FREQFORMER: Efficient Polyp Segmentation via Wavelet Transform

Xudong Zhou, Tianxiang Chen

Abstract: Polyp segmentation is crucial for detecting colorectal cancer. Despite some progresses, present methods struggle with shape-aware segmentation due to minimal gradient discrepancies between the polyp and background, causing blurry edges and lack efficiency due to feature map redundancy. We analyze the reason for the first challenge is that most methods either only rely on spatial domain features while ignoring frequency domain clues, or ignore interactions between the two domains. Frequency clues are sensitive to the minor gradient discrepancies in spatial domain, which helps shape-aware segmentation. Based on this, we propose Efficient Frequency-Space Interaction Attention (EFSIA) to interact spatial features with high and low frequency clues and further strengthen the interaction in the decoder by devising Fourier Cross Fusion Module (FCFM). To tackle the second challenge, we reduce the feature map redundancy with proposed EFSIA by removing similar feature maps and using efficient wavelet transform to recover them without accuracy drop. Experiments on Kvasir-SEG and CVC-ClinicDB prove the superiority of our method in accuracy and efficiency.