A patient who has lost part of the skull vault needs an implant that fits the defect exactly, and such implants are still drawn by hand in CAD by specialists, which makes each case slow and expensive. The task can be considered as a shape completion problem, so pretrained completion models, and now the first shape completion foundation models, look like a way to remove that manual step. We audited three publicly released pretrained shape completion models, two of them foundation models, applied zero-shot to 239 cranial defects, some carved from public skull datasets and some real post-craniectomy cases. Each output was compared with simply returning the patient's own remaining bone, and tested for whether the predicted material could be placed without intersecting the bone that is still there. On the 210 carved defects none of the three improved on doing nothing, and on the 29 real cases only one did so beyond chance. Yet not one output, from any model in any configuration, was implantable: none managed both to stay clear of the bone the patient still has and to place more material than doing nothing, where a task-specific volumetric model, trained for this task and selected on these cases, was implantable on all but one. The output that most convincingly resembles a skull is no more placeable than the rest, and at these point densities implant thickness cannot be read off the emitted points at all, which is a property of the output representation rather than of any one model. On this evidence the shape completion foundation models we could obtain cannot be trusted and, used as released, are not usable in clinical practice. Substantial research towards trustworthy foundation models for shape completion is still needed.
Figures & tables
Adv.
Pen.
Method
Params
Pretraining corpus
@2 mm
(mm)
Impl.
Ground truth
–
control
+0.802
0.35
239/239
No-op, unchanged
–
control
+0.000
0.35
239/239
Uniform resample
–
allocation-neutral
+0.169
–
–
Task-specific volumetric
–
trained on this task
+0.745
0.55
238/239
MSN-FM [ 25 ]
297 M
MedShapeNet, 100k
−0.156
6.60
0/239
Table 1: Medians over 239 skulls at the common 6144-point budget, deployment frame. Advantage is the per-skull gain in recall of the removed region at 2 mm over a no-op; implantable adds collision-freedom at 1 mm. The controls pass by construction, not by merit, and the task-specific row is ours and test-selected ( Sec. 3 ). 29 of the 239 are real craniectomy defects, separated out in Tab. 2 .
Figure 1: Three pretrained completion models, 239 skulls, equal 6144-point budget. (a) advantage over a no-op at 2 mm; (b) deepest penetration into retained bone, log scale, implantable count above each box. In (b): dotted rule, the 1 mm tolerance; dashed line, the task-specific reference, which is ours and test-selected ( Sec. 3 ); shaded band, the instrument’s floor.
Synthetic ( n=210 )
Clinical ( n=29 )
Clinical p
Model
Adv.
Better
Adv.
Better
Sign
Signed-rank
MSN-FM [ 25 ]
−0.162
0/210
−0.102
1/29
<0.0001
<0.0001
AdaPoinTr [ 26 ]
−0.042
29/210
+0.021
16/29
0.7111
0.0409
MGPC [ 14 ]
−0.139
4/210
+0.118
24/29
0.0005
<0.0001
Table 2: Advantage over a no-op at 2 mm, median per skull, by defect provenance. The clinical p -values are a two-sided sign test, counting AdaPoinTr’s three tied skulls against it, and a Wilcoxon signed-rank test, which drops them; they disagree for AdaPoinTr, whose clinical median is therefore not established. Every synthetic result is significant at p<10−24 under both. MUG500+ ships a manufactured implant, so its columns measure whether material went where the device goes.
Figure 2: Case B0001 of MUG500+, from outside (top) and in section (bottom). The patient’s own bone is grey, what each method adds is in colour, black marks material inside bone that is there. MSN-FM’s addition is granular; AdaPoinTr’s and MGPC’s read as solid caps, though only AdaPoinTr’s points lie on a surface. All three run into bone the patient still has. On this case two of them exceed the no-op; the medians in Tab. 1 do not.
