COLORECTAL polyp segmentation with a Boundary-Aware Attention-Based Segmentation framework (BAASNet) demonstrated improved Dice performance across nine publicly available datasets spanning five imaging modalities, according to the study abstract. BAASNet was developed to address challenges that can limit automated polyp segmentation, including image noise, complex textures, indistinct boundaries and diverse polyp morphologies.
Colonoscopy plays a central role in early colorectal cancer prevention, with precise identification of polyps supporting treatment planning and diagnostic accuracy. Segmentation models generate masks that encode clinically relevant structures, but the manual annotation required to develop such systems can be costly and time intensive.
BAASNet was designed to address these challenges through a boundary aware approach. The framework incorporated a boundary aware loss function intended to improve the delineation of polyp edges, an important consideration where boundaries may be difficult to distinguish from surrounding tissue.
The model was assessed on nine datasets, including two centre wise polyp detection benchmarks. On PolypDB, BAASNet achieved a mean Dice similarity coefficient of at least 89.60% across all five imaging modalities evaluated.
Across all benchmarks, the proposed model produced an average absolute improvement of approximately 3.3% in Dice compared with previous results. The reported gains varied between datasets, with improvements ranging from approximately 0.7% to 4.7% relative to the best previous results.
These findings suggest that incorporating boundary information into deep-learning-based polyp segmentation may help address some of the visual complexity encountered across different imaging settings. The evaluation across multiple datasets and modalities also provided evidence of the model’s generalisation capability.
The results position BAASNet as a potential approach for robust automated polyp segmentation in colonoscopy. Its performance across varied datasets and imaging modalities indicates potential for applications where consistent delineation of polyp structures is required.
The study also highlights the potential clinical value of reducing reliance on manual annotation through automated segmentation systems. By improving the identification of polyp boundaries, such systems could support diagnostic accuracy and treatment planning.
According to the reported findings, BAASNet could ultimately support real time deployment within automated colonoscopy workflows. Further assessment would be needed to establish its clinical utility, but the results demonstrate the potential of boundary aware deep learning for improving polyp segmentation across diverse imaging environments.
Naseem K et al. BAASNet: boundary-aware deep learning for accurate polyp segmentation in colonoscopy. Sci Rep. 2026;16:24170.
Featured Image: ShvedKristina on Adobe Stock
Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
