Skip to content

External model integration ledger

This ledger distinguishes automatic models from prompted tools and records whether a release can be reproduced. The records below describe historical integration runs. Current selection depends on runtime preflight, segmentation orientation evidence or separate report acceptance, and the enabled-model policy. Research use only; execution and geometry results do not establish clinical accuracy.

Input-storage correction (2026-10-15)

Stored DICOM orientation previously reached array-based models unchanged, so prone and feet-first series gave those models differently stored arrays than standard head-first-supine series. A synthetic matrix (fake lateral model) reversed a lateral label for HFP and FFS storage in every plane. Most public test fixtures converted from NIfTI are stored R,A,S or L,A,I, while the internal clinical fixtures are L,P,S.

The poor LNCTVSeg and HNLNL results on the public head-and-neck example (LNCTVSeg Dice 0.187551, 0.228038, 0.017221, 0.029241, 0, 0 for labels 1–6; HNLNL sided levels on both sides) were first attributed to this storage difference. GPU re-runs on 2026-10-15 showed otherwise: with the contract feeding L,P,S the LNCTVSeg result was unchanged. The cause was nnU-Net's mirror test-time augmentation, which averages left/right-flipped predictions of side-specific structures. Without mirroring, LNCTVSeg reached Dice 0.860, 0.858, 0.861, 0.896, 0.879, 0.888 on the full volume and passed the laterality audit; HNLNL 2D placed 14 of 22 sided components on the wrong side with mirroring and 1 of 17 without. Both now run without mirror augmentation. Fixture provenance remains the authors' example at the pinned source revision; this development-cohort case cannot establish independent accuracy.

A shared lossless input contract now supplies standard HFS storage for the acquisition plane: axial LPS, coronal LIP, sagittal PIR. Oblique images follow the closest-axis rule without resampling. Standard HFS arrays, affines and NIfTI files are left untouched. Both RTSTRUCT writers invert the storage transform against the original DICOM grid; prompt coordinates follow the image transform. Existing physical round-trip thresholds remain unchanged. There is no separate LNCTVSeg or HNLNL orientation shim. MRAnnotator's documented LAS input and LNQ's physical atlas requirements remain in place.

Models trained on other storage keep it explicitly. MRAnnotator receives L,A,S. The MONAI whole-brain UNesT bundle (monai_whole_brain_unest_mr, no Orientationd in its pipeline) receives R,A,S: with L,P,S input 55 of 61 left/right pairs were mirrored on the MNI152 template, and with R,A,S input a real T1 brain had 0 of 63 pairs swapped (H100, 2026-10-16). Labels are mapped back to the original grid exactly.

Exposure includes nnU-Net v1/v2 and other stored-array runners, their crop/fusion pipelines and prompted tools. Tasks with internal canonicalisation or registration retain that processing. The operations inventory describes all route families, inactive entries and the manager GPU matrix. Earlier outputs from non-HFS storage produced by exposed tasks should be regenerated. GPU storage-variant checks (HFS, HFP, FFS, FFP re-exports and the original series) gave identical physical predictions for CADS head-and-neck OARs and LTRC lung lobes (Dice 1.0, laterality passed in every variant) and for LNCTVSeg (Dice ≥ 0.9997). Software geometry evidence and local drafting do not establish clinical or manuscript readiness.

NVIDIA NV-Segment automatic tasks (2026-10-15)

Four automatic tasks are integrated behind separate installation, acceptance and orientation gates. Acceptance and orientation markers are kept in the model bundle, not in this repository. The existing vista3d_prompted_ct source, runner and acceptance remain unchanged.

Task Input Foreground labels Scope
nv_segment_ct_auto CT 117 NV-Segment-CT automatic selection
nv_segment_ctmr_ct_body CT 117 NV-Segment-CTMR CT_BODY
nv_segment_ctmr_mri_body MR 50 NV-Segment-CTMR MRI_BODY, TotalSegmentator-MRI classes
nv_segment_ctmr_mri_brain T1 MR 132 NV-Segment-CTMR MRI_BRAIN, LUMIR regions

The publisher's automatic CT selection excludes 15 of 132 CT classes, including the solid-tumour heads: IDs 2, 16, 18, 20, 21, 23–27 and 128–132. Renal cyst labels use “Lesion candidate”. This mode does not generate automatic tumour targets. Brain's stated 133 classes include background; label 255 is a foreground region and is preserved. All exported labels use uint16 and retain the publisher's sparse IDs.

Provenance and licences. NV-Segment-CT uses the NVIDIA Open Model License. Its checkpoint revision is afb51518689f71e6abb367ee6301b2cd0225c66a; weights/vista3d/NV-Segment-CT/model.pt is 871,970,895 bytes, SHA-256 c92bab26d00b4a5d89fa8a383900cdeb88302fd318e5e816df0bbec7106d9a1b. The source is pinned to f9f5f51b589e5dc9c23c453cf5138398e4084056 under sources/NV-Segment-CTMR/NV-Segment-CT.

NV-Segment-CTMR uses checkpoint revision 4fb8b4a6b2532be9f1c449a3726fe5440ab4213a. weights/vista3d/NV-Segment-CTMR/model.pt is 871,950,375 bytes, SHA-256 6c6d86a48a7a9d377ddb41d663d487cfedd6456df3b828ecd931e5cba5e71f99. The source is pinned separately to cb921f5c58837c0f42a713855d68b32af88e1cdd under sources/NV-Segment-CTMR-cb921f5c/NV-Segment-CTMR. Code is Apache-2.0. The weight licence, section 3.3 restricts use to noncommercial research or evaluation, with the stated NVIDIA/affiliate exception. CTMR weights do not use the CT Open Model licence. Source-file and label-map hashes for all modes are recorded in nv_segment_manifest.json and checked before inference.

Runtime and brain preprocessing. Segmentation uses envs/lnq (prepared with Python 3.11.14, PyTorch 2.13.0+cu130 and MONAI 1.6.0). The brain pipeline also needs runtime-envs/nv-segment (NumPy 1.26.4, SciPy 1.15.3, antspyx 0.6.1, Surfa 0.6.3 and pymedio 0.2.14), with its recorded package lock and lnq package linkage. These prepared runtime versions are provenance, not a claim of Pascal support. No packages or checkpoints are downloaded during a job.

Brain mode runs SynthStrip internally, then the publisher's N4/intensity-normalization and affine registration to the LUMIR template. It inverts the transform with nearest-neighbour labels and restores the exact original grid. Repairs are applied only to a job copy: MONAI point conversion, NaN background 0, strict checkpoint loading, uint16 export and exact inverse-grid dimensions. The output writer rejects noninteger, unknown or empty labels and grid mismatches.

SynthStrip 1.8 is extracted under runtime-homes/nv-segment/synthstrip-1.8; Docker is not needed for inference. Image digest: sha256:ebbc177221194371f16362513ace68312a22922bb581bdfa618ac7ff9c1d2c06. The standard checkpoint SHA-256 is 37417f802196186441aae3e7f385d94f8a98c64a88acaeaa2723af995c653e33; the script SHA-256 is fa18eb301568dc7c99bd9f69805545a7ce6d68e3fa3e1775c68053c9069cc4a1. The publisher offers weights under MIT or CC BY 4.0; cite Hoopes et al., NeuroImage 260 (2022), 119474.

MRI sequence scope. MRI_BODY selects the publisher's 50 TotalSegmentator-MRI classes. Sequence and field of view affect output; this integration establishes no sequence-specific performance. For brain, the pinned model card specifies standard T1, while the pinned source documentation also names T2 and FLAIR. The configured audit uses T1 only. Non-T1 brain acceptance remains unresolved.

GPU acceptance (2026-10-15). All four tasks passed the orientation audit and acceptance on public fixtures on an NVIDIA H100 (sm_90), and the validation sweep on a Quadro P6000 (Pascal sm_61, 24 GiB). RTSTRUCT ROI sets were identical on the two GPUs. The existing point-prompted vista3d_prompted_ct also passed on the P6000 with a point placed on the spleen of the public MSD spleen fixture.

Task Public fixture H100 inference RTSTRUCT ROIs P6000 job (upload to RTSTRUCT)
nv_segment_ct_auto MSD lung_053 10.6 s 100 498 s
nv_segment_ct_auto KiTS19 case00000 16.4 s 68 —
nv_segment_ctmr_ct_body KiTS19 case00000 13.0 s 69 661 s
nv_segment_ctmr_mri_body TotalSegmentator-MRI s0052 8.4 s 21 81 s
nv_segment_ctmr_mri_brain OpenNeuro ds002168 T1 4.9 s 132 regions 229 s

Peak H100 memory reached 21.3 GiB (CT automatic, KiTS19). Agreement with TotalSegmentator total on the H100 (Dice; model agreement, not accuracy): NV-Segment-CT automatic gave liver 0.98, spleen 0.97, kidneys 0.95–0.96, aorta 0.93–0.94, pancreas 0.91–0.94 and whole lungs 0.91–0.98 on both CT fixtures. CTMR CT_BODY was similar except pancreas 0.35 on the chest CT, where the pancreas is only partly in the field of view. CTMR MRI_BODY against the TotalSegmentator-MRI s0052 reference labels gave liver 0.92, left/right lung 0.89/0.93 and spleen 0.77. Left/right placement of paired organs and of all brain hemisphere pairs was correct.

