE MRI imaging data sets were analyzed using the voxel-by-voxel analysis developed in our laboratory.26 After contouring of the tumor, a DCE time–SI curve was generated for each voxel in the tumor as described in detail earlier.26 SI voxel histograms (Fig. 1) of the voxel population within the entire 3D tumor region were generated to derive the mean SI, median SI, variance, standard deviation, kurtosis of the SI distribution, and SI percentiles. On the basis of our earlier exploratory study in a different patient population,26 the 10th percentile of SI (SI[10%]) in the early therapy phase, representing the proportion of low-perfusion regions within the heterogeneous tumor, was the strongest predictor of tumor control. Residual 3D tumor volume also showed a promising correlation with tumor control.

SI percentiles in increments of 2.5% from SI(2.5%) to SI(40%) were generated to quantify the degree and proportion of the poorly perfused low-DCE components within the tumor. An SI percentile was defined as the SI value below which a given percentage of voxels within the histogram falls; for example, SI(10%) is the SI value below which the lowest 10% of voxels falls (Fig. 1). For volumetry, 3D volume measurement was performed for each sequential MRI by delineating the tumor on all the T2-weighted images as described earlier.24, 28

Outcome and Data Analysis

The outcome endpoint primary tumor recurrence (recurrence) was defined as tumor regrowth, persistence, or progression in the uterus or pelvis after completion of therapy. Cancer death (death) was defined as death from cervical cancer or cancer complications. Patients were evaluated after therapy at 3-month to 12-month intervals until death or data analysis. The median post-therapy follow-up of surviving patients was 4.1 years (range, 0.2-7.9 years).

The DCE MRI parameters identified in the SI voxel histogram and 3D volume, and the clinical prognostic factors, were analyzed for their discriminating capacity in predicting recurrence versus tumor control and death versus survival. Optimal SI percentile parameters across the range of SI percentiles, and optimal parameter thresholds, were determined by receiver operating characteristic (ROC) analysis. The threshold yielding the maximum Ji = {sensitivityi + specificityi − 1} for predicting disease recurrence, in which i represents each possible threshold, was selected. These same values were used as the threshold for cancer death sensitivity and specificity analysis. Sensitivity was calculated as the number of predicted positives divided by the number of true‒positives (7 recurrences and 14 deaths); specificity was calculated as predicted negatives divided by true‒negatives (55 recurrences and 48 deaths). Confidence intervals were calculated from the exact binomial distribution.

A combined clinical prognostic factor parameter categorized patients as unfavorable if they had at least 1 of lymph node involvement, stage III or IV disease, or nonsquamous histology. Patients not fitting any of these 3 criteria were categorized as favorable. Each clinical prognostic factor was weighted equally. To evaluate the combination of MRI parameters with and without clinical prognostic factors, thresholds were determined from logistic regression models for recurrence. The probability value generating the maximum {sensitivity + specificity − 1} for recurrence was selected as the optimal threshold. Probabilities of disease recurrence and death were estimated based on the combined SI(10%) and 3D volume thresholds.

Cox regression was used to confirm the predictive significance of MRI parameters for the time to disease recurrence and time to death outcomes. MRI parameters were evaluated as continuous variables; clinical prognostic factor parameters were dichotomous. A 2-sided α-level of .05 was used for the MRI parameters. No adjustments for multiplicity were made because of the confirmatory 1-sided nature of the tests.

RESULTS

Patient Characteristics

The distribution of stage, lymph node involvement, histology, and treatment variables is presented in Table 1. Tumor characteristics were equally distributed with respect to treatment variables. Ten patients had FIGO stage IB2 (bulky), 7 had stage IIA, 26 had stage IIB, 2 had stage IIIA, 14 had stage IIIB, and 3 had stage IVA tumors. The median age at diagnosis was 53.3 years (range, 29-75 years).

