Deep learning-based diatom taxonomy on virtual slides

Deep convolutional neural networks are emerging as the state of the art method for supervised classification of images also in the context of taxonomic identification. Different morphologies and imaging technologies applied across organismal groups lead to highly specific image domains, which need c...

Full description

Bibliographic Details
Published in:Scientific Reports
Main Authors: Kloster, Michael, Langenkämper, Daniel, Zurowietz, Martin, Beszteri, Bánk, Nattkemper, Tim W.
Format: Text
Language:English
Published: Nature Publishing Group UK 2020
Subjects:
Online Access:http://www.ncbi.nlm.nih.gov/pmc/articles/PMC7468105/
http://www.ncbi.nlm.nih.gov/pubmed/32879374
https://doi.org/10.1038/s41598-020-71165-w
Description
Summary:Deep convolutional neural networks are emerging as the state of the art method for supervised classification of images also in the context of taxonomic identification. Different morphologies and imaging technologies applied across organismal groups lead to highly specific image domains, which need customization of deep learning solutions. Here we provide an example using deep convolutional neural networks (CNNs) for taxonomic identification of the morphologically diverse microalgal group of diatoms. Using a combination of high-resolution slide scanning microscopy, web-based collaborative image annotation and diatom-tailored image analysis, we assembled a diatom image database from two Southern Ocean expeditions. We use these data to investigate the effect of CNN architecture, background masking, data set size and possible concept drift upon image classification performance. Surprisingly, VGG16, a relatively old network architecture, showed the best performance and generalizing ability on our images. Different from a previous study, we found that background masking slightly improved performance. In general, training only a classifier on top of convolutional layers pre-trained on extensive, but not domain-specific image data showed surprisingly high performance (F1 scores around 97%) with already relatively few (100–300) examples per class, indicating that domain adaptation to a novel taxonomic group can be feasible with a limited investment of effort.