Figure 3: How thin a plate can get before its two surfaces stop being distinguishable, at each point budget in play (ideal even resample, no model, 239 skulls per box). No budget the audited models emit or receive reaches 3 mm, so thickness is lost before a model is asked to recover it. That 3 mm is a design target rather than a law, and the clinical data break it: 11 of the 29 real MUG500+ implants (38 %) are thinner than 3 mm, the 29 having median 3.05 mm and mean 3.30±1.41 mm. The dashed box is AdaPoinTr’s other checkpoint ( Sec. 4 ).
This paper introduces LaS-Comp, a zero-shot and category-agnostic approach that leverages the rich geometric priors of 3D foundation models to enable 3D shape completion across diverse types of partial observations. Our contributions are threefold: First, \ourname{} harnesses these powerful generative priors for completion through a complementary two-stage design: (i) an explicit replacement stage that preserves the partial observation geometry to ensure faithful completion; and (ii) an implicit refinement stage ensures seamless boundaries between the observed and synthesized regions. Second, our framework is training-free and compatible with different 3D foundation models. Third, we introduce Omni-Comp, a comprehensive benchmark combining real-world and synthetic data with diverse and challenging partial patterns, enabling a more thorough and realistic evaluation. Both quantitative and qualitative experiments demonstrate that our approach outperforms previous state-of-the-art approaches. Our code and data will be available at LaS-Comp.
Weilong Yan, Haipeng Li, Hao Xu +4
National University of Singapore · University of Electronic Science and Technology of China · The Chinese University of Hong Kong +1
Recovering the full 3D spine anatomy from intraoperative ultrasound is an ill-posed inverse problem, as the complete structure must be inferred from incomplete and noisy observations. Acoustic occlusions and limited field of view create large unobserved regions, while view-dependent artifacts lead to variability in expert annotations of the visible anatomy. Current supervised ultrasound shape completion methods rely on synthetically generated incomplete-complete paired data to learn conditional mappings under a predefined distribution of simulated occlusions. However, real intraoperative occlusions do not necessarily follow this distribution, which can limit generalization to patient data. As a result, accurate and robust completion from noisy partial observations remains an unsolved problem. We propose a zero-shot shape completion framework that reconstructs the entire lumbar spine from partial ultrasound observations without relying on simulated training data. To accommodate unseen and irregular patterns of missing structures, we introduce blended completion, a mechanism that integrates the learned anatomical prior with incoming partial geometry at inference time. The method learns a generative diffusion prior over full anatomical shapes represented in an adaptive octree structure, enabling efficient modeling of the complete spine in a single forward pass. Validation on phantom and volunteer data shows that decoupling completion from a predefined corruption distribution improves generalisation under real occlusions, outperforming a fully supervised variant by 22% on HD95 completion error. Code and data are available at https://github.com/miruna20/ZODIAC.
Chair for Computer Aided Medical Procedures (CAMP) Technical University of Munich, Boltzmannstr. 3, 85748 Garching, Germany · Munich Center for Machine Learning (MCML), Munich, Germany · Konrad Zuse School of Excellence in Reliable AI (relAI), Germany
Medical point cloud completion is important for anatomical reconstruction and downstream clinical workflows, yet generative modeling in this setting remains insufficiently studied. We investigate completion through continuous-time generative modeling and introduce PCFM, a PTv3-backed flow matching approach for medical point cloud completion. We evaluate on SkullFix and SkullBreak, and additionally on the more recent Mandibular Defect dataset. We build strong baselines by adapting PTv3 to a deterministic encoder-decoder completion model and by instantiating diffusion completion (PCDiff) with both PVCNN and PTv3 denoisers. PCFM with PTv3 is competitive with the deterministic PTv3 baseline and achieves state-of-the-art generative performance across datasets, while requiring substantially fewer sampling steps than diffusion. At the best operating points, PTv3 also yields clear throughput gains, providing up to a 7× speed-up for PCFM compared to a PVCNN backbone. Finally, we study empirical scaling trends by varying model size and point cardinality, showing consistent gains with higher point resolution and informative trade-offs across model scales.
Kamil Kwarciak, Marek Wodzinski
Department of Measurement and Electronics, AGH University of Krakow, Krakow, Poland · Sano Centre for Computational Medicine, Krakow, Poland