Disease volumes as a marker for patient response in malignant pleural mesothelioma

Abstract

Background

The goal of this study was to create a comprehensive model for malignant pleural mesothelioma patient survival utilizing continuous, time-varying estimates of disease volume from computed tomography (CT) imaging in conjunction with clinical covariates.

Patients and methods

Serial CT scans were obtained during the course of clinically standard chemotherapy for 81 patients. The pleural disease volume was segmented for each of the 281 CT scans, and relative changes in disease volume from the baseline scan were tracked over the course of serial follow-up imaging. A prognostic model was built using time-varying disease volume measurements in conjunction with clinical covariates.

Results

Over the course of treatment, disease volume decreased by an average of 19%, and median patient survival was 12.6 months from baseline. In a multivariate survival model, changes in disease volume were significantly associated with patient survival along with disease histology, Eastern Cooperative Oncology Group performance status, and presence of dyspnea.

Conclusions

Analysis of the trajectories of disease volumes during chemotherapy for patients with mesothelioma indicates that increasing disease volume was significantly and independently associated with poor patient prognosis in both univariate and multivariate survival models.

introduction

Any image-based response evaluation method has two components; the first describes a protocol for making measurements, and the second describes how to classify patients into response categories once those measurements are available. Tumor response assessment with medical images has focused on reducing the dimensionality of the first component and discretization of the second component, beginning with the World Health Organization [1] bi-dimensional measurement technique that used the product of two linear measurements as a quasi-two-dimensional metric to assess tumor response across serial scans. Progressive disease (PD) was considered an increase of ≥25% from the minimum of previous tumor measurements, and partial response (PR) was considered a decrease of 50% or more from the baseline tumor measurement. Tumors not meeting the criteria for either PD or PR were classified as stable disease. The Response Evaluation Criteria In Solid Tumors, or RECIST, criteria were developed to simplify this measurement process to a single longest tumor diameter [2], and the threshold criteria were derived from a geometrical relationship between the cross-sectional area and the diameter of a sphere, leading to the current −30%/+20% RECIST classification criteria [3, 4]. The RECIST measurement technique and classification criteria are currently used across many diseases [5], although the standard measurement technique for mesothelioma is now defined by modified RECIST [6] and mesothelioma-specific classification criteria have been investigated [7, 8].

With advances in medical imaging technology, there has been a more recent drive to use full three-dimensional volume measurements for response assessment [4, 9, 10] and to track measurements over time using continuous rather than discretized response [3, 11]. Especially for diseases such as mesothelioma, where the morphological characteristics do not correspond to the spherical extrapolation of the RECIST classification criteria, disease volumes are a logical choice for tumor burden assessment. While the past history of response assessment has moved away from volumetric quantification for reasons such as ease of manual measurement, the techniques used for response assessment were always intended to identify tumor burden changes (i.e. volumetric changes), and the purpose of this study was to return to the direct evaluation of tumor volume for response assessment.

Numerous studies have investigated the use of tumor volume and tumor volume-related measurements for response assessment in patients with malignant pleural mesothelioma (MPM) using magnetic resonance imaging, computed tomography (CT), and fluorodeoxyglucose positron emission tomography (FDG-PET) imaging [12–18]. The main challenge in these volumetric studies is the segmentation of the complete tumor volume. Using FDG-PET imaging, the segmentation of MPM tumor volume is greatly facilitated by the FDG avidity of the tumor [15, 16, 19]. The segmentation of MPM in CT images is more problematic predominantly due to the similar appearance of neighboring structures on CT [20] and has required semiautomated tools [17, 18]. Frauenfelder et al. [17] used a linear shape-based interpolation technique, requiring manual contours on ‘every fourth or fifth slice’. Interobserver agreement of volumetric response classification was found to be much higher than that of manual modified RECIST response classification (general κ = 0.9 versus κ = 0.33, respectively) [17]. Liu et al. [18] utilized a combination of semiautomated techniques for volumetric MPM segmentation, and their analysis revealed changes in tumor volume (dichotomized as tumor growth versus tumor decrease) to be significantly associated with patient survival.

