Skip to content

GTRC-Net Total Tumour Burden Integration (gtrc_psma_petct_ttb, gtrc_fdg_petct_ttb)

1. Overview and Published Evidence

GTRC-Net (Global Threshold Regional Consensus Network) segments whole-body metastatic prostate cancer total tumour burden (TTB) on paired PET/CT examinations:

  • gtrc_psma_petct_ttb: PSMA-avid total tumour burden (GTRC_PSMA_TTB) on [68Ga]PSMA-11 or [18F]DCFPyL PET/CT (Dataset881_PSMA_PET).
  • gtrc_fdg_petct_ttb: FDG-avid metabolic total tumour burden (GTRC_FDG_TTB) on [18F]FDG PET/CT (Dataset882_FDG_PET).

Primary Reference and Cohorts

  • Publication: Jackson P, Buteau JP, McIntosh L, et al. Deep Learning for Automated Measures of SUV and Molecular Tumor Volume in [68Ga]PSMA-11 or [18F]DCFPyL, [18F]FDG, and [177Lu]Lu-PSMA-617 Imaging with Global Threshold Regional Consensus Network. J Nucl Med. 2025;66(11):1811–1817. PMID: 40935568. DOI: https://doi.org/10.2967/jnumed.125.270077.
  • Source repository: https://github.com/Peter-MacCallum-Cancer-Centre/GTRC-Net-Pretrained at pinned commit 9db02ddd569fd4e82a2caf8973aba6741d5cd4a8.
  • Public engineering fixture: Zenodo record 18150034 (train_0001.zip, CC BY 4.0).
  • Development and external test cohorts (mCRPC patients staged for [177Lu]Lu-PSMA-617 across the ProsTIC registry, LUCAB, LuPSMA, LuPARP, and PRINCE studies; only scans with quantifiable disease were included; exact values from Table 1, p. 1814):

| Task | Tracer | 5-Fold CV Cohort | External Holdout Test Cohort | Mean Volumetric DSC (CV / Holdout, Table 1) | Median Surface DSC (nnU-Net → GTRC-Net) | Biomarker Pearson r (Volume / SUVmean / SUVmax / VIP) | Liver Involvement Sens. / Spec. (Median Hepatic DSC) | |---|---|---:|---:|---|---|---| | gtrc_psma_petct_ttb | [68Ga]PSMA-11 (567) + [18F]DCFPyL (109) | 676 | 56 | 0.9365 (0.9333 Ga-68, 0.9553 DCFPyL) / 0.9520 | 0.85 → 0.96 | 0.98 / 0.98 / 0.96 / 0.97 | 73.7% / 98.4% (0.564) | | gtrc_fdg_petct_ttb | [18F]FDG | 390 | 56 | 0.8448 / 0.8447 | 0.80 → 0.94 | 0.94 / 0.88 / 0.91 / 0.96 | 81.8% / 98.1% (0.759) |

Patient-Specific SUV Threshold Contract

In JNM 2025;66(11):1811–1817 (Materials and Methods, Cohort Description and Ground Truth Reference), the authors state verbatim:

"Cases were delineated by nuclear medicine physicians and showed the extent of whole-body metastatic disease using either a fixed threshold (SUV of ≥3 for PSMA PET and SPECT imaging) or a variable liver-based threshold (1.5x mean + 2x SD of liver background for [18F]FDG PET imaging), consistent with the methods reported by Ferdinandus et al. (8), Buteau et al. (7), and Wahl et al. (20)."

(Note: the DEEP-PSMA Grand Challenge pages describe the FDG reference threshold as liver mean + 2 SD.) - gtrc_fdg_petct_ttb: There is no fixed FDG default. The user must enter the patient-specific FDG SUVbw threshold (0.1 to 100.0 g/mL) for each job via the portal UI (#gtrcFdgThreshold) or API (task_options.gtrc_fdg_petct_ttb.suv_threshold or fdg_suv_threshold). It is persisted in job.json with "source": "user-entered", passed to run_gtrc.py --suv-threshold, and recorded in the sidecar JSON. Jobs missing a valid FDG threshold fail closed with HTTP 400 / ValueError. - gtrc_psma_petct_ttb: Uses the published 3.0 g/mL threshold by default ("source": "published-default"), with optional per-job override in [0.1, 100.0] g/mL ("source": "user-entered"). The API preserves "published-default" only for gtrc_psma_petct_ttb at 3.0 and rejects "published-default" for any other PSMA value or any FDG threshold. - Implementation threshold range rationale (MIN_SUV_THRESHOLD = 0.1, MAX_SUV_THRESHOLD = 100.0): [0.1, 100.0] is an RTsegmentator implementation input-validation range (not a range stated in Jackson et al.). It prevents division by zero, negative numbers, NaN/Inf, or unscaled raw Bq/mL activity values (typically $10^3\text{–}10^5\text{ Bq/mL}$) accidentally entered in place of SUVbw, while accommodating any positive clinical PET SUVbw threshold.

