BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Weakly supervised deep learning methods for biomicroscopy - Thesis
  defense
DTSTART:20210324T180000
DTSTAMP:20260407T111516Z
UID:3c03233c2b9c444000578144758607fd1b5b663cef274de49b36bdf7
CATEGORIES:Miscellaneous
DESCRIPTION:Adrian Shajkofci\nDear colleagues\,\nYou are invited to my pub
 lic thesis defense "Weakly supervised deep learning methods for biomicrosc
 opy" that will take place on Zoom. The event will take place on :\n\n24 Ma
 rch 2021 6.00pm\nhttps://epfl.zoom.us/j/83689334674\nID de réunion : 836 
 8933 4674\n\nThe presentation will be in both french and english\, dependi
 ng on the attendance.\nThe abstract is at the end of the email !\n\nAbstra
 ct :\nOptical microscopy\, an invaluable tool in biology and medicine to o
 bserve and quantify cellular function\, organ development\, or disease mec
 hanisms\, requires constant trade-offs between spatial\, temporal\, and sp
 ectral resolution\, invasiveness\, acquisition time\, and postprocessing e
 ffort. Deep learning technologies have enabled multiple applications that 
 are transforming our day-to-day routines\, including the way we approach m
 icroscopy. Yet despite the ever-increasing computational power\, it is oft
 en the lack of labeled training data that is the limiting factor for wide 
 adoption in this domain. Annotating data is often a lengthy and expensive 
 task\, since it involves tedious work\, generally by skilled experts.\n\nI
 n this thesis\, I explored “weakly supervised” learning methods target
 ed at a variety of applications to enhance microscopy images and extract p
 hysical information from a single image. The specificity of these “weakl
 y supervised” methods is the fact that they use very little prior inform
 ation about the image in order to keep the effort to annotate training dat
 a as low as possible. Specifically\, I reduced the dimensionality of the l
 earning problem by targeting the experiment towards estimating the paramet
 ers of a spatially-variant Point Spread Function (PSF) model using a Convo
 lutional Neural Network (CNN)\, which does not require instrument or objec
 t-specific calibration. Using such a model permitted to simulate realistic
 ally accurate training data that could be generalized\, once the model was
  trained\, to real microscopy images. I extensively benchmarked different 
 network architectures\, training datasets and simulation modalities toward
 s the optimal PSF prediction performance and robustness to image degradati
 on.\n\nStarting from the estimated PSF model parameters\, I developed a va
 riety of applications\, such as a semi-blind spatially-variant deconvoluti
 on method for image deblurring and enhancement\, a robust and fast microsc
 opy auto-focus\, a method for the estimation of the object surface from a 
 single 2D image\, and a method for the estimation of the object velocity i
 n a fluid\, all of them with minimal need for a priori knowledge about the
  optical setup.\n 
LOCATION:https://epfl.zoom.us/j/83689334674
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR