p-brain 3.1.2

Automated quantitative DCE-MRI of cerebral perfusion, microvasculature, and blood–brain-barrier permeability. Point it at a subject's scans and it produces the full derivatives tree — T1/M0 mapping, arterial-input-function extraction, signal-to-concentration, and pharmacokinetic and diffusion modelling — as voxel-, tissue-, and parcel-level maps. No notebook, GUI, or server.

Install · Quickstart · Example run · How it works · Data layouts · Commands · Assist · Configuration · Principles · Citation

p-Brain (the p is for perfusion and permeability) is a cross-platform command-line tool for quantitative DCE-MRI. It serves two purposes. As shipped, it is a validated, ready-to-run pipeline you point at real scanner data to get quantitative maps (Ki, CBF, MTT, CTH, Ktrans, FA, and others). It is also a framework you extend: each step — kinetic model, AIF, segmentation backend, or a whole stage — is a self-contained plug-in, so adding your own means writing a single file with no changes to the core. Drop a model into pbrain/models/, call it with --models yourmodel, and it runs on every subject, is aggregated to every anatomical level, is written as NIfTI/CSV/JSON, and is given diagnostics automatically.

Install

pip install p-brain
pbrain --help

Optional extras pull in heavier dependencies only when you need them:

pip install "p-brain[cnn]"        # TensorFlow — the CNN arterial-input-function (default AIF)
pip install "p-brain[metal]"      # Apple-Silicon GPU for the TF stages (pins TF 2.16 + tensorflow-metal)
pip install "p-brain[diffusion]"  # dipy — the diffusion track (DTI, DKI, CSD, tractography)
pip install "p-brain[dicom]"      # pydicom — DICOM input
pip install "p-brain[all]"        # everything at once

On Apple Silicon, [metal] is how the CNN AIF and SynthSeg run on the GPU — Apple's tensorflow-metal plugin only supports TensorFlow ≤ 2.16, so this pins the matched pair. Install it into a fresh environment (a newer TensorFlow already present can't be safely downgraded in place). On Linux, GPU works through [cnn] with a CUDA build of TensorFlow. Without a GPU backend, --device mps/cuda cleanly falls back to CPU (identical results, only slower).

Runs on Linux, macOS, and Windows with Python 3.10–3.12. From a clone, pip install -e ".[dev]" installs it in editable mode.

Quickstart

pbrain setup            # interactive: install extras, offer to fetch weights and data
pbrain fetch-weights    # CNN weights           (Zenodo 10.5281/zenodo.15697443)
pbrain fetch-data       # example subject sub-01 (Zenodo 10.5281/zenodo.20826857, ~99 MB)

pbrain fetch-data downloads a real subject and prints a ready-to-run, weights-free command. Then, on your own data:

# One flag on raw Philips PAR/REC: inputs are auto-discovered by protocol name
pbrain run --subject-dir /data/20230403x2

# A whole study: each sub-directory is a subject, run in parallel
pbrain run-cohort --cohort /data/patients /data/controls --workers 4

Example run

On an interactive terminal, pbrain run opens a live cockpit — the brain draws itself in, then an analysis panel tracks the pipeline while log lines scroll above it and a status line shows where it is:

  ⢀⣴⣿⣿⣿⣷⣦⡀
  ⠈⢻⣿⣿⣿⣿⣿⠿   p-Brain
  ⠀⠀⠉⠉⠹⣿⠟⠀   perfusion & permeability

  reading DCE … dcm2niix
  DCE shape · 256×256×10 · 250 frames
  fitting T1 map … median 1642 ms
╭──────────────────────── analysis ────────────────────────╮
│  ✓ load            DCE 256×256×10·250, IR 8 TIs           │
│  ✓ t1_m0           T1 median 1642 ms                      │
│  ✓ signal_to_conc  saturation_recovery                    │
│  ◆ aif             extracting … ROI 704 vox · peak 3.2 mM │
│  ○ tissue_roi      synthseg                               │
│  ○ kinetic         patlak · tikhonov                      │
╰───────────────────────────────────────────────────────────╯
 ⠴ p-Brain · aif · 22s · CNN rICA, 3 slices        auto ⇥

Piped or redirected (| tee, > log, --quiet), the cockpit is suppressed and you get plain, greppable log lines instead — the numeric results are identical either way.

