Image Segmentation
MLX
English
native-mlx
apple-silicon
medical-imaging
ct
kidney
segmentation
research
Instructions to use MedMLX/unest-renal-ct-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use MedMLX/unest-renal-ct-mlx with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download MedMLX/unest-renal-ct-mlx --local-dir unest-renal-ct-mlx
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Download reference-comparison.json from MedMLX/unest-renal-ct-mlx: direct link, hf CLI and curl.
- Browser
- Download file 8.04 kB
-
https://huggingface.co/MedMLX/unest-renal-ct-mlx/resolve/main/reference-comparison.json
- Command line
-
hf download hf://MedMLX/unest-renal-ct-mlx/reference-comparison.json
-
curl -L -o reference-comparison.json https://huggingface.co/MedMLX/unest-renal-ct-mlx/resolve/main/reference-comparison.json
8.04 kB
| { | |
| "schema_version": "medmlx.reference.v1", | |
| "model": "unest", | |
| "date": "2026-10-12T16:48:01.119295+00:00", | |
| "case": { | |
| "dataset": "AMOS22 (validation)", | |
| "case_id": "amos_0041", | |
| "input_sha256": "80b1001c8227d5c5df84f88344fa73ca8ab68334a6e0f9c733d4a4e28d8c42ad" | |
| }, | |
| "original": { | |
| "code": "MONAI/renalStructures_UNEST_segmentation", | |
| "version_or_commit": "0.2.7; MONAI 1.4.0, PyTorch 2.4.0", | |
| "precision": "float32; evaluator amp=false", | |
| "command": "cd <paper-repo>/work/reference/unest/bundle && <paper-repo>/work/reference/unest/.venv/bin/python -m monai.bundle run evaluating --meta_file configs/metadata.json --config_file configs/inference.json --bundle_root <paper-repo>/work/reference/unest/bundle --device cpu --datalist '[\"<repositories>/research/radnn/work/nnunet-next-20261001/public-ct/amos_0041/image.nii.gz\"]' --output_dir <paper-repo>/work/reference/unest/out '--dataloader#num_workers' 0 '--handlers#0#map_location' cpu", | |
| "source_files_sha256": { | |
| "scripts/__init__.py": "b3d76349dea4bd588f9c5bcc475d4381ddf42efe2863a15a3c9af86592592ead", | |
| "scripts/networks/__init__.py": "b3d76349dea4bd588f9c5bcc475d4381ddf42efe2863a15a3c9af86592592ead", | |
| "scripts/networks/nest/__init__.py": "d9dde32a3a5ac3578b5fa8c1fecab9b7af7311a852017bef1e3016c98e7bfe53", | |
| "scripts/networks/nest/utils.py": "c4beaa898e7a846de407b82378f5f3f54516718147c03da021e05b278f884b9a", | |
| "scripts/networks/nest_transformer_3D.py": "3bc882fc948301b5a57b6919088a4d8d9a11d931add1b6765435213887dbfefd", | |
| "scripts/networks/patchEmbed3D.py": "c115045843b844119aff508498c65627b2e97e32dab1d026bc0f9c8ba00f8d42", | |
| "scripts/networks/unest.py": "530e4caed71a34152263759eae5f4f1429bd3bcd167cdcb724d85f6136c5199d", | |
| "scripts/networks/unest_block.py": "dec389f37991f07c0803f85e8d3041c86f597c583fb08b0aa9179fae61b7faa1", | |
| "configs/inference.json": "b09b12efbce60ebbc16edb50ef4529be061709c8c6f127aa1df4c9fa3f3e4563", | |
| "configs/metadata.json": "1f31c4cf047e3e8ec5d2dc9a9b638728514b5bf2a7ba95cbb6a1140ab34dddb7", | |
| "configs/multi_gpu_train.json": "b99950a7b4564b03048edb0571601e4a91bbcfa761f00dc12464cb3f9a7fb48f", | |
| "configs/train.json": "5f4ef1ce140605712b5301d4dd7b2c80dffcf970cfef4589d9d03b4e166b3d4a" | |
| }, | |
| "device": "cpu", | |
| "checkpoint_sha256": "8928e88771d31945c51d1b302a8448825e6f9861a543a6e1023acb9576840962" | |
| }, | |
| "mlx": { | |
| "run_record": "results/runs/unest.json", | |
| "output_path": "<paper-repo>/work/runs/unest/amos_0041/out/mask.nii.gz", | |
| "output_sha256": "69d2db5a95a89392cda3853bee2eb8bb268626e7bedec5856d80473b2bdda8c3" | |
| }, | |
| "comparison": { | |
| "metric": "minimum per-label mask Dice", | |
| "value": 1.0, | |
| "per_label": { | |
| "0": { | |
| "dice": 1.0, | |
| "original_voxels": 21395717, | |
| "mlx_voxels": 21395717, | |
| "intersection_voxels": 21395717 | |
| }, | |
| "1": { | |
| "dice": 1.0, | |
| "original_voxels": 81477, | |
| "mlx_voxels": 81477, | |
| "intersection_voxels": 81477 | |
| }, | |
| "2": { | |
| "dice": 1.0, | |
| "original_voxels": 15224, | |
| "mlx_voxels": 15224, | |
| "intersection_voxels": 15224 | |
| }, | |
| "3": { | |
| "dice": 1.0, | |
| "original_voxels": 3390, | |
| "mlx_voxels": 3390, | |
| "intersection_voxels": 3390 | |
| } | |
| }, | |
| "threshold": { | |
| "min_dice": 0.99, | |
| "scope": "every label present in either mask, including background" | |
| }, | |
| "result": "pass", | |
| "differing_voxels": 0, | |
| "geometry": { | |
