MedMLX
/

flexray-mlx

@@ -17,38 +17,35 @@ pipeline_tag: image-segmentation
 base_model: VictorButoi/flexray
 ---
-# FleXray Radiograph Segmentation
-Native MLX port of [FleXray](https://github.com/VictorButoi/FleXray).
-Segment anatomical structures in radiographs locally on Apple Silicon.
-- Released `flux0375` member
-- Lung segmentation in the quick start
-- Single-member inference; noncommercial weights
-MLX implementation by [MedMLX](https://github.com/MedMLX).
-[GitHub](https://github.com/MedMLX/flexray-mlx) · [Usage](https://github.com/MedMLX/flexray-mlx#quick-start) · [Upstream](https://github.com/VictorButoi/FleXray)
-**Requires the companion `flexray-mlx` runtime.** Follow the quick start below.
-## Requirements
-Apple Silicon Mac with Metal GPU access, Python 3.13 or 3.14 and [uv](https://docs.astral.sh/uv/).
-The weights are included here; checkpoint conversion is not needed.
 ## Quick start
 ```bash
 uv venv --python 3.13
 source .venv/bin/activate
-uv pip install "flexray-mlx @ https://github.com/MedMLX/flexray-mlx/releases/download/v0.2.1/flexray_mlx-0.2.1-py3-none-any.whl"
 uv tool install huggingface-hub
 hf download MedMLX/flexray-mlx --local-dir models
 ```
-Authenticate with `hf auth login` if repository access is required.
-Run this example in the same environment, replacing the input paths with your own:
 ```python
 from flexray_mlx.medmlx import run
@@ -63,72 +60,10 @@ result = run(
 print(result["outputs"])
 ```
-Only `flux0375` is packaged. Keep `ensemble=False`; the retained upstream
-ensemble descriptor lists five members, but the other four are not included.
-## Input and output
-Input is one radiograph: a raster image supported by Pillow or a single-frame
-2D DICOM. The example segments the lungs.
-Each run writes:
-- `source_image_label.png` — 0/1 mask at source resolution
-- `source_image.png` — 8-bit grayscale input
-- `probabilities.npy` — float32 probabilities for all 61 labels
-- `prediction.json` — label, threshold and member metadata
-## Technical details
-The companion runtime is [v0.2.1](https://github.com/MedMLX/flexray-mlx/releases/tag/v0.2.1), source `5a0ef5f36f40ef606819a48b5869d381386909a2`, with `medmlx-core==0.2.3`. All 206 flagship tensors exactly match the original publisher's pinned `6b63ddf09cfdff99aea494e3738cf243a4e22cdb` bundle used for that release. The ensemble descriptor, flagship configuration, label schema and preprocessing metadata are also byte-identical. The installed public bundle admission accepts this Hub layout with `ensemble=False`, without loading MLX. No weights or runtime assets were changed by this card update.
-The current fresh installed-wheel suite passed 122 tests with zero skips using independently recorded, seeded upstream numerical fixtures. Separate real released-weight API runs checked flagship execution and the complete external five-member ensemble with TTA; the latter does not imply the missing four members are included here. Those execution/output checks are separate from released-checkpoint cross-framework agreement and clinical accuracy. The following one-case reference comparison is historical and retains its original revision and settings.
-Historical native result: **success**. Original-code comparison: **pass**, `minimum_per_label_dice` = `0.9999704176268988`.
-Paper uses flux0375 only; the prepared bundle supports `ensemble=False`. The unchanged upstream ensemble descriptor lists all five members, but only the flagship member is packaged. Original comparison passes. Montgomery radiograph panels are excluded from submission/distribution pending image-sharing clearance.
-These are recorded results for one configuration, not a new benchmark or an
-estimate of clinical accuracy. The [native run](https://huggingface.co/MedMLX/flexray-mlx/blob/main/native-run.json)
-and [original-code comparison](https://huggingface.co/MedMLX/flexray-mlx/blob/main/reference-comparison.json)
-retain the commands, precision, geometry and output checks.
-<details>
-<summary>Checkpoint provenance and recorded configuration</summary>
-Recorded configuration: **FleXray flux0375 single member, label 'lungs', threshold 0.5, no TTA**.
-Evaluated package revision: `d768bb15d04ee677ecec0c469fd3170256b345df`.
-Recorded input: NLM Montgomery County CXR / MCUCXR_0001_0.
-Original publisher: VictorButoi.
-Original implementation: [upstream source](https://github.com/VictorButoi/FleXray).
-Source revision or version: `bf6d01426fd3ae012aff9253ec9c9952c58bc960`.
-Original checkpoint: [pinned checkpoint](https://huggingface.co/VictorButoi/flexray/resolve/6b63ddf09cfdff99aea494e3738cf243a4e22cdb/members/flux0375/model.safetensors).
-The released safetensors snapshot is loaded directly, without retraining.
-The [replay harness](https://github.com/MedMLX/Med-MLX/blob/main/benchmarks/README.md)
-describes how to reproduce the recorded configuration with local paths.
-</details>
-## Intended use
-Research and education. Outputs require human review and are not intended for
-clinical diagnosis or treatment decisions.
-## License
-Weights: **CC-BY-NC-4.0** ([license](https://huggingface.co/MedMLX/flexray-mlx/blob/main/LICENSE)).
-Upstream FleXray code is **MIT**. Original MLX port contributions are
-[Apache-2.0](https://github.com/MedMLX/flexray-mlx/blob/main/LICENSE), with retained MIT and BSD notices for FleXray, PyTorch and TorchVision adaptations in
-[THIRD_PARTY_NOTICES.md](https://github.com/MedMLX/flexray-mlx/blob/main/THIRD_PARTY_NOTICES.md). These code terms do not replace the noncommercial weight license above.
-Retain the [attribution and notices](https://huggingface.co/MedMLX/flexray-mlx/blob/main/NOTICE) when redistributing
-the bundle or a conversion:
-- Retain the applicable license, copyright and attribution notices; identify the upstream checkpoint, authors, and MLX conversion changes.
-- Credit the creator and source, link the applicable CC license, retain supplied notices, and identify conversion changes.
-- Label converted weights noncommercial and preserve the applicable noncommercial limitation.

