Pipeline Progress & Blockers

Status as of 2026-03-23

§1 Route A Pipeline Status

Table 1. Route A pipeline steps

#StepStatusOutputDate
1DICOM → NIfTI conversionDone114 patient directories in ReMIND_nifti/2026-03-16
2First-surgery patient filteringDonefirst_surgery_patients.json (62 patients)2026-03-16
3aRegistration V1 (baseline)Doneregistration_results_v1.csv (204 rows, 6.9% conv.)2026-03-16
3bRegistration V2 (+mask, 3k iter)Doneregistration_results_v2.csv (114 rows, 56.1% conv.)2026-03-17
3cA/B test: optimizer configsDoneab_test_20260317_075447.csv (70 rows, 7 configs)2026-03-17
3dRegistration V3 (2 mm dil.)Doneregistration_results_v3.csv (114 rows, 73.7% conv.)2026-03-17
3eRescue passesDone84/114 valid after QC exclusion (73.7%)2026-03-18
4Pseudo-label generationDone84 labeled volumes → 44,502 2D slices (43 patients)2026-03-17
5nnU-Net data formattingDoneDataset001_iUS (43 patients)2026-03-18
6nnU-Net trainingBlocked
7Evaluation on RESECT-SEGPending

§2 Route B Pipeline Status

Table 2. Route B pipeline steps

#StepStatusOutputDate
1Virtual sweep generation (v3)Done1,090 sweeps across 110 cases2026-03-18
2Best-effort deep tumor fixDone4 cases verified (v3_deep_fix/)2026-03-18
32D→3D NIfTI conversionDone1,878 MRI + 1,090 seg volumes (33.7 GB)2026-03-18
4MMHVAE data normalisationPending
5MMHVAE inferenceBlocked
6Synthetic iUS → nnU-Net trainingPending
7Evaluation on RESECT-SEGPending

§3 Current Blockers

Critical   GPU / nnU-Net Environment Issue — Blocks Route A Step 6

Remote machine (x99-debian) has CUDA 12.2 / Driver 535.247.01 with PyTorch 2.5.1+cu124. GPU driver communication anomaly: nvidia-smi works but PyTorch reports RuntimeError: No CUDA GPUs are available. Root cause is a driver state issue requiring machine reboot.

Impact: Cannot start Route A nnU-Net training until GPU access is restored.

Mitigation: Reboot remote machine. After reboot, verify with python -c 'import torch; print(torch.cuda.is_available())', then start 5-fold cross-validation training (estimated 60–120 hours).

Blocked   MMHVAE Pre-trained Weights Unavailable — Blocks Route B Step 5

The MMHVAE repository (github.com/ReubenDo/MMHVAE) documents pre-trained weights at pretrained/mmhvae_f0/models/CP_main_1000.pth but provides no download link. No GitHub Release exists.

Impact: Cannot run MMHVAE inference to synthesise iUS images. Options: (1) contact the author (R. Dorent) to request weights, (2) train from scratch (~72h GPU time on RTX 2080 Ti), or (3) try older MHVAE checkpoint if available.

Status: Email draft prepared (email_draft_dorent.md). Verify whether the email has actually been sent.

Resolution priority: GPU driver fix is higher priority — it unblocks Route A training immediately with existing data. MMHVAE weights are needed for Route B but have a longer timeline regardless (normalisation step still pending).

§4 Expected DSC Performance Roadmap

Fig. 1. DSC targets and literature references

Fig. 1: Expected Dice Similarity Coefficient (DSC) ranges for each route, based on published literature. Route A baseline (pseudo-label with registration noise): 0.58–0.62 (Faanes et al., 2025). Route B (MMHVAE synthesis): 0.74 (Dorent et al., 2025 TPAMI). Route A+B combined target: 0.84 (expert inter-rater agreement ceiling from ReMIND dataset). Supervised baseline on RESECT-SEG = 0.73 (Dorent et al. 2025).

Table 3. Literature DSC references

MethodDatasetDSCSource
nnU-Net + noisy pseudo-labels (registration)ReMIND (55 patients)0.58–0.62Faanes et al. 2025
MMHVAE synthesis → SegResNetRESECT-SEG0.74Dorent et al. 2025 (TPAMI)
Supervised baseline (w/ manual iUS labels)RESECT-SEG0.73Dorent et al. 2025 (TPAMI)
Expert inter-rater agreementReMIND0.84Juvekar & Dorent et al. 2024