BaggedUNet: Deep Machine Vision approach for Polyps Segmentation in Gastrointestinal Tract

Polyps segmentation is one of the key medical challenges in the gastrointestinal (GI) tract. Polyps segmentation provides the early-stage diagnosis of polyps which may lead to colon cancer in the GI tract. Deep learning models such as U-Net can segment polyps with good performance. But individual deep learning models may suffer from generalization problems. Deep ensemble learning combines the power of both deep and ensemble learning so that the final combined model has better generalization ability. In this paper, a bagging based U-Net architecture (BaggedUNet) is proposed to improve the polyps segmentation in GI-Tract. Our proposed BaggedUNet model trains several lighter U-Net architectures. Decisions from various models are then combined using majority voting. The proposed method is compared with recent deep learning architectures: U-Net and ResUNet++. The evaluation of models is performed using quantitative metrics including Dice coefficient and mean Intersection over Union (mIoU). The proposed BaggedUNet architecture was able to achieve 3 %-9 % improvement on different evaluation metrics on two publicly available datasets for polyps segmentation.