Skip to content

LNQ2023 atlas adapter

run_lnq_atlas.py runs source revision e48a74f3223e5d9ef06f79fb08282450c29f7a2d. The five-fold archive is installed at weights/lnq-atlas-20261015/. Its provenance.json records archive SHA-256 ae0735e8444179deb51114dbf3fc469c854758f06849d69747985bbbaa3fecba and a SHA-256 for every file returned by required_files(). Folds 0–4 passed strict CPU loading with no missing or unexpected keys. 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). The task stays off unless that marker and the orientation marker pass.

The adapter requires:

  • All five checkpoint_best.pth files, plans.json, dataset.json, and the six atlas images under weights/lnq-atlas-20261015/weights.
  • nnU-Net 2.2.1 and dynamic-network-architectures 0.2, in runtime-envs/lnq-atlas-ref. The adapter calls nnUNetTrainer.build_network_architecture(plans_manager, dataset_json, configuration_manager, num_input_channels, enable_deep_supervision=False). Output heads come from the label manager. determine_num_input_channels must return 3. The predictor is constructed with perform_everything_on_gpu=True.
  • A separate TotalSegmentator 1.5.7 runtime. Set RTSEG_LNQ_TOTAL_PYTHON, or pass --total-python. Registration uses the v1 organ IDs. A v2 mask is not a substitute.
  • Task 256 fast weights at weights/totalsegmentator-v1.5.7 (override RTSEG_LNQ_TOTAL_WEIGHTS or pass --total-weights). The launcher refuses to start if TOTALSEG_WEIGHTS_PATH is unset, the pinned trainer directory nnUNetTrainerV2_ep8000_nomirror__nnUNetPlansv2.1 is missing or ambiguous, or model_final_checkpoint.model, model_final_checkpoint.model.pkl, or plans.pkl is missing or does not match its pin. Inference does not download weights.

run_total_v1.py isolates Path.home() in the child process because TotalSegmentator 1.5.7 ignores TOTALSEG_HOME_DIR. It turns usage telemetry off in that private configuration. Large nnunet_tmp_ files stay in the supplied project directory. Multiprocessing sockets use the short TMPDIR.

The adapter keeps the published steps: CT, atlas distance map and probability map; registration and lung crop; five-fold average of restored foreground softmax probabilities; atlas threshold 0.15–0.30 from get_probability_threshold_probmap with t=0.3; lung convex hull; removal of components whose equivalent ellipsoid diameter is below 5 mm on any axis. Input is reoriented to DICOM LPS. Oblique CTs are rejected. The last step resamples the mask onto the original CT with nearest-neighbour interpolation. Upstream process.py returns the mask on the registration grid and does not do that resample. The GPU harness applies the same resample before Dice.

The output is one ROI, Mediastinal_Lymph_Node_Candidate: a visible mediastinal lymph-node candidate, not an elective nodal region or a station map. No remaining candidate completes with no RTSTRUCT. The exporter reports empty prediction: no RTSTRUCT created, and the portal records that as a completed task with no contour.

This task is routed to the H100 worker. On the P6000 portal (compute capability 6.1), keep lnq_atlas in REMOTE_GPU_TASKS. Preflight probes runtime-envs/lnq-atlas-ref/bin/python on the execution host. An ordinary sm_XY cubin covers a device with the same major capability and a minor capability greater than or equal to Y. The reference lock's torch 2.0.1+cu118 build lists sm_60, so that cubin covers compute capability 6.1. That is a compatibility rule, not a completed P6000 inference run. sm_90a is not treated as sm_90. The acceptance marker stores the GPU name and compute capability. An H100 marker does not enable the task on a P6000. envs/lnq is not used for this task.

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. Unless a result is marked otherwise, the paper evaluates on the LNQ test set: 100 patients and 861 reference nodes. The ablation that uses all modifications, a binarization threshold of 0.2 and no minimum-diameter filter had Dice 0.6033 and found 57% of reference nodes (LN-found 0.5654), compared with 27% (LN-found 0.2749) when training on CT only. The challenge submission, a five-model ensemble with threshold 0.3 and a 5 mm minimum diameter, had Dice 0.5690 and ASSD 6.8976 mm on that same test set. The authors state that post-processing fit on the small validation set did not carry over to the test set. The installed adapter uses the submission settings (threshold 0.3, 5 mm), not the 0.2 ablation.

python run_lnq_atlas.py --input ct.nii.gz --output prediction.nii.gz \
  --work NEW_PRIVATE_WORK_DIRECTORY --source PINNED_LNQ_SOURCE \
  --weights EXTRACTED_WEIGHTS --manifest provenance.json \
  --total-python TOTAL_V1_ENV/bin/python \
  --total-weights WEIGHTS/totalsegmentator-v1.5.7

The network contract, set in the adapter and recorded in deterministic_contract.json, is CUBLAS_WORKSPACE_CONFIG=:4096:8, nnUNet_compile=false, and, after nnUNetPredictor.__init__, cudnn.deterministic=True with cudnn.benchmark off. nnU-Net 2.2.1 does not set cudnn.benchmark in that constructor. The recorded value is the string off. The task-256 contract is separate. TotalSegmentator 1.5.7 restores nnUNetTrainerV2_ep8000_nomirror with deterministic=False and fp16=True, so the trainer turns cudnn.benchmark on. predict_cases_fastest does not clear that flag. run_total_v1.py sets deterministic cuDNN and benchmark off after load_model_and_checkpoint_files returns, then writes total_deterministic_contract.json. Mixed precision, autocast, fastest mode, TTA off, one resampling thread and the published 3 mm resample stay as released. Thread caps are set in that child process only.

The acceptance harness runs the portal job twice and requires identical task-256 masks and identical final masks before the reference comparison. It keeps both pre-threshold foreground probability maps on the original CT grid. Their maximum absolute difference, and the probability values at mask voxels that differ, are diagnostics. The pass criterion remains exact mask agreement.

The reference process uses lnq-upstream-ref (nnU-Net v2.2.1, commit 84512916ff61ff027245ac175071bf7e0abbc29b, the last release that still provides training.dataloading.data_loader_2d and perform_everything_on_gpu). Production uses runtime-envs/lnq-atlas-ref, built from that environment's lock. The reference imports the released predictor unmodified. The harness does not patch torch, cuDNN, or the predictor in the adapter process. Agreement compares the mask run_task.py passes to the RTSTRUCT exporter with the reference mask on the CT grid. Foreground Dice must be exactly 1. A failure writes no marker.

Remaining reference adaptations, and why they do not change the network numerics: storage paths are redirected; the reference reuses the adapter's TotalSegmentator mask and runs registration again (that stage is not an independent comparison); the reference output is resampled with the same nearest-neighbour call the adapter uses to leave the registration grid (grid only, default value 0, uint8). torch.load(..., weights_only=False) loads the trusted checkpoints. Torch 2.0.1 accepts that argument. It does not change tensor values.

Sources: pinned repository, public share, MELBA paper, repository licence, CC BY-NC-ND 4.0.