Master's Thesis : Cell segmentation in whole-slide cytological images

Abstract

The digital pathology is a field of medicine that leverage computer aided techniques to view, manage and analyse microscope acquired images, commonly called whole-slide images. The field has been rapidly evolving in the recent years thanks to novel research around AI based techniques for automated diagnosis. A popular sub-field of the automated diagnosis is the segmentation (or detection) of relevant regions in a whole-slide image.

This thesis deals with the problem of cell detection using segmentation in whole-slide cytological images as part of an automated diagnosis system for the thyroid cancer. The work is conducted within Cytomine R&D project team from ULiège. Among the challenges is the implementation of a segmentation algorithm using novel deep learning methods while dealing with incomplete training data.

The proposed solution comprises of a U-Net network for the segmentation along with an iterative data improvement method for incomplete data completion. The implementation also achieves the required level of modularity and scalability for the subsequent integration in the ULiège Cytomine instance which was almost complete.

Promising results were obtained thus demonstrating the abilities of the U-Net and the data improvement method. However, inaccuracies remained mainly due to false positives and all available data weren't used because their incompleteness couldn't be fully addressed. Further work should enhance the training set building, for instance using active learning, and especially pay attention to the diversity and completeness of the data used.