Small spurious labels occur: on the chest CT, NV-Segment-CT drew left/right femur (277/1,068 voxels) and hip (4/327 voxels) labels, and MRI_BODY drew a 489-voxel brain label in a torso MRI. Review every contour in its anatomical context. These technical results do not establish clinical or manuscript readiness.

Orientation fixtures and acceptance. The configured public DICOM fixtures are MSD lung for CT automatic, KiTS19 case00000 for CTMR CT body, TotalSegmentator-MRI s0052 for MRI body, and the existing hippunfold-ds002168-T1w/dicom fixture for brain. The latter contains 256 DICOM files and derives from the public OpenNeuro ds002168 dataset; the publisher's separate brain NIfTI example is not used as a substitute for a DICOM round trip. The audit checks execution, source-grid conversion, physical contour round trips and model identity. It does not assess segmentation accuracy. See operations for marker generation and local/remote activation.

NVIDIA NV-Reason-CT draft reports (2026-10-15)

Task nv_reason_ct_report produces JSON, Markdown and HTML for diagnostic chest/abdomen CT; it does not produce RTSTRUCTs. Publisher and license: OpenMDW-1.1, revision 386b93e034983f6c1fc841a43833a1b6a0cd9c13. The checkpoint SHA-256 is 655c15e59dfa1957f7d5f475b1702d689fb7107160f21256028391bb5f6e6ba5 (10,644,167,648 bytes).

Technical acceptance of this version on an H100 (2026-10-15, fast coverage): a chest report for public MSD lung CT with total segmentation; an abdomen report for public KiTS19 case 00000 in a report-only job, using the segmentation-guided crop after the publisher's lung-based crop fell back; no AI text for a pelvic CT without a supported region; and HTTP 400 for a CT series tagged as CBCT. Both generated reports' crops contained 100% of the segmented regional organs. Fast coverage took 156–270 s, model loading 19–23 s and warm-cache generation 57–64 s for about 400 tokens; a chest job took about 7 minutes. Peak model-process allocation was 11.2 GiB; plan for about 12 GiB, and the worker requires 14 GiB free before loading. A mixed job (total on a local Pascal GPU, report routed over SSH to an H100 worker) also completed in a production-like deployment. Greedy decoding is deterministic within one process but not bit-identical across processes, so repeated runs can word a report differently.

Known limitations:

  • Without anatomical gating, the model describes organs outside the field of view. It can generate confident findings from unreadable input. Coverage masks can also be wrong.
  • On public KiTS19 case 00000, it described a 7.8 mL (about 2.6 cm) renal tumour as a hypodensity “too small to characterize”. Segmentation checks do not validate diagnostic findings.
  • CBCT is refused. A negative CBCT header check does not establish diagnostic image quality.
  • Use an sm_80-class or newer CUDA GPU. A 24 GiB Pascal GPU could not run the model: bf16 SDPA exhausted memory, fp16 failed in cuDNN conv3d, and disabling cuDNN also exhausted memory. Per-task SSH routing supports a separate compatible worker.

TotalSegmentator total --fast (3 mm) is the default coverage mode, without an ROI subset; RTSEG_NV_REASON_CT_COVERAGE_MODE=full restores full resolution. On three test series (chest, abdomen, pelvis), fast and full coverage gave identical region decisions and per-organ statuses; fast took 140–270 s and full 223–378 s on the H100. Report availability uses validation/nv_reason_ct_report.json, independently of segmentation orientation markers. Radiologist review is required; technical acceptance and local drafting do not establish clinical or manuscript readiness.

Spinal Cord Toolbox 7.3 (2026-08-13)

  • The official SCT 7.3 runtime and all 15 requested catalog-task checkpoint sets are installed.
  • Fourteen single-series tasks are represented in catalog/sct.json. The T1+T2 tumor/edema/cavity task is visible but blocked by its registered two-series input contract.
  • sct_spinalcord passed H100 inference, non-empty RTSTRUCT conversion, source-SOP validation, model-specific DICOM identity, and patient-space orientation audit. The other installed tasks remain acceptance-gated until suitable sequence-specific public fixtures produce non-empty predictions; installation is not presented as proof.
  • The complete runtime is a checksum-protected read-only SquashFS image on NFS at runtime-homes/sct/sct-7.3.squashfs, mounted at runtime-homes/sct/sct_7.3. The 17 GB worker-local staging tree was removed after mounted-runtime validation.

GOUHFI 2.0 (2026-08-13)

  • Official tag 2.0.0 source and both public Zenodo model archives were downloaded directly to NFS and matched the publisher's MD5 values. Each task uses all five folds.
  • gouhfi_v2_brain_mr emitted all 35 released labels; its strictly validated RTSTRUCT has 35 non-empty ROIs and references 78 images from the selected series.
  • gouhfi_v2_dkt_mr emitted all 62 bilateral DKT labels; its strictly validated RTSTRUCT has 62 non-empty ROIs and references 67 selected-series images.
  • The portal path applies brain extraction and the official LIA/intensity conformation, runs official inference and postprocessing, restores results to native DICOM geometry, and applies the model-specific DICOM identity fields. Both orientation audits passed. This is a research model, not a clinically approved contouring system.

SIAM v3 (2026-08-13)

  • The current SIAM source and official pred_DS108_LcsfP_Ano five-fold archive are installed on NFS; the 5.1 GB archive matches Zenodo MD5 0c3ca6d0730a3dcb924e079af04474c3.
  • siam_v3_head_mr_ct ran all five folds on the H100 and produced all 17 checkpoint labels. The 4,704,902-byte RTSTRUCT contains 17 non-empty ROIs, references 91 selected-series images, carries model-specific DICOM identity, and passed patient-space orientation validation.
  • SIAM accepts MRI contrasts and CT, but the acceptance fixture was T1 MRI. Its anomaly output is retained as Anomaly_Candidate; it is not presented as a definitive tumour, lesion diagnosis, GTV or CTV.

FastSurferVINN (2026-08-13)

  • The official stable source and all three VINN view checkpoints are installed on NFS. The shared CUDA runtime needs only a small NFS-hosted h5py overlay.
  • fastsurfervinn_t1_mr ran axial, coronal and sagittal inference on the H100, emitted 95 non-empty cortical/subcortical labels, and produced a 1,704,354-byte RTSTRUCT referencing 77 selected-series images.
  • The wrapper requests native orientation, converts the upstream MGZ result to NIfTI without changing its affine, and nearest-neighbour restores labels to the original DICOM grid. Model-specific identity and strict orientation audits passed.
  • The model expects a whole-head T1-weighted MRI. Outputs are research anatomy labels and do not define glioma, metastasis, GTV or CTV boundaries.

TractSeg 2.10 (2026-08-13)

  • Official source and all nine public pretrained archives are installed and checksum-indexed on NFS (1.3 GB total).
  • The released peaks fixture ran successfully through the official CUDA path: all 72 tract channels were non-empty with 237,552 foreground voxels.
  • The portal entry remains disabled with status requires_diffusion_peaks_or_dwi_gradients. TractSeg requires a 9-channel CSD/bedpostX peaks NIfTI, or raw 4D DWI plus matching b-values and b-vectors. A selected scalar 3D DICOM series cannot satisfy that contract, so no false RTSTRUCT acceptance is claimed.

BraTS Toolkit legacy ensemble (2026-08-13)

  • Official source commit 86e7fa89b48e965b6c23151931f88e074930f75e is installed on NFS. All legacy image artifacts are stored under images/brats-toolkit; model preflight requires every one before reporting the family installed.
  • The MIC-DKFZ image is an intact 2018 CUDA 9/PyTorch 0.4 environment. Its embedded CUDA kernels cannot execute on the H100, although the runtime sees the device. A documented compatibility patch maps its released checkpoint to CPU and makes the otherwise read-only /params tree writable through an NFS bind mount.
  • Released fold-0 inference on Brats18_TCIA10_639_1 passed. Output shape and affine exactly matched the 240x240x155 source, with labels 1, 2 and 4 all non-empty (103,221; 79,248; and 1,009 voxels respectively). Output SHA-256 is ca13836cbeda6fff78546985dabf62fc9f0b37b75c5e6af381b8a1661d642150.
  • This is not selectable: all toolkit algorithms require skull-stripped, co-registered FLAIR, T1, T1c and T2 NIfTI inputs in SRI-24 space. The portal currently selects one DICOM series, so silently substituting or guessing the other channels would be unsafe.