How it works

Ten self-contained stages, each cached to the derivatives tree and re-runnable in isolation:

 load ──▶ t1_m0 ──▶ signal_to_conc ──▶ aif ──▶ tissue_roi ──▶ normalisation
   │      (T1/M0)   (signal → mM)     (CNN /   (SynthSeg      (baseline,
   │                                   auto)    parcels)       Gd leak)
   ▼
 kinetic ──▶ diffusion ──▶ summary ──▶ diagnostics
 (patlak,    (DTI/DKI/     (voxel ·    (per-map QC
  tikhonov,   CSD, tracts)  tissue ·    overlays)
  tofts…)                   parcel)

Every stage is a plug-in resolved at run time; pbrain list shows all plug-points, pbrain list models drills into one. Results are written at three levels (voxel, tissue-region, parcel) in NIfTI, CSV, and JSON, with a diagnostics montage per map.

Inputs are normalised first. The load stage converts whatever you point it at — Philips PAR/REC or DICOM (via dcm2niix), NIfTI passed straight through — into the 4-D NIfTI the rest of the pipeline runs on. Every downstream stage sees the same canonical volume, so the analysis is source-format-agnostic.

_{Voxelwise Ki, vb, CBF, MTT and CBV from a single automated pbrain run.}

Data layouts

pbrain run <path> works out what you pointed it at — one subject or a whole folder — and which on-disk convention it follows, then runs accordingly:

pbrain run /data/20230403x2        # one subject → runs it
pbrain run /data/patients          # a folder of subjects → fans out as a cohort
pbrain layout /data/patients       # preview what it detects (read-only, no run)

Three layouts are recognised out of the box:

Detection is deterministic and tries single-subject first, so a subject you've already run once (it has a pbrain/ output tree) isn't read as a one-subject cohort.

Don't know the layout, or which file is which? --assist works it out. Point it at a folder no built-in convention recognises and a local model reads the file tree — names only, no pixel data — to propose which folders are subjects and which files are the DCE, T1, and IR. You accept the proposal or correct it in plain words ("the T1 is the MP-RAGE, not the FLAIR"). The confirmed layout is frozen to pbrain.layout.toml at the data root, so later runs read it directly — point-at-anything convenience with a reproducible pipeline. More in Assist.

Commands

Assist

Running a subject usually means knowing which series is the DCE, which is the T1, where the IR sits, and how your folders are arranged. --assist lifts that requirement — an optional local-model layer that reads your acquisition parameters and file tree and proposes the mapping, so you can run a subject, or a whole unfamiliar archive, without hand-writing a single path.

pbrain run --subject-dir X --assist

What it does:

• Finds your inputs for you. From the scan headers it proposes which series is the DCE, T1, and IR; for a folder no built-in layout recognises, it reads the file tree (names only) and proposes which folders are subjects and which files are each input. You accept it, or correct it in plain words — "the T1 is the MP-RAGE" — and it re-proposes. A confirmed layout is frozen to pbrain.layout.toml, so later runs read it directly.
• Explains a failure. When a stage errors, it turns the traceback into a plain-language cause and a concrete fix.
• Writes the boilerplate. It summarises each stage's QC in a sentence and drafts a Methods paragraph from the run's provenance, ready to edit for a paper.

Two things keep it safe to leave on:

• It reads text, not data. The model sees scan headers, run provenance, QC statistics, and error messages — not voxel values — and it computes no result. The maps come from the deterministic pipeline. A suggestion that would change a run, such as which series is the DCE, takes effect only after you confirm it, and is written to the manifest so a re-run reproduces the same numbers.
• It's local and opt-in. Assist talks only to a model running on your machine via Ollama. With none configured, the assist features stay off and the run is unchanged. First use walks you through installing Ollama and pulling a model sized to your hardware, and asks before it downloads anything.

