This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
RARE-UNet: Resolution-Aligned Routing Entry for Adaptive Medical Image Segmentation
Created by
Haebom
Author
Simon Winther Albertsen, Hjalte Svaneborg Bj{\o}rnstrup, Mostafa Mehdipour Ghazi
Outline
RARE-UNet is a resolution-aware multi-scale segmentation architecture that performs accurate segmentation without performance degradation even at low-resolution data. Its key design elements include multi-scale blocks integrated at various encoder depths, a resolution-aware routing mechanism, and consistency-based learning to align multi-resolution features with high-resolution representations. On two benchmark brain imaging tasks for hippocampus and tumor segmentation, it achieves the highest average Dice scores (0.84 and 0.65) across all resolutions, compared to the original UNet, multi-resolution augmented UNet, and nnUNet, while maintaining consistent performance and significantly reduced inference time even at low resolutions. This demonstrates the efficiency and scalability of RARE-UNet. The code is available at https://github.com/simonsejse/RARE-UNet .
Takeaways, Limitations
•
Takeaways:
◦
Provides high-accuracy segmentation results even from low-resolution medical image data.
◦
Maintains consistent performance across inputs of varying resolutions.
◦
Inference time is significantly reduced compared to existing models.
◦
It demonstrates applicability to various medical image segmentation tasks.
•
Limitations:
◦
Currently, only evaluations have been performed on brain imaging data, so generalization performance to other types of medical imaging data requires further study.
◦
Hyperparameter tuning may be required to optimize for specific resolutions.
◦
Additional experiments using larger datasets may be needed.
