Advanced AI tools for cell and tissue segmentation of multiparametric histology images

High-dimensional imaging, especially Imaging Mass Cytometry (IMC), has greatly enhanced analytical capabilities of cellular and tissue exploration. Cell segmentation, a crucial step for such analysis, has concomitantly improved with the development of deep learning tools, which typically require extensive annotated datasets to reach a high level of efficacy. In this thesis, we present a new cell segmentation tool based on a Bayesian Convolutional Neural Network capable of estimating output uncertainty. We also introduce an "offline" active learning framework leveraging uncertainty to enhance the efficiency of the annotation process, allowing for improved model training with fewer annotated images. The model and framework are tested on an independent IMC dataset, composed of multiple proprietary and public images, obtained from heterogeneous tissues. Our results demonstrate that the uncertainty estimates help to identify segmentation errors, allowing manual correction. Employing these estimates in our active learning framework allows for efficient and guided training of the model on selected images, achieving results comparable to those of models trained on a considerably higher number of images. Moreover, our Bayesian model surpasses a similar non-Bayesian baseline on our IMC dataset, showcasing the advantages of Bayesian networks on small datasets. The proposed tool holds promise for biomedical imaging analysis, offering a practical solution to the annotated training dataset bottleneck while enhancing model reliability and applicability across small but varied datasets.