Multi-Scale Position-Aware Cell Nucleus Mask Attention for Tumor Budding Detection

Tumor budding refers to the presence of a single tumor cell or a small group of tumor cells that bud from the invasive front of the primary tumor, which is an independent predictor of survival in stage II cancers and plays a vital role in the prognosis of cancer. However, it is challenging for pathologists to manually locate and identify tumor budding from large-scale whole slide images. The development of artificial intelligence in computer vision makes it possible to detect tumor budding from the whole slide images automatically. In this paper, we explore the detection of tumor budding using an object detector based on deep learning. Considering that cell topological information is essential for tumor budding detection, we explore cell topological knowledge as a multi-scale position-aware cell nucleus mask attention to improve the representation learning for tumor budding detection. To introduce the nucleus mask information, we use the pre-trained U-Net model on a pan-cancer dataset. The experimental results show that our proposed model works well in tumor budding detection comparing with existing methods, and the cell topological information is important in representation learning for tumor budding detection.