nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation (1809.10486v1)

Abstract: The U-Net was presented in 2015. With its straight-forward and successful architecture it quickly evolved to a commonly used benchmark in medical image segmentation. The adaptation of the U-Net to novel problems, however, comprises several degrees of freedom regarding the exact architecture, preprocessing, training and inference. These choices are not independent of each other and substantially impact the overall performance. The present paper introduces the nnU-Net ('no-new-Net'), which refers to a robust and self-adapting framework on the basis of 2D and 3D vanilla U-Nets. We argue the strong case for taking away superfluous bells and whistles of many proposed network designs and instead focus on the remaining aspects that make out the performance and generalizability of a method. We evaluate the nnU-Net in the context of the Medical Segmentation Decathlon challenge, which measures segmentation performance in ten disciplines comprising distinct entities, image modalities, image geometries and dataset sizes, with no manual adjustments between datasets allowed. At the time of manuscript submission, nnU-Net achieves the highest mean dice scores across all classes and seven phase 1 tasks (except class 1 in BrainTumour) in the online leaderboard of the challenge.

• The paper presents a self-configuring approach that eliminates the need for manual tuning of U-Net parameters across diverse medical imaging datasets.
• It leverages dynamic 2D/3D architectures and a U-Net Cascade to optimize data preprocessing, training protocols, and inference strategies.
• Results from the Medical Segmentation Decathlon demonstrate superior Dice scores and enhanced generalizability across various imaging modalities.

nnU-Net: Self-Adapting Framework for U-Net-Based Medical Image Segmentation

The paper "nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation" presents an advanced framework that refines the application of U-Net architectures in the domain of medical image segmentation. This work addresses the common challenge of needing extensive, problem-specific architectural modifications for effective segmentation, as highlighted by the diverse requirements across medical imaging datasets.

Background and Motivation

The U-Net architecture is a prominent tool in medical image segmentation, lauded for its simplicity and efficiency since its introduction in 2015. However, tailoring U-Nets to new datasets requires handling various parameters, such as architectural specifics and pre-processing steps, which significantly influence performance. This paper introduces the nnU-Net framework, which aims to auto-configure these parameters, minimizing the need for manual adjustments and improving generalizability across different tasks.

Methodology

nnU-Net operates on a foundation of 2D and 3D U-Nets and a U-Net Cascade. It discards recent architectural innovations like residual or dense connections, shown to be sometimes overfitted to limited scenarios. Instead, it emphasizes optimizing non-architectural factors, such as data pre-processing, network training protocols, and inference strategies.

Network Architecture

The nnU-Net encompasses:

• 2D and 3D U-Nets: These versions adjust dynamically based on dataset characteristics. The 2D U-Net might be suboptimal for 3D data but provides robustness against anisotropic resolutions, as seen in several datasets.
• U-Net Cascade: This model generates coarse segmentations that are further refined, addressing performance issues in standard 3D U-Nets when processing large images due to GPU memory limits.

Dynamic Configuration

The framework automatically determines appropriate network configurations by analyzing dataset attributes like median image shape. For example, input patch sizes and pooling strategies are adapted to work within hardware constraints while preserving spatial context.

Preprocessing and Training

nnU-Net incorporates:

• Resampling: Aligning images to a common resolution to facilitate network learning of spatial semantics.
• Normalization: Differentiating strategies for CT and MRI images to maintain intensity consistency.
• Data Augmentation: Applying transformations like elastic deformations, which are consistent across datasets to enhance training robustness.

Results

Tested in the Medical Segmentation Decathlon challenge, nnU-Net achieved the highest mean Dice scores across all classes in many datasets. It particularly excelled in adapting to varying entities and modalities without manual interventions.

Implications and Future Work

The results underscore nnU-Net's capability to generalize across diverse datasets, an essential feature for scalable medical imaging solutions. The research highlights the underestimated impact of non-architectural optimizations compared to structural network modifications.

Future avenues could include refining model selection heuristics to predict the best-fit architecture before training commences, thus speeding up deployment. There's also scope for further ablation studies on components like activation functions and augmentation parameters, enhancing the robustness and adaptability of the nnU-Net framework.

In conclusion, this research advances the discourse on medical image segmentation by presenting a comprehensive, self-configuring, and adaptable framework that efficiently leverages the foundational strengths of U-Net architectures.