Licence Status

Neither the upstream GitHub repository nor the weight files state a software or weight licence ("No licence stated by the authors"). Per owner decision on 2026-10-22, these terms are accepted and recorded in catalog/gtrc.json and sidecar provenance.

2. Paired PET/CT Input Contract and QIBA SUVbw Conversion

Both tasks register "modality": "PT", "requires_ct": true, and "requires_preflight": true in catalog/gtrc.json: 1. Reference series (job/input): Attenuation- and decay-corrected 3D PET series (Modality == "PT"). Exported RTSTRUCT contours reference the PET SOP Instance UIDs and physical grid. 2. Paired series (job/input_ct): Co-registered 3D CT series (Modality == "CT") in the same StudyInstanceUID and FrameOfReferenceUID. - Every CT slice is inspected (inspect_ct_series_metadata) to verify that all slices share one StudyInstanceUID and FrameOfReferenceUID matching the PET series. - If an upload contains multiple CT series matching the PET series's StudyInstanceUID and FrameOfReferenceUID, the job fails closed with HTTP 400 unless the user explicitly selects the companion CT series (#gtrcCtCompanion / ct_series_key). - Before running inference, gtrc_models.check_physical_overlap enforces: (a) 3D physical LPS bounding-box overlap fraction >= 0.25, (b) CT physical axial (Z, Inferior–Superior) extent covering the PET axial (Z) extent within max(5.0 mm, pet_spacing_z, ct_spacing_z) (AXIAL_COVERAGE_MIN_TOLERANCE_MM = 5.0), and (c) 100% of PET voxels with SUVbw >= suv_threshold lying inside the actual sitk.sitkLinear interpolator support of the CT volume (decided directly from the resampled all-ones CT support image using the same identity transform, sitkLinear interpolator, and 0.0 outside default as the upstream CT resampling in preprocess_pet_ct, without a narrower [0, size - 1] index restriction) so no above-threshold PET voxel receives the -1000 HU outside-buffer default value.

Vendor-Neutral PET-to-SUVbw Converter (pet_to_suvbw.py)

pet_to_suvbw.py implements the RSNA Quantitative Imaging Biomarker Alliance (QIBA) SUVbw profile (FDG-PET/CT Technical Committee SUVbw Pseudocode; DICOM PS3.3 C.8.9.1 / C.8.9.2 / C.7.6.16.2.11):

$$\text{decay_factor} = 2^{-(t_{\text{ref}} - t_{\text{inj}}) / T_{1/2}}, \quad \text{SUV}{\text{bw}}\;[\text{g/mL}] = C$$}}\;[\text{Bq/mL}] \times \frac{W_{\text{patient}}\;[\text{kg}] \times 1000}{A_{\text{injected}}\;[\text{Bq}] \times \text{decay_factor}

