M3FPolypSegNet++: Adaptive Gaussian frequency decomposition with multi-task learning for polyp segmentation

Colorectal cancer remains one of the most prevalent and lethal malignancies globally, primarily due to challenges in timely detection and removal of polyps. Recent CNN- and Transformer-based frameworks have already reported encouraging polyp segmentation accuracy on public benchmarks. Yet no prior approach provides an adaptive, computationally efficient mechanism that jointly leverages frequency cues and global context, leaving robustness under heterogeneous clinical conditions largely unaddressed. To bridge this gap, we design M3FPolypSegNet++ by coupling a learnable Gaussian Adaptive Frequency Atrous Spatial Pyramid Pooling (GAF-ASPP) module with a Transformer backbone, unifying context-rich and frequency-aware features within a lightweight framework suitable for real-time deployment. The GAF-ASPP module adaptively separates feature maps into distinct low- and high-frequency components via a learnable Gaussian filter. Unlike the traditional box-filter with static weight approaches (M3FPolypSegNet), the GAF-ASPP dynamically highlights critical frequency bands according to the input image, effectively reducing artifacts and enhancing segmentation performance. Furthermore, we integrate a Transformer-based encoder, enabling more comprehensive global context modeling and improved representation of long-range dependencies. Extensive experiments on multiple benchmark datasets demonstrate that M3FPolypSegNet++ consistently outperforms state-of-the-art methods across both seen and unseen clinical settings.