Table 1. Patient Characteristics
Patient CharacteristicsAll Patients, % (No.)aRT DosebRT DurationbPrimary RecurrenceDisease- Specific Survival
≥85 Gy<85 GyPc<55 Days≥55 DaysPcHR (95% CI)PdHR (95% CI)Pd
  • RT indicates radiotherapy; Gy, grays; HR, hazard ratio; 95% CI, 95% confidence interval; FIGO, International Federation of Gynecology and Obstetrics.

  • Combined clinical prognostic factors consist of stage (I-II vs III-IV), lymph node involvement (uninvolved vs involved), and histology (squamous vs nonsquamous). Unfavorable factors are characterized by any of the following: lymph node involvement, stage III to IV, or nonsquamous histology. Favorable factors are characterized by none of these.

  • a

    Numbers in parentheses are patient numbers.

  • b

    Treatment variables include RT dose and RT duration as significant treatment-related factors influencing RT outcome. Statistical significance in the groups with a treatment duration of <55 days versus ≥55 days is contrary to expectation, with less recurrences noted in the protracted treatment group.

  • c

    P value was derived using the chi-square test.

  • d

    P value was derived from Cox univariate analysis.

FIGO stage   .850  .748 .49 .25
 I-II69% (43)9% (6)60% (37) 32% (20)37% (23) 1.7 1.9 
 III-IV31% (19)5% (3)26% (16) 13% (8)18% (11) (0.4-7.6) (0.6-5.4) 
Lymph nodes   .705  .449 .28 .03
 Uninvolved73% (45)11% (7)62% (38) 31% (19)42% (26) 2.3 3.2 
 Involved27% (17)3% (2)24% (15) 14% (9)13% (8) (0.5-10.2) (1.1-9.1) 
Histology   .185  .983 .07 .004
 Squamous82% (51)10% (6)72% (45) 37% (23)45% (28) 4.0 4.9 
 Nonsquamous18% (11)5% (3)13% (8) 8% (5)10% (6) (0.9518.1) (1.7-14.4) 
Recurrence   .262  .002
 No89% (55)11% (7)78% (48) 34% (21)55% (34)     
 Yes11% (7)3% (2)8% (5) 11% (7)0% (0)     
Cancer death   .404  .102
 No77% (48)10% (6)67% (42) 30% (19)47% (29)     
 Yes23% (14)5% (3)18% (11) 15% (9)8% (5)     
Overall death   .577  .301
 No74% (46)10% (6)64% (40) 31% (19)43% (27)     
 Yes26% (16)5% (3)21% (13) 15% (9)11% (7)     

Individual MRI Parameters

The overall results indicated that low-DCE patterns, indicative of poor tumor perfusion and hypoxia, and high tumor volumes in the very early treatment phase (at 2-2.5 weeks of therapy) were associated with unfavorable therapy outcome, confirming the MRI parameters identified in our exploratory study.26 Among the DCE parameters, the heterogeneity-based SI percentile in the early treatment phase at 2 to 2.5 weeks of therapy provided the best prediction of outcome. ROC analysis across the range of SI percentiles demonstrated that SI(10%) had the highest discriminating capacity (Fig. 2).

Figure 2.

Receiver operator characteristic (ROC) analysis of signal intensity (SI) percentiles is shown. ROC analysis was performed to determine the best SI percentile (SI[%]) for the discrimination of disease recurrence (solid blue curve) and cancer death (dashed red curve). Area under the curve values of SI(2.5%) to SI(30%) are above 0.70, thereby supporting the usefulness of the SI parameter across a range of percentiles associated with low dynamic contrast-enhancement and indicative of poor perfusion and hypoxia. Consistent with the results of the exploratory study,26 SI(10%) shows the best discriminating capacity.

