#### multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels --- ## Class activation mapping - basic idea: train classification network and leverage the semantic information from the deeper layers to achieve WSSS. - As training process of classfication goes further, network tends to differentiate objects by the most discriminative features (especially in histopathology image) --- - model architecture  --- - progressive dropout attention  --- - experimental results  --- - experimental results  --- - ablation studies  --- - cutout data augmentation  --- ## Future work - data augmentation - associating classification by a multi-task learning strategy - enhancing both tasks with information exchange - cross-task to improve each other - noise-correction for pseudo masks ---