Generating annotation data of tissue images

Abstract

Currently, there is interest in applying machine learning techniques to analyse digital pathology images automatically. Machine learning techniques often rely on training with a ground-truth image input. The quality and amount of training data determines the quality of the detector, as expressed in the rate of true and false positives, and robustness against variations in the appearance of the input images. The present application proposes to obtain image data of the same sample before and after at least one re-staining step (firstly with a structure-revealing stain, and secondly with a bio marker revealing stain). Sections of the first and second image data having a good registration relationship are chosen, along with the probability of detecting a desired candidate object (such as nucleus) and the probability of the bio marker revealing stain being present annotation data suitable for training a machine learning algorithm on the first and/or the second image data is provided.