Univariate Cox regression indicated that both SI(10%) and tumor volume significantly correlated with both time to disease recurrence and time to death. Lower SI(10%) and higher volume were associated with higher hazard ratios for recurrence (P = .04 and P = .03, respectively) and death (P = .03 and P = .01, respectively) (Table 2). SI(10%) and volume remained significant predictors in a 3-parameter model containing SI(10%), volume, and the combined clinical prognostic factors for recurrence (hazard ratio [HR], 4.3; 95% confidence interval [95% CI], 1.3-15.1 [P = .02] and HR, 2.7; 95% CI, 1.2-6.0 [P = .03], respectively) and for death (HR, 2.6; 95% CI, 1.2-5.6 [P = .01] and HR, 2.2; 95% CI, 1.3-3.7 [P = .003], respectively). The correlation between SI(10%) and 3D volume was .03, indicating minimal or no overlap in the information provided by each.

Table 2. Univariate Hazard Ratio Estimates for Primary Tumor Recurrence and Cancer Death
ParameterPrimary Tumor RecurrenceDeath From Cancer
HR (95% CI)aPHR (95% CI)aP
  • HR indicates hazard ratio; 95% CI, 95% confidence interval; SI(10%), 10th percentile of signal intensity; 3D, 3-dimensional; FIGO, International Federation of Gynecology and Obstetrics.

  • a

    HR estimates and 95% CIs were derived using univariate Cox regression models for time to primary tumor recurrence and time to cancer death.

  • b

    For a decrease of 1 standard deviation (decrease of 0.5).

  • c

    For an increase of 1 standard deviation (increase of 24).

  • d

    Stage III to IVA disease versus stage I to II disease.

  • e

    Yes versus no.

  • f

    Nonsquamous versus squamous.

SI(10%)b2.6 (1.0-6.5).041.9 (1.0-3.5).03
3D volumec1.9 (1.1-3.5).031.9 (1.2-2.9).01
FIGO staged1.7 (0.4-7.6).491.9 (0.6-5.4).25
Lymph node involvemente2.3 (0.5-10.2).283.2 (1.1-9.1).03
Histologyf4.0 (0.9-18.1).074.9 (1.7-14.4).004

Sensitivity and Specificity of Parameters

Threshold values for SI(10%) and 3D volume were explored. The thresholds to discriminate patients with favorable versus unfavorable outcome were equivalent to the lowest SI(10%) and the highest 3D volume values with no recurrence event. This corresponds to the goal of maximizing the sensitivity in predicting patients with unfavorable outcome, because these patients are the target group requiring change in therapy strategy to improve outcome. On the basis of this method, the discriminatory thresholds were 2.0 for SI(10%) and 22 cm3 for 3D volume.

With these thresholds, the early therapy SI(10%) <2.0 predicted recurrence and death with sensitivities of 100% and 79%, respectively (Table 3). The 3D volume threshold of ≥22 cm3 predicted recurrence with the same sensitivity but lower specificity than SI(10%), and death with higher sensitivity, but lower specificity (Table 3). The SI(10%) threshold of 2.0 in the current patient population was close to the threshold of 2.5 in our previous exploratory study.26 When applied to our current patients, the prior threshold of 2.5 demonstrated sensitivities similar to the 2.0 threshold, but inferior specificities for predicting disease recurrence and death (18% and 19%, respectively) (Table 3). The mean SI, averaging the SI values over the entire tumor region without heterogeneity assessment, had lower specificities (33% and 33%, respectively) than the heterogeneity-based SI(10%) (56% and 58%, respectively) in predicting recurrence and death (Table 3).

Table 3. Discriminating Capacity of MRI Parameters, Combined CPF Parameters, and Combined MRI/CPF Parameters to Predict Primary Tumor Recurrence and Cancer Death
Type of ParameterParametersPrimary Tumor RecurrenceaDeath by Cancera
Sensitivity (95% CI)Specificity (95% CI)Sensitivity (95% CI)Specificity (95% CI)
  • MRI indicates magnetic resonance imaging; CPF, clinical prognostic factor; 95% CI, 95% confidence interval; SI(10%), 10th percentile of signal intensity; mSI, mean signal intensity; 3D, 3-dimensional.

  • a

    Sensitivity and specificity estimates (in percentage) for comparing the discrimination of logistic regression models.

