WaveSeq: Frequency-Aware Fusion with Sequential State-Space Consistency for Polyp Segmentation

Polyp segmentation in endoscopic images remains challenging because of low-contrast boundaries and subtle texture gradients that differ markedly from those in natural images. Large-scale pretrained segmentation models such as the Segment Anything Model (SAM) provide strong semantic priors but underrepresent the high-frequency boundary cues that are critical for polyp delineation. We propose WaveSeq, a framework that adapts a frozen pretrained encoder through three synergistic components. A parallel wavelet branch decomposes features into frequency sub-bands to recover boundary details suppressed by the encoder, while a gated cross-attention module fuses the wavelet and encoder streams through learnable gates that balance the two signals. Crucially, stacked 2D state-space blocks complement the encoder's global yet unordered attention with directional context propagation along spatial scan paths. Their input-dependent selective mechanism leverages the fused semantic and frequency cues to determine what context to retain in the recurrent hidden state, thereby enforcing coherent predictions along continuous polyp contours. Across five polyp benchmarks, WaveSeq achieves state-of-the-art mDice and mIoU, demonstrating the effectiveness of joint frequency-sequential adaptation.