#30 · ML / AI methods

Simulation and self-supervision drive deep learning for noisy microscopy data

The common thread is machine learning applied to microscopy data where acquisition quality or missing ground truth limits analysis. TomoSegNet trains a cryo-electron tomography membrane segmentation model solely on data simulated from a biophysical membrane model. A self-supervised neural field jointly estimates spatially variant optical aberrations and reconstructs 3D structure from a single blurred image. A paired high/low-SNR fluorescence dataset of 324 images across 12 sub-datasets benchmarks three deep denoising models and a tile-stitching approach for extreme low-SNR restoration. The Cell Behavior Video Classification Challenge separately benchmarks 35 computer vision methods on time-lapse videos, comparing tracking-based features with end-to-end deep models. What is new is the reliance on simulation and self-supervision to supply training signal in place of experimental annotation, alongside standardized benchmarking of competing approaches.