Abstract
Recent advances in deep learning models such as U-Net Transformer (UNETR) and Swin U-Net Transformer (SwinUNETR) have significantly improved tumor segmentation accuracy and local structural sensitivity in medical imaging. Notably, OrgUNETR model demonstrated that predicting both organs and tumors enhances tumor segmentation performance by leveraging their anatomical relationships. However, those models still struggle to capture the complex and irregular shapes of tumors that result from structural constraints or physiological pressure exerted by surrounding organs. These irregular morphologies often manifest as lobulated, or infiltrative margins, containing high-frequency edge information that is challenging to represent using convolution network or transformer based architectures alone. To address this, we propose FreqDualNet, a frequency-aware extension of SwinUNETR and OrgUNETR, which integrates three types of frequency-awareness modules into the network. These modules leverage spectral information through Fourier-based transformations and high-frequency residual refinement, improving the model’s ability to detect sharp boundaries and shape complexity. Our model showed a performance gain of approximately 12.75% in tumor segmentation accuracy on dataset KiTS19. (0.5072 vs 0.4499) By directly utilizing the morphological differences between organs and tumors, FreqDualNet improves both segmentation accuracy and generalizability, with strong potential to assist tumor diagnosis, prognosis evaluation, and understanding tumor-organ interactions. For code details, you can visit
https://github.com/JungroLee/FreqDualNet.