  • b

    The SI(10%) threshold of <2.0 versus ≥2.0 was determined by receiver operating characteristic analysis in the current patient study population. The SI(10%) threshold of the independent patient population in the previous exploratory study24 was 2.5. Using the threshold of 2.5, the sensitivities and specificities were 100% (59%-100%) and 18% (9%-31%), respectively, for tumor recurrence, and 93% (66%-100%) and 19% (9%-33%), respectively, for death.

  • c

    ≤3.25 vs >3.25.

  • d

    ≥22 cm3 vs <22 cm3.

  • e

    Probability threshold from logistic regression model, based on both SI(10%) and 3D volume.

  • f

    International Federation of Gynecology and Obstetrics stage III to IV disease versus stage I to II disease.

  • g

    Involved lymph nodes versus uninvolved lymph nodes.

  • h

    Nonsquamous cell (adenocarcinoma or undifferentiated carcinoma) versus squamous cell carcinoma.

  • i

    Unfavorable (defined as having any lymph node involvement or stage III-IV disease or nonsquamous histology) versus favorable (defined as having none of these).

  • j

    Probability thresholds derived from logistic regression were identified separately for each CPF risk group: unfavorable (any lymph node involvement or stage III-IV disease or nonsquamous histology) and favorable (none of these).

MRISI(10%)b100 (59-100)56 (42-70)79 (49-95)58 (43-72)
mSIc100 (59-100)33 (21-47)86 (57-98)33 (20-48)
3D volumed100 (59-100)42 (29-56)93 (66-100)46 (31-61)
Combined MRI parametersa,e100 (59-100)67 (53-79)93 (66-100)48 (33-63)
CPFStagef43 (10-82)71 (57-82)43 (18-71)73 (58-85)
Lymph node statusg43 (10-82)75 (61-85)50 (23-77)79 (65-90)
Histologyh43 (10-82)85 (73-94)43 (18-71)90 (77-97)
Combined CPF parametersi71 (29-96)51 (37-65)79 (49-95)54 (39-69)
Combined CPF/ functional MRI assayCombined MRIa,j and combined CPF100 (59-100)71 (57-82)93 (66-100)60 (45-74)

Comparison of the 2-Week to 2.5-Week Time Point With Pretherapy MRI Parameters

MRI parameters at the early therapy time point were compared with pretherapy parameters using Cox regression. Pretherapy SI(10%) was not found to be significantly associated with time to tumor recurrence (P = .06) or death (P = .08). Pretherapy tumor volume was associated with death (P = .02) but not recurrence (P = .16). In the sensitivity analysis, pretherapy tumor volume was found to be equivalent to early therapy volume for the prediction of recurrence, with a sensitivity of 100% and a relatively low specificity of 42% in both the pretherapy and early therapy MRI. However, pretherapy volume was inferior in predicting death, with a sensitivity of only 79%, compared with 93% for the early therapy volume. Similarly, the SI(10%) in early therapy demonstrated better prediction. The sensitivities of the early therapy SI(10%) in predicting recurrence and death were 100% and 79%, respectively, compared with 71% and 43%, respectively, for the pretherapy SI(10%).

Combined Heterogeneity DCE and 3D Volumetry Parameter (Combined MRI Parameter)

The early therapy MRI-parameters, SI(10%) and 3D-volume, were combined to derive a combined MRI-parameter. It improved the outcome prediction over that of the individual MRI parameters alone, resulting in a sensitivity of 100% and specificity of 67% for predicting tumor recurrence, and a sensitivity of 93% and specificity of 48% for predicting death (Table 3).

