AI-assisted annotation of large and multimodal imaging datasets

Abstract

The annotation of histological images through different stains is an important task for diagnosis of diseases such as cancer, but it is also very time-consuming. Despite its repetitive nature, doing such annotations for a same tissue in several stains still requires specialized skills to be done. Nevertheless, the usage of computer vision and machine learning techniques may be used to reduce the time needed to perform this task. This thesis will try to reduce the time needed by developing methods allowing to use the annotation from one tissue image and transfer it to its other modalities (i.e. images with other stains). The annotations considered are freehand polygons, delimiting the area of interest, and up to 25 stains per tissue are available in the dataset used. To achieve such a transfer of annotation, several global feature-based and pixel-based registrations of the whole images are compared. Afterwards, local registrations are performed following the global ones to enhance the results. Those local adjustments are done either using a second time the best techniques from the global registration or by performing a local segmentation using a deep neural network. From those techniques, only the featurebased methods lead to honorable results, with a second local adjustment achieving either a much better or much worse registration. Even though the performances are not sufficient to reliably perform the fully automatic transfer of annotations, the feature-based methods may be used to give an estimate and reduce the interaction required from the annotator.