CEREBRUM-7T and DeepBrainSeg (2026-08-13)

  • CEREBRUM-7T source and its included 14.9 MB HDF5 checkpoint are installed and checksum-indexed. The published inference performs voxelwise z-scoring using training-cohort mean and standard-deviation volumes, but those two volumes are absent from the repository/release. The model is therefore visible with the precise missing_voxelwise_training_normalization blocker; invented scalar normalization is not accepted as reproduction.
  • The catalog's stale mobarakol/DeepBrainSeg URL was corrected to the actual koriavinash1/DeepBrainSeg upstream. Source, BraTS fixtures and all four 0.0.1 checkpoints are on NFS; all four state dictionaries load successfully in the current PyTorch runtime. It remains input-contract blocked because inference requires registered T1, T1ce, T2 and FLAIR volumes plus a brain mask.
Release Input Download state Execution state
HNLNL 2D+3D ensemble H&N planning CT Complete 2D+3D five-fold release downloaded MIRROR TTA REMOVED (2026-10-15); re-validation required: nnU-Net mirror test-time augmentation mixed left and right levels (laterality audit above), so predictions now use --disable_tta. Earlier: found and isolated a 516-slice 1.5-mm planning H&N CT. The exact portal branch was smoke-tested through all ten folds, probability-level 2D+3D averaging, published component postprocessing, RTSTRUCT serialization and cleanup. All 20 classes were nonempty and the strictly valid 20-ROI RTSTRUCT references 125 source images; the multi-gigabyte temporary probabilities were removed afterward. The exact numeric label order was recovered from LevelNameList in the paper's Europe PMC DataSheet_2.zip supplement: Ia, bilateral Ib/II/III/IVa/IVb/V, VIa/VIb/VIIa, and bilateral VIIb/VIII.
Choi TMLI whole-body/planning CT Existing five-fold Zenodo release; voxel-space geometry contract and anatomical review workflow GEOMETRY AND ONTOLOGY UNCONFIRMED (2026-10-15): released training metadata use dummy 1 mm spacing and 219–275 × 512 × 512 arrays; the paper reports mean 1.113 × 1.113 mm pixels and 5 mm slices. The previous true-spacing input introduced about fivefold cranio-caudal scale mismatch. The runner now resamples HU at the paper scale, pads to at least 512 × 512, tests eight explicit LPS-axis flip variants, supplies a dummy identity-LPS/1 mm header, and restores labels to the original CT grid. GPU screening on an internal whole-body CT (2026-10-15; 2.7–11 min per fold after the geometry fix, previously 34–57 min) found no coherent variant: best mean inter-fold Dice 0.291 (variant 000) and 0.250 (110), and the five-fold ensemble failed the coherence screen. The intensity offset (+1024) matches the released intensity statistics. Default variant stays unconfirmed; the label map is not sent for clinical review while predictions are incoherent. The review sheet starts with component-based anatomical summaries and generates a reviewer CSV. Task remains disabled. See geometry and review workflow.
CL-Net 235-class checkpoint CT Existing archive matches the current public release; no download repeated NODAL CHECKPOINTS ABSENT (2026-10-15): HF API revision a18539d2a8473c0eca1b4e6160523b1de12c65f3, ckpt.zip 985,573,570 bytes (939.916 MiB), SHA-256 d6ab11b67236eb647e8f84c2daebbe33eb242693b9e49da85de706e42ec63c0d. All seven decoder bpr_scores.json maps contain no LN labels; Task1007/Task1008 checkpoints remain absent. Source configuration includes 33 named nodal BPR entries and decoder index ranges, but no matching public nodal weight package. The 986 MB listing is the same archive, not a newer nodal release.
ISRT CT-only (ResUNet, SegResNet, SwinUNETR) planning CT All 9 weights downloaded; all three ensembles portal-integrated PASS: each three-checkpoint ensemble produced a nonempty mask and valid RTSTRUCT: ResUNet 152,010 foreground voxels/112 referenced images; SegResNet 38,659/114; SwinUNETR 236,098/128
ISRT PET fusion variants registered planning CT + PET1, optionally PET2 All 36 weights downloaded; four PET1-only ablation ensembles portal-integrated TECHNICAL PASS, DOMAIN WARNING: same-Frame-of-Reference CT/PSMA-PET with BQML converted to SUVbw×1000. PET1 early/deform (48,177 voxels), early/rigid (13,996), late/deform (21,746), and late/rigid (29,898) each produced a valid one-ROI RTSTRUCT. The fixture lacks a distinct planning CT, so the AC CT was supplied for both CT channels; it is also outside the pediatric Hodgkin/FDG domain. The PET1+PET2 releases remain input-blocked because no longitudinal PET2 is available
Pediatric thoracic lymphoma contrast CT Full nnU-Net v1 export downloaded, MD5 verified, extracted, and portal-integrated PASS: nonempty mask; valid one-ROI RTSTRUCT references 192 source images. Tested source is low-dose AC CT rather than pediatric contrast CT, so protocol warning is shown
SegRap 2023 Task 2 paired noncontrast + contrast H&N CT All five folds downloaded TECHNICAL PASS, INPUT WARNING: the wrapper was repaired to use TotalSegmentator 2.17 for body cropping (upstream silently continued after its pinned TotalSegmentator failed). All five folds produced GTVp 34,720 and GTVnd 536 voxels; valid 2-ROI RTSTRUCT/56 refs. Only one appropriate planning H&N CT was available, so it was deliberately duplicated into native and contrast channels; this validates execution/geometry but not paired-contrast clinical performance
UWLAIR HNTS-MRG pre/mid-RT T2 MRI Full 28.5 GB, 20-model Dropbox ensemble downloaded, integrity-tested, and extracted PASS: public HNTS-MRG24 patient 101. Pre-RT produced GTVp 3,529 and GTVn 21,767 voxels; valid 2-ROI RTSTRUCT/33 refs. Mid-RT produced GTVn 15,327 voxels; valid 1-ROI RTSTRUCT/29 refs. The single-series pre-RT 10-model ensemble is portal-integrated as uwlair_hntsmrg_pre_t2; a portal job completed with GTVp/GTVn RTSTRUCT and HTTP download. The source fixture is NIfTI-only, so RTSTRUCT geometry was validated against explicitly synthetic, geometry-faithful MR DICOM references. Mid-RT remains validation-only because it requires paired pre-RT image/mask registration inputs, not one selected series
WltyBY HNTS-MRG T2 MRI Source cloned; all five Task 1 and all five Task 2 weights downloaded and reconstructed into runnable nnU-Net exports PASS: public HNTS-MRG24 patient 101. Task 1 produced GTVp 3,252 and GTVn 21,322 voxels; valid 2-ROI RTSTRUCT/32 refs. Task 2 (mid-RT T2 + registered pre-RT T2 + registered pre-RT mask) produced GTVn 13,311 voxels; valid 1-ROI RTSTRUCT/27 refs. RTSTRUCT geometry was tested against synthetic, geometry-faithful MR DICOM references because the public source is NIfTI/MHA only
elitap HNTS-MRG T2 MRI Source cloned; pre-RT and mid-RT checkpoints downloaded PASS: pre-RT produced GTVp 2,982 and GTVn 22,897 voxels (valid 2-ROI RTSTRUCT/33 refs); mid-RT produced GTVn 13,738 voxels (valid 1-ROI RTSTRUCT/27 refs). The repository omitted the checkpoint's required 3d_fullres_oversample_4c plan, recovered exactly from checkpoint metadata. Geometry was tested against synthetic DICOM references because the public fixture is NIfTI/MHA only
STU-Net HNTS-MRG T2 MRI Both Task 1 B and S release checkpoints downloaded and deserialized successfully REPRODUCTION BLOCKED BY RELEASE PACKAGING: repository contains only raw nnU-Net-v1 training checkpoints. It omits task plans/properties, preprocessing configuration, inference wrapper, fold/export metadata, and a compatible trained-model directory; these cannot be recovered from either checkpoint. The generic STU-Net repository supplies architecture code but not the HNTS task preprocessing. Running with invented plans would not validate the released models and was intentionally rejected
lnq-segmenter inguinal-v1 CT Complete, checksum verified; portal-integrated fold 0 PASS: complete 702-slice CT produced 494 candidate voxels (max probability 0.999660). Valid one-ROI RTSTRUCT references 35 source images.
lnq-segmenter abdominopelvic-v1 CT Complete, checksum verified; portal-integrated fold 0 PASS: complete 702-slice CT produced 106 candidate voxels (max probability 0.997274). Valid one-ROI RTSTRUCT references 16 source images.
lnq-segmenter axillary-v1 CT Complete, checksum verified; portal-integrated fold 0 PASS: complete 702-slice CT produced 704 candidate voxels (max probability 0.999568) and a strictly valid one-ROI RTSTRUCT. This demonstrates that the prior empty five-fold result was ensemble/fixture-sensitive rather than a broken checkpoint.
lnq-segmenter mediastinal-v1 CT Complete, checksum verified PASS: five-fold inference on a geometry-preserving thoracic crop found 92 candidate voxels; restored full geometry and produced a valid one-ROI RTSTRUCT with 14 contour-image references
Bouget CT_Lungs / UNet-SW32 / AGUNet / AGUNet-APG contrast CT Source, exact 1.0 GB archive and isolated Python 3.7/TensorFlow 1.14 environment installed CT_Lungs executes; UNet-SW32 and AGUNet each produced valid RTSTRUCTs (65 and 128 referenced images). APG executes with explicit zero priors and produced a valid RTSTRUCT (38 refs), but clinically meaningful priors cannot be generated because that model is unreleased. Upstream automatic CT_Lungs handoff also has a filename bug (pred_ vs labels_)
Raidionics Bouget deployment contrast CT Lungs + lymph-node ONNX weights downloaded and portal-integrated PASS: nonempty mask; valid one-ROI RTSTRUCT references 127 source images from the 702-slice DICOM CT
LNQ2023 probabilistic atlas CT Five-fold archive installed; H100 acceptance passed 2026-10-22 The public archive (3,870,838,958 bytes, SHA-256 ae0735e8444179deb51114dbf3fc469c854758f06849d69747985bbbaa3fecba) is installed at weights/lnq-atlas-20261015/ with five checkpoint_best.pth files, six atlas images, plans and provenance.json. Folds 0–4 passed strict CPU loading with no missing or unexpected keys (fold 0 75ab8a6070b73abc6779cafed9104b9ca00326440a48ef792afd1196322cd127, fold 1 4a7a66e2770c317e2b1f34a149f4d0ea1a3831159d164b0808fa36bd02030c72, fold 2 bfb9c0c3eb48ec899cbeaf765e02bb91db9a7f8bfcbc5827fd78e4b20ff3124e, fold 3 5a02b9f2aec78571e6a348bd518899c60b921e30355d2a6a89d244cc5df3f0c9, fold 4 9d927c4a1be964ca3023804ccecaa70882be7f596e4d9be8ec464898802a2523). run_lnq_atlas.py keeps registration, foreground-probability averaging and the published post-processing, then resamples onto the original CT. It requires nnU-Net 2.2.1 and dynamic-network-architectures 0.2 in runtime-envs/lnq-atlas-ref, and TotalSegmentator 1.5.7 with pinned task-256 weights, checkpoint pickle and plans pickle, and refuses to download them. The task is routed to the H100 through REMOTE_GPU_TASKS. An acceptance marker recorded on the H100 does not enable the task on a P6000 (compute capability 6.1). H100 acceptance passed on 2026-10-22 on a public CT (MSD Task06 lung_053): two production runs and the released predictor gave identical masks (17,234 voxels). No remaining candidate completes with no RTSTRUCT. Code licence CC BY-NC-ND 4.0. Archive weight terms are not stated. Owner accepted these terms on 2026-10-22. Engelson et al., MELBA 2024, sections 4.1–4.3 and Tables 2 and 4, on the 100-patient, 861-node test set: the ablation (threshold 0.2, no minimum-diameter filter) had Dice 0.6033 and found 57% of reference nodes, versus 27% with CT-only training. The challenge submission (threshold 0.3, 5 mm minimum diameter) had Dice 0.5690 and ASSD 6.8976 mm on that same test set. The installed adapter uses the submission settings.
CompAI LNQ Model 7 thoracic CT GitLab source and exact 249 MB LFS checkpoint downloaded; Apache-2.0 PASS; PORTAL ENABLED: release-specific lung normalization and Model 7 checkpoint produced 494 candidate voxels in standalone validation. The exact portal branch was then smoke-tested on the 735-slice CT: TotalSegmentator lung cropping, custom nnU-Net preprocessing, inference, full-geometry restoration and strict RTSTRUCT conversion all passed, yielding one ROI with 76 source-image references. PyTorch 2.6 release-checkpoint loading is handled explicitly and temporary data are cleaned.
DBDMP LNQ CT Source and 298 MB Google Drive checkpoint downloaded PASS; PORTAL ENABLED: deeper inspection found that public crop2lung assigns z/y/x voxel indices directly as an x/y/z physical origin—twice and without spacing/direction. The repaired geometry-preserving implementation on the 735-slice whole-body CT produced 44,550 raw foreground voxels, of which 44,524 met the release intensity gate and 44,271 survived official component refinement. The strictly valid RTSTRUCT has one ROI and 64 referenced CT images. PyTorch 2.6 checkpoint loading was also made explicit for this trusted release. The checkpoint names an unavailable trainer (nnUNetTrainer_LNQ4Tversky3), so the compatible VNetv2 implementation is selected from its documented inference export.
WCODE-PIA LNQ CT Source cloned Weights are distributed through Baidu; download/access remains blocked outside that service
LN-Seg-FM variants H&N CT Source plus 4.5 GB Google Drive weights for original nnU-Net, ResEnc-M, ResEnc-L, and SwinUNETR downloaded PASS: all four fold-0 automatic checkpoints ran on the 516-slice planning H&N CT and produced nonempty visible-node masks: original 2,649, ResEnc-M 5,881, ResEnc-L 2,561, and SwinUNETR 4,033 voxels. Each produced a valid one-ROI RTSTRUCT referencing respectively 51, 61, 60, and 89 source images. All four are portal-integrated; visible job (private job identifier withheld) completed and returned four downloadable RTSTRUCTs. The Swin release required its documented nnU-Net 2.5.1/MonAI 1.3.0 runtime, isolated from the shared environment. These are visible-node candidate contours, not named elective neck levels
PAM CT/MR/PET + user prompt Source plus both public checkpoints downloaded; portal-integrated PASS: interactive DICOM slice viewer supplies a positive point or 2D box. On HNTS-MRG24 T2 MR, PAM propagated 83,487 foreground voxels over 51 slices. A geometry bug in the upstream tutorial was corrected by restoring origin and direction, and the resulting one-ROI RTSTRUCT passed strict validation. A portal job completed and is downloadable.
BiomedParse v2 supported text target on CT/MR/PET Runner and offline contracts implemented Weights available; per-device GPU acceptance and final bundle installation required
SAT3D CT/MR/PET + 3D prompt Source plus model and critic checkpoints downloaded PASS; PORTAL ENABLED: repaired the public release's stale build_sam3D_swin alias, isolated its optional timm/torchvision dependency issue, and added prompt-centred 128³ inference with positive/negative point support. A T2 H&N MRI WebUI job produced 25,017 foreground voxels across 27 slices and a strictly valid one-ROI RTSTRUCT with 27 referenced source images. Requires at least one positive 3D point.
SAM-Med2D CT/MR/PET slices + box/point prompts Source plus 2.56 GB official checkpoint downloaded; portal-integrated PASS, SLICE-WISE: fixed the upstream eager torchvision/NMS dependency and PyTorch 2.6+ checkpoint-loading incompatibility. A box on MR slice 41 produced 1,720 foreground pixels and a valid one-ROI/one-referenced-image RTSTRUCT. A portal job completed. The portal explicitly states that users must prompt every desired slice; no unvalidated 3D propagation is implied.

