Image Segmentation
MLX
Safetensors
English
native-mlx
apple-silicon
medical-imaging
x-ray
segmentation
noncommercial
research
Instructions to use MedMLX/flexray-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use MedMLX/flexray-mlx with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download MedMLX/flexray-mlx --local-dir flexray-mlx
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Download reference-comparison.json from MedMLX/flexray-mlx: direct link, hf CLI and curl.
- Browser
- Download file 2.54 kB
-
https://huggingface.co/MedMLX/flexray-mlx/resolve/main/reference-comparison.json
- Command line
-
hf download hf://MedMLX/flexray-mlx/reference-comparison.json
-
curl -L -o reference-comparison.json https://huggingface.co/MedMLX/flexray-mlx/resolve/main/reference-comparison.json
2.54 kB
| { | |
| "schema_version": "medmlx.reference.v1", | |
| "model": "flexray", | |
| "date": "2026-10-12", | |
| "case": { | |
| "dataset": "NLM Montgomery County CXR", | |
| "case_id": "MCUCXR_0001_0", | |
| "input_sha256": "754d3ef3986e8682624f2e1c2cd14aacd3a503869f965a38235b6b07465e12da" | |
| }, | |
| "original": { | |
| "code": "https://github.com/VictorButoi/FleXray", | |
| "version_or_commit": "bf6d01426fd3ae012aff9253ec9c9952c58bc960", | |
| "device": "cpu", | |
| "precision": "float32", | |
| "checkpoint_sha256": "716ce3a8fb2a6abe22b2f2188fd596ddc4642b273fb0d336d629448154fc6d24", | |
| "command": "<paper-repo>/work/reference/flexray/.venv/bin/python -I -B <paper-repo>/benchmarks/reference/flexray.py" | |
| }, | |
| "mlx": { | |
| "run_record": "results/runs/flexray.json", | |
| "output_path": "<paper-repo>/work/runs/flexray/out/source_image_label.png", | |
| "output_sha256": "b15043cce30dd30b470f153c7722f903405965aa04010258e8b039a258bd57a9", | |
| "additional_outputs": { | |
| "probabilities": { | |
| "path": "<paper-repo>/work/runs/flexray/out/probabilities.npy", | |
| "sha256": "7e2c354963f329e6bf3001586d732b8ff6dd1e02f12c0f32bbe898b7cdd39f75" | |
| } | |
| } | |
| }, | |
| "comparison": { | |
| "metric": "minimum_per_label_dice", | |
| "value": 0.9999704176268988, | |
| "per_label": { | |
| "0": { | |
| "dice": 0.9999866929450898, | |
| "original_voxels": 13564053, | |
| "mlx_voxels": 13564414 | |
| }, | |
| "1": { | |
| "dice": 0.9999704176268988, | |
| "original_voxels": 6101787, | |
| "mlx_voxels": 6101426 | |
| } | |
| }, | |
| "threshold": 0.99, | |
| "differing_voxel_count": 361, | |
| "probabilities_diagnostic": { | |
| "cosine_similarity": 0.9999999999978441, | |
| "max_abs_diff": 1.3709068298339844e-05 | |
| }, | |
| "result": "pass" | |
| }, | |
| "notes": "Unmodified hash-admitted original UNet, preprocessing and InferenceRunner; flux0375 only, lungs channel 59, sigmoid, threshold 0.5, no TTA. The PNG source-grid export follows the package's original fixture recipe: threshold, crop padding, then nearest resize. Dice includes both binary labels; all 61 probability channels are also compared as a diagnostic. No source_sha256 exists in the run record: it records the original released safetensors directly; member_safetensors_sha256 is its original-source identity. The checkpoint directory moved under research/radnn. Raw outputs, source archive, dependency versions and log: work/reference/flexray/. Montgomery image remains local under its recorded restriction. One-case reproducibility comparison; no accuracy or performance qualification.", | |
| "failure": null | |
| } | |