Combining frequency transformer and CNNs for medical image segmentation

Image segmentation is one of the most challenging and difficult tasks in digital image processing. It has many medical applications such as cancerous tumors segmentation, organ segmentation, or abnormalities segmentation. Recent techniques combining convolution-based models and transformers are proposed for automatic medical segmentation tasks. These techniques achieve good results but require much time and resources. In this paper, we propose a new model to segment medical images which combines CNNs and frequency transformers in a parallel way to minimize the number of parameters and to reduce computation time. This work presents a powerful model, composed of two main branches, able to learn global-local feature interactions which are currently in a medical image. The first branch based on Frequency Transformer (FT) employs Fourier Transform instead of multi-head attention to capture global dependencies. While a no-deeper convolutional neural network (CNN) is employed to get rich local information. With a small number of parameters, the proposed model was tested on many public medical image databases and achieves state-of-the-art results for lesion/tumor segmentation tasks.