BGnet: Background determination for arbitrary point-spread-function images with deep neural nets
Dr. Leonhard Möckl, Stanford University
Leuchs-Russell Auditorium, A.1.500, Staudtstr. 2
Background fluorescence, especially when it exhibits distinct spatial features, is a primary factor for reduced image quality in optical microscopy. This is particularly detrimental when analyzing single molecules for 3D localization microscopy or single-molecule tracking. In my talk, I will introduce BGnet, a deep neural network with U-net-type architecture, as a general background estimation method. BGnet is capable to rapidly determine arbitrary background shapes from images of arbitrary point-spread-functions (PSFs) with excellent accuracy. The resulting background-corrected PSF images, both for simulated and experimental data, lead to a substantial improvement in localization precision. Finally, I will demonstrate that the improved localization precision directly translates to higher quality of super-resolution reconstructions of biological structures.