Local modality fixtures

A private secondary fixture pool was used only after DICOM-level modality, anatomy, geometry, and Frame of Reference checks. The available MR fixture was pelvic T1 and the PET fixture was whole-body PET. They establish MR/PET ingestion and inspection, but are not valid substitutes for head-and-neck T2 MRI or for a model requiring registered planning CT plus PET. Public HNTS-MRG24 patient 101 supplies appropriate pre/mid-treatment T2 MRI in NIfTI/MHA form. Multi-series and multi-modality uploads remain separated into explicit series selections; registration-dependent models require compatible Frame of Reference/geometry or a deliberate registration workflow.

RTSTRUCT acceptance gate

Every automatic output must satisfy all of the following:

  1. prediction shape and affine match the geometry-preserving model input;
  2. every nonzero label has an explicit clinical/research ROI name;
  3. the result is DICOM RT Structure Set Storage;
  4. the output contains at least one nonempty ROI and contour set;
  5. every referenced SOP Instance UID belongs to the selected source DICOM series.

An empty candidate-detection mask is reported as “no candidates detected”; contours are never fabricated merely to create a file.

Both RTSTRUCT converters now position polygons on each referenced DICOM slice using its own position, orientation and row/column spacing, in the inference reader's GDCM order. This corrects rt-utils' unequal-spacing swap and uniform-stack assumption. The tested non-square outputs previously failed the physical geometry gate. A square-pixel gapped fixture passed the old affine-based gate despite a 1.600000 mm source-plane error; the new gate checks the acquired planes directly. The Dice, plane-distance and bounding-box thresholds remain unchanged. TotalSegmentator uses the same converter. Synthetic orientation and overlap tests cover this change; they do not assess model accuracy.

Nonempty labels with no valid polygon are omitted with the ROI name and voxel count in job and download notes. SCT empty masks complete through the standard no-structures path. HD-CTBET's writable directory mirror keeps training logs outside its read-only bundle. SegVol rejects combined point and box prompts through its catalog contract and shared HTTP validation.

