Abstract
XNet introduces a wavelet-based X-shaped unified architecture for fully- and semi-supervised biomedical segmentation. So far, however, XNet still faces limitations, including performance degradation when images lack high-frequency (HF) information, underutilization of raw images, and insufficient fusion. To address these issues, we propose XNet v2, a low- and high-frequency complementary model. XNet v2 performs wavelet-based image-level complementary fusion, using fusion results along with raw images as inputs to three different sub-networks to construct consistency loss. Furthermore, we introduce a feature-level fusion module to enhance the transfer of low-frequency (LF) and HF information. XNet v2 achieves state-of-the-art in semi-supervised segmentation while maintaining competitive results in fully-supervised learning. Extensive experiments on three 2D and two 3D datasets demonstrate the effectiveness of XNet v2.
Introduction
In medical imaging, the scarcity of high-quality labeled data makes semi-supervised learning (SSL) a vital area of research. The original XNet architecture leveraged wavelet transforms to decompose images into low-frequency (LF) and high-frequency (HF) components, processing them in parallel branches. However, this approach suffered from three primary drawbacks: dependency on high-frequency details (leading to failure in "blurry" datasets like skin lesions), neglect of the original raw image information, and insufficient feature fusion between branches. XNet v2 addresses these "frequency-blind" limitations to provide a more universal and robust framework for biomedical image segmentation.
Methodology
XNet v2 employs a triple-network architecture comprising a Main network (M), an LF network (L), and an HF network (H).
Image-Level Complementary Fusion
To prevent model degradation in datasets lacking HF information, XNet v2 introduces image-level complementary fusion. Instead of pure components, the LF branch receives $LF + \alpha HF$ and the HF branch receives $HF + \beta LF$. This ensures that the HF branch remains active and stable even when the high-frequency signal is weak, effectively turning it into a specialized edge-enhancer that still understands base shapes.
Triple-Branch Consistency and Fusion
The model utilizes Cross Pseudo Supervision (CPS) across the three branches. The Main network (M) processes the raw image directly, while the L and H branches process the fused frequency components. Predictions from the frequency branches serve as pseudo-labels for the Main branch and vice versa. Furthermore, asymmetric feature fusion modules are introduced: shallow features are fused for HF-Main interaction to capture textures, while deep features are used for LF-Main interaction to capture global semantics.
Experiments and Results
The effectiveness of XNet v2 was evaluated across diverse 2D (GlaS, CREMI, ISIC-2017) and 3D (P-CT, LiTS) datasets, representing a wide variety of imaging modalities.
As shown in the experimental results, XNet v2 consistently outperforms previous state-of-the-art models. On the 3D P-CT dataset, XNet 3D v2 achieved a Dice score of 80.21%, a significant improvement over the original XNet 3D (75.67%) and other models like MC-Net+. In the 2D ISIC-2017 skin lesion dataset—a known challenge for the original XNet—the inclusion of raw image information and complementary fusion allowed XNet v2 to achieve superior results, demonstrating its enhanced universality.
Conclusion
XNet v2 successfully addresses the limitations of its predecessor by integrating raw image data and implementing multi-level frequency fusion. By treating frequency decomposition as a principled form of perturbation and ensuring cross-branch information flow, it achieves state-of-the-art performance in semi-supervised biomedical segmentation. While the triple-branch architecture adds complexity during training, the model remains efficient for inference by utilizing only the Main branch.
Code is available at: https://github.com/Yanfeng-Zhou/XNetv2
Summary updated