ECCV 2020 Daily - Thursday

The same philosophy of leveraging intrinsic statistics in bioimages to compensate for missing ground-truth annotations is taken to the extreme when it comes to Cryo-EM. Here, the task is to infer the structure of extremely small particles (e.g., individual proteins) from many noisy electron microscopy (EM) projections. In Multi-CryoGAN: Reconstruction of Continuous Conformations in Cryo-EM Using Generative Adversarial Networks , Gupta et al . use a GAN to transfer the classical pose and conformation estimation into a distribution matching problem: A distribution of latent variables is learnt that represents pose and conformation of the imaged particles. Synthetic 2D images are then produced through a forward model to produce realistic projections that are indistinguishable to a discriminator from real measurements . The whole model is trained end-to-end in an unsupervised manner . Being in the intersection between computer vision, machine learning, and life science, contributions to the field of BioImage Computing are not just academic proof of concepts: successful methods are not only improving over the previous state of the art, but are also easily usable by experimentalists. This point was clearly made by several invited speakers: 2 Workshop 2 DAILY T h u r s d a y Wei Ouyang from the KTH Royal Institute of Technology presented (amongst many other topics) ImJoy , a computational platform for the deployment of deep learning solutions . With ImJoy, experimentalists have access to state-

RkJQdWJsaXNoZXIy NTc3NzU=