Frequency-Based Hybrid Method Bridging CNN and Transformer for Medical Image Segmentation

Various CNN-based approaches have been proposed to automatically generate promising segmentation results in clinical diagnosis, but these models tend to have a bias towards local feature extraction, limiting their ability to capture high-level global information such as shape. Additionally, CNN predictions are often biased towards textural information at the expense of other local information, which is an influential yet undervalued point requiring separate processing. This paper presents a new approach to address these biases with a hybrid CNN-Transformer architecture for medical image segmentation. The proposed method effectively utilizes multi-level representations of the transformer and CNN to preserve both global and local senses, and introduces a novel_Frequency Transform (FT) module to tackle texture bias. The FT module elegantly bridges the transformer and CNN, transforming multi-level features into an aggregated frequency domain rich in both high-level shape information and low-level local information, thereby reducing bias towards textural details. Experimental results show the effectiveness of the proposed framework. Experimental results demonstrate that our proposed framework can achieve state-of-the-art segmentation results on public medical image datasets.