Does the application of diffusion weighted imaging improve the prediction of survival in patients with resected brain metastases? A retrospective multicenter study
Brain metastases are common in clinical practice. Many clinical scales exist for predicting survival and hence deciding on best treatment but none are individualised and none use quantitative imaging parameters. A multicenter study was carried out to evaluate the prognostic utility of a simple diffusion weighted MRI parameter, tumor apparent diffusion coefficient (ADC).
A retrospective analysis of imaging and clinical data was performed on a cohort of 223 adult patients over a ten-year period 2002–2012 pooled from three institutions. All patients underwent surgical resection with histologically confirmed brain metastases and received adjuvant whole brain radiotherapy and/or chemotherapy. Survival was modelled using standard clinical variables and statistically compared with and without the addition of tumor ADC.
The median overall survival was 9.6 months (95% CI 7.5–11.7) for this cohort. Greater age (p = 0.002), worse performance status (p < 0.0001) and uncontrolled extracranial disease (p < 0.0001) were all significantly associated with shorter survival in univariate analysis. Adjuvant whole brain radiotherapy (p = 0.007) and higher tumor ADC (p < 0.001) were associated with prolonged survival. Combining values of tumor ADC with conventional clinical scoring systems such as the Graded Prognostic Assessment (GPA) score significantly improved the modelling of survival (e.g. concordance increased from 0.5956 to 0.6277 with Akaike’s Information Criterion reduced from 1335 to 1324).
Combining advanced MRI readings such as tumor ADC with clinical scoring systems is a potentially simple method for improving and individualising the estimation of survival in patients having surgery for brain metastases.
KeywordsBrain metastasis Cerebral metastasis Diffusion MRI DWI Biomarkers Survival modelling Personalised medicine
Apparent diffusion coefficient
Akaike’s Information Criterion
Diffusion weighted MRI
Graded prognostic assessment
Recursive partitioning analysis
Whole brain radiotherapy
Brain metastases (BM) are an increasing clinical challenge causing significant morbidity and mortality . Multiple treatments are available including neurosurgical resection, radiotherapy, radiosurgery, chemotherapy and immunotherapy. The variety of available treatments can make it difficult to formulate a patient-specific treatment plan. To help, simple prognostic models based on clinical information were developed, including the Recursive Partitioning Analysis (RPA) scale  and Graded Prognostic Assessment (GPA) score . Recent schemes have added primary cancer (the disease-specific GPA ) and biological information such as receptor pathway status and biochemical parameters [5, 6, 7, 8] but no study has added advanced quantitative imaging measures to try and improve these models.
Although commonly used, the performance status - which is the major factor in these scores - is subjective and can vary between visits due to confounders such as corticosteroid use. There is a need for objective, non-invasive biomarkers that reflect the intrinsic biologic behaviour of the tumor. Diffusion weighted imaging (DWI) is a rapidly obtained sequence in clinical practice and is an accepted part of standard brain tumor imaging . Single center studies have identified apparent diffusion coefficient (ADC) values within the BM as a particularly promising marker; tumor ADC correlated with survival and recurrence after surgical resection  and survival after radiosurgery , whilst ADC changes at the tumor edge may indicate a more locally aggressive phenotype .
Therefore, it is logical to add tumor ADC to traditional clinical scores like RPA, GPA to determine whether this improves the prediction of survival. This would be of immense clinical value in helping to select the appropriate treatments for the individual patient based on their prognosis – personalised medicine – without the need for further invasive tests or procedures. Also as a proof of concept for applying an advanced MRI biomarker in this additive way, we are perhaps opening up the possibility of including it in future larger studies.
This study retrospectively pooled clinical and radiological data of patients with BM from three institutions and compared survival models using just the standard clinical factors with those using standard clinical factors plus an imaging biomarker, the tumor ADC. This allowed an evaluation of whether the imaging marker could improve the prognostication of overall survival in patients with surgically resected BM. In addition to providing sufficient numbers for statistical analysis, the collaborative nature of the project across three different institutions will help in establishing the external validity of the results, assess any differences in acquiring the imaging marker between large centers and identify obstacles that may hinder its implementation in clinical practice.