Explicit header checks enforced across all ordered slices: - Units (0054,1001) and RealWorldValueMappingSequence (0040,9096): - Validated on every slice. - "BQML": converted to float32 SUVbw using the verified radiopharmaceutical and patient-weight tags below. - Explicit body-weight SUV (Units == "GML" with SUVType (0054,1006) equal to "BW" or absent, or Units == "SUVBW", or Units == "SUV" with SUVType == "BW", or Units paired with a verified UCUM g/ml{SUVbw} mapping): retained in float32 SUVbw (conversion_mode = "already_suvbw", suv_scale_factor = 1.0) and never converted twice. - Every RealWorldValueMappingSequence item must carry a 1-item UCUM MeasurementUnitsCodeSequence (0040,08EA). LUTExplanation (0028,3003) and LUTLabel (0040,9210) are allowed as descriptive text on linear mappings; only actual lookup-table data (LUTData (0028,3006) or RealWorldValueLUTData (0040,9212)) is rejected as a nonlinear LUT. For integer PixelData, RealWorldValueFirstValueMapped (0040,9216) and RealWorldValueLastValueMapped (0040,9211) are required (DICOM PS3.3 Table C.7.6.16-12b) and verified against each slice's actual decoded stored pixels and SmallestImagePixelValue/LargestImagePixelValue; out-of-range pixels or contradictory extrema are rejected. Linear RWVM RealWorldValueSlope/RealWorldValueIntercept must agree with RescaleSlope/RescaleIntercept within RWVM_SLOPE_RTOL = 1e-5 and bound the resulting SUV error across the verified stored-value domain within RWVM_MAX_SUV_ERROR_TOL = 1e-4 g/mL. - Non-body-weight SUV types ("BSA", "LBM", "LBMJAMES", "LBMJANMA", "IBW"), non-activity units ("CNTS", "CPS", "PROPCNTS", etc.), missing/non-UCUM MeasurementUnitsCodeSequence, nonlinear LUT data, or unapplied/contradictory linear mappings are rejected on every slice with a ValueError. - Rescale Tags: - Requires explicit RescaleSlope (0028,1053) > 0 and finite RescaleIntercept (0028,1052) on every slice (missing tags are rejected rather than defaulted). - Decay and Image Corrections (Units == "BQML"): - Requires DecayCorrection (0054,1102) == "START" and CorrectedImage (0028,0051) containing both "DECY" and "ATTN" on every slice. - Dose, Half-Life, Weight, and QIBA Scan-Start Timing (Units == "BQML"): - Requires PatientWeight (0010,1030) in (0, 500] kg, RadionuclideTotalDose (0018,1074) >= 10,000 Bq (an explicit heuristic floor MIN_PLAUSIBLE_DOSE_BQ = 10_000.0 to reject accidental MBq/mCi entries while acknowledging that DICOM specifies Bq without a formal lower bound), and RadionuclideHalfLife (0018,1075) > 0 s, consistent across slices. - Parses explicit DICOM DT timezone suffixes (&ZZXX), cross-checks them against TimezoneOffsetFromUTC (0008,0201) when present, retains each slice's acquisition UTC offset (s_acq_tz) in every timestamp comparison (including SeriesTime (0008,0031) vs AcquisitionTime/AcquisitionDateTime), converts timezone-aware timestamps to UTC, and rejects any unresolved aware/naive or conflicting timezone offset combination (such as an explicit AcquisitionDateTime or RadiopharmaceuticalStartDateTime DT offset paired with a naive SeriesTime or RadiopharmaceuticalStartTime TM without TimezoneOffsetFromUTC). - Resolves the decay reference scan-start timestamp (_resolve_qiba_scan_start) across all slices following the QIBA profile while requiring consistency across all present acquisition timing records (SeriesDate/SeriesTime, AcquisitionDate/AcquisitionTime, FrameReferenceTime + ActualFrameDuration, GE (0009,100D) with private creator (0009,0010) == "GEMS_PETD_01", and Siemens (0071,1022) with private creator (0071,0010) == "SIEMENS MED PT"), rejecting conflicting evidence (> 2.0 s discrepancy or scan start later than AcquisitionTime + 1.0 s) rather than preferring one source. - Midnight and Date Crossing Guard: Never guesses a +86400 s rollover. Requires paired date+time for both scan start and injection whenever dates differ across tags or times cross midnight.

3. Published Inference Method and Static Blockers Fixed

Role of rigid.tfm, threshold.json, and totseg_24.nii.gz in Upstream Source

Inspection of all upstream scripts (00_copy_deep_psma_training_data.py through 06_Train_Consensus_Evaluator.py, GTRC.py, GTRC_Infer.py, gtrc_utils.py) and the Zenodo train_0001 archive confirmed: - rigid.tfm (inter-study rigid transform between the PSMA CT and FDG CT in the DEEP-PSMA dataset) and totseg_24.nii.gz (TotalSegmentator organ labels) are never read in either training or inference. - threshold.json ({"suv_threshold": 2.999862} in train_0001/PSMA/threshold.json; {"suv_threshold": 2.9650267106980666} in train_0001/FDG/threshold.json) is read only during training data preparation (00_copy_deep_psma_training_data.py), where 01_Preprocess_Dataset.py defines training class 2 (norm) as (PET / suv_threshold >= 1.0) & (TTB == 0). - At inference time (GTRC.py / GTRC_Infer.py), normal physiologic uptake (label 2) is predicted directly by the 3-class nnU-Net (0=background, 1=ttb, 2=norm), and the SUV threshold is supplied via --suv-threshold (resolve_job_suv_threshold): - gtrc_psma_petct_ttb: 3.0 g/mL ("source": "published-default", or "user-entered" when overridden in [0.1, 100.0]; acceptance uses train_0001/PSMA/threshold.json = 2.999862). - gtrc_fdg_petct_ttb: required per-job patient-specific threshold ("source": "user-entered", finite in [0.1, 100.0], no default; acceptance uses train_0001/FDG/threshold.json = 2.9650267106980666).