The goal of this study was to create a comprehensive model for MPM patient survival utilizing time-changing estimates of disease volume from CT imaging in conjunction with clinical covariates such as patient age, sex, performance status (PS), white blood cell count, blood platelet count, histological cellular subtype, presence of chest pain, and weight loss [12, 21–25]. Continuous, time-varying measurements of disease volume were hypothesized to be a significant predictor of MPM patient survival, both as a single covariate and in conjunction with other clinical covariates.

patients and methods

patient cohort

Imaging and clinical data from 81 patients were obtained from a prospective study involving FDG-PET and CT imaging of MPM [26]. All patients were over 18 years old with histologically or cytologically confirmed MPM and had not received prior chemotherapy or radiotherapy. Patient accrual occurred from late 2003 to 2010, and the original study was approved by the local institutional Human Research Ethics Committee at Sir Charles Gairdner Hospital in Perth, Western Australia, with patients providing written informed consent. The subsequent analysis of the HIPAA (Health Insurance Portability and Accountability Act)-compliant data was approved by both the originating institution's Human Research Ethics Committee and the Institutional Review Board at The University of Chicago, where the analyses for the present study were carried out. The original study was not a treatment study, so patients were treated as clinically indicated. Combination chemotherapy consisted of cisplatin or carboplatin with gemcitabine or pemetrexed (Eli Lilly, Indianapolois, IN). Palliative radiotherapy was used when indicated, and a single patient had undergone previous pleurectomy/decortication (Table 1).

Table 1.

Description of the patient cohort used in this study

Sex (n)
Male	68
Female	13
Age at diagnosis (years)
Median	66
Range	41–80
Chemotherapy (n)
Carboplatin/pemetrexed	7
Cisplatin/pemetrexed	42
Cisplatin/gemcitabine	32
Histology (n)
Epithelioid	60
Sarcomatoid	7
Biphasic	14
T stage (n)
T1	19
T2	22
T3	25
T4	15
N stage (n)
N0	25
N1	3
N2	40
N3	13
M stage (n)
M0	70
M1	11
IMIG stage (n)
I	13
II	5
III	36
IV	27
Known asbestos exposure (n)
Yes	75
No	6
Chest pain (n)
Yes	50
No	31
Shortness of breath (n)
Yes	67
No	14
ECOG performance status (n)
0	38
1	38
2	5
Talc pleurodesis (n)
Yes	30
No	51
Weight (kg)
Median	75
Range	52–121
Height (cm)
Median	172
Range	155–189
Smoking status (n)
Never	36
Past	39
Present	6
Pleurectomy/decortication (n)
Yes	1
No	80

imaging

Patients were imaged using helical CT, with a baseline scan performed up to 1 month before the first cycle of chemotherapy and follow-up scans performed throughout their treatment regimen (typically after the first cycle, then every two cycles thereafter). CT staging was carried out by a thoracic radiologist or a medical oncologist experienced in mesothelioma imaging according to the Union for International Cancer Control tumor-node-metastasis (TNM) staging system (2002). Pathologic staging was not carried out.

A total of 281 thoracic CT scans were used in this study, with a median of four scans per patient. Ten patients had only a baseline scan with one follow-up scan, while 30 patients had three scans total, 34 patients had four scans total, and seven patients had five scans total. The median interval between scans was 49 days. One hundred ninety-seven scans had been performed on General Electric scanners (HiSpeed CT/i, n = 105; LightSpeed Pro 16, n = 2; or LightSpeed VCT, n = 90), and 84 scans had been performed on Philips Brilliance 64-slice scanners. At least 135 of the scans had been performed with contrast media.

Slice thickness and reconstruction kernels were matched across scans for each patient. Slice thicknesses for the different scans used in this study then were 0.63 mm (n = 6), 1 mm (n = 18), 1.25 mm (n = 36), 2.5 mm (n = 97), or 5 mm (n = 124). In-plane voxel dimensions ranged from 0.54 to 0.87 mm, and all reconstructed axial images had an in-plane matrix size of 512 by 512 pixels. The kVp setting for the scans was predominantly 120 kVp (n = 273), with 100 kVp (n = 2), and 140 kVp (n = 6) also used. Reconstruction kernels fell into two broad categories, with ‘lung’ kernels (Philips ‘L’ and General Electric ‘lung’ kernels) used for 185 scans and ‘standard’ kernels (Philips ‘B’ and General Electric ‘chest’, ‘soft’, and ‘standard’ kernels) used for the remaining 96 scans.

