Volumetric functional imaging faces trade-offs between calcium trace quality, imaging speed, and laser power. This paper introduces a supervised deep-denoising method (NIDDL) to address these limitations. NIDDL boasts a small memory footprint, fast training and inference (50-70ms), high accuracy, and generalizability, requiring only small, non-temporally-sequential training datasets (~500 image pairs). The method is demonstrated on whole-brain imaging, large-field-of-view imaging in freely moving animals, and recovering complex neurite structures in C. elegans, enabling faster and longer-term imaging experiments.

Shivesh Chaudhary, Sihoon Moon, Hang Lu