File size: 2,757 Bytes
30bbb7a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | {
"schema_version": "medmlx.reference.v1",
"model": "segresnet-brats",
"date": "2026-10-12",
"case": {
"dataset": "MSD Task01 BrainTumour (training)",
"case_id": "BRATS_001",
"input_sha256": {
"t1c_path": "96abcdbb541c4af37f03bef466eee5d11a3f91b9a1ea10b1a4a22cb38b7b3978",
"t1_path": "740b0adfa8de83fcbdfa5cd5f14689f6d947120bfbf057fe82954c07477bb173",
"t2_path": "8e2de5df0abd8b25ac1136c6570a6d03f2255107ef1265145b606f07acd1f1af",
"flair_path": "7571148ce691dccff5a84109091f43385b7c069a99cd7a967b2002aaeec64cdd"
}
},
"original": {
"code": "MONAI / MONAI model-zoo brats_mri_segmentation",
"version_or_commit": "MONAI 1.6.0 (eccefc57550b111ed781d82249dfe77872a0e918); model-zoo db9eeeef7ba64ce2ced0230d9bd4c524f69b7f0c",
"device": "cpu",
"precision": "float32 (AMP disabled on CPU)",
"checkpoint_sha256": "860ccb3f1c21c99d0410ad8a1ac4ef6b8fab60cec0a503b0ba42675741a750ae",
"command": "<paper-repo>/work/reference/segresnet-brats/.venv/bin/python -I -B <paper-repo>/benchmarks/reference/segresnet-brats.py"
},
"mlx": {
"run_record": "results/runs/segresnet-brats.json",
"output_path": "<paper-repo>/work/runs/segresnet-brats/out/brats_mri_segmentation_mask.nii.gz",
"output_sha256": "ce7f2ca25eeb2e1dea09e748dfdf1ec666a16296bc523a0b537608575ed5dcd6"
},
"comparison": {
"metric": "minimum_per_label_dice",
"value": 1.0,
"per_label": {
"0": {
"dice": 1.0,
"original_voxels": 8791633,
"mlx_voxels": 8791633
},
"1": {
"dice": 1.0,
"original_voxels": 27348,
"mlx_voxels": 27348
},
"2": {
"dice": 1.0,
"original_voxels": 74690,
"mlx_voxels": 74690
},
"4": {
"dice": 1.0,
"original_voxels": 34329,
"mlx_voxels": 34329
}
},
"threshold": 0.99,
"differing_voxel_count": 0,
"result": "pass"
},
"notes": "Pinned original MONAI CPU SegResNet and full bundle preprocessing, SlidingWindowInferer and postprocessing. Channel order T1gd,T1w,T2w,FLAIR is exactly the recorded MLX order; case.input_sha256 contains each recorded role's checksum. Original model.pt was found in the recorded HF snapshot 370f7f9d062745fbac445e7fe6d6616d35df04ec and matches source_sha256. Sigmoid, inverse transform, threshold 0.5 and original BraTS label precedence 4 > 1 > 2 used. Dice covers every label including background; all MRI grids and both mask affines/shapes match. AMP is disabled for CPU float32 inference. Raw outputs, source configuration, dependency versions and log: work/reference/segresnet-brats/. Training-exposed MSD case; this is reproducibility evidence, not an accuracy study.",
"failure": null
}
|