14 March 2025 in IEEE Transactions on Instrumentation and Measurement
Automated polyp segmentation of colonoscopy images is crucial in clinical settings, providing indispensable data for diagnosis and surgical procedures. Deep convolutional neural networks (CNNs) have shown promise in this domain, yet existing methods often fail to adequately address interactions between multilevel features, leading to suboptimal results. To overcome these limitations, we propose the attention-duided asymmetric multiscale polyp segmentation network (AAPCNet), a novel framework designed to effectively capture comprehensive semantic information for accurate polyp segmentation.AAPCNet leverages the Res2Net-50 backbone with a split-aggregation strategy embedded in Bottle2neck blocks to extract rich multilevel features. To enhance contextual understanding, we introduce the deep aggregation and fusion module (DAFM), which employs large-sized dilated and asymmetric convolutions to capture multiscale information, addressing the challenges posed by polyps of varying sizes. Furthermore, the spatial contextual fusion module (SCFM) utilizes spatial and channel-wise attention mechanisms to refine features by emphasizing polyp-specific details while suppressing irrelevant background information. The innovation of our lightweight yet effective decoder lies in its unique architecture, which integrates a residual block between two SCFM modules, enabling feature refinement, enhanced polyp details, precise localization, and accurate boundary delineation while suppressing noise. This architecture achieves superior segmentation performance and outperforming state-of-the-art CCN-based models in both in-domain and out-of-domain datasets. Comprehensive experiments demonstrate that AAPCNet consistently achieves a favorable balance between accuracy and computational efficiency. Our codes and results are publicaly available at https://github.com/Mkhan143/AAPCNet.
15 July 2025 in Image and Vision Computing
21 March 2026 in Neurocomputing
1 January 2026 in Fractals
1 January 2026 in Fractals
1 January 2026 in Fractals