w02 alpha06
About
Summary
This algorithm predicts rigid reduction transformations for peripelvic fracture fragments in PENGWIN Task 3. It uses a point-cloud transformer adapted from the official AssemblyTransformer baseline. The model takes a fractured pelvis fragment mesh as input and outputs one 4x4 rigid transformation matrix for each fragment.
The submitted version is trained on clinical fracture data using coordinate reconstruction supervision with an additional fragment-level pose regularization term. A conservative pose scaling post-processing step is applied to improve robustness on validation cases.
Mechanism
The algorithm targets pelvic fracture reduction planning from segmented peripelvic fracture fragment meshes.
Input fragments are loaded from the OBJ mesh and grouped by fragment identifier. Surface points and normals are sampled from each fragment and passed to a transformer-based point-cloud assembly network. The network predicts restored coordinates for the sampled points. For each fragment, a rigid transformation is then recovered from the input points and predicted restored points using SVD-based rigid registration.
All predicted transforms are expressed relative to the first sacrum fragment, which is used as the fixed reference fragment. The final output is a JSON file containing one homogeneous 4x4 transformation matrix per fragment.
The submitted container includes the model weights internally and does not require a separate Grand Challenge model upload.
Interfaces
This algorithm implements all of the following input-output combinations:
Validation and Performance
The method was evaluated on a held-out clinical validation split of 17 cases from the available training data. Metrics were computed using the Task 3 evaluation protocol.
| Model | Trans error (mm) ↓ | Rot error (deg) ↓ | TRE (mm) ↓ | CD_raw (mm) ↓ | PA (%) ↑ |
|---|---|---|---|---|---|
| Clinical baseline, coordinate loss only | 6.030 ± 2.586 | 8.70 ± 7.39 | 4.591 ± 1.868 | 4.769 ± 1.456 | 58.6 |
| Clinical baseline + pose loss, w=0.2 | 5.835 ± 2.600 | 8.66 ± 7.09 | 4.613 ± 2.026 | 4.787 ± 1.620 | 60.6 |
| Submitted version, w=0.2 + conservative alpha=0.6 post-processing | 5.988 ± 2.950 | 8.94 ± 7.33 | 4.661 ± 1.979 | 4.882 ± 1.551 | 64.9 |
The submitted version was selected because it gave the best validation Part Accuracy while keeping translation, rotation, TRE, and Chamfer Distance close to the stronger continuous-metric baselines. The alpha=0.6 post-processing scales relative predicted fragment poses toward the identity transform, which reduced over-correction on several clinical validation cases.
Challenge Performance
| Date | Challenge | Phase | Rank |
|---|---|---|---|
| July 3, 2026 | PENGWIN2026 | Task 3: PENGWIN-Reduction-Preliminary | 4 |
Uses and Directions
This algorithm was developed for research purposes only.