Physics-based machine learning for subcellular segmentation in living cells

Abstract Segmenting subcellular structures in living cells from fluorescence microscope images is a ground truth (GT)-deficient problem. The microscopes’ three-dimensional blurring function, finite optical resolution due to light diffraction, pixel and the complex morphological manifestations of all contribute GT-hardness. Unsupervised segmentation approaches are quite inaccurate. Therefore, manual relying on heuristics experience remains preferred approach. However, this process tedious, given countless present inside single cell, generating analytics across large population or performing advanced artificial intelligence tasks such as tracking greatly limited. Here we bring modelling deep learning nexus for solving GT-hard problem, improving both accuracy speed segmentation. We introduce simulation-supervision approach empowered by physics-based GT, which presents two advantages. First, GT resolves Second, computational relevant physical aspects assists models compensate, great extent, limitations physics instrument. show extensive results small vesicles mitochondria diverse independent living- fixed-cell datasets. demonstrate adaptability microscopes through transfer learning, illustrate biologically applications automated motion analysis.