disease segmentation

For each scan, the pleural disease was segmented using a semiautomated method [27] modified with a semiautomated shape-based interpolation [28] for hemithorax segmentation that requires seeding of the approximate hemithorax boundary on selected CT sections. The semiautomated hemithorax segmentation is shown in Figure 1. After the pleural disease was identified as the region between the semiautomated hemithoracic segmentation and the automated lung segmentation as outlined in Sensakovic et al. [27], a watershed segmentation [29] was used to split the disease segmentation into three-dimensional regions based on morphology and spatial proximity; this process allowed manual editing to remove entire three-dimensional subregions that may have been erroneously included in the disease segmentation. Segmentation editing was performed without knowledge of patient outcome by an observer (ZEL) trained in thoracic anatomy. Pleural disease segmentation is shown in Figure 1. To calculate pleural disease volume for each scan, a pixel-counting technique was used [30].

Figure 1.

(A) Semiautomated hemithorax segmentation. This axial section did not contain a seeded contour, so the dashed contour is an interpolation from other manual axial contours. (B) Pleural disease segmentation from the semiautomated hemithorax segmentation. Because the hemithorax contour only needed to encompass the identifiable pleural disease, manual editing of pleural disease segmentations was generally minimal (restricted to the removal of the partial volume artifacts adjacent to the aortic arch in this axial section).

statistical analysis

Pleural disease volume measurements were modeled using scaled logarithmic transforms of relative changes from baseline, known as the specific growth rate (SGR) [11]. The definition of the SGR metric is

(1)

where V(t) denotes the volume measurement at an arbitrary time point and t₀ indicates the time of baseline scanning (times were modeled as fractional years). Patient survival was measured from the pretreatment baseline scan until either patient death or administrative censoring. Clinical covariates available for each patient included sex, histology, weight, height, body surface area, TNM staging, International Mesothelioma Interest Group (IMIG) staging, smoking status, asbestos exposure, presence of chest pain, presence of dyspnea, weight loss prior to diagnosis, Eastern Cooperative Oncology Group (ECOG) PS, talc pleurodesis status before chemotherapy, blood hemoglobin levels, white cell count, platelet count, and age at diagnosis. All covariates were fixed at their baseline values in the survival models. In concordance with prior prognostic models for mesothelioma [24, 25], some covariates were dichotomized as follows: ECOG PS level 0 versus level 1 or 2, epithelioid histology versus other, N stage N0 versus other, IMIG stages 1, 2, or 3 versus stage 4, blood platelet count ≤400 versus >400 per nl, blood hemoglobin ≤140 versus >140 g per liter, and blood white cell count ≤8.3 versus >8.3 per nl.

Survival modeling was carried out using Cox proportional hazards (PH) models [31, 32] with potentially time-varying covariates [33], and models were constructed using a forward selection process [34]. Survival model performance was assessed using Heagerty's C^τ [35], which is scaled from 0 to 1 since it is derived from receiver operating characteristic analysis. C^τ is especially useful for survival models with time-varying covariates; C^τ = 0.5 indicates no prognostic ability, and C^τ = 1.0 indicates perfect prognostic ability.

To obtain estimates of model predictive performance for novel cases [34], two methods were used. First, a leave-one-out cross-validation (LOOCV) was used in which each of the 81 patients was omitted one at a time. A multivariate survival model that consisted of the terms included in the full-cohort model was fit to the remaining 80 patients, and the fit model then was applied to the left-out patient. The process was repeated until each of the 81 patients had been excluded from the training cohort. A value of C^τ was calculated from these LOOCV-based survival predictions, called C^τ_cv.

The second validation procedure required repeated random subsampling of the patient cohort and was intended to generate a confidence interval (CI) for model performance. For each subsampling iteration, a model was trained on two-thirds of the patient cohort (n = 54, randomly selected) and tested on the remaining one-third (n = 27) of the patient cohort. The final multivariate survival model derived from the full patient cohort was fit to the training cohort, and the fit model was applied to the testing cohort. The repeated random subsampling simulation consisted of 1000 iterations, and the mean performance metric value will be referred to as C^τ_sub. All analyses were carried out using Revolution R Enterprise (version 4.3, based on R version 2.12; Revolution Analytics, Palo Alto, CA) [36].

results

patients and overall survival

Table 1 presents patient characteristics for the 81 MPM patients included in this study. Median survival for the entire patient cohort was 12.6 months (2.5–97.5 percentiles, 10.6–14.9 months; range 1.7–60 months). Of the 81 patients, there were 76 observed deaths, while the remaining 5 patients were censored after a median duration of 35 months. The overall Kaplan–Meier survival curve is shown in Figure 2.