Additional catalog integrations (2026-08-10)

  • The nine CADS cads-model-open_v1.0.0 checkpoints (tasks 551–559) are installed under the CC BY-SA 4.0 model license and exposed as separately selectable CT tasks. Together they contain 167 fixed labels spanning organs, vertebrae, vessels, bones and muscles, ribs, RT OARs, brain tissues, head-and-neck OARs and body regions. Tasks 557 and 558 retain their official prerequisite inference/postprocessing (553, and 552+553 respectively). Every checkpoint passed DICOM-to-GPU-inference-to-RTSTRUCT validation on abdominal, thoracic or head CT; outputs contained 6–22 non-empty ROIs and referenced only the selected source series. Archive hashes are recorded in models/manifests/cads-open.json, with exact RTSTRUCT hashes and counts in models/validation/cads_open_*.json.
  • Nine CADS research-weight tasks, cads_research_551_organs through cads_research_559_body_regions, use the same datasets and the same predict_images.py -license research entry point. Weights are CC BY-NC-SA 4.0 (non-commercial, share-alike). Code is Apache-2.0. Owner accepted these terms on 2026-10-22. The CADS README says research-weight performance is “nearly identical to the reference model in the reported evaluation.” cads/LICENSE_MODEL says: “Our comparative evaluation showed that its performance is nearly identical to that of the reference model across the reported test subsets.” The same files say open-weight performance may be degraded for some detailed head-and-neck organ-at-risk structures. Research and open dataset.json labels match. The research catalogue keeps the dataset.json spacing (right psoas major, white matter, and the other spaced names in tasks 556 and 557). Research and open checkpoint_final.pth files differ in bytes and in tensor hashes. plans.json is not byte-identical. Task 557 3d_fullres patch size differs (research [160, 192, 192], open [128, 192, 192]). Tasks 557 and 558 still run their published dependency tasks under the research licence. The portal refuses any other licence and refuses to download a missing checkpoint. H100 acceptance passed for all nine tasks on 2026-10-22 on public fixtures. Weights: weights/cads/research/Dataset55x_*.
  • Both official CLIP-Driven UniversalModel checkpoints are installed and selectable: U-Net and Swin UNETR. The U-Net produced all 32 released organ, tumor and cyst ROIs; Swin UNETR produced 28 non-empty ROIs on the same abdominal CT. Both referenced all 167 source images and passed strict RTSTRUCT validation. The release's MONAI 0.9 inversion restores arrays to the original voxel grid but leaves a stale 1.5-mm RAS affine; the runner replaces that header with the original model-input affine after direct-order spatial verification against CADS (liver Dice 0.9765). Exact provenance and evidence are in models/manifests/universal-model.json and models/validation/universal_model_*.json.
  • All four official MRAnnotator fold-0 checkpoints are installed and enabled: abdomen (8 labels), shoulder/knee (10), pelvis/prostate (4), and spine (22). Each completed DICOM-to-GPU-inference-to-RTSTRUCT validation after the release-required LAS reorientation and restoration to source geometry. The pelvic test used an official Medical Segmentation Decathlon T2 prostate case; the shoulder/knee acceptance exercised the shoulder portion of that combined checkpoint. Provenance and exact output hashes are recorded in models/manifests/mrannotator.json and models/validation/mrannotator_*.json.
  • All four PYCAD Model Zoo releases are installed and enabled: vertebrae CT, skull CT, cardiac MRI and knee-joint MRI. Each official fold-0 checkpoint produced a nonempty, strictly valid RTSTRUCT. The cardiac release documents its classes only as h1/h2/h3, so the portal preserves those names and warns that the anatomical ontology is unavailable. The knee test proves execution and geometry on an out-of-domain MRI fixture, not knee accuracy. Archive and output hashes are in models/manifests/pycad-model-zoo.json and models/validation/pycad_*.json.
  • The official standalone SynthSeg v1 checkpoint is installed and enabled as synthseg_v1_brain. It produced all 31 released non-background brain labels on the TemplateFlow MNI152 T1 template and a strictly valid RTSTRUCT with 80 referenced images. SynthSeg emits a canonical 1-mm segmentation, so the runner restores it to the selected DICOM geometry with nearest-neighbour resampling before contour conversion. Its exact TensorFlow 2.2/CUDA 10.1 environment is incompatible with the H100's CUDA 12 stack and is therefore explicitly CPU-isolated with eight threads. Provenance is in models/manifests/synthseg.json.

  • The MARS brainstem five-fold nnU-Net release is installed and enabled for whole-brain T1 MRI. Its three label identities—mesencephalon, pons and medulla oblongata—were verified from the upstream volume-extraction script, and all three were nonempty in a strictly valid RTSTRUCT. The companion MARS-WMH five-fold archive and source are also installed, but remain disabled because the official pipeline requires separate FLAIR and T1 series plus registration of T1 into FLAIR space; duplicating one selected series would not satisfy that contract. Checksums and provenance are recorded in models/manifests/mars-brain-models.json.

  • lungmask_r231, lungmask_ltrc_lobes, and lungmask_ltrc_lobes_r231 are installed from the official v0.0 release, checksum-verified, and enabled. All three completed DICOM-to-RTSTRUCT acceptance tests on thoracic CT. The strict validator reported respectively 2, 5, and 5 non-empty ROIs, with all contour image references belonging to the source DICOM series. Exact hashes and result sizes are recorded in models/manifests/lungmask.json.

