DeepWonder3D skips ill-posed 3D reconstruction, extracts neurons ~10x faster on TB-scale imaging

bravo_abad · x · 2026-09-07

Many scientific pipelines retain intermediate steps only out of tradition—and those steps can be ill-posed. In multiview calcium imaging, 3D volume reconstruction introduces errors from scattering and noise before biological analysis even starts.

DeepWonder3D (Chen et al.) works directly on multiview projections: 3D U-Nets handle denoising, background removal and neuronal extraction, then temporal agreement across views is combined with a physics-based optical model to recover each neuron's 3D position.

On terabyte-scale cortical recordings, it extracts thousands of neurons in hours, running 10x faster than reconstruction-based baselines. The lesson: don't make ML solve a bad intermediate problem.

Original post →

More from Research

Research channel →