MGM Reduction V2

---

About

Summary

We extend the official Task 3 baseline with residual coordinate prediction and parameter-efficient clinical adaptation using LoRA and a residual input adapter. The model was pretrained on simulated data and fine-tuned on a balanced mixture of simulated and clinical cases. At inference, predictions from 25 independent surface samplings are aggregated for each fragment.

Mechanism

1. Task¶

Task 3: Pelvic Fracture Reduction Planning

2. Team name¶

MultiGeoMed

3. Authors¶

Wojciech Szymanski, Gniewosz Drwiega , Marek Wodzinski

4. Affiliations¶

• Sano – Centre for Computational Personalised Medicine
• AGH University of Krakow

5. Contact author and email address¶

Wojciech Szymanski szymanski@agh.edu.pl

6. Full title or name of the algorithm¶

Residual Coordinate Assembly Transformer

7. Method description¶

Our method extends the official AssemblyTransformer baseline in three ways. First, its coordinate head predicts a residual displacement relative to each input point and is initialized as an exact identity mapping. Second, the simulation-pretrained backbone is adapted to clinical data using rank-32 LoRA updates to all intra- and inter-fragment attention projections, while the original backbone remains frozen. Fine-tuning also updates the coordinate head and a zero-initialized 6 -> 64 -> 384 residual input adapter for raw coordinates and normals. Third, inference is repeated for 25 independent surface samplings, and the resulting rigid transformations are aggregated fragment-wise in SE(3). The remaining network structure, rigid pose extraction, and iterative reduction procedure follow the official baseline.

8. Main technical contributions and components¶

1. Identity-initialized residual coordinate regression. The model learns point displacements while retaining an exact identity mapping at head initialization.
2. Parameter-efficient clinical adaptation. Rank-32 LoRA updates are added to all intra- and inter-fragment attention projections, while the simulation backbone remains frozen.
3. Residual clinical input adapter. A zero-initialized 6 -> 64 -> 384 adapter jointly uses raw coordinates and normals to correct the frozen input representation without disturbing the pretrained initialization.
4. Validation-guided full-data refit. The stopping epoch is chosen on a non-overlapping held-out split and then reused in a fresh refit that exposes the model to all 170 official clinical cases.
5. Rigid-transform sampling ensemble. Twenty-five independent mesh samplings are reduced iteratively, then aggregated separately for each fragment using robust transformed-centroid statistics and quaternion averaging.

9. Complete pipeline¶

1. Apply the official baseline OBJ parsing, surface sampling, and geometric normalization.
2. Process the points with the simulation-pretrained AssemblyTransformer augmented by the residual coordinate head, clinical input adapter, and LoRA attention updates.
3. Use the baseline rigid pose extraction and iterative reduction procedure to obtain one transformation per fragment.
4. Repeat the complete inference procedure for 25 independently seeded surface samplings.
5. Aggregate the resulting rigid 3D transformations separately for each fragment, normalize them relative to the first sacral fragment, and write reduction-poses-matrices.json.

10. Use of external data¶

No external data were used.

11. Use of externally pretrained models¶

No externally pretrained model was used.

12. Preprocessing¶

The standard baseline preprocessing was retained. The only additional training preprocessing was the one-time conversion of clinical OBJ meshes into cached point clouds. Ground-truth transformations were applied to obtain assembled target coordinates and normals, and 5,000 surface points per anatomical bone were allocated equally among that bone's fragments.

Inference used the unchanged baseline volume-weighted surface sampler and geometric normalization. The sampler was rerun independently 25 times to produce the sampling ensemble.

13. Data augmentation¶

The official baseline augmentation pipeline was retained and applied to both simulated and cached clinical geometry. The only changed parameter was the synthetic fragment-dropout probability, increased from 0.10 during simulation pretraining to 0.15 during fine-tuning. This operation was not applied to clinical cached samples.

No additional normal corruption, contact erosion, empirical pose mixture, pose curriculum, representation-sampling augmentation, or real clinical pose replay was used (real_pose_ratio=0).

14. Training and validation strategy¶

Stage 1 — simulation baseline initialization¶

• The official baseline was trained on the simulated dataset. Its zero-based epoch 914 checkpoint initialized all subsequent clinical adaptation runs.