  • mrsegmentator and mrsegmentator_body_comp are installed with the upstream checksum-aware downloader and enabled. The base model produced 21 non-empty ROIs on thoracic CT. The MRI-only body-composition release was tested on an in-domain upper-abdominal/thoracic MRI fixture from the official TotalSegmentator MRI dataset and produced five non-empty tissue ROIs. Both outputs passed strict RTSTRUCT geometry/reference validation; provenance and result sizes are in models/manifests/mrsegmentator.json.
  • totalspineseg is installed from both official r20260730 model archives and enabled. The two-stage pipeline produced 16 non-empty, anatomically labelled spine/cord/canal ROIs on cervical T2 MRI and a strictly valid RTSTRUCT referencing 69 source images. The shared runtime pins Kornia 0.7.4 because upstream auglab 20260109 is incompatible with Kornia 0.8.3. Archive hashes and the acceptance result are recorded in models/manifests/totalspineseg.json.
  • All ten exposed MOOSE clinical-CT groups are installed from the official release registry and enabled: body, cardiac, digestive, lungs, muscles, organs, peripheral bones, ribs, vertebrae, and body composition. Every group produced a non-empty, strictly valid RTSTRUCT; together the tests exercised 117 output ROIs. The body-composition pipeline also uses the upstream fast vertebra model as its localization dependency. Per-task results and checkpoint-tree hashes are recorded in models/manifests/moose.json.
  • The BAMF LiTS and Cervix-Cancer-CBCT-Segmentator nnU-Net v2 archives were downloaded from Zenodo, verified against both published MD5 values and local SHA-256 hashes, and run as full five-fold ensembles. BAMF produced seven non-empty RTSTRUCT ROIs on abdominal CT. The cervix release produced all four checkpoint labels and a valid RTSTRUCT on pelvic CT; this proves execution and geometry only, because an appropriate public CBCT acceptance fixture was not supplied by the release. Exact upstream label strings are retained (the ambiguous CP and tumsomething names are not silently reinterpreted).
  • mrisegmentator 0.4.2 and its official five-fold Box checkpoint are installed and portal-integrated. On the 208-slice MRI fixture it produced 16 non-empty structures from its 62-class label map; the RTSTRUCT references all 208 source images and passed strict validation. The upstream CLI help currently says 68 structures, but its distributed dataset.json contains 62 foreground labels, so the portal reports the checkpoint-owned value.
  • vibesegmentator dataset 100 is installed from all three official release archives/folds and runs through an explicit offline model root. Its 72-class output produced 33 non-empty ROIs on the 208-slice MRI fixture. The resulting 37,750,432-byte RTSTRUCT references all source images and passed strict validation. Source commit, checkpoint-tree hash, and output hash are recorded in models/manifests/vibesegmentator.json.
  • dentalsegmentator is installed through the MOOSE-compatible distribution of the public DentalSegmentator checkpoint. All five checkpoint-owned labels (upper skull, mandible, upper/lower teeth and mandibular canal) were nonempty on the head CT fixture. Its 2,930,026-byte RTSTRUCT references 126 source images and passed strict validation.
  • The legacy nnU-Net v1 Task004_Hippocampus five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. It was tested on an official Medical Segmentation Decathlon T1 MRI converted to geometry-faithful research DICOM and produced both anterior and posterior hippocampal ROIs in a valid RTSTRUCT.
  • The legacy nnU-Net v1 Task006_Lung five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. The full portal branch was tested on official MSD training CT lung_053, converted to a 252-slice CT DICOM series with signed HU storage. It produced one non-empty lung-tumor ROI, referenced 77 source images, and passed strict RTSTRUCT validation. The temporary 5.0 GB model and 9.2 GB fixture archives were removed after their checksums and the retained artifacts were recorded.
  • The legacy nnU-Net v1 Task002_Heart five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. On official MSD cardiac MRI la_007, converted to a 130-slice research DICOM series, it produced one left-atrium ROI and a strictly valid 159,588-byte RTSTRUCT that references 76 source images.
  • The legacy nnU-Net v1 Task009_Spleen five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. Official MSD CT spleen_19 produced one non-empty spleen ROI and a strictly valid 71,702-byte RTSTRUCT referencing 18 source images.
  • The legacy nnU-Net v1 Task003_Liver five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. On a 167-slice abdominal CT fixture it produced a non-empty liver ROI and a strictly valid 872,904-byte RTSTRUCT referencing 110 source images. The checkpoint's tumor label was empty on this non-tumor-specific fixture and was not fabricated.
  • The legacy nnU-Net v1 Task007_Pancreas five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. On the abdominal CT fixture it produced a non-empty pancreas ROI and a strictly valid 217,706-byte RTSTRUCT referencing 57 source images. The tumor label was empty on this fixture and was not fabricated.
  • The legacy nnU-Net v1 Task008_HepaticVessel five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. It produced a non-empty hepatic-vessel ROI and a strictly valid RTSTRUCT on the abdominal CT fixture. The tumor label was empty and was not fabricated.
  • The legacy nnU-Net v1 Task027_ACDC five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. A cardiac MRI technical fixture produced all three checkpoint labels and a strictly valid 1,173,466-byte RTSTRUCT. The fixture is not an ACDC short-axis cine case, so the portal retains the model's sequence-domain warning.
  • The legacy nnU-Net v1 Task024_Promise five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. It completed the full portal path and produced a structurally valid prostate RTSTRUCT on a whole-torso MRI fixture. This is an execution/geometry test only: the fixture is not PROMISE12 prostate T2 MRI, and the portal explicitly preserves that sequence requirement.
  • The legacy nnU-Net v1 Task055_SegTHOR five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. On a 179-slice thoracic CT it produced all four checkpoint labels (esophagus, heart, trachea and aorta) and a strictly valid 1,037,898-byte RTSTRUCT that references all source images.
  • The legacy nnU-Net v1 Task017_AbdominalOrganSegmentation BTCV five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. Its exact 13-label ordering comes from the official nnU-Net dataset-conversion source. It produced 12 non-empty ROIs and a strictly valid 2,726,606-byte RTSTRUCT referencing all 167 source images; only the gallbladder label was empty on this fixture.
  • The distinct legacy nnU-Net v1 Task029_LITS five-fold 3D-full-resolution ensemble is installed from its own checksum-verified public archive. It produced a non-empty liver ROI and a strictly valid RTSTRUCT referencing 110 source images. Its tumor label was empty on this fixture and was not fabricated; it remains a separate task from Task003_Liver.
  • The legacy nnU-Net v1 KiTS archive is installed under its actual upstream identifier, Task048_KiTS_clean (older catalogs and the conversion script call the dataset Task040). On public KiTS19 case_00000, both the kidney and kidney-tumor labels were non-empty. The resulting 1,421,284-byte RTSTRUCT references the complete 611-image source series, contains contours on 264 images, and passed strict validation. This test also exposed and fixed the research NIfTI-to-DICOM helper's handling of files whose stored slice axis is not axis 2; it now canonicalizes from the NIfTI affine before export.
  • The legacy nnU-Net v1 Task010_Colon five-fold 3D-full-resolution ensemble is installed from the checksum-verified public model-zoo archive. On public Medical Segmentation Decathlon CT colon_053, it produced a non-empty colon tumor ROI and a strictly valid 45,168-byte RTSTRUCT referencing the complete 137-image source series, with contours on eight images.
  • The MONAI Model Zoo spleen_ct_segmentation bundle version 0.6.1 is installed with its weights, TorchScript export, training/evaluation configs, metadata and license. The portal executes the bundle-owned preprocessing, sliding-window inference and inverse transforms through MONAI 1.6.0. On MSD spleen_19, it produced a non-empty spleen ROI and a strictly valid 69,122-byte RTSTRUCT referencing the complete 51-image source series, with contours on 18 images.
  • The MONAI Model Zoo multi_organ_segmentation DiNTS bundle version 0.0.6 is installed with its model, TorchScript export, architecture-search code, configs and metadata. A narrowly scoped compatibility override permits the checksum-pinned official NumPy architecture pickle under PyTorch 2.6 while retaining safe checkpoint loading elsewhere. It produced six non-empty ROIs (artery, portal vein, liver, spleen, stomach and pancreas) and a strictly valid 1,240,254-byte RTSTRUCT on the abdominal CT fixture. Gallbladder was empty on this fixture and was not fabricated.
  • The MONAI Model Zoo pancreas_ct_dints_segmentation bundle version 0.5.2 is installed and uses the same checksum-scoped DiNTS compatibility path. On the abdominal CT fixture it produced non-empty pancreas and pancreatic-tumor ROIs and a strictly valid 244,048-byte RTSTRUCT with contours on 59 of the 167 source images. This proves execution and geometry; target accuracy still requires pancreas-protocol local validation.
  • The MONAI Model Zoo pediatric_abdominal_ct_segmentation DynUNet bundle version 0.4.6 is installed with its training/inference configuration and A100 TensorRT export. It produced all three labels (liver, spleen and pancreas) and a strictly valid 1,248,946-byte RTSTRUCT on the abdominal CT fixture. This is an execution/geometry test on an adult scan, not pediatric performance validation; the portal preserves the pediatric domain warning.
  • The MONAI Model Zoo renalStructures_UNEST_segmentation bundle version 0.2.7 is installed. On public KiTS19 case_00000, all three renal substructures—cortex, medulla and pelvicalyceal system—were non-empty. Its strictly valid 4,768,052-byte RTSTRUCT references the complete 611-image source series and contains contours on 352 images.
  • The complete MONAI renalStructures_CECT_segmentation bundle version 0.2.3 is also retained and checksum-recorded, but remains disabled. Inspection of its actual inference config proves that it requires co-registered arterial, venous and excretory CT volumes as three channels; the current one-series portal contract cannot satisfy that requirement without fabricating phases.
  • The MONAI Model Zoo swin_unetr_btcv_segmentation bundle version 0.5.8 is installed. Its MONAI 1.4-era config is adapted per job by removing the now-obsolete SwinUNETR.img_size constructor argument; the checksum-recorded upstream bundle remains unchanged. The CPU-postprocessing flag also moves the multi-channel inversion to CPU (see the decision table below). Before this CPU change, all 13 checkpoint labels were non-empty on the abdominal CT fixture, producing a strictly valid 3,272,662-byte RTSTRUCT with the complete 167-image source series referenced.
