Our paper "MedPCFM: Improving Medical Point Cloud Completion by Integrating Point Transformers and Flow Matching" (Kamil Kwarciak, Marek Wodziński) was accepted and presented at AI4M3D, the workshop on Artificial Intelligence for Medical 3D Vision, held on September 9, 2026 in Malmö, Sweden as part of ECCV 2026.
Medical point cloud completion supports anatomical reconstruction, but generative approaches to it remain little studied. MedPCFM treats completion as continuous-time generative modeling, introducing PCFM, a flow matching method built on a Point Transformer V3 backbone, evaluated on the SkullFix, SkullBreak, and newer Mandibular Defect datasets. PCFM with PTv3 performs comparably to a deterministic PTv3 baseline while achieving the best generative results across all three datasets, and needs far fewer sampling steps than diffusion-based alternatives, giving up to a 7x throughput gain over a PVCNN backbone. A scaling study across model size and point count also shows consistent gains at higher point resolution.