Abstract
Deep learning models for medical image segmentation often struggle with domain shift when deployed across different clinical sites. FAN-Net addresses this by adaptively normalizing the low-frequency amplitude components of the Fourier spectrum, which primarily encode style information, while preserving structural phase components. Combined with a domain classifier and gradient adversarial training, the model achieves state-of-the-art performance on the ATLAS stroke lesion dataset, demonstrating superior generalization to unseen clinical domains.
Introduction
The clinical utility of automated stroke lesion segmentation is often hindered by "domain shift"—performance degradation when a model trained on data from one hospital is applied to another. This shift stems from variations in MRI hardware, protocols, and imaging settings, which alter the visual "style" (contrast, brightness, noise) of the scans. Traditional spatial-domain normalization methods often fail to preserve the subtle semantic details necessary for identifying complex stroke lesions. FAN-Net addresses this by operating in the frequency domain, isolating style from anatomy.
Methods
The proposed Fourier-based Adaptive Normalization Network (FAN-Net) builds upon a standard U-Net architecture by integrating two key components designed to decouple anatomical structure from site-specific style.
Fourier-based Adaptive Normalization (FAN) Module
The FAN module operates in the frequency domain, leveraging the property that the amplitude component of a Fourier-transformed image contains low-level style statistics, while the phase component encodes high-level semantics like edges and shapes. The module adaptively scales and shifts ($\gamma$ and $\beta$) the low-frequency amplitude components to produce a domain-invariant representation.
Domain Adversarial Training
To further ensure domain invariance, a Domain Classifier (DC) branch is attached. Using a Gradient Reversal Layer (GRL), the network is trained to maximize segmentation accuracy while simultaneously minimizing the classifier's ability to identify the source site. This adversarial process effectively strips domain-specific "fingerprints" from the features fed into the U-Net.
Experimental Setup
The model was evaluated using the ATLAS (Anatomical Tracings of Lesions After Stroke) v1.2 dataset, comprising 229 T1-weighted MR images from 9 different clinical sites. A "leave-one-site-out" cross-validation protocol was employed to simulate real-world deployment to unseen clinics.
Results and Discussion
FAN-Net achieved a state-of-the-art Dice score of 0.5591, outperforming standard U-Net (0.4712) and nnU-Net (0.5047). Notably, the model demonstrated a high Recall (0.5762), indicating its effectiveness in detecting small or subtle lesions that baseline models frequently overlook.
Ablation studies confirmed that the combination of FAN and DC provides the best performance, with a Dice score of 0.5098 on specific test splits compared to 0.4396 for the baseline. Furthermore, sensitivity analysis showed that adjusting only a small portion of the low-frequency spectrum ($\alpha = 0.1$) is optimal; excessive modification of high-frequency components leads to the loss of critical tissue boundary information.
Conclusion
FAN-Net provides a lightweight and effective solution for cross-domain medical image segmentation. By isolating style in the frequency domain, it preserves anatomical integrity while achieving robust generalization. Future work will explore multi-modal MRI data and more complex non-linear normalization techniques in the Fourier space.
Summary updated