Abstract

The segmentation of histopathological images is an important problem in the field of medical image processing. However, the high cost of manual annotation and the lack of large-scale annotated data are important factors that restrict the application of deep learning methods in this field. To overcome these challenges, we propose a two-stage weakly supervised semantic segmentation model based on pathological tissue relationships. Our framework leverages the potential relationships between various tissues in histopathological images through a similar Graph Parsing Attention Mechanism to improve segmentation performance. At the segmentation stage, we validate the effectiveness of our cyclic pseudo-mask strategy for denoising and segmentation, and further enhance segmentation performance through multi-resolution supervision. Our model exhibits advanced performance on both BCSS and LUAD histopathology datasets, demonstrating the superiority of our framework. The contribution of our paper lies in the introduction of prior knowledge about the potential relationships between tissues into the weakly supervised semantic segmentation domain, which realizes high-quality histopathological image segmentation on small sample datasets. Moreover, we propose novel strategies such as cyclic pseudo-masks and multi-resolution supervision to improve segmentation performance. Our framework has significant application value and theoretical significance, providing accurate diagnostic support for doctors.