Authors:
Dominika Petríková
1
;
2
;
Ivan Cimrák
1
;
2
;
Katarína Tobiášová
3
and
Lukáš Plank
3
Affiliations:
1
Cell-in-fluid Biomedical Modelling & Computations Group, Faculty of Management Science and Informatics, University of Žilina, Slovak Republic
;
2
Research Centre, University of Žilina, Slovak Republic
;
3
Department of Pathology, Jessenius Medical Faculty of Comenius University and University Hospital, Martin, Slovak Republic
Keyword(s):
Hematoxylin, Eosin, Ki67, Clustering, Neural Networks, Digital Pathology.
Abstract:
The Ki67 positive cell score assessed by immunohistochemistry (IHC) is considered a good biomarker of cell proliferation in determining therapeutic protocols. Manual estimation of Ki67 scores has several limitations as it is time consuming and subject to inter-rater variability. Moreover, the IHC staining is not always available. This could potentially be addressed by using neural network models to predict Ki67 scores directly from hematoxylin and eosin (HE) stained tissue. However, neural networks require large well-annotated datasets, the creation of which is often a laborious process requiring the work of experienced pathologists. Such database containing images of HE stained tissue with Ki67 labels is currently not available. In this paper, we propose a semi-automated dataset generation approach to predict Ki67 scores from pairs of HE and IHC slides with minimal assistance from experts. Using a sample of 15 pairs of whole slide images stained by HE and IHC methods, we proposed a
workflow for generating HE patches with Ki67 labels using image analysis methods such as clustering and tissue registration. From the IHC images processed by the aforementioned methods, we estimated the percentage of Ki67 positive cells in each patch. To verify the validity of the proposed approach we automatically assigned Ki67 labels to HE patches from manually annotated HE - Ki67 pairs. To illustrate the potential of neural network for assigning the Ki67 label to HE patches, we trained a neural network model on a sample of three whole slide images, which was able to classify Ki67 positivity ratio of tissue from HE patches into two Ki67 labels.
(More)