Abstract
Medical Image Segmentation (MIS) includes diverse tasks, from bone to organsegmentation, each with its own challenges in finding the best segmentationmodel. The state-of-the-art AutoML-related MIS-framework nnU-Net automates manyaspects of model configuration but remains constrained by fixed hyperparametersand heuristic design choices. As a full-AutoML framework for MIS, we proposeAuto-nnU-Net, a novel nnU-Net variant enabling hyperparameter optimization(HPO), neural architecture search (NAS), and hierarchical NAS (HNAS).Additionally, we propose Regularized PriorBand to balance model accuracy withthe computational resources required for training, addressing the resourceconstraints often faced in real-world medical settings that limit thefeasibility of extensive training procedures. We evaluate our approach acrossdiverse MIS datasets from the well-established Medical Segmentation Decathlon,analyzing the impact of AutoML techniques on segmentation performance,computational efficiency, and model design choices. The results demonstratethat our AutoML approach substantially improves the segmentation performance ofnnU-Net on 6 out of 10 datasets and is on par on the other datasets whilemaintaining practical resource requirements. Our code is available athttps://github.com/automl/AutoNNUnet.