Focal-Unet: Unet-like Focal Modulation for Medical Image Segmentation

Medical image segmentation plays a critical role in healthcare, and transformer-based U-shaped architectures have emerged as powerful alternatives to traditional CNN-based models. However, these transformer models still face persistent challenges, such as blockiness and cropped edges in segmentation masks, due to the patch partitioning operation inherent in their design. To address these issues, we propose Focal-UNet, a novel architecture that integrates a focal modulation mechanism to improve segmentation performance. Using an asymmetric encoder-decoder structure, Focal-UNet aggregates local and global features effectively. This design leverages the broad receptive field of transformers while maintaining the local feature emphasis characteristic of CNNs, striking a crucial balance between global and local information. Our results demonstrate that Focal-UNet outperforms SwinUNet, one of the most advanced transformer-based architectures for medical image segmentation, achieving a 1.68% improvement in the DICE score and a 0.89% reduction in Hausdorff Distance on the Synapse dataset. This performance gain is particularly impressive, as Swin-UNet is widely regarded as a state-of-the-art model. Additionally, Focal-UNet demonstrates its robustness in low-data settings, achieving a 4.25% higher DICE score on the NeoPolyp dataset, further showcasing its potential for real-world applications where annotated data is limited.