Abstract: nnU-Net Revisited

Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier‐Hein, Paul F. JägerInformatik 2025 · DOI · OpenAlex

Abstract

Recent advancements in 3D medical image segmentation have been dominated by the introduction of novel architectures, specifically Vision Transformers (ViTs) and State-Space Models like Mamba. These models are frequently presented as superior successors to the traditional Convolutional Neural Network (CNN)-based U-Net. This paper revisits the nnU-Net framework to investigate whether these architectural shifts provide genuine performance benefits. By benchmarking CNN, Transformer, and Mamba-based models under identical conditions, we find that properly scaled CNNs continue to achieve state-of-the-art results. Our analysis suggests that the perceived superiority of newer models often stems from unfair comparisons and a lack of rigorous validation, pointing to a significant "innovation bias" in the field.

Introduction

In the domain of 3D medical image segmentation—critical for tasks such as tumor identification and organ delineation in CT and MRI scans—the nnU-Net framework has long served as the benchmark for performance. Its success demonstrated that a meticulously configured U-Net architecture, combined with automated preprocessing and training strategies, is exceptionally difficult to outperform. Despite this, the research landscape has recently been flooded with claims that newer paradigms, such as self-attention-based Transformers and Selective State Space Models (Mamba), have finally surpassed the U-Net. This study seeks to verify these claims by leveling the playing field and re-evaluating the strengths of "old-school" CNNs.

The Validation Crisis

A central argument of this work is that the medical imaging community is facing a "validation crisis." Many papers claiming architectural breakthroughs rely on shaky empirical foundations. Specifically, new models are often compared against weak or unoptimized baselines, tested on limited datasets, or granted disproportionate computational resources. This creates an illusion of progress where architectural novelty is prioritized over robust, reproducible performance gains.

Methodology: A Level Playing Field

To ensure a fair comparison, we conducted a rigorous benchmarking study across three major architectural paradigms:

  1. CNN-based: Traditional U-Nets utilizing modern blocks like ResNet and ConvNeXt.
  2. Transformer-based: Models incorporating self-attention mechanisms.
  3. Mamba-based: The latest generation of models using Selective State Space layers.

Crucially, all architectures were integrated into the nnU-Net framework. This approach ensures that the "secret sauce" of nnU-Net—its automated data preprocessing, augmentation, and training configurations—is applied uniformly. Furthermore, we focused on scaling, allowing CNN-based models to utilize modern hardware capacities fully, rather than restricting them to the smaller configurations typical of older studies.

Key Results and Insights

Our findings provide a critical perspective on the current state of architectural research:

  • Resilience of CNNs: Despite being viewed as "older" technology, CNN-based U-Nets (particularly those using ResNet or ConvNeXt blocks) remain the top performers for 3D medical segmentation tasks.
  • Framework vs. Architecture: The results indicate that the training framework—how data is handled and how the model is trained—is often a more significant determinant of performance than the specific arrangement of layers.
  • The Impact of Scaling: Much of the reported superiority of Transformers and Mamba models can be attributed to increased model size or better hardware utilization. When CNN-based U-Nets are scaled to comparable levels, the performance gap typically vanishes or reverses in favor of the CNN.
  • Debunking Novelty Claims: Many recent claims of architectural superiority failed to hold up when subjected to the rigorous validation standards of the nnU-Net framework.

Conclusion

The pursuit of architectural novelty for its own sake has led to an "innovation bias" that obscures the true capabilities of existing methods. This study demonstrates that the most effective recipe for state-of-the-art performance remains a CNN-based U-Net integrated into the nnU-Net framework and scaled for modern hardware. We call for the medical imaging community to adopt more stringent validation standards, moving away from the race for novelty and toward robust, reproducible benchmarking that accounts for computational budgets and baseline strength. Scientific progress must be measured by actual performance gains, not just the complexity of the underlying architecture.

Summary updated

BibTeX
@inproceedings{isensee2025abstract,
  author = {Fabian Isensee and Tassilo Wald and Constantin Ulrich and Michael Baumgartner and Saikat Roy and Klaus Maier‐Hein and Paul F. Jäger},
  title = {{Abstract: nnU-Net Revisited}},
  booktitle = {Informatik},
  year = {2025},
  doi = {10.1007/978-3-658-47422-5\_32},
}

Loading related content...

Figure