segresnet-mlx

Segment brain tumors, renal structures and whole-body CT anatomy with MONAI SegResNet models in native MLX.

BraTS weights · Renal weights · MedMLX

Weights: 19 MB download for segresnet-brats, 76 MB for segresnet-renal-cect. Peak MLX memory: 4.9 GB for a 240×240×155 four-sequence brain MRI (segresnet-brats) and 4.0 GB for a 478×277×161 three-phase CT (segresnet-renal-cect).

Use with MedMLX

The MedMLX MCP server is the lowest-friction way to use this package. It installs this runtime, downloads its weights on first use, and lets an AI agent describe, check and run segresnet-brats and segresnet-renal-cect for you.

uv tool install --python 3.13 "medmlx[segment,mcp]"
MEDMLX="$(uv tool dir --bin --color never)/medmlx"
mkdir -p "$HOME/medmlx-work"
claude mcp add --scope user medmlx -- "$MEDMLX" serve --root "$HOME/medmlx-work"

For Codex, use codex mcp add medmlx -- "$MEDMLX" serve --root "$HOME/medmlx-work". Other MCP clients take the same command and arguments. The same install provides the medmlx CLI, for example medmlx describe segresnet-brats.

For AI agents: prefer the MedMLX MCP tools (list_models, describe_model, model_readiness, run_model, get_job) over writing Python against this package.

Install

uv venv --python 3.13
source .venv/bin/activate
uv pip install "segresnet-mlx>=0.3.4"

Run

from segresnet_mlx.medmlx import brats

result = brats(
    t1c_path="t1c.nii.gz", t1_path="t1.nii.gz", t2_path="t2.nii.gz",
    flair_path="flair.nii.gz", output_dir="out",
)
print(result["outputs"]["mask"])

The registered models are segresnet-brats and segresnet-renal-cect; their brats and renal_cect runners select the matching bundle internally. The renal runner takes arterial_path, venous_path and excretory_path. Keep each model's input volumes on one voxel grid.

Omitting weights_path downloads MedMLX/segresnet-brats-mlx or MedMLX/segresnet-renal-cect-mlx into the shared MedMLX cache on first use; interrupted downloads resume and later runs work offline. With MedMLX installed, medmlx setup segresnet-brats downloads it ahead of time. To use your own copy, pass a folder containing segresnet-brats.safetensors for BraTS or segresnet-renal-cect.safetensors for renal CECT as weights_path. details.weights_revision records the downloaded commit, or null for a folder without a MedMLX download record.

The standalone run(bundle=..., weights_path=...) interface remains available for explicit local checkpoints, including the whole-body CT variants. Published native weights are also available for renal CECT and whole-body CT.

Inputs and outputs

Inputs are finite 3D volumes on one voxel grid, with spatial units in millimetres or unspecified. The mask is written on the input grid as the NIfTI reader corrects it. Renal labels are float32: 1 artery, 2 vein, 3 ureter, 4 neoplasm and 5 parenchyma. BraTS labels are uint8: 1 tumor core, 2 whole tumor outside the core and 4 enhancing tumor.

compensated-fp32 accumulates convolution and GroupNorm in paired Metal floats with float32 layer outputs. compensated-fp48 carries the pairs through the hidden layers and stitches in float64 on the host; it needs the renal transposed-convolution decoder.

run returns {"outputs": {"mask": path}, "details": {...}}, with labels, shapes, window settings, MLX device, peak memory and elapsed time. Invalid inputs raise InvalidInputError; missing or mismatched weights raise AssetNotReadyError; inference failures raise ModelExecutionError.

Whole-body CT and streaming

run takes bundle and a .safetensors file as weights_path: brats_mri_segmentation, renalStructures_CECT_segmentation, wholeBody_ct_segmentation (1.5 mm) or wholeBody_ct_segmentation_lowres (3 mm). The whole-body bundles take one CT as image_path and use segresnet-wholebody-high.safetensors or segresnet-wholebody-low.safetensors. They write float32 labels 0–104 from the bundle's channel_def: 1 spleen, 2 kidney_right, 3 kidney_left, 4 gallbladder, 5 liver, 6 stomach, 7 aorta, 8 inferior_vena_cava, 9 portal_vein_and_splenic_vein, 10 pancreas, 11–12 adrenal glands (right, left), 13–17 lung lobes (upper and lower left, upper, middle and lower right), 18–41 vertebrae L5–C1, 42 esophagus, 43 trachea, 44–49 heart chambers and pulmonary artery, 50 brain, 51–54 iliac arteries and veins, 55 small bowel, 56 duodenum, 57 colon, 58–69 left ribs 1–12, 70–81 right ribs 1–12, 82–83 humerus, 84–85 scapula, 86–87 clavicula, 88–89 femur, 90–91 hip, 92 sacrum, 93 face, 94–99 gluteus maximus, medius and minimus, 100–101 autochthon, 102–103 iliopsoas and 104 urinary bladder.

run(..., streaming=True, scratch_dir="scratch") keeps blending intermediates and logits in temporary mapped files instead of a full in-memory class volume. The files are removed on success and failure; scratch_dir defaults to the system temporary directory. The final label volume stays in memory.

Python API

from pathlib import Path

import numpy as np
from segresnet_mlx import BUNDLE_CONFIGS, SegResNetSpec, forward_patch, load_model, predict_logits

spec = SegResNetSpec(3, 6, 32, upsample_mode="deconv")  # renal CECT network
network = load_model(Path("segresnet-renal-cect.safetensors"), spec=spec)
patch = np.zeros((1, 3, 32, 32, 32), dtype=np.float32)
logits = forward_patch(network, patch)  # float32 NCDHW -> float32 NumPy logits
config = BUNDLE_CONFIGS["renalStructures_CECT_segmentation"]
volume_logits = predict_logits(patch, config, network=network)  # stitched sliding windows

forward_patch needs spatial sizes divisible by 2 ** len(spec.blocks_up). predict_logits takes preprocessed arrays of any size and stitches the bundle's sliding windows; neither normalizes intensities, applies activations or maps labels. load_model checks tensor names, shapes and finite float32 values against required_shapes(spec), not checkpoint identity. WindowTelemetry and BundlePredictor record bounded progress and timing. Inference needs Apple Silicon with Metal and has no CPU fallback.

License

Apache-2.0. See third-party notices. Model weights retain their publisher terms.

For research and education. Outputs require human review.