Comparative Analysis of Self-Supervised Learning Techniques for Electron Microscopy Images

BIO Web of Conferences, vol. 129, pp. 10037

Abstract

Deep learning has revolutionized a wide array of tasks across differentdomains, including electron microscopy (EM) image analysis, by leveraginglarge labeled datasets for training. However, the scarcity of such labeleddatasets in EM necessitates the exploration of alternative methods. Self-supervised learning (SSL) emerges as a promising approach to leverageunlabeled data, featuring techniques such as, e.g., masked image modeling(MIM) — which predicts missing parts of the input data, as well as contrastivelearning — which learns by distinguishing between similar and dissimilar pairsof data. This study aims to investigate the impact of these SSL techniques onEM images, providing a case study on the effectiveness of leveragingunlabeled data in a domain where labeled datasets are limited and expensiveto create.

Authors 2

1. Bashir Kazimi

   Forschungszentrum Jülich

   Affiliation as printed

   Forschungszentrum Jülich, Institute for Advanced Simulation – Materials Data Science and Informatics (IAS-9), Aachen, Germany

2. Stefan Sandfeld Aachen Chair of Materials Data Science and Informatics

   RWTH Aachen University · Forschungszentrum Jülich

   Affiliation as printed

   Forschungszentrum Jülich, Institute for Advanced Simulation – Materials Data Science and Informatics (IAS-9), Aachen, Germany

   RWTH Aachen University, Chair of Materials Data Science and Informatics, Aachen, Germany