Abstract
Deep learning models for medical image segmentation often struggle with domain shift, where performance drops when applied to data from different scanners or clinical sites. This study investigates the robustness of the nnU-Net framework in the context of the Multi-Centre, Multi-Vendor & Multi-Disease Cardiac Image Segmentation (M&Ms) Challenge. We demonstrate that combining heavy data augmentation with Batch Normalization significantly improves domain generalization. Our approach achieved 1st place in the challenge, showing that well-engineered standard architectures can outperform complex domain adaptation methods.
Introduction
In clinical practice, semantic segmentation models must handle variations in image appearance caused by different MRI scanner vendors and imaging protocols. This "domain shift" often leads to brittle performance. While specialized domain adaptation techniques exist, they frequently require unlabeled target data or complex training procedures. This work explores whether simple, customizable pipeline components within the nnU-Net framework can achieve high generalizability across diverse cardiac MRI datasets.
Methods
The proposed solution leverages the nnU-Net framework with key modifications:
- Heavy Data Augmentation (DA): Beyond standard rotations, we implemented aggressive variations in brightness, contrast, Gaussian noise, and simulated low resolution. Crucially, elastic deformations and gamma transformations were used to simulate the stylistic variations of different scanners.
- Normalization Layers: We compared standard Instance Normalization (IN) with Batch Normalization (BN), hypothesizing that BN better captures features invariant to single-image styles when paired with diverse data.
- Ensembling: Stability was maximized using an ensemble of five 2D and five 3D nnU-Net models.
Experimental Setup
The study used the M&Ms Challenge dataset. Training data consisted of 150 annotated cases from Vendor A (Siemens) and Vendor B (Philips). Testing was performed on 200 patients, including "unseen" domains from Vendor C (GE) and Vendor D (Canon). Evaluation focused on the Left Ventricular blood pool (LV), Left Ventricular Myocardium (LVM), and Right Ventricular blood pool (RV).
Results and Discussion
Our method ranked first in the M&Ms Challenge, maintaining high Dice scores even on unseen vendors (C and D), with performance drops limited to approximately 1-2%.
The Impact of Normalization and Augmentation
Experimental results showed that Batch Normalization combined with "Heavy DA" significantly outperformed the baseline nnU-Net configuration. Interestingly, we observed an asymmetry in domain shift: models trained on Vendor A generalized better to Vendor B than vice versa.
Data Saturation and Generalization
We investigated how much target domain data is needed to bridge the gap. Performance gains were found to saturate at approximately 30% of the target domain data (roughly 15 patients), suggesting that massive target datasets are not always necessary for effective adaptation.
Conclusion
This study demonstrates that robust domain generalization in cardiac MRI segmentation can be achieved through strategic engineering of data augmentation and normalization layers. The success of this approach in the M&Ms Challenge highlights the effectiveness of well-tuned, standard architectures over more complex, specialized domain adaptation methods. Future work will focus on understanding the underlying causes of domain-specific performance variations.
Summary updated