In September, the CAIMed research center was represented at the international AutoML Conference in New York: Prof. Dr. Steffen Oeltze-Jafra and Prof. Dr. Marius Lindauer (both CAIMed), together with Jannis Becktepe and Leona Hennig, presented their new paper Auto-nnU-Net. The work introduces a novel framework for the automated configuration of AI models in medical image segmentation \u2013 a key field for applying artificial intelligence in healthcare.<\/p>\n\n\n\n
While the well-known nnU-Net framework already automates many aspects of model setup, it remains limited by fixed hyperparameters and design heuristics. Auto-nnU-Net extends this concept into a full AutoML system for medical image segmentation, integrating Hyperparameter Optimization (HPO), Neural Architecture Search (NAS), and Hierarchical NAS (HNAS) to improve model accuracy and efficiency automatically.<\/p>\n\n\n\n
A major innovation is the Regularized PriorBand algorithm, which balances model accuracy and computational cost \u2014 a crucial factor in clinical environments where computing resources are often limited.<\/p>\n\n\n\n
Across diverse datasets from the Medical Segmentation Decathlon, Auto-nnU-Net achieved substantial performance gains: it outperformed nnU-Net on six out of ten datasets while maintaining realistic computational demands.<\/p>\n\n\n\n