For the combined MRI parameter, the estimated probability of death was 53% in patients with both MRI parameters beyond the thresholds (SI[10%]<2 and 3D volume ≥22 cm3), 13% with 1 parameter beyond threshold, and 0% (upper confidence limit, 28%) with neither parameter beyond threshold (Table 4). The probability of recurrence with both MRI parameters beyond thresholds was 37%. Kaplan-Meier estimate of actuarial 5-year local control rate (ie, freedom from recurrence) was 58%, 100%, and 100%, respectively, for both, 1, and neither parameter beyond the threshold. The disease-specific and overall survival rates were 44%, 84%, and 100%, respectively, for both, 1, and neither parameter beyond the threshold; and 44%, 81%, and 90%, respectively.

Table 4. Combined MRI Parameters: Estimated Probabilities (Shown as Percentage) of Primary Tumor Recurrence and Cancer Death
MRIMRI Parameter DescriptionNo. of PatientsPrimary Tumor Recurrence, % (95% CI)Death From Cancer %, (95% CI)
  • MRI indicates magnetic resonance imaging; 95% CI, 95% confidence interval; SI(10%), 10th percentile of signal intensity; 3D, 3-dimensional.

  • a

    Unfavorable indicates that both MRI parameters were beyond the threshold.

  • b

    Intermediate indicates that 1 MRI parameter was beyond the threshold.

  • c

    The probability of disease recurrence in patients with at least 1 MRI parameter beyond the threshold is 0% by definition because of the selection of thresholds to produce 100% sensitivity for recurrence.

  • d

    Favorable indicates no MRI parameters beyond the threshold.

UnfavorableaSI(10%) <2.0 and 3D volume ≥22 cm31937 (16-62)53 (29-76)
IntermediatebSI(10%) <2.0 and 3D volume <22 cm3 or SI(10%) ≥2.0 and 3D volume ≥22 cm3320c13 (4-29)
FavorabledSI(10%) ≥2.0 and 3D volume ≤22 cm3110c0 (0-28)

Combined Clinical Prognostic Factor/MRI Predictive Assay

The combination of all the clinical prognostic factors and the MRI-parameters was evaluated. For the individual clinical prognostic factors (stage, lymph node status, and histology), sensitivities were ≤50% (Table 3). The combined clinical prognostic factors were dichotomized into unfavorable (any 1 of: stage III-IV disease, involved lymph nodes, or nonsquamous histology) versus favorable (none of these) categories.

Sensitivity/specificity analysis demonstrated that the combination of the MRI parameters with clinical prognostic factors further increased the discriminating capacity. The clinical prognostic factor/MRI predictive assay predicted all recurrences (sensitivity, 100%) and improved specificity to 71% over the combined MRI parameter (67%) and combined clinical prognostic factor (51%) alone. It predicted cancer death with the same sensitivity as the combined MRI parameter alone (93%), and improved the specificity to 60% compared with the combined MRI parameter alone (48%) and combined clinical prognostic factor (54%) alone (Table 3). Probabilities of tumor recurrence and death were highest (42% and 67%, respectively) in patients with both unfavorable MRI parameters and clinical prognostic factors, and were 29% and 29%, respectively, in patients with unfavorable MRI parameters despite favorable clinical prognostic factors (Table 5).

Table 5. Combined MRI and CPF Parameters (CPF/MRI Predictive Assay): Estimated Probabilities (Shown as Percentage) of Primary Tumor Recurrence and Cancer Death
MRIPrimary Tumor Recurrence, %Death From Cancer, %
CPF UnfavorableCPF FavorableCPF UnfavorableCPF Favorable
  • MRI indicates magnetic resonance imaging; CPF, clinical prognostic factor.

  • a

    Unfavorable indicates that both MRI parameters were beyond the threshold.

  • b

    95% confidence interval (95% CI), 15-72.

  • c

    95% CI, 4-71.

  • d

    95% CI, 35-90.

  • e

    Intermediate indicates that 1 MRI parameter was beyond the threshold.

  • f

    The probability of recurrence in patients with only 1 MRI parameter beyond the threshold is 0% by definition because of the selection of thresholds to produce 100% sensitivity for recurrence.

  • g

    95% CI, 4-41.

  • h

    95% CI, 0-34.

  • i

    Favorable indicates no MRI parameters beyond the threshold.