Figure 2.

Overall survival for the patient cohort in this study.

Across all patients, the mean pleural disease volume at baseline was 1511 ± 1065 (range 225–5287 ml). At the first follow-up scan, the mean disease volume had reduced to 1397 ml, with geometric mean change from baseline of −11%. By the end of treatment, the geometric mean change in disease volume from baseline was −19%. The semiautomated techniques used in this study (hemithoracic segmentation and final pleural disease contour editing) required 10–20 min of active user intervention per case on average.

univariate survival analysis

Covariates predictive for survival in univariate Cox PH models at the α = 0.10 level are shown in Table 2. All clinical covariates found to be predictive for survival in this study have been reported in previous prognostic models for patients with mesothelioma. Continuous, time-varying measurements of the pleural disease volume on serial CT scans [modeled according to equation (1)] may now be included among covariates that are significantly associated with patient survival. The model indicates that larger disease SGR values are associated with a larger hazard; therefore, patients with disease growth have worse prognosis than patients with disease shrinkage, and patients with substantial growth have worse prognosis than patients with minimal growth.

Table 2.

Factors predictive for survival (from baseline imaging) in univariate Cox PH models, including hazard ratios and 95% confidence intervals (CI)

Variable	Hazard ratio	95% CI	P-value
N stage
N0	1	–	–
N1+	1.58	0.95–2.61	0.076
M stage
M0	1	–	–
M1	1.98	1.03–3.81	0.040
Dyspnea
No	1	–	–
Yes	1.94	0.98–3.82	0.056
Weight loss
Continuous (in kg)	1.05	1.00–1.10	0.054
Talc pleurodesis
No	1	–	–
Yes	0.67	0.41–1.08	0.098
ECOG performance status
0	1	–	–
1 or 2	1.56	0.98–2.46	0.059
Histology
Epithelioid	1	–	–
Other	2.26	1.34–3.83	0.0023
Blood platelet count
≤400 per nl	1	–	–
>400 per nl	1.58	0.97–2.57	0.067
Disease volume
Continuous, SGR	1.26	1.11–1.42	0.00031

Disease volume was modeled as the continuous specific growth rate (SGR) from baseline

multivariate survival analysis

The final multivariate survival model included the covariates disease volume SGR (hazard ratio [HR] = 1.31, P = 0.00045), histology (HR = 2.28, P = 0.0029), dyspnea (HR = 3.20, P = 0.0020), and ECOG PS (HR = 1.69, P = 0.029; Table 3). No two-way interactions between covariates in this final multivariate model were significant. The performance of the final multivariate model on the full patient cohort (after training on the full patient cohort) was calculated as C^τ = 0.690. From the LOOCV, the survival model performance from the final multivariate model was C^τ_cv = 0.661. The 1000 random subsampling iterations yielded a mean model performance of C^τ_sub = 0.660 with a 95% CI of 0.556–0.746.

Table 3.

Factors predictive for survival in the final multivariate Cox PH model, including hazard ratios and 95% confidence intervals (CI)

Variable	Hazard ratio	95% CI	P-value
Disease volume
Continuous, SGR	1.31	1.13–1.53	0.00045
Histology
Epithelioid	1	–	–
Other	2.28	1.33–3.93	0.0029
Dyspnea
No	1	–	–
Yes	3.20	1.53–6.71	0.0020
ECOG performance status
0	1	–	–
1 or 2	1.69	1.05–2.72	0.029

Disease volume was modeled as the continuous specific growth rate (SGR) from baseline.

discussion

The goal of this study was to create a survival model for MPM patients using continuous, time-varying, image-based measurements of pleural disease volume modeled as specific growth rate from baseline. Disease volume extracted from semiautomated pleural disease segmentations proved to be a significant predictor of patient survival in conjunction with clinical covariates.

Two essential aspects of this study were the ability to model changes in disease volume through time and the modeling of continuous (not discretized) changes in disease volume. In the only other study using changes in CT disease volume as a prognostic covariate for patients with MPM [18], analysis was limited to a single follow-up time point, and changes in disease volume from baseline were discretized as ‘volume decrease’ or ‘volume increase’. In the current study, all available follow-up time points were used, and changes in disease volume were treated continuously. In the work by Liu et al., hypothetical patients #2 and #3 in Figure 3 would be interpreted identically after the first follow-up volume quantification, and subsequent updates to the volume trajectory would be ignored. In the present study, however, the distinct disease volume trajectories across all time points for all three patients in Figure 3 would be considered as continuous changes from baseline. In the present patient cohort, fitting a model similar to Liu et al. in which the only covariate was discrete volume change after the first follow-up scan yielded C^τ = 0.521 (compare with C^τ = 0.690 for the full time-varying model in this study).

