The REMBRANDT Public Dataset: What it Means for Brain Cancer Research and Treatment

Historically, radiology has primarily supported disease diagnosis by providing information about tumor anatomy.1 Today, imaging has evolved to assist not only with improving diagnoses, but it has also become a recognized tool in the assessment and confirmation of disease, allowing oncologists to select appropriate treatment even when a new pathologic specimen is unavailable.2

The emerging field of radiogenomics is creating a new world of collaboration among medical disciplines.2 Integrating clinical imaging studies (i.e., radiomics) with genomics studies by correlating imaging features (e.g., phenotypes) and molecular markers is already advancing precision medicine and patient outcomes.2

When radiomics is used in conjunction with genomics, underlying physiologic and subcellular processes can be identified that not only help direct treatment decision-making but support new treatment discovery.2 However, extracting quantitative features from images created across various sites that can be used for comparison and study has posed a significant challenge, in part, due to the lack of standardized procedures for image acquisition and analysis.1


In August 2018, the brain cancer data collection known as REMBRANDT (REpository for Molecular BRAin Neoplasia DaTA) was made publicly available to researchers and clinicians around the world through the Georgetown Database of Cancer (G-DOC).3 This dataset is unique in that it is one of only two collections that contains both genomic data as well as linked diagnostic treatment, including imaging studies and outcomes data.3 Most big data collections contain only one of these, either genomic information or clinical data.3

REMBRANDT genomic data is linked with corresponding diagnostic brain imaging scans, RNA information, and more than 13,000 points of outcomes data.3 The companion image collection includes pre-surgical magnetic resonance scans from 130 patients in the REMBRANDT database.4  

Combined datasets like REMBRANDT with its imaging data give researchers the opportunity to analyze and integrate genetic information with outcomes such as overall survival, time to progression, or drug response, among other outcomes, whenever and with whoever they are working.2,4 Clinicians and investigators can research the link between radiological phenotype and tissue genotype because of G-DOC’s extensive clinical, gene, and expression data are for the same deidentified patient cases as the images.5

To make interpreting images across various sites possible, a standardized scoring system was developed.

Standardizing collaboration

REMBRANDT includes a feature set known as VASARI (Visually Accessible Rembrandt Images) that functions as a standardized scoring system designed to overcome variations due to multisite image sources.2,5 Its aim is to produce comprehensive and reproducible magnetic resonance imaging (MRI) interpretations by validating data received from different medical institutions that make collaboration among and across teams possible.2 There are 30 semantic descriptors of brain tumor imaging features separated into groups based on the extent of tumor resection, alterations in the area of the lesion, morphology of the lesion margin, morphology of the lesion substance, and lesion location.2

Standardized imaging descriptions of pathology are critical to advancing the new collaboration required in radiogenomics as well as precision medicine.2 VASARI can be used for grading MRIs allowing imaging experts around the world to collaborate, interpret scans, and support clinical decision making, hypothesis testing, and trial design.2

In focus: Glioblastoma

Glioblastoma multiforme (GBM) is a complex highly malignant brain tumor with a median survival of 14 months.1,6 Due to its characteristic intra-tumoral and inter-patient heterogeneity it’s difficult to acquire a representative biopsy.1 As a result, several teams of researchers have used radiogenomics approaches to investigate various methods for reliably interpreting imaging features.1

A key molecular biomarker of GBM involves assessing the methylation status of the promoter region of the methylguanine methyltransferase (MGMT) gene.6 It is important because of its association with longer survival when methylated.6 Radiogenomics investigators, including medical physicists and radiologists, were then able to pinpoint magnetic resonance texture features that could be used to classify regions of the tumor as methylated or unmethylated.6 This provides a noninvasive correlated image phenotype that can be used as an important biomarker in preoperative GBM tumors.6

GBM patients with specific invasive imaging signatures such as ependymal involvement, invasion of deep white matter tracts, and tumor extension across the midline were scored with VASARI.2  Analysis of imaging features included the ability to identify those patients with a substantially decreased overall survival, as well as additional changes in other important regulator genes.2

In a study of 416 patients with GBM where researchers used the VASARI feature set of 30 semantic features to standardize radiological assessments, semantic features were confirmed to be strongly related to clinical outcomes after researchers found necrosis and tumor enhancement were a clear and reliable outcome predictor.1 A team of neuroradiology experts concurred in their findings as well that the VASARI features comprehensively quantify the phenotype.1 Relevant and clinically meaningful biological information was acquired from reproducible measurements and assessments.1

Two strategies for developing imaging features in GBM that were identified by radiogenomics investigators include using existing features and developing new phenotypes.1 For the first one, they used the standardized semantic features scored by radiologists - proportion edema, proportion necrosis, proportion enhancing, lesion size, tumor location, and definition of enhancing margin.1 The second approach involved extracting fully computational derived features using imaging and statistical methods.1

A team of investigators that wanted to evaluate the predictive value of VASARI grading in patients with astrocytoma concluded that the scoring system should be used as an adjunct to biopsy.7 Their data illustrated that magnetic resonance features, including noncontrast enhancing tumor proportion, edema proportion, and enhancement quality provided particularly precise and detailed tumor information.7

Over the past 50 years, radiology, imaging research, and the imaging sciences have evolved to serve a central role as indispensable clinical tools in modern healthcare.6 Imaging experts believe continued progress that improves patient lives will require ongoing collaboration in the numerous emerging areas of opportunity provided by precision medicine.6 This is because achieving the large-scale quantitative imaging necessary to support large research sample sizes is a key component that can only happen with collaboration among multiple stakeholders, including radiologists, imaging scientists, medical physicists, biologists, statisticians, and bioinformaticians.6 Powering large multisite studies requires standardizing imaging procedures and harmonizing data acquisition methods as critical next steps.


  1. Towards precision medicine: from quantitative imaging to radiomics. Journal of Zhejiang University SCIENCE B Accessed 11/15/2018
  2. Magnetic Resonance Spectroscopy, Positron Emission Tomography and Radiogenomics—Relevance to Glioma. Frontiers in Neurology Accessed 11/15/2018
  3. Georgetown Offers Brain Cancer Data for Precision Medicine Research. Health IT Analytics Accessed 11/15/2018
  4. The REMBRANDT study, a large collection of genomic data from brain cancer patients. Nature Accessed 11/15/2018
  5. REMBRANDT. Cancer Imaging Archive Accessed 11/15/2018
  6. Promoting Collaborations Between Radiologists and Scientists. Academic Radiology Accessed 11/15/2018
  7. Potential Utility of Visually AcceSAble Rembrandt Images (VASARI) Assessment in Brain Astrocytoma Grading. Journal of Computer Assisted Tomography Accessed 11/15/2018