Live controls

During an interactive run:

• Review mode. --mode auto (default) runs straight through. --mode verify opens a browser review at each decision checkpoint so you can confirm the suggested choice; --mode manual additionally lets you correct it. The checkpoints:

  • AIF — the input-function curve and its vessel ROI on the DCE slice (with a motion-correction toggle and the candidate vessels); confirm, drag the max voxel, or draw the ROI yourself.
  • Baseline — the pre-contrast baseline point on the first-peak region of the mean curve; confirm or drag it.
  • Tissue segmentation — the parcellation mask across slices; confirm, or draw exclusion regions to cut artefacts out of the tissue mask.
  • Per kinetic model — the model's own plot (Patlak plot; tissue-curve fit for the residue/compartment models) with its parameter values. In manual mode some models expose editable fit parameters (e.g. Patlak's fit-window and regression) that re-fit live on confirm. Any model gains a review for free by defining review().
  • Diffusion — a scalar-map summary (median FA/MD/… + the primary map's central slice) for every diffusion model.

  Reject any review to stop the run; ⇥ Tab cycles the mode live, shown in the status line, and the run timers pause while a review is open. Everything is served locally (no data leaves your machine) and your choice is recorded in the manifest, so a re-run reproduces it.

• Ctrl-C stops cleanly at any point; completed stages are cached, so the next run picks up where it left off.

Configuration

Flags override a config file, which overrides defaults.

pbrain run --config study.toml --subject-dir X
pbrain run --subject-dir X --models patlak,tikhonov --opt models.patlak.regression=ols

--config takes TOML (built-in) or YAML (pip install pyyaml). Any plug-in option is settable with --opt <plugin>.<key>.<opt>=<value> or in the file. Acquisition parameters (flip angle, TR) are read from the scan sidecar when present and can be overridden per run.

Add your own

Every plug-point (kinetic models, AIF extractors, signal-to-concentration, segmentation backends, diffusion models, whole stages) is a single file that registers itself. A kinetic model is a dataclass declaring what it accepts, what it produces, and an extract/fit method:

# pbrain/models/my_model.py — then: pbrain run --models my_model
@dataclass(frozen=True, slots=True)
class MyModel:
    key = "my_model"
    produces = {"ki_map": np.ndarray}
    def fit(self, conc, aif, t_s, **opts) -> dict: ...

It is then aggregated to every level, written in every format, and given diagnostics with no further wiring.

Models

Set any option with --opt models.<key>.<opt>=<value>. Defaults are what you get without setting anything.

The list is meant to be extended — see Add your own. The diffusion track (--diffusion) adds FA, MD, and tractography via dipy.

Principles

Plug-in architecture. Kinetic models, AIF extractors, signal-to-concentration methods, segmentation backends, and stages are self-registering modules. Adding one is a single file; the core is unchanged.

The optional model reads text, not data. Input mapping, QC summaries, methods text, and error explanations come from scan headers and run provenance, not voxel values, and the model computes no result. A suggestion that would alter a run takes effect only once you confirm it, and is recorded in the manifest.

Runs are reproducible. Inputs, plug-in choices, and parameters are recorded per run; an --assist-resolved layout is written to pbrain.layout.toml. The same inputs and settings produce the same outputs.

One input representation. load converts PAR/REC, DICOM, or NIfTI to a 4-D NIfTI; every downstream stage operates on that volume.

Output is separable from computation. The interactive cockpit is TTY-gated; a redirected or --quiet run emits plain text with identical numeric results.

Requirements

• Linux, macOS, or Windows with Python 3.10–3.12
• dcm2niix on PATH for PAR/REC and DICOM conversion
• Optional: TensorFlow ([cnn]), dipy ([diffusion]), Ollama (assist)

Citation

If p-Brain contributes to your work, please cite the accompanying paper (Tireli et al.) and this repository. See LICENSE for terms.

MIT · Built by Edis Devin Tireli · Functional Imaging Unit, Rigshospitalet · University of Copenhagen