Synthetic & Open Data for iUS Tumor Segmentation
This repository documents the data, intermediate results, and quality control outputs for three technical routes to train intraoperative ultrasound (iUS) brain tumor segmentation models without manual iUS annotations. Route A transfers MRI tumor labels to iUS space via rigid registration. Route B synthesises realistic iUS from MRI using MMHVAE (Dorent et al., 2025). Route A+B combines both approaches via two-stage training.
All data derives from the ReMIND dataset (Juvekar et al., 2023) — 114 patients, 41 GB DICOM, publicly available from TCIA. Of these, 62 are first-surgery patients suitable for registration; 110 have sufficient anatomy for virtual sweep simulation. The entire pipeline from raw DICOM to training-ready data was executed over 2026-03-16 to 2026-03-18 on a single RTX 2080 Ti machine.
Table 1. Technical Route Comparison
The three routes differ in how training labels are obtained and whether the model sees real or synthetic iUS. Route A is the simplest baseline (direct label transfer with registration noise); Route B eliminates registration error by synthesising iUS from MRI; Route A+B leverages both real and synthetic data.
| Route A (Baseline) | Route B (Core Method) | Route A+B (Innovation) | |
|---|---|---|---|
| Pipeline | MRI → Rigid Reg → Pseudo-label → nnU-Net | MRI → Virtual Sweep → MMHVAE → nnU-Net | A pre-train → B fine-tune |
| Label quality | Noisy (3–5 mm registration error) | Precise (zero registration error) | Complementary |
| Training images | Real iUS | Synthetic iUS | Both |
| Data scale | 44,502 slices / 43 patients | 1,090 sweeps / 110 cases | Combined |
| DSC reference | 0.58–0.62 (Faanes 2025) | 0.74 (Dorent 2025) | Target: 0.84 |
| Status | Data ready; training blocked (GPU) | Sweeps done; MMHVAE blocked (weights) | Pending both routes |
Table 2. Data File Index
All quantitative results are stored as flat CSV or per-case JSON files. This index lists the primary data files referenced throughout the dashboard; each is linked to the relevant detail page where its contents are analysed.
| File | Rows | Key Columns | Route |
|---|---|---|---|
registration_results_v1.csv | 204 | patient_id, translation_magnitude_mm, mi_improvement, stop_condition, status | A |
registration_results_v2.csv | 114 | + convergence, mask_coverage | A |
registration_results_v3.csv | 114 | + convergence, mask_coverage | A |
ab_test_20260317_075447.csv | 70 | config (7 optimizers × 10 patients), converged, stop_class | A |
sweep_metadata.json × 110 | ~10 sweeps each | tumor_diameter_mm, C2_dist_to_tumor, tumor_pixels, saved_slices | B |
conversion_*.csv | ~1090 | case_id, modalities, volume_shape, seg_nonzero | B |
Table 3. Literature DSC References
Published Dice Similarity Coefficient (DSC) values for iUS tumor segmentation methods relevant to this project.
| Method | Dataset | DSC | Source |
|---|---|---|---|
| nnU-Net + noisy pseudo-labels (registration) | ReMIND (55 patients) | 0.58–0.62 | Faanes et al. 2025 |
| MMHVAE synthesis → SegResNet | RESECT-SEG | 0.74 | Dorent et al. 2025 (TPAMI) |
| Supervised baseline (w/ manual iUS labels) | RESECT-SEG | 0.73 | Dorent et al. 2025 (TPAMI) |
| Expert inter-rater agreement | ReMIND | 0.84 | Juvekar & Dorent et al. 2024 |