W-PolypBox: Exploring bounding box priors constraints for weakly supervised polyp segmentation

Accurate polyp segmentation is crucial for the diagnosis and treatment of colorectal cancer, as it enhances detection rates and improves patient prognosis. However, challenges remain due to indistinct boundaries, multi-scale variations, and similarities to adjacent tissues. The high cost of pixel-level annotation has also resulted in a shortage of annotated data, exacerbating segmentation difficulties and limiting the clinical applicability of existing methods. Bounding box annotations provide valuable prior knowledge, providing relatively accurate semantic and location information at a lower annotation cost. To address these challenges, we propose W-PolypBox, a weakly supervised model for polyp segmentation that incorporates bounding box prior constraints. Our model starts with the development of C-Net, a cascade decoding network designed to extract high-quality polyp features. We then introduce PolypBox, a component that transforms fully supervised methods into weakly supervised ones by leveraging only bounding box annotations during training. In PolypBox, we define an uncertain bounding box regression loss to restrict polyp predictions within the box. An embedding consistency loss is introduced to ensure consistency across embeddings, followed by a fore/background matching loss to enforce similarity between pixels within the box and the mixed fore/background prototype. Finally, a neighborhood pixel consistency loss is designed to maintain region connectivity. We evaluated W-PolypBox on five public polyp datasets. Results show it outperforms other stateof-the-art weakly supervised methods and matches fully supervised performance. This indicates the proposed approach's superior feasibility for widespread clinical use.