Learned differentiable Poisson reconstruction enables 4D intraoperative OCT on low-speed systems

Affordable four-dimensional intraoperative optical coherence tomography (4D iOCT) is hindered by reliance on costly swept-source OCT systems and memory-intensive voxel-based frameworks. We present Learned Differentiable Poisson Reconstruction (LDPR), a geometry-aware framework that converts raw B-scans into sparse point clouds, extracts features via a point-voxel dual representation enabling fast spatial access with minimal memory overhead, and reconstructs intraoperative scenes from generated point-normal pairs using a differentiable Poisson solver and GPUaccelerated Marching Cubes. Running on a 20kHz spectral-domain OCT platform, LDPR maintains peak GPU memory below 1.5GB and delivers real-time 4D visualization at 15.6Hz while preserving fine instrument and tissue details. Experiments with retinal scissors, forceps, and a flex loop confirm continuous, speckle-robust reconstructions. By unifying sparse sensing with differentiable geometry processing, LDPR overcomes the memory, noise, and continuity limitations of voxel-based methods and enables 4D iOCT in resource-limited surgical settings.