nnUNet meets pathology: bridging the gap for application to whole-slide images and computational biomarkers

J. Spronck, T. Gelton, L. van Eekelen, J. Bogaerts, L. Tessier, M. van Rijthoven, L. van der Woude, M. van den Heuvel, W. Theelen, J. van der Laak and F. Ciompi

Medical Imaging with Deep Learning 2023.


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 applied to the field of histopathology, likely due to inherent limitations that nnUNet has when applied to the pathology domain "off-the-shelf". We introduce nnUNet for pathology, built upon the original nnUNet, and engineered domain-specific solutions to bridge the gap between radiology and pathology. Our framework includes critical hyperparameter adjustments and pathology-specific color augmentations, as well as an essential whole-slide image inference pipeline. We developed and validated our approach on non-small cell lung cancer data, showing the effectiveness of nnUNet for pathology over default nnUNet settings, and achieved the first position on the experimental segmentation task of the TIGER challenge on breast cancer data when using our pipeline "out-of-the-box". We also introduce a novel inference uncertainty approach, which proved helpful for the quantification of the tumor-infiltrating lymphocytes biomarker in non-small cell lung cancer biopsies of patients treated with immunotherapy. We coded our framework as a workflow-friendly segmentation tool and made it publicly available.