 base_model: VictorButoi/flexray
 ---
+<img src="https://huggingface.co/MedMLX/flexray-mlx/resolve/main/banner.png" alt="MedMLX — Segment. RADIOGRAPHS / FLEXRAY" width="100%">
+# FleXray · Radiographs
+Segment anatomical structures in radiographs with the released FleXray flux0375 member.
+[GitHub](https://github.com/MedMLX/flexray-mlx) · [Files](https://huggingface.co/MedMLX/flexray-mlx/tree/main) · [Docs](https://github.com/MedMLX/flexray-mlx/blob/main/README.md#quick-start)
+| Input | Output |
+| :--- | :--- |
+| Radiograph | Anatomical segmentation |
+**Apple Silicon · Metal · Python 3.13 / 3.14**
 ## Quick start
 ```bash
 uv venv --python 3.13
 source .venv/bin/activate
+repo="https://github.com/MedMLX/flexray-mlx"
+release="$repo/releases/download/v0.2.1"
+uv pip install "$release/flexray_mlx-0.2.1-py3-none-any.whl"
 uv tool install huggingface-hub
 hf download MedMLX/flexray-mlx --local-dir models
 ```
+Replace the example input paths with your own files.
 ```python
 from flexray_mlx.medmlx import run
 print(result["outputs"])
 ```
+Includes the `flux0375` member. Keep `ensemble=False`.
+## Use & license
+For research and education. Not validated for clinical use.
+Weights: [CC-BY-NC-4.0 · noncommercial](https://huggingface.co/MedMLX/flexray-mlx/blob/main/LICENSE). [Notices](https://huggingface.co/MedMLX/flexray-mlx/blob/main/NOTICE).

banner.png ADDED Viewed

banner.svg ADDED Viewed