Considering the high resolution of Oral Squamous Cell Carcinoma(OSCC)pathological image,we propose to simply draw curves to cover up all foreground regions,generating coarse labels with back-ground around involved in.The overall time of coarse annotation is about one sixth the time of professional precise annotation.Considering the fact that background labels are clean and the foreground labels are noisy,multiple background regions are used to establish templates to filter out the mislabelled foreground pixels and refine the foreground segments.Specifically,Background Similarity-aware Mean Teacher(BS-MT)is advanced,which explores background representations with high diversity and robustness and guides the student model to generate high-quality foreground labels via background pixel filtering.On the privately-es-tablished OSCC-CA dataset,our method significantly outperforms U-Net by 2.5％and other widely-used weakly-supervised methods by more than 1.1％in terms of mIoU.
关键词
弱监督分割/病理图像分割/背景相似度感知平均教师
Key words
weakly-supervised segmentation/pathological image segmentation/background similarity-a-ware mean teacher
维普
万方数据