Preprocessing, nnU-Net v2, and Subregion Consensus Postprocessing

  1. CT Resampling and Threshold Division:
  2. CT is resampled to the PET physical grid using sitk.ResampleImageFilter() (SetReferenceImage(pt_suv), identity transform, linear interpolation, SetDefaultPixelValue(-1000)).
  3. Channel 0 (gtrc_0000.nii.gz) is pt_rs = pt_suv / suv_threshold (NoNormalization in plans.json).
  4. Channel 1 (gtrc_0001.nii.gz) is ct_rs in HU (CTNormalization: PSMA clip [-1000, 920], mean 32.071888, SD 212.490005; FDG clip [-1000, 1039], mean 67.151413, SD 188.757507).
  5. 5-Fold 3D nnU-Net v2 with Deterministic cuDNN:
  6. Runs nnUNetv2_predict -d <881|882> -c 3d_fullres across all 5 folds (fold_0..4/checkpoint_final.pth), producing gtrc.nii.gz (0=background, 1=ttb, 2=norm).
  7. Because nnUNetPredictor.__init__ sets torch.backends.cudnn.benchmark = True, run_gtrc.py wraps nnUNetPredictor.__init__ via DETERMINISTIC_NNUNET_ENTRYPOINT so torch.backends.cudnn.benchmark = False and torch.backends.cudnn.deterministic = True are enforced after nnU-Net initializes its predictor, and writes the child process's effective torch/cuDNN flags to nnunet_runtime_flags.json.
  8. Documented H100 train_0001 cuDNN policy deviation: On H100 (sm_90) train_0001/PSMA, a controlled repeatability experiment (repeatability-20261022/controlled-comparison.txt) established that deterministic cuDNN (cudnn.deterministic=True, cudnn.benchmark=False, 4 runs with 0 intra-group differing voxels) and unmodified autotuned nnUNetv2_predict (cudnn.benchmark=True, 3 runs with 0 intra-group differing voxels) each reproduce their own raw nnU-Net label predictions identically while differing from each other by 1 raw nnU-Net label voxel, with 0 differing subregion or final TTB voxels. On train_0001/FDG (acceptance_h100_r1b), deterministic reference vs autotuned runner differed by 2 raw nnU-Net label voxels with 0 differing subregion or final TTB voxels (same mechanism expected, not separately ablated). Both the production runner and the upstream reference wrapper in acceptance_gtrc.py enforce deterministic cuDNN.
  9. Acceptance contract by mode: reference mode gives the runner and unmodified GTRC_Infer.py the same NIfTI inputs and requires exact agreement of raw nnU-Net labels, subregions and the final TTB mask. DICOM mode reads PET back from int16 DICOM, which carries the declared SUV encoding error (at most 0.000385 SUVbw on train_0001), so raw nnU-Net labels and subregions can differ at a few boundary voxels; those counts are recorded as diagnostics. The exported TTB mask must still agree exactly with reference mode, and the RTSTRUCT round trip must return it. On the H100 (acceptance_h100_r2) DICOM mode differed from reference mode by 1 (PSMA) and 4 (FDG) raw nnU-Net label voxels and 10 (PSMA) and 0 (FDG) subregion voxels, with 0 differing TTB voxels.
  10. Watershed Candidate Subregion Partitioning:
  11. Threshold mask: tar = (pt_ar >= 1.0).astype('int8') (voxels with SUV >= suv_threshold).
  12. Calls gtrc_utils.create_subregion_labels(pt_rs, tar, local_maxima_threshold=0.16666, sphere_radius=8.0):
    • Builds an 8.0 mm physical sphere footprint on the PET voxel spacing (np.meshgrid with indexing='ij' followed by np.swapaxes(sphere, 0, 2)).
    • Detects local maxima via skimage.morphology.h_maxima(qs_ar, 0.16666, sphere), labels peaks with 6-connected scipy.ndimage.label, and partitions tar via skimage.segmentation.watershed(-qs_ar, numbered_peaks, mask=threshold_array).
    • Preserved upstream indexing quirk: Above-threshold connected components without an h_maxima peak (missing_peaks_ar) are assigned labels_ws[missing_peaks_ar_numbered == (i+1)] = n_peaks + i for i in range(n_missing) and cast to sitkInt16. Because i starts at 0, the first peakless component receives label n_peaks (merging with peak n_peaks when n_peaks >= 1, or remaining 0/background when n_peaks == 0). run_gtrc.py preserves this exact upstream function so outputs match unmodified GTRC_Infer.py bit-for-bit.
  13. Six Subregion Features, 5-Fold Keras MLP Voting, and Final TTB Re-painting:
  14. For each candidate subregion 1..total_subregions, computes [total_volume, suv_max, suv_mean, ct_hu_mean, pred_ttb_overlap, pred_norm_overlap] (where suv_max and suv_mean are computed on pt_rs = pt_suv / suv_threshold, matching GTRC_Infer.py).
  15. Loads fold_0..4/consensus_model.keras on CPU (CUDA_VISIBLE_DEVICES="-1", tf.config.set_visible_devices([], "GPU")). Each model votes positive if score > 0.5; a subregion is accepted if votes > 2.5 (strict majority ≥3 of 5).
  16. Final re-painting sets output_ar[(pred_ttb_ar > 0) & (tar > 0)] = 1.