Study design, setting and participants
Demographic and clinical summary (n = 223)
Count (% of total)
Status extracranial disease
No evidence of disease
Number of brain metastases
1.5 - 2
2.5 - 3
3.5 - 4
1.5 - 2
2.5 - 3
3.5 - 4
WBRT after neurosurgery (complete resection)
Imaging acquisition and analysis
MRI data were processed at each individual institution, then MRI results along with clinical data were anonymized and statistical analysis performed at one institution by one statistics researcher with expertise in survival analysis and modelling (DH at LIV). Overall survival (OS) from initial diagnosis of BM to death was calculated, censored at the last recorded clinical contact. In univariate analysis differences in OS were examined, using log rank tests, based on each of the factors listed in Table 1. Data were stratified by center to account for potential confounding differences in treatment, MRI acquisition parameters and follow up between institutions. A Cox proportional hazards model was fitted for each of GPA, RPA, WBRT and ADC separately and then all combinations of these variables. The goal was to examine which model best described overall survival. In multivariate analysis, the best fitting Cox proportional hazards models were selected using variable reduction by Akaike’s Information Criterion (AIC) , which is a way of measuring concordance (how well the model fits the data) whilst penalising for extra parameters. Specifically, a low AIC value indicates a more accurate model. Statistical analysis was performed using R version 3.21 (R Core Team, 2013).
Clinical outcomes for this population
Median overall survival was 9.6 months (95% CI 7.7–11.7). A total of 187 deaths were observed in our cohort (83.9%) during a median follow up of 9.6 months (interquartile range = 4.5 to 17.3 months, minimum follow up = 7 days, maximum follow up = 8 years). Greater age (hazard ratio [HR] for death = 1.02, 95% CI 1.01–1.04, p = 0.002), worse KPS (HR = 1.03, 95% CI 1.02–1.05, p < 0.0001) and uncontrolled extracranial disease (partial, progressive or synchronous vs. complete response, HR = 2.44, 95% CI 1.47–4.06, p < 0.0001) were all significantly associated with shorter survival in univariate analysis stratified by institution. A likelihood ratio test from a univariate cox model (stratified by institution) with number of BM as the independent predictor did not show a significant relationship between number of metastases and time to death (p = 0.156). A likelihood ratio test from a univariate cox model (stratified by institution) with cancer type as the independent predictor did not show a significant relationship between cancer type and time to death (p = 0.0573). GPA (I vs. II p = 0.02, I vs. III p = 0.001, I vs. IV 0.154) and RPA (class I vs. II, p = 0.0016, I vs. III p = 0.02) categories were significantly associated with differences in survival in univariate analysis. Disease specific-GPA data was not significant on non-stratified analysis (p = 0.114), therefore it was not considered further. Adjuvant WBRT (bearing in mind this was a historical series) was administered in 144/223 of patients (64.6%) and was associated with significantly longer survival (p = 0.007) when stratified by center; therefore, it was included in the subsequent multivariate analyses.
Comparison of models for predicting overall survival in brain metastases
Graded Prognostic Assessment (GPA)
Recursive Partitioning Analysis (RPA)
GPA + Tumor ADC
RPA + Tumor ADC
GPA + WBRT + Tumor ADC
RPA + WBRT + Tumor ADC
Survival modelling for lung cancer cases alone (n = 115)
Graded Prognostic Assessment (GPA)
Recursive Partitioning Analysis (RPA)
GPA + Tumor ADC
RPA + Tumor ADC
GPA + WBRT + Tumor ADC
RPA + WBRT + Tumor ADC
This multicentre retrospective study suggests that for a specific population of surgically resected brain metastases, an advanced quantitative MRI based biomarker – the tumor ADC - improves the prediction of overall survival when added to the standard, widely used clinical indexes such as recursive partitioning analysis (RPA) and graded prognostic assessment (GPA) score.
What new information this adds to the field
In this study, we have demonstrated that even in a multi-institutional series using standard of care clinical MRI techniques, there is a significant association of tumor ADC with length of survival after resection and improvement of existing clinical scores for predicting survival by incorporating tumor ADC values. Predicting prognosis is critical in patients with BM, not only to guide management discussions between clinicians and patients, but also to stratify patients for randomised trials. The ADC values from three international institutions were similar in range and variance to the values reported in the literature [12, 15, 16]. This cohort is representative of routine clinical practice with respect to demographics, proportion of primary cancer types and overall survival [1, 17]. There was also a clear separation of patient survival based on the standard RPA and GPA indices, lending further validity to our patient cohort. Finally the effect we observed persisted even when just lung cancer cases were analysed, suggesting that the tumor ADC is not simply a surrogate of primary.
