Show abstract - PDF - Reviews

Image segmentation is at the core of many tasks in medical imaging, including the engineering of computational biomarkers. While the self-configuring nnUNet framework for image segmentation tasks completely shifted the state-of-the-art in the radiology field, it has never been adapted to overcome its limitations for application on the pathology domain. Our study showcases the potential of nnUNet in computational pathology and bridges the gap that currently exists in utilizing nnUNet for pathology applications. Our proposed nnUNet for pathology framework has demonstrated its significance and potential to shift the state-of-the-art in the computational pathology field, as seen from the exceptional first-place segmentation ranking on the TIGER challenge\'s experimental leaderboard 1. Our framework includes critical hyperparameter adjustments and pathology-specific color augmentations, as well as an essential WSI inference pipeline and a novel inference uncertainty approach that proves helpful for biomarker development. We release the code of our accurate and workflow-friendly segmentation tool to promote and foster growth within the computational pathology community.
Hide abstract