Classification of visibility in multi-stain microscopy images
Jonathan Ganz, Christof Bertram, Robert Klopfleisch, Samir Jabari, Katharina Breininger, Marc Aubreville
Annotating mitotic figures (MF) in hematoxylin and eosin (H&E) stained slides is an error- prone task that can lead to low inter-rater concordance. Immunohistochemical staining against phospho-histone H3 (PHH3) can lead to higher concordance but, at the same time, to generally higher mitotic figure counts. By annotating MF in PHH3-stained specimen and transferring them to an H&E- re-stained version of the same slide, the high specificity of PHH3 can be used to create high-quality data sets for H&E images. Since considerably more MF can be recognized only in PHH3, this in turn leads to the introduction of label noise. To overcome this problem, we present an attention-based dual-stain classifier which is designed to discriminate MF based on their visibility in H&E. Additionally, we present a data augmentation approach that focuses especially on presenting a large variability of cell pairs to the attention network. The combination of the two methods leads to a weighted accuracy of 0.740 in discriminating H&E-identifiable from non-identifiable MF. Therefore, by automatically discriminating the visibility of MF in H&E slides, PHH3-guided annotation can be used to generate a more reliable ground truth for MF in H&E.
Wednesday 6th July
Poster Session 1.1 - onsite 15:20 - 16:20, virtual 11:00 - 12:00 (UTC+2)