Guide the user through brain extraction (skull stripping) in the Neurodesk
environment: identify the input, present tool options without bias, write a
SLURM-compliant bash script, submit it, and hand off to QC on completion.
Run this as a sub-flow of the steps below — still present the tool menu (Step 3) and let the
user choose; the choice selects the recipe rather than a module load:
-
Identify the input image
Read the invocation arguments or current user request for a NIfTI path. If
none is given, ask the user:
"Please provide the path to your input NIfTI image (e.g., sub-01/anat/sub-01_T1w.nii.gz)."
Once a path is provided, verify it exists:
ls "<input_path>"
If missing, stop and report the exact path that was not found. Do not proceed.
Check the file extension ends in .nii or .nii.gz. If not, warn the user
that the file may not be a valid NIfTI and ask them to confirm.
Check that the input is a 3D volume, not a 4D timeseries:
python3 -c "import nibabel as nib; img = nib.load('<input_path>'); d4 = img.shape[3] if len(img.shape) > 3 else 1; print('dim4=' + str(d4))"
If dim4 > 1, warn the user:
"This appears to be a 4D volume (dim4 = N). Brain extraction expects a 3D
anatomical image. Please confirm you intended this file, or select a
single-volume NIfTI before proceeding."
Do not proceed past Step 1 until the user confirms the correct input.
-
Identify the output location
If an output prefix was given as the second argument, use it. Otherwise,
default to the same directory as the input with _brain appended before the
extension (e.g., sub-01/anat/sub-01_T1w_brain). Confirm with the user or
infer silently if the context is clear.
Determine the output directory and set OUT_DIR for use in the script.
-
Present tool options
Load references/tool-comparison.md from this skill directory and present
the comparison table to the user. Include columns: Tool, Method, Speed,
Template Required, GPU Optional, Best For.
After presenting the table, ask:
"Which tool would you like to use? If you'd like a recommendation based on
your data type or constraints, let me know."
Do not express a preference unprompted. If the user asks for a recommendation:
- Default anatomy T1w, no special constraints → FSL BET (fast, universal)
- Difficult anatomy, low contrast, or high accuracy required → FreeSurfer mri_synthstrip or HD-BET
- Multi-modal (T2, FLAIR, DWI) → mri_synthstrip or SynthStrip
- Need maximum accuracy and have time → ANTs antsBrainExtraction
- Functional / EPI data → AFNI 3dSkullStrip
Wait for the user to confirm a tool before proceeding.
After the user selects a tool, verify it is available on this cluster and
capture the exact installed version:
module spider <tool_name>
Parse the output for the version string matching the tool reference. If a
newer version is available, note it to the user. If the reference version is
absent, ask the user which available version to use. Update the module load
line in the script to match the confirmed version before proceeding.
-
Write the SLURM script
ANTs special case (check first, before writing anything): If the user
selected ANTs, immediately verify the template exists at the path specified
in references/ants-brain-extraction.md:
ls "<template_path>"
If it does not exist, pause and offer to add a DataLad or wget download step
for the OASIS-30 template. Do not proceed to write the ANTs script until the
template is confirmed present.
Load the reference file for the chosen tool:
- FSL BET →
references/fsl-bet.md
- ANTs →
references/ants-brain-extraction.md
- FreeSurfer mri_synthstrip or SynthStrip →
references/synthstrip.md
- HD-BET →
references/hd-bet.md
- AFNI 3dSkullStrip →
references/afni-3dskullstrip.md
Determine the next available script number with zero-padded two-digit numbering:
LAST=$(ls analysis_[0-9][0-9]_*.sh 2>/dev/null | sort | tail -1)
N=$([ -z "$LAST" ] && echo "01" || printf "%02d" $(( ${LAST:9:2} + 1 )))
Use $N as the script number (e.g., 01, 02, 03).
Ask the user:
"Does your cluster require --partition or --account in the SLURM header?
If so, please provide the values; otherwise I'll omit them."
Write a script named analysis_<N>_brain_extraction.sh containing:
- Full SLURM header with resource values from the tool reference (add
--partition / --account if provided by the user)
mkdir -p logs/ and mkdir -p "$OUT_DIR"/module load <tool>/<pinned-version> (exact version confirmed in Step 3)- The extraction command with recommended default parameters
- A clear comment above each command block explaining what it does
Show the complete script to the user and ask:
"Does this script look correct? Reply 'yes' to submit or request changes."
-
Submit the job
After the user confirms the script is correct, ask:
"Would you like to record this analysis with DataLad for provenance? If
so, I can add the appropriate DataLad provenance step. Reply 'yes' to use
DataLad, or 'no' to submit directly."
If yes, use the local DataLad workflow available in the current environment
before proceeding. If no, submit directly:
sbatch analysis_<N>_brain_extraction.sh
Capture and report the job ID. Tell the user:
"Job submitted — ID:
Monitor: squeue -j <jobid>
Logs: tail -f logs/brain_extraction_<jobid>.out"
-
Wait for completion and hand off to QC
Tell the user to notify you when the job finishes. Once they confirm,
first verify the job exited cleanly:
sacct -j <jobid> --format=JobID,State,ExitCode --noheader
If State is not COMPLETED or ExitCode is not 0:0, do not proceed to QC.
Instead show the last 30 lines of the error log:
tail -30 logs/brain_extraction_<jobid>.err
Report the failure to the user and stop.
If the job completed successfully, verify the expected output file exists:
ls "<expected_mask_path>"
If the output is absent, stop and instruct the user to check the SLURM log
for errors before proceeding.
Once both checks pass:
- Identify the brain mask output path from the tool reference (it varies by tool)
- Invoke the brain-extraction-qc workflow with the original image and brain mask:
Run brain-extraction-qc <original_input> <brain_mask> <output_prefix>
