Localization-Enhanced Voting-based Ensemble of Semantic Segmentation Models for Cervical Polyp Segmentation

In this paper, we propose an automated cervical polyp detection method based on semantic segmentation and ensemble learning. As polyps may cause cancer both in the colon and cervical area, it is crucial to detect them at an early stage. Our solution combines U-Net, DeeplabV3+, and PSPNet segmentation models using the combination of the predicted regions to enhance the segmentation accuracy of the polyp and its boundaries. To evaluate the proposed ensemble-based system, we measured the accuracy performance compared to standard ensemble methods and other state-of-the-art segmentation techniques as our solution outperforms them both measured by the Jaccard and Dice scores.