Data Classification

Introduction

Classification is the process of consistently identifying and labeling different types of data. Each modality represented in Flywheel, such as MR, PET, CT, or EEG, has its own, possibly multi-dimensional classification scheme. As a result, the first step in classifying data is to identify its modality (or instrument type), followed by its detailed classification.

Classification is typically performed by a gear (for example, the DICOM Classifier), which automatically runs based on the Gear Rules for a project. Accurate and detailed classification is key in triggering further automated data processing. This article lists the intent, measurement, and feature aspects of classifications for the MR modality as well as some examples for we determine classification.

Note

Prefer videos?

We have a collection of videos showing you how to get the most out of Flywheel. See these for more information:

• Advance UI- Data Management: Classifications

Instruction Steps

Modalities in Flywheel

In order to accommodate domain-specific classification of data from different modalities, Flywheel has developed an extensible, multi-dimensional classification system. We call each dimension an aspect of classification. Many data types will be classified according to just a single aspect.

The Classifier gear uses the DICOM modality tag for classification, and all modalities in Flywheel are listed using the standard DICOM abbreviations. For example "MR" stands for magnetic resonance and "OPT" stands for ophthalmic tomography. Classifications can be adjusted and extended on a per-site basis. Check your local site for the most accurate information.

MR Data

The ever-increasing diversity in MR data makes its classification particularly complex. MR data is currently classified according to the following aspects, which are all multi-select.

Intent

Intent refers to the overall purpose of the acquisition and allows the system to distinguish between data acquired with similar parameters that are intended for different purposes in a downstream process.

• Localizer
• Shim
• Calibration
• Fieldmap
• Structural
• Functional
• Screenshot
• Non-image

Measurement

Measurement refers to the physical MR properties being measured by the pulse sequence used. By default the measurement is parsed out of the series description, such that a series description="T1_2mm_Whole_Brain" will have a measurement of "T1" added to the classification for that series.

• B0
• B1
• BOLD
• CEST
• CSI
• Diffusion
• EPSI
• Fingerprinting
• MRA
• MT
• PD
• Perfusion
• Phoenix
• Susceptibility
• SVS
• T1
• T1rho
• T2
• T2*

Feature

Features refer to the pulse sequence properties and control variable/options used at the time of scan. By default these features are parsed out of the series description, such that a series description="T1_1mm_ISO_MPRAGE" will have a feature "MPRAGE" added to the classification for that series.

• 2D
• 3D
• AAscount
• Compressed-Sensing
• Contrast-Agent
• Control
• DCE
• Derived
• DKI
• DSC
• DSI
• DTI
• EPI
• EPISTAR
• FA
• FAIR
• FAIREST
• FLAIR
• Gradient-Unwrapped
• HARDI
• In-Plane
• iVASO
• Label
• M0
• Magnitude
• MIP
• MP2RAGE
• MPRAGE
• Multi-Band
• Multi-Echo
• Multi-Flip
• Multi-Shell
• Navigator
• NODDI
• PASL
• pCASL
• Phase
• Phase-Contrast
• Phase-Reversed
• PICORE
• PRESS
• QC
• QSM
• Quantitative
• Resting-State
• RMS
• SBRef
• Singlerep
• SPGR
• Spin-Echo
• Spiral
• Steady-State
• STEAM
• SWI
• Task
• TOF
• TRACE
• Transmit-Reference
• Uniform
• VASO
• WASSR
• Water-Reference

Custom Classifications

The Classifier gear does not always recognize custom pulse sequence acquisition names or labels that are not descriptive enough. In cases where the classifier gear is not assigning the desired classifications for acquisitions, customized rules can be attached to the project to parse the SeriesDescription to help the Classifier gear. See our article on custom classifications for more information.

Examples

MR DICOM classification is implemented in the dicom-mr-classifier gear, which extracts DICOM headers and stores them as metadata on the DICOM file object. The DICOM Series Description constitutes the primary input to classification.

The following table lays out some of the most common strings used for determining classification. For example, if the dicom-mr-classifier gear encounters the label Task_Stroop, it will classify the data as “Functional”. Labels are not case sensitive.