  • The MONAI Model Zoo wholeBody_ct_segmentation bundle version 0.2.7 is installed with both official 1.5 mm and 3 mm checkpoints. Full-resolution inference completed but its 105-channel softmax exceeded the 48 GB GPU; the bundle-provided low-resolution checkpoint completed the uncropped 735-slice whole-body CT. It produced 102 non-empty ROIs and a strictly valid 18,842,824-byte RTSTRUCT. Only face and gallbladder were empty and were not fabricated. Label names are loaded directly from bundle metadata.
  • The MONAI Model Zoo ventricular_short_axis_3label bundle version 0.3.5 is installed with both its state-dict and TorchScript checkpoints. On the 130-slice MSD cardiac MRI fixture it produced all three checkpoint labels (LV pool, myocardium and RV pool). The 1,974,990-byte RTSTRUCT references all source images and passed strict validation.
  • The MONAI Model Zoo wholeBrainSeg_Large_UNEST_segmentation bundle version 0.2.7 is installed. Its metadata defines 132 foreground labels (not 133 foreground structures). A cropped MSD hippocampus T1 technical fixture produced 38 non-empty ROIs and a strictly valid 121,882-byte RTSTRUCT. This proves bundle execution and DICOM geometry only; full-brain anatomical performance remains subject to an in-domain, full-field-of-view validation.
  • The MONAI Model Zoo prostate_mri_anatomy bundle version 0.3.6 is installed and enabled. Its actual metadata confirms a single-channel MRI contract and two foreground zones. Both central gland and peripheral zone were non-empty on a 208-slice whole-torso MRI technical fixture, producing a strictly valid 2,000,172-byte RTSTRUCT. This is an execution/geometry test, not validation against the model's Prostate158 3T T2 acquisition domain.
  • The complete MONAI brats_mri_segmentation bundle version 0.5.4 is retained and checksum-recorded but disabled. Its config and metadata require four co-registered inputs in exact T1c, T1, T2 and FLAIR order. The current portal selects one DICOM series, so enabling it would require fabricating or misidentifying channels.
  • HD-BET 2.0.1 and its automatically distributed release-2.0.0 checkpoint are installed with the runtime home and weights redirected to the NFS model bundle. On an 82-slice head-and-neck T2 MRI fixture it produced a non-empty brain mask and a strictly valid 370,834-byte RTSTRUCT. The contour references 28 source images; no geometry or image outside the selected series is used.
  • HD-CTBET 1.1 is installed from upstream commit 78da973b with its official fold-0 CT checkpoint. Upstream's exporter compared a 3-D array directly to a three-element target shape and failed with a broadcasting error; the retained one-line patch compares seg_old_size.shape, with the exact patch and hashes recorded in models/manifests/hd-ctbet.json. On a 142-slice head CT it produced a non-empty brain ROI and a strictly valid 598,312-byte RTSTRUCT referencing 92 source images.
  • LungTumorMask v1.3.1 is installed from upstream commit d1bca4b8 with the released dc_student.pth checkpoint stored on NFS. Its 2021 MONAI imports are no longer available, so the compatibility runner preserves the released 49,117,073-parameter network (all 52 checkpoint tensors match) and delegates lung localization to the installed lungmask R231 model. On the 179-slice thoracic CT fixture it produced a non-empty lung-tumor candidate ROI and a strictly valid 84,664-byte RTSTRUCT with contours referencing 45 source images. This proves execution and geometry, not tumor accuracy on the technical fixture.
  • TrackRAD Track'n'Treat source and its checksum-verified 468 MB averaged MedSAM2 checkpoint are installed on NFS. It remains deliberately disabled: upstream inference requires sagittal 2D cine-MRI plus an explicit first-frame target mask and returns a temporal MHA mask sequence, whereas the portal's current contract is a static DICOM series producing RTSTRUCT.
  • MedSAM2 source and all six public WangLab checkpoints are stored on NFS. The FLARE25 RECIST CT-lesion checkpoint loads strictly and is portal-integrated for a positive/negative point set or 2-D box on one axial seed slice. On the released FLARE23 public fixture, box propagation scored Dice 0.9098 and single-point propagation scored Dice 0.9073. The exact DICOM branch on MSD lung case 53 produced a one-ROI RTSTRUCT referencing 92 source images with zero physical bounding-box delta. The model is restricted to research and education use, and every propagated lesion contour requires review.
  • The separate ImprintLab Medical-SAM2 GPL release and its official 38.9M- parameter tiny checkpoint are also installed. All 468 checkpoint tensors load strictly. Box propagation ran on the public FLARE lesion technical fixture (Dice 0.4613, explicitly out of its BTCV domain), and the complete DICOM branch produced a one-ROI RTSTRUCT referencing 73 source images with zero physical bounding-box delta. It remains clearly distinguished from the higher-performing WangLab CT-lesion checkpoint in the model picker.
  • nnInteractive 2.5.1 source commit bbe12fdccc876cb2d4e0a47133811e362608e000 and the official v1.0 fold-0 checkpoint are installed on NFS. The checkpoint SHA-256 is b3ac4421f85457bbd1aa0d87f5e67bcb7bc8e2ce6b824b6ac45077cc5d630ea9 and its model license is CC BY-NC-SA 4.0 (distinct from the Apache-2.0 source license). A positive point on the public MNI T1 fixture produced a non-empty source-grid mask and a strictly valid, orientation-audited RTSTRUCT with model-specific DICOM identity. The portal supports positive and negative points and the checkpoint's required 2-D box interaction.
  • VISTA3D/NV-Segment-CT source commit f9f5f51b589e5dc9c23c453cf5138398e4084056 and Hugging Face revision afb51518689f71e6abb367ee6301b2cd0225c66a are installed on NFS. The 871,970,895-byte checkpoint has SHA-256 c92bab26d00b4a5d89fa8a383900cdeb88302fd318e5e816df0bbec7106d9a1b. A positive-point run on the public MSD spleen CT fixture produced a non-empty 3-D mask and a one-ROI RTSTRUCT referencing 18 source images; strict orientation validation passed. The runner records and applies a one-line MONAI 1.6 tensor compatibility repair without modifying upstream source.
  • VoxTell source commit 8ef0332aec5d5dff925d6f6370386c00334b57ab and Hugging Face revision 0809ac94c5ed594198bf4e85d63245cf222464d7 are installed on NFS with the official v1.1 checkpoint, its 14,194-label embedding bank, and the Qwen3-Embedding-4B fallback for unseen free text. The prompt spleen produced 101,225 foreground voxels on the public MSD fixture and a one-ROI RTSTRUCT referencing 19 source images. Direct source-grid conversion matched the ROI physical bounding box exactly.
  • SegVol source commit 5768c9795794a327a2629ce7f137ab03d5ea8378 and Hugging Face revision 516701d5bef430604a13f5db7ce896149d075663 are installed on NFS. Its 723,726,667-byte checkpoint has SHA-256 500f2758a8f989339b2b2baf09a819169bc87549795193d3cfe505726ac0b399. The exact released MONAI 0.9 architecture loads all 180,890,064 parameters; a documented one-line seed-range repair makes MONAI 0.9 compatible with NumPy 2 without changing model computation. Text-prompted spleen inference achieved Dice 0.9801 on the publisher's fixture. The same prompt on the MSD DICOM fixture produced a one-ROI RTSTRUCT referencing 18 source images with zero per-ROI physical bounding-box delta. Text, point and 3-D box prompts are exposed by the portable runner.
  • UniverSeg source commit 833a0c34c65e38d675e21bd48ddec6797cc03259 and its sole official release checkpoint are installed on NFS. The 4,748,449-byte OpenRAIL++-M research checkpoint has SHA-256 62c98c48c08bc05f84ed19187844d5eb7e6b7a48d2f690f2033fd8de04543884; all 1,182,785 parameters load strictly. With 64 support image-mask pairs it achieved Dice 0.9101 for label 1 on the publisher-linked OASIS fixture. The model is visible but disabled because its class is defined by uploaded 2-D support images and masks, an input type the current portal does not accept. Reusing a fixed unrelated support set would not implement few-shot contouring.
  • HippUnfold source commit 544fdb6e37eb42567b1ad3c7f4901a180bdb19b8 and both catalog checkpoint releases are staged on NFS. The ADNI T1 archive is 2,452,807,680 bytes (SHA-256 55be492ccc65b4fd60bdd18dde2e63c0c5676060cb950b2bbbd2fd12ea57ebf8) and contains all five nnU-Net-v1 folds. The Maguire T2 archive is 1,138,683,578 bytes (SHA-256 ee4e28d246cb1acb86df6b99be50c66b9d762356385643092877a9723a552de7) and contains all five nnU-Net-v2 folds. Both are visible as acceptance-pending until the matching development runtime produces native-space subfields and a strictly validated, orientation-audited RTSTRUCT. Limited-field T2 input may require a separately registered T1 series and must not be treated as a valid single-series case.
  • HD-GLIO-AUTO source and checksum-verified v2 checkpoint are installed, and its published inference code confirms a mandatory native T1, post-contrast T1, T2 and FLAIR input set. The pipeline then performs HD-BET brain extraction and rigid registration before tumor inference. It remains visible but disabled until multi-series selection is implemented; supplying one series four times would be invalid.
  • All twelve MATTO-GBM TRANSCAN archives match their published Zenodo MD5 checksums and are extracted on NFS. The single-channel enhancing-tumor model is enabled with the publisher's HD-BET and 0.1/99.9-percentile preprocessing. Its five-fold best-checkpoint ensemble produced a non-empty enhancing-tumor ROI and a strictly valid 271,356-byte RTSTRUCT on the axial T1c fixture, referencing 44 source images. The other eleven models remain visible as an installed family while their FLAIR, SUV PET, derived-mask, atlas-orientation, or eyes-masking contracts are implemented and tested.
  • MultiTalent V1 source commit 0bc7b2c32e31cb00008354f2420b65b85ef09578 and the complete official Zenodo Task100 archive are installed on NFS. The 4,634,439,232-byte archive matches publisher MD5 a86617d4c71ef7b97d9eaf931df390cd and contains standard U-Net and residual-encoder checkpoints for all five folds plus the all-data checkpoints. The release implements 47 independent, overlapping dataset-specific heads, not the newer repository branch's 117-class claim. The standard and residual all-data checkpoints have SHA-256 03851e77590fcd89b96320b666068971a484593ee948573feae6f88dc3c43fd0 and 506b91701e0957d10b67c0375275053a1f4364d3febe81289f34b5a1237bb775. Both loaded and ran on the H100. Their overlap-preserving RTSTRUCTs contained 47 and 44 non-empty ROIs respectively, referenced all 51 source CT images, had zero per-ROI physical bounding-box delta, and carry distinct model names in StructureSetLabel, StructureSetName and SeriesDescription. The Zenodo record is non-commercial and the constituent public datasets may impose additional terms.
  • HECKTOR 2020 winning-solution source commit 5ebc6774d178139b52abebbe926b4cafabe3e29d and all eight official PET/CT ensemble checkpoints are installed on NFS. The 696,677,684-byte Dropbox archive has SHA-256 1666bad2c4355bb059a1339e6fc605f65838da5bd931139030c24ccab52d02d7. After removing the publisher's DataParallel module. prefix where present, every 310-tensor state dictionary loads strictly into the released FastSmoothSENormDeepUNet_supervision_skip_no_drop architecture. The family remains disabled because published inference requires registered PET and CT plus a patient-specific 144 mm physical head-and-neck bounding box. That bounding-box CSV is part of the HECKTOR challenge data, whose official page currently says access cannot be granted while hospital agreements are being extended. Guessing a crop or running a pelvic PET fixture would not prove valid inference.

