Combining Bottom-Up, Top-Down, and Smoothness Cues for Weakly Supervised Image Segmentation | Synapse