Point‐Implicit Geometry Learning Enables Resource‐Efficient 4D Intraoperative OCT

ABSTRACT Four‐dimensional (4D) intraoperative optical coherence tomography (iOCT) provides volumetric visualization of surgical tool–tissue interactions, but its clinical use remains constrained by costly swept‐source hardware and memory‐intensive voxel reconstructions. Here, we identify a geometry‐limited imaging regime in which surface curvature and noise, rather than sampling density alone, primarily determine the achievable fidelity of volumetric reconstruction. We introduce Point‐Implicit Geometry Learning (PIGL), a framework that represents sparsely sampled OCT data as oriented point sets and reconstructs continuous surfaces through a differentiable implicit field. This representation follows a curvature‐ and noise‐aware sampling criterion, enabling stable surface recovery at sampling pitches up to four coarser than the conservative optical‐Nyquist pitch, while corresponding to approximately a twofold relaxation relative to the practical sampling requirement of voxel‐centric reconstruction in the same scenario. Implemented on a 20‐kHz spectral‐domain OCT system, PIGL achieves interactive 4D updates (15.6 Hz) with 35 lower GPU memory than voxel‐centric baselines, while preserving temporally coherent instrument–tissue geometry. By demonstrating the feasibility of geometry‐aware reconstruction on accessible OCT hardware, this work offers a practical pathway toward resource‐efficient 4D intraoperative OCT.