Source synchronization and interpretation of evidence

The source now includes cached catalogs, resumable uploads, per-task process-tree termination, read-only bundle mirrors, TotalSegmentator auxiliary-label filtering and the CPU prostate SBRT crop tool. Direct source-grid RTSTRUCT conversion and independent physical round-trip checks replace the earlier orientation workaround. See operations for limits and configuration.

The captured deployment returned 253 tasks, 182 enabled. These counts describe that installation; fresh installations must pass their own preflight and orientation gates. All model-specific warnings and numerical results above are retained. Private fixture/job identifiers and private artifact hashes are withheld without relabeling the associated measurements as public-fixture validation. Public fixture maps and external private override configuration are documented in operations.

The dated SCT statements above describe successive historical states. Current task definitions are in catalog/sct.json, with additional recorded validation in models/validation/. Neither a copied marker nor this source synchronization establishes clinical or manuscript readiness.

LNCTVSeg NPC grouped nodal CTVs (2026-10-15)

lnctvseg_nnunet_v1 exports IDs 1–6 unchanged with the published names below. The checkpoint has eight output channels (background plus IDs 1–7); class 7 is undocumented by the authors and is not exported. It is never relabelled or merged into a named group. Raw predictions retain every class for audit. Every job records the omitted class 7 voxel count and volume in conversion notes, job details and the result ZIP. Execution remains gated on GPU acceptance with the current manifest digest.

The pinned author repository is GPL-3.0. The author-distributed weights have no separate licence in the archive or pinned repository. The dataset, version 3, is CC BY 4.0. These are separate from the article licence. Citation: Luo et al., Scientific Data 11, 1085 (2024).

The archive is 861,358,205 bytes, SHA-256 7612f33556d163e26226bed7064c73212323d488d5c6010399307f778a40b246. The manifest pins its inspected files. The nnU-Net metadata specifies nnUNetTrainerV2, 3d_fullres, stage 1, one CT channel, a 48 × 192 × 192 patch and 3 × 1.1005859375 × 1.1005859375 mm spacing in array z/y/x order. Despite the CT input, the saved normalization is nonCT: per-case z-score over the cropped, resampled volume, without a nonzero mask or percentile clipping. Intensity properties are absent. The flat checkpoint was trained with fold all; inference stages a writable job-local model/all layout, links the immutable checkpoint, and copies metadata with local paths. No postprocessing file is supplied. Strict CPU loading passed in nnU-Net 1.7.1 / PyTorch 2.5.1 with no missing or unexpected keys.

Published ID RTSTRUCT ROI (GPU acceptance required)
1 LN CTV R II-III-Va
2 LN CTV L II-III-Va
3 LN CTV R IV-Vb
4 LN CTV L IV-Vb
5 LN CTV R Ib
6 LN CTV L Ib

These are NPC-specific grouped elective nodal CTVs, not individual levels or visible-node masks. The numeric map comes from the pinned README and agrees with patient-space laterality of its example reference masks. The paper calls the lower-neck group IV+Vb+Vc in parts of the text; export names retain the repository's IV+Vb labels. Contouring follows the participating institutions' NPC conventions, including modified boundaries. Training data include contrast-enhanced and non-contrast CT. Grouped masks cannot recover their constituent levels.

SwinUNETR loads strictly in the existing MONAI 1.6.0 runtime: 159 tensors, one input channel, eight output channels and feature size 48. This does not identify the LNCTVSeg training image size, MONAI version, spacing, orientation or intensity transforms. The linked 3D UX-Net code uses 96³ windows and recommends MONAI 0.9.0, but supplies preprocessing only for FeTA, FLARE and AMOS. Those settings cannot be assigned to this checkpoint. SwinUNETR therefore remains informational. 3D UX-Net and RepUX-Net also remain informational: no release-matched implementation and LNCTVSeg preprocessing configuration are installed, and strict loading was not run. All three checkpoints have eight output channels; no heavy environment was added.

The public lnctvseg-example-hn-ct/dicom fixture contains 200 derived DICOM slices from the author-provided contrast-enhanced example, with full neck coverage. NIfTI→DICOM→NIfTI preserved every HU value and the affine exactly. The six reference components passed the body-midline laterality check; their volumes were 92.9898, 96.0721, 49.4617, 56.9312, 8.3018 and 7.3030 mL in label order. The case belongs to the development cohort. Its source hashes and licence are recorded in models/manifests/lnctvseg-fixture.json and the installed fixture's SOURCE.md.

A CPU smoke completed in 321.50 s on a 287 × 203 × 85 voxel crop containing all six reference groups. The crop was selected using the reference bounding box plus margins; no input downsampling was used. Native release resampling and normalization were applied, with mirror augmentation disabled. The native folder predictor could not create multiprocessing semaphores on the worker, so the same v1 preprocessing, prediction and export functions ran serially. Same-ID Dice for labels 1–6 was 0.872640, 0.869314, 0.874993, 0.899687, 0.886700 and 0.899765 (mean 0.883850). The prediction also contained 3,563 voxels / 19.1328 mL of undocumented class 7. These cropped development-cohort overlaps do not resolve that class or establish full-volume performance.

The operator GPU audit runs every integrated LNCTVSeg task and HNLNL. It records same-ID Dice against the pinned reference, including the separate volume of omitted ID 7, from raw predictions. It retains predictions, checks the six-label RTSTRUCT round trip and patient-space laterality, and records elapsed time and sampled process VRAM. No GPU run or acceptance marker was produced during this implementation. Local loading and fixture results do not establish clinical or manuscript readiness.

Additional nodal weight checks (2026-10-15)

  • Wen et al., Medical Physics 2024, doi:10.1002/mp.17330: the indexed article describes CMU-net for six pelvic elective regions (160 cases; 120 training, 40 testing). No trained checkpoint was identified in the DOI/title/author/model-name searches across GitHub, Zenodo and Hugging Face. The publisher page remained behind browser security verification, so the full text and supplements were not verified. The related open-access Wen pelvic-CTV paper is a different article and does not establish availability of this model.
  • Buhl et al., Radiotherapy and Oncology 2024, doi:10.1016/j.radonc.2024.110567: accessible full text describes four nnU-Net v1 models, by laterality and internal-mammary extension to intercostal space 3 or 4. It links one 3.03 MB DOCX supplement, whose retrieval returned HTTP 403. No trained checkpoint was identified in the paper or author/DBCG searches across GitHub, Zenodo and Hugging Face. A future release would need the four checkpoint ensembles, exact class maps, preprocessing, postprocessing and a model licence.

These are bounded availability checks, not proof that no release exists elsewhere.

MONAI postprocessing device decisions

Inspection of the catalog-selected inference configs on 2026-10-15 gives the following decisions. Channel counts include background. postprocess_on_cpu: true inserts EnsureTyped(keys="pred", device="cpu") immediately before inversion and sets Invertd.device to CPU in a job-local compatibility config. Softmax remains before inversion; argmax remains after it. Network/inferer devices and bundle files remain unchanged. YAML loading uses PyYAML in the worker environment.

Catalog task Channels Postprocessing order relevant to inversion CPU flag
monai_spleen_ct 2 Softmax → inversion → argmax On
monai_multi_organ_ct 8 Softmax → inversion → argmax On
monai_pancreas_dints_ct 3 Softmax → inversion → argmax On
monai_pediatric_abdominal_ct 4 Softmax → inversion → argmax On
monai_renal_unest_ct 4 Softmax → inversion → argmax On
monai_swin_unetr_btcv_ct 14 Softmax → inversion → argmax On
monai_whole_body_ct 105 Softmax → argmax → nearest-neighbour inversion of one label channel Off
monai_ventricular_short_axis_mr 4 Softmax → Invertd → argmax; preprocessing has no spatial transform to invert Off
monai_whole_brain_unest_mr 133 Softmax → Invertd → argmax; preprocessing has no spatial transform to invert Off
monai_prostate_anatomy_mr 3 Argmax → connected-component filtering → inversion of one label channel Off

The six enabled pipelines invert interpolated multi-channel predictions on the original grid. Whole-body CT and prostate anatomy MRI already reduce them to one label channel before inversion. Cardiac MRI and whole-brain MRI perform loading, channel setup and intensity/type conversion without spatial resampling or orientation changes, so their Invertd calls do not perform spatial inversion. The change does not reduce inference or softmax memory. CPU inversion uses host RAM and may change numerical results or wall time; GPU off/on agreement remains to be measured with the comparison script and operations procedure. Existing model-domain and acceptance restrictions still apply.

The container also installs Ubuntu Jammy libxrender1, which supplies the libXrender.so.1 required by OpenCV in the CERR prostate runtime. This library addition requires an image rebuild.

BiomedParse v2 (biomedparse_prompted) uses the pinned upstream 3D inference method at e02096c03af0d79c6994ffc2d60a49eeb0361e1f. A single CT, MR or PET series and one documented target produce a prompt-named ROI and .dcm.biomedparse.json with slice existence scores, decisions and voxel count. No detected object is a valid empty result. Supported targets are recorded in catalog/biomedparse-prompts.json, derived from the v2 README and verbatim 3D example prompts; the v1 target_dist.json is not used. Unsupported targets and multiple [SEP] prompts are rejected.

Code: Apache-2.0. Weights: CC BY-NC-SA 4.0, for research and development only, not clinical decisions. Owner accepted these terms on 2026-10-22. This task selects one target; it does not define an elective nodal atlas. Published v1 and BoltzFormer benchmark results are distinguished from the released v2 volumetric checkpoint in the catalogue card. See the upstream v2 README, Nature Methods and BoltzFormer. GPU memory use and model agreement remain unmeasured locally.