  • j

    95% CI, 0-71.

  • k

    95% CI, 0-37.

Unfavorablea 42b29c67d29c
Intermediatee  0f 0f17g7h
Favorablei  0f 0f 0j0k

DISCUSSION

Cervical cancer is a leading health problem worldwide. Standard chemoradiation therapy fails in approximately one‒third of patients with stage IB2-IVA disease.29, 30 The classic prognostic factors, FIGO stage,1 lymph node status,2 and histology,3 have been well established to guide therapy selection.31 However, they have been criticized as suboptimal in predicting outcome in the individual patient.4, 5 Specifically, the gold standard clinical prognostic factors have not accounted for heterogeneity of tumor responsiveness to the specific ongoing therapy, and have not been able to reliably predict failure early during the treatment course, when therapy adjustments are still feasible by radiation dose intensification, changes in the concurrent therapy, or addition of novel clinical trial therapies.

Our research was intended to develop an early predictive assay for ultimate outcome by augmenting essential information from both the established clinical prognostic factors and from predictive parameters for the effectiveness of treatment. We took a new approach to investigate parameters that reflect the effective delivery of cytotoxic agents and oxygen to the heterogeneous tumor, and, after delivery to the tumor, to assess the subtle early response to the specific treatment regimen. Our current study validates, with a different patient population and refined imaging methods, 2 independent MRI parameters as outcome predictors: the functional DCE parameter SI(10%), first discovered in our exploratory pilot study,26 and the morphologic 3D volume parameter. SI(10%), reflecting essential information on chemotherapy and oxygen delivery to the tumor, and residual 3D volume, assessing early treatment response, predicted ultimate treatment outcome at an ultra-early time point, within 2 to 2.5 weeks of the initiation of therapy.

Functional MRI: Heterogeneity-Based DCE Parameter SI(10%)

The DCE parameter SI(10%), which provides a heterogeneity-based quantification of poorly perfused regions within the tumor, was confirmed as a sensitive independent predictor of recurrence and cancer death (Fig. 1) (Tables 2 and 3). This finding is in keeping with our premise that the tumor perfusion status during therapy, which reflects the effectiveness of delivery of cytotoxic agents and oxygen to the tumor, critically influences the success of chemotherapy and RT. Poorly perfused cell populations, represented by low-DCE regions within the heterogeneous tumor, likely contribute critically to treatment failure. In cervical cancer, there is ample indirect evidence that poor tumor blood supply adversely affects treatment response,14, 15, 17, 18 and that hypoxia-imparted radioresistance results in tumor recurrence and decreased survival.13, 16 Similarly, the effectiveness of chemotherapy delivery to the tumor critically depends on adequate tumor perfusion.9, 10 DCE MRI characterizes microvascular structure and function,32, 33 indirectly assesses perfusion and oxygenation status,20, 34, 35 and correlates with microvessel density20 and with direct oxygen measurements22 in cervical cancer.

SI(10%), as developed in our laboratory, specifically assessed the intratumoral heterogeneity of perfusion, a well-known phenomenon in solid tumors11, 12 that has been most challenging to assess in clinical patients. Our voxel-analysis identifies, separates, and analyzes the heterogeneous low-perfusion (low-DCE) components within the tumor.11 The clinical importance of heterogeneity assessment is shown by our result that SI(10%) was far superior in predicting treatment failure than the mean SI that averages DCE over the entire tumor region (Table 3).

Morphologic MRI: 3D Volume and Combined MRI Parameter

Whereas the functional parameter SI(10%) likely reflects oxygen and chemotherapy delivery to the tumor during therapy as a fundamental precondition for tumor cell kill by a specific cytotoxic regimen, the improved volume measurement by 3D volumetry can now detect the actual early subtle response to the specific treatment in the individual patient. Irregular configurations and nonlinear shrinkage of the individual tumor cannot be appreciated by the traditional unidimensional or bidimensional tumor measurements or by physical examination. With the much higher precision of MRI23 and our measurement methodology, high residual 3D volume assessed in early therapy (2-2.5 weeks into treatment), independently predicted recurrence and death with much better precision (Tables 2 and 3). By combining both, SI(10%) and 3D volume, assessing the effective delivery and the early therapy response measure of the cytotoxic effect, respectively, the accuracy of the ultra-early outcome prediction further improved over the individual MRI parameters alone (Table 4). Such a new combined MRI predictor has not been applied in clinical care for cervical cancer.

