SwinDeepLab:Swin Transformer-based DeepLab v3+ for Gland Segmentation.

Gland is an important tissue structure in the human body, but its epithelial layer has a high regeneration frequency, which makes it susceptible to glandular cancer. The morphological characteristics of glands are often used by doctor to judge the degree of malignancy in glandular cancer. Accurate glandular segmentation in histological images is an essential step for obtaining morphological data. However, the significant variations in glandular appearance at different levels of malignancy and the low contrast of background staining make it challenging to segment glands precisely. Additionally, the process of creating pathological images is time-consuming and laborious. In this paper, we collected a dataset of pathological images of gastric mucosa and annotated the gland (GMG). We proposed a gland segmentation model based on the DeepLab framework and the Swin Transformer to solve this problems. We tested the proposed method on our own glandular dataset GMG and a public dataset GlaS. The experimental results demonstrated that our approach achieves high accuracy and efficiency in the field of glandular segmentation.