Keywords: Transmission electron microscopy, Mitochondria segmentation, Intensity correction, Intensity standardization
TL;DR: In this work we developed and analyzed intensity correction and standardization methods for training a neural network on electron microscopy data
Abstract: Intensity of acquired electron microscopy data is subjected to large variability due to the interplay of many different factors, such as microscope and camera settings used for data acquisition, sample thickness, specimen staining protocol and more. In this work, we developed an efficient method for performing intensity inhomogeneity correction on a single set of combined transmission electron microscopy (TEM) images and demonstrated its positive impact on training a neural network on these data. In addition, we investigated what impact different intensity standardization methods have on the training performance, both for data originating from a single source as well as from several different sources. As a concrete example, we considered the problem of segmenting mitochondria from EM data and demonstrated that we were able to obtain promising results when training our network on a large array of highly-variable in-house TEM data.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: both
Primary Subject Area: Segmentation
Secondary Subject Area: Validation Study
Source Latex: zip