HistoSegResT: A Weakly Supervised Learning Method for Histopathology Image Segmentation

The morphology of glands, such as size and contour, has been used routinely by pathologists to diagnose the malignant degree of several adenocarcinomas in the process of pathological diagnosis. To automatically segment the gland regions, fully supervised segmentation algorithms require labor-intensive and time- consuming labeling at the pixel level. In this paper, we propose a weakly supervised learning method HistoSegResT (HSRT), which only uses image-level labels (i.e., malignant and benign) to complete histopathology image segmentation. In HSRT, the structure of CNN is used to extract the underlying features of the image, and the self-attention mechanism of the transformer is used to encode the long-range dependencies in histopathology images. In addition, a reconstruction loss is designed to discover the most integrated region of the object. A series of experiments show that the proposed HSRT method outperformed existing state-of-the-art methods with the same level of supervision on the GlaS dataset and can effectively relieve under-activation and over-activation of generated CAMs.