A dual-consistency semi-supervised learning method for histopathology image segmentation

Segmenting tumor regions from the whole-slide images (WSIs) is crucial for quantitative characterizing of the tumor microenvironment and supporting clinical diagnosis. Compared with classification tasks, segmentation requires pixel-level annotations at gigapixel resolution, which makes manual labeling particularly costly. To address the issues, we propose a dual-consistency semi-supervised learning method for histopathological image segmentation, including hierarchical consistency (HC) and coarse-fine grained consistency (CFGC) regularizations. HC enforces prediction alignment across multiple levels in the decoding, encouraging the network to learn coherent representations and improving robustness when training with unlabeled data. Furthermore, CFGC establishes consistency between coarse-grained and fine-grained segmentation results through a multi-scale convolution module (MSC) that enhances feature propagation from the encoder to the decoder. The proposed experimental results demonstrate that the proposed method based on HC and CFGC for semi-supervised histopathology image segmentation can finely segment cancerous regions with little annotation data, assisting pathologists in diagnosis.