Registration and Segmentation of Double Stained Histological Images | CESCG 2023 Posters

Our work aims to pre-train a neural network on a pre-text task in an unsupervised fashion using histopathology data and further, using only a few annotated regions of interest from whole slide images, train the aforementioned network to perform downstream tasks, such as classification of patches or segmentation of chosen biological structures. To enhance the reliability of the model’s predictions and to better understand which features influenced the decision, the interpretability method is proposed.
We propose a self-supervised method, using unique double-stained (hematoxylin & eosin and immunohistochemistry p63 staining) dataset of tissue taken from the breast. Our solution is composed of two stages. In the first phase, the autoencoder is trained to roughly determine the presence of myoepithelial cells from H&E, which are clearly visible on p63 stained images as brown cells on a mostly blue background. To do so, pseudo segmentation masks are obtained by computing the Euclidean distance of colour of a myoepithelial cell in L*a*b space with the rest of the p63 stained image. This information is valuable itself, as the aforementioned cells are not easy to determine from H&E staining and p63 staining is much more expensive and not as usual as H&E. Channel-wise occlusion in multiple colour spaces is performed to inspect the colour and structural dependencies of the model’s predictions. The encoder part of the trained autoencoder is taken and fine-tuned on annotated images in the second stage. We assume that utilizing the pre-trained encoder reduces the network’s requirement regarding the number of annotated images.
The student research has been presented in form of a poster by Nina Masaryková from FIIT STU Bratislava.