From a technical point of view, we selected tumor ADC, since single center studies of BMs have previously shown an association of higher tumor ADC with improved survival or delayed recurrence after surgery or radiosurgery [10, 12, 15, 16, 18], and this measurement is simple to obtain in clinical practice. It would have been better if the MRI data could have been uploaded centrally and then analysis of ADC maps performed centrally, perhaps by two researchers with some reliability analysis. However, this is not what would happen in clinical practice if this measurement were adopted and placing ROIs over the tumor on clinical workstations is something radiologists could reasonably do in the existing clinical workflow without the need of specialised software. To maximize the generalizability of our study results, real, retrospective data obtained at different institutions in different countries were utilized and as a result details about tumor volume, location and size were not available to be analysed as confounders. In general, as these were resected tumors they were all likely to be large or else they would have had SRS and in accessible rather than deep locations (again, that would likely have favoured SRS). Tumor volume may affect ADC and this is why the method explicitly uses multiple areas of reading and avoids necrosis, which is likely to be more common and centrally located in larger tumors. For the same reasons, retrospective information about adjuvant chemotherapy regimes were not available universally, although whole brain radiation was documented and analysed. Chemotherapy in general is poorly effective against BM although it may have affected overall survival which reflects systemic disease. Extracranial disease control was able to be analysed as it is a potential important confounder, and since this was a historical series, transformative treatments like immunotherapy which would certainly have had a huge impact on overall survival - e.g. for melanoma cases - were not in use. Primary cancer type did not seem to influence survival in this cohort therefore differences in treatment of the different primary cancers (e.g. radioresistant vs. radiosensitive types or targeted agents) did not seem be a confounder.
Biological significance and future directions
Although in stroke cases ADC has been shown to vary with MRI coils, vendors and field strength  DWI data appears to be comparable across vendors and institutions [20, 21]. Practical measures such as using multiple readers, standard derivations of measures and even cloud based post-processing platforms for central processing of raw data are all likely to minimise variation in future studies. Nonetheless, even the current, widely used GPA and RPA scales can suffer from inter-observer variability, such as the subjective nature in which KPS is determined. Tumor ADC is widely studied in other solid organ cancers as a biomarker of survival  although paired values for primary and BM have not been reported, it would be interesting to determine which influence overall survival more. High tumor ADC has been correlated with lower tumor cellularity , reduced extracellular matrix density  and greater degree of tumor differentiation  in BM and any of these could reasonably be surrogates of improved survival. We do not suggest that quantitative imaging biomarkers alone will replace the traditional clinical factors that have demonstrated utility across large numbers of patients over many years and changes in treatment modalities. Rather, we propose that quantitative imaging biomarkers be added to these clinical indices towards improving and personalising prognostication. Finally, the treatment paradigms in BM are currently undergoing great change with neo-adjuvant and adjuvant cavity SRS, immunotherapy and prophylactic chemotherapy for asymptomatic micrometasteses all being reported. This may make clinical models predicting prognosis obsolete. The solutions will be to re- apply these models to each new cohort (e.g. does pre-op tumor ADC improve survival modelling for surgery + cavity SRS patients too?), which is exactly how they were refined and developed originally and to embrace the opportunity to incorporate these individualised, non-invasive biomarkers provided by advanced MRI techniques (e.g. by testing MRI features as a measure of immunotherapy response as has been tried recently with post contrast features in a radiomics approach ).
We have demonstrated that diagnostic MRI scans including DWI sequences contain significant biological information which can be incorporated with standard clinical parameters to improve the prediction of overall survival in patients with surgically resected BM and better inform clinical decision making.
RZ wrote the manuscript with assistance from YJC, DMH, SW and SC. RZ, YJC, ASB and DH gathered data and performed statistical analysis. HP, MP, MDJ and SM devised study concept and made alterations to manuscript. All authors read and approved the final manuscript.
RZ was supported by the Medical Research Council, UK (MR/L017342/1) and the Royal College of Surgeons, England. DH is supported by the Medical Research Council, UK.
Ethics approval and consent to participate
Ethical principles were adhered to in line with the principles of the Declaration of Helsinki. LIV patients were part of an approved tumour biobank (National Research Ethics Service # 11/WNo03/2). VIE study was approved by the local and institutional ethics committees (ethics committee protocol number 641/2011). PENN data was gathered in the course of routine clinical care and local IRB approval was obtained for retrospective analysis.
Consent for publication
The authors declare that they have no competing interests.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.