Static Blockers Fixed

  1. GTRC_Infer.py line 27 hardcoded nn_predict_exe = r"C:\GTRC\GTRC_venv\Scripts\nnUNetv2_predict.exe" and used os.system(); run_gtrc.py invokes nnUNetv2_predict from the isolated gtrc environment via subprocess.run(..., check=True).
  2. GTRC_Infer.py lines 157/167 referenced data/LuPSMA/... while the directory on disk is data/LUPSMA/....
  3. nnunet_config_paths.py created directories relative to the working directory and ../GTRC_baseline/ at import time; run_gtrc.py confines nnUNet_raw and nnUNet_preprocessed inside the job workspace and points nnUNet_results to the read-only bundle weights.

4. Outputs, Sidecar Schema, and Empty-Result Handling

  • RTSTRUCT ROI: GTRC_PSMA_TTB (gtrc_psma_petct_ttb) or GTRC_FDG_TTB (gtrc_fdg_petct_ttb), exported on the reference PET grid via convert_validate_rtstruct.py.
  • Declared Output Artifacts (OUTPUT_ARTIFACTS):
  • <task>_RTSTRUCT.dcm (RTSTRUCT on reference PET grid)
  • <task>_RTSTRUCT.dcm.lesions.json (biomarker, connected-component, subregion, threshold, and provenance sidecar)
  • <task>_RTSTRUCT.dcm.ttb_mask.nii.gz (binary NIfTI TTB mask on reference PET grid)
  • Naming every declared output artifact with the <task>_RTSTRUCT.dcm prefix guarantees automatic transfer by REMOTE_OUTPUT_GLOBS (f"{task}_RTSTRUCT.dcm*") on routed remote GPU jobs. The intermediate <task>_pet_suvbw.json conversion metadata stays in the scratch work directory and is embedded in lesions.json["job_context"]["pet_suvbw_conversion"].
  • Empty Predictions: If 0 voxels are predicted as TTB, gtrc_models.run_task copies <task>_RTSTRUCT.dcm.lesions.json and <task>_RTSTRUCT.dcm.ttb_mask.nii.gz to job/output/ and exits with "empty prediction: no RTSTRUCT created", which app.py (including routed SSH retrieval) records as a valid empty result (NO_STRUCTURES_FOUND.txt + sidecars in rtstruct-results.zip).

5. Out-of-Scope Note: Lu-PSMA SPECT/CT (Dataset883_LUPSMA_SPECT)

In Jackson et al. (J Nucl Med 2025;66(11):1811–1817, Table 1, p. 1814), the [177Lu]Lu-PSMA-617 SPECT/CT model was evaluated on 477 scans in 5-fold cross-validation (mean DSC 0.9696), with no external SPECT holdout test cohort. Dataset883_LUPSMA_SPECT and data/LUPSMA/subregion_metrics/fold_0..4/consensus_model.keras are included in the bundle layout proposal (LUPSMA_WEIGHT_FILES in gtrc_models.py), but no portal task is enabled because: 1. Portal upload assessment currently ingests CT, MR, and PT single-frame slice series, not multi-frame NM (Nuclear Medicine) SPECT tomograms. 2. Quantitative Lu-PSMA SPECT in GTRC_Infer.py requires input voxel values calibrated in quantitative SUV (Bq/mL with camera sensitivity calibration factor, decay correction, injected activity, and patient weight, or vendor quantitative SPECT SUV maps), whereas clinical NM DICOM exports frequently store uncalibrated counts (CNTS).