Figure 3.

Disease volume trajectories for three hypothetical patients.

While the use of MPM tumor volume as a prognostic covariate was motivated over a decade ago [37] and other studies have investigated correlations between discrete changes in MPM disease volume and patient survival [16, 18], this study is the first to model changing disease volumes using the SGR metric and combine clinical covariates with continuous, time-varying disease volumes in a multivariate survival model. Although not reported in the results, a model based only on disease volume yields C^τ = 0.610, while C^τ = 0.690 for the final multivariate model that included baseline clinical covariates. Therefore, time-dependent disease volume changes and clinical covariates at baseline are both crucial components of the final multivariate survival model.

The indication that non-epithelioid histology, presence of dyspnea, and non-zero ECOG PS are associated with poor prognosis is consistent with previous prognostic models for MPM patients [22]. These factors are routinely available as part of the diagnostic workup of patients with MPM. The presence of dyspnea and ECOG PS can change over time and could be treated as time-varying in the multivariate prognostic model; however, these data are not typically available after the baseline assessment of each patient.

The modeling of disease volume SGR as a continuous and time-varying quantity has clear advantages. First, any attempt to discretize disease volume changes into two or more classes will result in somewhat arbitrary class distinctions. Liu et al., for example, discretized response into two classes: disease growth or shrinkage. As another example, Frauenfelder et al. [17] used an extrapolation of the three-class RECIST classification criteria to volumetric changes based on a spherical geometry. The modeling of changes in disease volume as time-varying is more complicated but allows arbitrary volumetric trajectories to be modeled for each patient. In many published modeling studies, tumor measurements are acquired at exactly two time points, baseline and ‘follow-up’. Follow-up typically is defined after a set number of chemotherapy cycles, but what happens before or after that event? If a patient gets four follow-up scans during treatment, the researchers have updated information about how the patient's disease volume has changed at multiple time points since baseline. The modeling approach used in this study takes advantage of the updated information, whereas fixed follow-up models cannot.

Pleural disease segmentation is not necessarily the same as segmenting specifically mesothelioma tumor. For cases in which pleural effusion is present and spatially mixed with tumor, the separation of tumor from effusion can be difficult [27]. Pleural effusion regions were manually excluded only if they were localized and spatially independent of tumor regions. True tumor volume may have more prognostic value than pleural disease volume, although both pleural effusion and tumor negatively impact patients' quality of life. The feasibility of pleural disease segmentation in the clinical workflow is unclear, and therefore, the routine clinical use of disease volumes for MPM must be investigated. Pleural disease/tumor segmentation may be avoided by models that incorporate changing lung volume as a surrogate for tumor volume change.

Increasing the patient cohort may reveal associations between new covariates and survival, potentially improving the performance of the survival model. Validation of the model in a larger independent patient cohort remains as a future study. Finally, the patients in this study were treated with cytotoxic chemotherapy; treatment-specific models could be developed for individual patient cohorts, since the model derived in this study may not be directly applicable to patient cohorts receiving biologically different treatments.

In summary, continuous volumes derived from semiautomatic segmentations of pleural disease in patients with MPM are prognostically significant. Disease volume (specifically the ‘specific growth rate’) is allowed to change over time, allowing for survival prediction from arbitrary disease volume trajectories. In a final multivariate model, disease histology, dyspnea, and ECOG PS were also significant predictors of patient survival.

funding

This work was supported by The University of Chicago Comprehensive Cancer Center; the Raine Medical Research Foundation; the US National Institutes of Health (grant numbers T32EB002103 and R01CA102085); the Simmons Mesothelioma Foundation; the Kazan Law Firm's Charitable Foundation; the National Health and Medical Research Council, Australia; and the Cancer Council Western Australia.

disclosure

The authors would like to acknowledge the following possible conflict of interest: S.G. Armato III receives royalties from computer-aided diagnosis technology licensed through The University of Chicago. All remaining authors have declared no conflicts of interest.