Comparison With the Exploratory Study

The present confirmatory study validates SI(10%) as a predictive parameter, and further refines the SI(10%) threshold. The threshold value of 2.0 in our current patient population, separating favorable from unfavorable outcome, approximates the SI(10%) of 2.5 from our exploratory study.26 The small difference between the current threshold and the pilot study is not unexpected, and may be in part related to the larger patient number and improved imaging technique of the current study population. However, the optimal threshold must be further refined through future study.

Combined MRI/Clinical Prognostic Factor Predictive Assay

SI(10%) and 3D volume were superior to clinical prognostic factors in predicting therapy outcome (Tables 2 and 3). This does not invalidate gold standard clinical prognostic factors, but opens opportunities to augment their prognostic abilities with the MRI parameters. Our study explored this synergistic paradigm as a novel noninvasive MRI/clinical prognostic factor predictive assay by using the MRI predictors as adjunctive independent information. This further improved prediction of treatment failure to sensitivities of 100% and 93%, respectively, for recurrence and death. With the MRI/clinical prognostic factor predictive assay, probabilities of recurrence and death (42% and 67%, respectively) were highest in patients with both unfavorable clinical prognostic factors and unfavorable MRI parameters, but were also substantial in those with unfavorable MRI parameters despite favorable patient group based on clinical prognostic factors (29% and 29%, respectively) (Table 5). This suggests that unfavorable MRI parameters identified a high-risk subgroup within the favorable clinical prognostic factor category. Despite the small size of the subgroups in this exploratory analysis, these trends provide direction for further study to validate the predictive assay and use it clinically.

Although, to the best of our knowledge, the current study has the largest patient population and longest follow-up time of its kind published to date, the confidence intervals of our results were wide. However, the high sensitivity of 93% to 100% in predicting unfavorable outcome suggests that the combined clinical prognostic factor/MRI predictive assay is a promising tool to effectively identify up front most patients who will ultimately fail therapy. In this regard, the sensitivity to predict treatment failure may be the most important criteria. With the poor salvage options after failed first-line therapy in cervical cancer, such timely prediction is critical. The early availability of the predictive assay within 2 to 2.5 weeks of the initiation of therapy can serve to further personalize treatment strategies. This may profoundly impact therapy success 3-fold. First, early prediction of ultimate failure of ongoing standard therapy provides a window of opportunity to implement more intense treatment in those whose high risk of failure justifies the toxicity of more aggressive therapy. Second, prediction of success may avoid unnecessarily intense or experimental therapies in patients with low risk of failure, thereby reducing morbidity and healthcare cost. Third, incorporation of such refined predictive assays into clinical trial designs can streamline patient selection and improve efficacy of studies investigating novel therapies.

Our results support that this novel noninvasive approach is achievable in a 45-minute MRI that can be implemented in community clinical settings to allow broad patient access.

In conclusion, the results of the current study confirm the clinical feasibility and effectiveness of 2 independent MRI parameters, quantifying heterogeneous tumor perfusion and early therapy-specific 3D volume response, for predicting recurrence and death in patients with cervical cancer. The combined predictive powers of the pretreatment clinical prognostic factors and the independent early therapy MRI parameters can be used as a novel predictive assay for ultimate outcome of the specific ongoing treatment regimen, and may have profound impact to personalize therapy strategies in cervical cancer.

CONFLICT OF INTEREST DISCLOSURES

Supported by the National Institutes of Health (contract grant number: RO1 CA 71,906).

Ancillary