Stage 2 — method development on the original clinical split¶

• The 170 clinical cases were manually divided into 140 development cases and 30 cases reserved for independent local inference evaluation.
• Within the 140-case development directory, the loader used the first 90% (126 cases) for training and the remaining 10% (14 cases) for validation.
• mix_balanced produced 126 clinical and 126 randomly selected simulated samples per training epoch. Validation was deterministic and contained 14 clinical plus 14 simulated samples.
• The model was initialized from simulation epoch 914 and fine-tuned for the complete 300-epoch schedule. The development run used 1 GPU, batch size 8, accumulation 2, and therefore an effective batch size of 16.
• The best checkpoint on this development protocol was epoch 180. The separate 30-case set was used to compare complete inference pipelines, not to update network weights.

Stage 3 — stopping-epoch selection on all available clinical cases¶

• After the architecture and fine-tuning recipe were fixed, the manually separated directories were merged into one 170-case directory.
• clinical_train_use_all=false created a new deterministic sorted split of 153 clinical training cases and 17 non-overlapping clinical validation cases.
• Training remained a 50/50 balanced mixture: 153 clinical and 153 sampled simulated cases per epoch.
• Because val_data_type was left unset, validation inherited data_type=mix_balanced: it contained 17 clinical and 17 simulated samples. Thus, V2 stopping-epoch selection used combined clinical and simulation validation, not clinical-only validation.
• The run used 2 GPUs, batch size 8 per GPU, accumulation 2 (effective batch size 32), and the same 300-epoch OneCycle schedule. Checkpoint epoch 182 was selected from the held-out validation trajectory for the final refit.

Stage 4 — final full-data refit¶

• A fresh fine-tuning run was initialized again from simulation checkpoint 914; it did not continue from the Stage 3 LoRA weights.
• clinical_train_use_all=true exposed all 170 clinical cases to training. Each epoch contained all 170 clinical cases and 170 sampled simulated cases.
• All model, augmentation, optimizer, and 300-epoch scheduler settings were retained. The run used 2 GPUs, batch size 8 per GPU, accumulation 2 (effective batch size 32).
• The submission checkpoint was saved immediately after zero-based epoch 182, matching the stopping point selected in Stage 3. Any validation values in this full-data run were diagnostic only because the clinical validation subset overlapped the training data.

This four-stage procedure first established a simulation-trained initialization, kept the 30 manually held-out cases independent during method development, estimated a stopping point after all 170 official clinical cases became available to the development split, and then used every official case in the final parameter refit without selecting a checkpoint on overlapping validation data.

15. Loss function¶

The baseline point-wise assembled-coordinate MSE was retained without changes. No fragment-balanced term, direct translation or rotation loss, rigid loss, sampling-consistency loss, teacher loss, or other auxiliary loss was used.

16. Base network architecture¶

The 12-layer, 384-dimensional official AssemblyTransformer backbone was retained. Three modifications were introduced:

1. An identity-initialized residual coordinate head, Linear(384,384) -> SiLU -> Linear(384,3, bias=False), predicts a displacement added to each input coordinate.
2. A zero-initialized residual input adapter, Linear(6,64) -> GELU -> Linear(64,384), adapts raw coordinates and normals to the clinical domain.
3. Rank-32, alpha-64 LoRA updates are applied to the QKV and output projections of both attention operations in every transformer layer.

The original backbone remains frozen. Only the LoRA matrices, residual input adapter, and coordinate head are optimized, corresponding to approximately 1.94 million of 37.63 million parameters (5.17%).

17. Ensembling during inference¶

The method used one network checkpoint and a 25-member sampling ensemble; it did not ensemble different trained models.

For each member, a deterministic seed derived from base seed 42 and the case ID produced a new mesh-surface sampling. The complete iterative inference pipeline was run independently for that sampling. Before aggregation, each prediction was expressed relative to the first sacral fragment. For every fragment:

• rotations were averaged with a sign-invariant quaternion mean;
• the transformed fragment centroid was aggregated with the coordinate-wise median;
• translation was reconstructed so that the averaged rotation mapped the original centroid to the median transformed centroid.

The aggregated transforms were normalized once more relative to the first sacral fragment. No rotation-angle scaling and no centroid-displacement scaling were applied in the submitted method.

18. Public code repository¶

TBD

---

Interfaces

This algorithm implements all of the following input-output combinations:

Validation and Performance

---

Challenge Performance

Date	Challenge	Phase	Rank
Aug. 18, 2026	PENGWIN2026	Task 3: PENGWIN-Reduction-Preliminary	9
Aug. 18, 2026	PENGWIN2026	Task 3: PENGWIN-Reduction-Final Test	4

Uses and Directions

This algorithm was developed for research purposes only.

Warnings

Common Error Messages