FDDA-Unet: A Frequency-Domain-based Lightweight Network Used for Medical Image Segmentation

Enhancing segmentation performance is the goal of several medical image segmentation techniques. However, these methods often come with a significant increase in the number of parameters and computer complexity, which limits their practical application, especially on edge devices. To address this issue, a lightweight segmentation network, namely Frequency-Domain Dual Attention U-net (FDDA-Unet), was proposed. The proposed network was mainly composed of a Frequency-Domain Dual Attention (FDDA) Block and a Double Depthwise Separable Convolution (DDSC) Block. The proposed model can acquire frequency-domain mixed attention thanks to the FDDA Block, which can enhance feature extraction capabilities and lead to improved model performance in the lightweight model. In addition, the DDSC Block located at the decoder of the model also can reduce the parameters and computations of the network since it processes the size of the feature map and its channels separately. The experimental results demonstrated that the proposed FDDA-Unet outperformed other segmentation models in terms of performance with fewer parameters, computations, and Dice similarity coefficient. Consequently, the FDDA Block and DDSC Block have been employed in the proposed model that achieves competitive results in medical image segmentation tasks.