MedPCFM: Improving Medical Point Cloud Completion by Integrating Point Transformers and Flow Matching

Medical point cloud completion is critical for reconstructing anatomy damaged by trauma, tumor resection, or congenital defects. While generative models like diffusion have shown promise, they are often computationally expensive. This paper investigates whether flow matching, combined with the high-performance Point Transformer V3 (PTv3) architecture, can provide a more efficient and accurate approach to completing missing medical point cloud data.

The authors propose PCFM (Point Cloud Flow Matching), a generative model that treats the completion of missing anatomy as a continuous-time transport process from a simple noise distribution to the target geometry. The model uses PTv3 as a backbone to process point sets efficiently via serialized attention. The researchers compare PCFM against deterministic encoder-decoder baselines and diffusion-based models (PCDiff) across three medical datasets: SkullFix, SkullBreak, and a Mandibular Defect dataset. They also conduct a systematic scaling study to evaluate how model size and point density affect reconstruction accuracy.

PCFM achieves state-of-the-art generative performance across all tested datasets, outperforming existing diffusion-based methods in both accuracy and computational efficiency. By replacing traditional PVCNN backbones with PTv3, the authors report up to a 7x speed-up in throughput. The study also identifies clear scaling trends, showing that increasing point cardinality consistently improves reconstruction quality, whereas increasing model capacity beyond a certain point yields diminishing returns.

This work bridges the gap between high-resolution point cloud modeling and efficient generative completion in clinical settings. By reducing the number of sampling steps required for high-quality reconstruction, PCFM makes generative modeling more practical for time-sensitive surgical planning and patient-specific implant design, where both geometric precision and computational speed are essential.