PointReduceNet based Peripelvic Fracture Reduction Planning
About
Summary
The pipeline predicts one rigid transform per fragment directly from the input meshes. It is a compact point-cloud network, deliberately lighter than the challenge AssemblyNet baseline, chosen for speed and robustness under a tight time budget
Mechanism
Input and preprocessing: a single OBJ file per case, with fragments identified by group ID (1 to 100 sacrum, 101 to 200 left ilium, 201 to 300 right ilium). Each fragment surface is sampled to a fixed number of points with normals, and the whole assembly is centred on its mean and scaled to a unit frame. - Per-fragment encoder: a PointNet turns each fragment point cloud into a shape descriptor, to which a bone one-hot and the fragment centroid are appended. - Global reasoning: a Transformer self-attention layer lets the fragments attend to each other, so the model reasons about how the pieces fit together rather than each in isolation, the way a surgeon mentally reassembles the puzzle. - Peripelvic Fracture Reduction 2 / 3 - Pose decoding: a per-point head predicts, for every point, its position in the reassembled bone; a rigid transform is then recovered from these predicted positions with a Kabsch (SVD) fit. - Anchor convention: the first sacrum fragment is fixed to the identity and every other pose is expressed relative to it, matching the challenge evaluation.
Interfaces
This algorithm implements all of the following input-output combinations:
Validation and Performance
Challenge Performance
| Date | Challenge | Phase | Rank |
|---|---|---|---|
| Aug. 19, 2026 | PENGWIN2026 | Task 3: PENGWIN-Reduction-Preliminary | 41 |
| Aug. 19, 2026 | PENGWIN2026 | Task 3: PENGWIN-Reduction-Final Test | 22 |
Uses and Directions
This is a research tool. It works as a first automatic pass whose output is reviewed and corrected by a qualified operator, for research on fracture reduction and fixation planning studies, dataset pre-annotation, and challenge benchmarking. It is not a certified medical device. It must not be used on its own for diagnosis, for treatment decisions, or for guidance during surgery, and it should not be pushed outside the setting it was built for, such as non-CT images, anatomy other than the pelvis, or scans and protocols that differ a lot from the training data.
Warnings
Performance is only established on data that looks like the PENGWIN training set. Results can drop on out-of-distribution inputs: unusual resolution or slice thickness, contrast-enhanced scans, metal or motion artifacts, orthopedic hardware, pediatric anatomy, or cases with tumors or prior surgery. The predicted pose is only as good as the network point cloud. The Kabsch step cannot fix a wrong prediction, it only projects it onto the nearest rigid motion. Because training scatters fragments synthetically, a mismatch between the synthetic displacement distribution and real clinical displacement can cause systematic errors, such as over-rotation. We mitigate this by calibrating the augmentation, but residual bias can remain. Usual failure modes are a fragment placed at the right position but with a rotation offset, and small or thin fragments that are harder to align. Heavily comminuted cases with many small fragments are the most difficult.