MEG, start to finish.

Twenty-three FIF recordings from an Elekta system, two subjects, three sessions, including the empty-room recordings taken before and after each visit. MEG is the modality where the most metadata is already in the file, so this page is largely about what BIDS Manager should be left alone to do, and about the handful of things it genuinely cannot know.

Two subjects, three sessions, twenty-three files.

The folder uses Elekta's usual convention: a subject folder, then a date-named folder per visit, then the recordings.

MEG is the easy modality, and it is worth knowing why

A FIF file is rich. It carries the channel list with types, the sampling rate, the acquisition filters, the head position, the digitised points, the coil frequencies and the measurement date. mne-bids reads all of that and writes it into the sidecar, the channels table and the coordinate system files without being asked.

The practical consequence is that BIDS Manager deliberately does less here. It does not overwrite the manufacturer mne-bids derives from the format, and it does not duplicate the head-localisation fields the channels already answer. What it asks for is the short list that genuinely is not in the file, and nothing more.

Sessions come out of the path.

When the path between the subject folder and the recording contains a date-shaped token, YYMMDD or YYYY-MM-DD, each date is treated as a separate session, ordered by date.

• Several dates for one subject becomes one session per date. sub_us04rt22 has two, so its recordings carry ses-1 and ses-2.
• A single date is not a session in any meaningful sense, so no session entity is written. sub_ye07us06 is the example.

Where the path has no date at all, the recording date inside the file is used instead, which is the same rule the MRI side applies to study dates. Either way, look at the ses column before converting. It is inferred, and inference is exactly the kind of thing worth a glance.

Scan the folder.

Create a project, point Raw input at the top of the tree, and press Scan. Twenty-three rows, one per .fif, no skips.

The task and run entities come out of the filenames cleanly on this dataset: task_driving_run_01.fif decomposes into task-driving and run-1 with no help from you. That is worth checking rather than assuming, because it is a filename convention rather than a standard, and the next lab's convention will differ.

The manufacturer is inferred from the format rather than a header: .fif means MEGIN or Elekta, .ds means CTF, .con and .sqd mean KIT, .kdf means KRISS. It is shown as a read-only suggestion and is not written into the sidecar, because mne-bids writes the manufacturer itself and overwriting a value derived from the file with one guessed from an extension would be a step backwards.

Read the inventory.

On a CTF or 4D system the recording is a folder, not a file: .ds for CTF, and a set of numbered files for 4D and BTi. The scanner treats the folder as one recording, which is what you want, and the Editor's tree gives it a folder icon so it is obvious at a glance.

Empty-room recordings.

Empty-room data is not a task recording, and the standard has a specific place for it: the task label noise, with the timing carried by an acq entity. On this dataset task_emptypre and task_emptypost become task-noise with acq-pre and acq-post.

Getting this right matters more than it looks. An empty-room recording named as if it were a task will be analysed as one by any pipeline that selects on task, and a noise recording averaged into a condition is a hard error to find later.

Associating the noise recording with the session

The MEG section of the dataset metadata carries an associated empty-room field, which is one of the small number of things mne-bids cannot derive: it records which noise recording belongs to which session. Fill it if your analysis will use it. It sits alongside the other two fields in the same category, the dewar position and any subject-artefact description, which are equally facts about the session rather than the file.

The short list MEG actually needs.

Open Dataset metadata. The MEG section is deliberately small, because most of what BIDS wants is already in the FIF file. These are the fields that are not:

Convert.

Press Run conversion. All twenty-three recordings convert. Each row goes to mne-bids, which writes the FIF in its native format, without transcoding, alongside a channels table, the coordinate system files describing where the sensors and the head were, and the JSON sidecar.

Two MEG-specific behaviours are worth knowing. The standard-montage step that EEG runs is skipped, because MEG sensor positions come from the device geometry rather than from a named cap layout. And the conversions are run one at a time rather than in parallel, because the underlying writer is not safe to run concurrently on the same tree; DICOM and physio conversion stays parallel.

Look at the signal.

Open the dataset in the Editor and click a converted recording. The metadata card comes first, then Load signal opens the interactive viewer.

MEG needs channel-type handling that EEG does not, and the viewer has it. Magnetometers and gradiometers can be shown separately or together, and CTF axial gradiometers are handled as their own type rather than being lumped in. Without that, a magnetometer trace and a gradiometer trace share an axis whose scale suits neither.

Check the empty-room recording too

The power spectrum of a noise recording is the most useful single view in an MEG dataset. It tells you the noise floor of the room on that day, it shows the mains peak and its harmonics, and a sensor that has drifted or is failing stands out from its neighbours immediately. Comparing the before and after recordings from one session takes about a minute and occasionally saves a study.

Validate.

Press Validate dataset. On this dataset: 50 clean files, 24 warnings, no errors. As with EEG, the warnings are TODO placeholders in recommended fields: the dataset-level ones about your study, plus the MEG recommended fields that neither the file nor the template supplied, ManufacturersModelName and SoftwareVersions among them.

If you want the deeper pass, turn on Deep checks. It reads file contents rather than only names and metadata, which catches a truncated recording that is otherwise indistinguishable from a healthy one.

Real numbers from this dataset.

From the command line.

# 1. Create the dataset and its project.
bidsmgr-create ~/bids/elekta --name "Elekta MEG sample"

# 2. Scan. 23 rows; sessions inferred from the date folders.
bidsmgr-scan ~/raw/MEG_Elekta_sample_data \
    --project ~/bids/elekta --line-freq 50

# 3. Convert. 23 of 23.
bidsmgr-convert --project ~/bids/elekta

# 4. Fill what can be filled, mark what cannot.
bidsmgr-metadata --project ~/bids/elekta --fill-todos

# 5. Validate. 50 ok / 24 warn / 0 err.
bidsmgr-validate ~/bids/elekta

Every flag is documented in the CLI reference.

What usually goes wrong with MEG.

Next

The EEG tutorial covers the modality where the file holds the least, which is the mirror image of this one. The multimodal tutorial shows MEG converting in the same pass as MRI, PET and EEG.

The full GUI walkthrough → CLI reference