Semi-supervised pathological image segmentation via cross distillation of multiple attentions and Seg-CAM consistency

Lanfeng Zhong, Xiangde Luo, Xin Liao, Shaoting Zhang, Guotai Wang

Abstract: Highlights•A novel framework, CDMA+, is introduced for semi-supervised pathological image segmentation, leveraging diverse attention mechanisms for generating predictions for unlabeled images.•A Cross Decoder Knowledge Distillation method is proposed for robust and efficient learning from noisy pseudo labels.•A novel Seg-CAM consistency by employing an auxiliary classification head and introducing a consistency constraint between the CAM and segmentation results.•Extensive experiments show CDMA+ achieves encouraging results and outperform eight state-of-the-art methods on two public pathological image segmentation datasets.