| "shape": [ | |
| 512, | |
| 512, | |
| 82 | |
| ], | |
| "affine_max_abs_diff": 0.0 | |
| }, | |
| "original_output_path": "<paper-repo>/work/reference/unest/out/image/image_trans.nii.gz", | |
| "original_output_sha256": "950cf267574172ec8fbd175049f409c536f74e1b64ec2f06620b272fea8e59d4", | |
| "retained_cuda_vs_saved_mlx": { | |
| "metric": "minimum per-label mask Dice", | |
| "value": 0.9998524856173477, | |
| "per_label": { | |
| "0": { | |
| "dice": 0.9999999298925124, | |
| "original_voxels": 21395718, | |
| "mlx_voxels": 21395717, | |
| "intersection_voxels": 21395716 | |
| }, | |
| "1": { | |
| "dice": 0.9999815897835572, | |
| "original_voxels": 81476, | |
| "mlx_voxels": 81477, | |
| "intersection_voxels": 81475 | |
| }, | |
| "2": { | |
| "dice": 0.9999671581989557, | |
| "original_voxels": 15225, | |
| "mlx_voxels": 15224, | |
| "intersection_voxels": 15224 | |
| }, | |
| "3": { | |
| "dice": 0.9998524856173477, | |
| "original_voxels": 3389, | |
| "mlx_voxels": 3390, | |
| "intersection_voxels": 3389 | |
| } | |
| }, | |
| "threshold": { | |
| "min_dice": 0.99, | |
| "scope": "every label present in either mask, including background" | |
| }, | |
| "result": "pass", | |
| "differing_voxels": 4, | |
| "geometry": { | |
| "shape": [ | |
| 512, | |
| 512, | |
| 82 | |
| ], | |
| "affine_max_abs_diff": 0.0 | |
| }, | |
| "original_output_path": "<paper-repo>/work/runs/unest/diagnosis/original_amos_0041_pred.nii.gz", | |
| "original_output_sha256": "5d9eb307a7e8dc6a6d2aaf24eae75667dfd383349b9b18a09c7d6f515ebf0836" | |
| }, | |
| "cpu_vs_retained_cuda": { | |
| "metric": "minimum per-label mask Dice", | |
| "value": 0.9998524856173477, | |
| "per_label": { | |
| "0": { | |
| "dice": 0.9999999298925124, | |
| "original_voxels": 21395717, | |
| "mlx_voxels": 21395718, | |
| "intersection_voxels": 21395716 | |
| }, | |
| "1": { | |
| "dice": 0.9999815897835572, | |
| "original_voxels": 81477, | |
| "mlx_voxels": 81476, | |
| "intersection_voxels": 81475 | |
| }, | |
| "2": { | |
| "dice": 0.9999671581989557, | |
| "original_voxels": 15224, | |
| "mlx_voxels": 15225, | |
| "intersection_voxels": 15224 | |
| }, | |
| "3": { | |
| "dice": 0.9998524856173477, | |
| "original_voxels": 3390, | |
| "mlx_voxels": 3389, | |
| "intersection_voxels": 3389 | |
| } | |
| }, | |
| "threshold": { | |
| "min_dice": 0.99, | |
| "scope": "every label present in either mask, including background" | |
| }, | |
| "result": "pass", | |
| "differing_voxels": 4, | |
| "geometry": { | |
| "shape": [ | |
| 512, | |
| 512, | |
| 82 | |
| ], | |
| "affine_max_abs_diff": 0.0 | |
| }, | |
| "original_output_path": "<paper-repo>/work/reference/unest/out/image/image_trans.nii.gz", | |
| "original_output_sha256": "950cf267574172ec8fbd175049f409c536f74e1b64ec2f06620b272fea8e59d4" | |
| } | |
| }, | |
| "notes": "Same full amos_0041 input and checkpoint, unchanged bundle preprocessing, 96^3 windows, sw_batch_size=4, overlap=0.5 and postprocessing. Only device, datalist, output location, DataLoader workers (0), and checkpoint map_location=cpu are overridden. The initial run failed because the checkpoint stored CUDA tensors and the bundle loader omitted map_location; checkpoint-device-attempt.log is retained. The retained diagnosis covers this case/settings but used CUDA, so it is reused as secondary evidence and a CPU reference is executed to meet the contract. amos_0041 was selected after the amos_0008 checkpoint failure; this is a post-failure demonstration case, not a prospectively selected accuracy sample. Recorded radnn input/checkpoint paths moved to GitHub/research/radnn; hashes match.", | |
| "failure": null, | |
| "verification": { | |
| "inputs": { | |
| "image": { | |
| "path": "<repositories>/research/radnn/work/nnunet-next-20261001/public-ct/amos_0041/image.nii.gz", | |
| "sha256": "80b1001c8227d5c5df84f88344fa73ca8ab68334a6e0f9c733d4a4e28d8c42ad" | |
| } | |
| }, | |
| "mlx_outputs": { | |
| "mask": { | |
| "path": "<paper-repo>/work/runs/unest/amos_0041/out/mask.nii.gz", | |
| "sha256": "69d2db5a95a89392cda3853bee2eb8bb268626e7bedec5856d80473b2bdda8c3" | |
| } | |
| }, | |
| "checkpoint": { | |
| "path": "<repositories>/research/radnn/checkpoints/monai/model-zoo/renalStructures_UNEST_segmentation/models/model.pt", | |
| "sha256": "8928e88771d31945c51d1b302a8448825e6f9861a543a6e1023acb9576840962" | |
| } | |
| } | |
| } | |