Learning Objectives

1. Necessity and challenges of big data in MRI

2. Different ways to collect big data in MRI

3. By example: MRI in the Rhineland Study

Overview: What is Big Data?

Over the past years there has been an increasing interest in large-scale imaging studies of the general population or specific patient groups. A common goal of these efforts is the development of reliable imaging-based biomarkers that indicate an increased risk of certain diseases, which may result in specific preventive or therapeutic measures or the longitudinal assessment of therapeutic relevance. Due to the large variability of neurological diseases and age-related effects, as well as the typically associated small biophysical changes in brain tissue at early (pre-symptomatic) stages, large sample sizes are required in order to draw neuroscientifically relevant conclusions. A pioneering example of large-scale MRI in a prospective cohort study is given by the Rotterdam Study, which successfully collects neuroimaging data in the general population for more than 20 years [1]. Large-scale MRI studies are challenging with respect to logistics and technology of data acquisition, archiving and processing. Thus, the term Big Data has become increasingly popular in the MRI community. The five V’s of Big Data – Volume, Velocity, Variety, Veracity, and Value – perfectly describe the challenges of MRI in large-scale MRI studies: High-throughput exams produce large amount of images at high data rates (Volume and Velocity), which are inherently multimodal and the subsequent analysis results in a plethora of different data types (Variety). However, the quality of MR images varies strongly even if acquired with (presumed) identical setup, e.g. due to subject-motion or scanner drifts (Veracity). Finally, hypothesis-driven analysis of the data and mining of the data by means of multivariate classification and feature extraction may provide novel insights and imaging-based biomarkers, which may have an impact on future healthcare (Value).

Concepts: How to acquire Big Data?

Different approaches exist to gather large-scale MRI data. Retrospective imaging studies perform meta-analysis on massive amounts of data across many different studies and sites; a prominent example is the ENIGMA study [2]. Instead, a prospective study allows to collect data, acquired with identical set-up and methods tailored to the study questions. A leading example was given by the Human Connectome Project (HCP), which applies state-of-the-art MRI technology in 1000 subjects [3]. Another approach is given by multi-centric prospective studies, where dedicated imaging protocols are harmonized across many different sites and scanner types, aiming at increased sample size at cost of decreased acquisition efficiency and quality. A successful example of this approach is the Alzheimer’s Disease Neuroimaging Initiative (ADNI) [4]. Recently there have been increasing efforts to utilize advanced MRI technology in very large-scale population studies by means of dedicated centers. For example the recently started UK-Biobank study aims at brain MRI of 100,000 subjects [5]. A similar concept, with the additional goal of longitudinal examination, is implemented by the Rhineland Study, which is carried out by the DZNE (Bonn, Germany) [6]. The Rhineland Study investigates aging, in particular of the human brain and related neurological disorders, across the adult lifespan. The main study starts in March 2016 and is carried out in three dedicated study centers of which two are currently operational. All centers are identically equipped, each with a Siemens Prisma MRI scanner. The Rhineland Study will include up to 30,000 subjects, aged 30 years or over at first visit, and reexamination every three years. The emphasis is on quantitative measures, including one-hour MRI examination of brain structure and function.

Methods: MRI in the Rhineland Study

This lecture briefly introduces the Rhineland Study and summarizes the concepts of its MR protocol. Forward-looking and robust sequences were implemented, tailored for the advanced scanner hardware. The 40min core protocol consists of sequences to acquire T1-weighted, T2-weigthed, SWI (QSM), resting state fMRI, and multidimensional diffusion-weighted contrast in the whole brain with high isotropic resolution, respectively. Also, carotid artery MRI and body fat quantification with the Dixon technique is performed in every subject. Additionally, a 10min “free protocol” with varying sequences addresses additional research questions such as perfusion or metabolism in subgroups of the cohort. A dedicated workflow for image analysis was implemented, consisting of a centralized archiving and compute system as well as automated (minimal) processing pipelines for each modality. A tailored daily quality assurance routine was developed, including dedicated MR phantoms to quantify scanner performance across study centers. The MRI protocol was implemented, tested, retested, and refined during a thorough pilot phase in 2015. Results will be presented in the lecture.

Outlook

Finally, the presentation will address current practical and conceptual problems of big data in large-scale MRI studies, e.g. the difficulty to store raw data, centralized vs. mesh computing, and the overarching problem of freezing software, hardware, and protocols in long-term studies. Some suggestions for future research directions are made to overcome current limitations.