Figure 3.

Deep learning methods for segmentation and detection of colonoscopy polyps: method design of all top teams for detection and segmentation tasks are depicted. Each network is fed with an input image and the output prediction is then either directly predicted or an ensemble of network is used for prediction. The description of backbone and nature of these networks is provided in Table 1. Each output prediction for detection task is bounding box prediction with class label polyp while for segmentation are the pixel-wise classification where polyp classes are provided with label 1 and background pixels as label 0. At bottom, an overlay on the original image, bounding box prediction (in black) and segmentation prediction (in red) are also shown.