PathTGSeg: The pathologic image segmentation of breast cancer via template matching and graphic calculation

Pathologic image analysis is important for providing fundamental references for the clinical diagnosis of breast cancer. Although many methods have achieved outstanding performance in the pathologic image segmentation of breast cancer, there are still two issues limiting further development in this task. First, diverse and complex appearances exist within the observed scope for the same type of breast cancer. Second, inconsistent distribution and ambiguous borders of cancer tissues also pose challenges to obtaining accurate diagnoses of tumor regions. Therefore, to address these issues, this work proposes a template and graphic visual transformer network to perform pathologic image segmentation of breast cancer. The template visual transformer introduces a novel segmentation pattern that leverages templates selected from typical cases of breast cancer to locate cancer lesions by comparing the similarity between the templates and latent regions. Meanwhile, the graphic visual transformer builds graphical features to describe the distribution relationships of cancer lesions more accurately. Extensive experiments conducted on BIS5k and other zero-shot datasets show that our method not only achieves more robust performance on pathologic images of breast cancer when pretrained on data of the same type, but also demonstrates great potential on zero-shot pathologic images.