Attention-enhanced U-Net based network for cancerous tissue segmentation

Abstract

Cancerous tissue segmentation is a key step in further refining the identification of cancer cell aggregation regions after cell segmentation, which is crucial for early diagnosis, precise staging and personalized treatment strategies for cancer. However, there are relatively few researchers in this field, highlighting the need for further exploration. This paper has proposed an automated cancer segmentation method based on Attention-Enhanced U-Net, fusing two key features, color and density. The method is mainly divided into two steps: cell segmentation and cancer segmentation. On both Multi-Organ Nuclei Segmentation and Triple-Negative Breast Cancer datasets, the cell segmentation results achieved 73.91% and 77.51% F1-Score, Mean Intersection over Union scores of 63.55% and 67.30%, and Dice Similarity Coefficient of 72.41% and 77.52%, and which are better than other deep learning models. We also tested cancer segmentation on images from the pathology library, achieving a Dice Similarity Coefficient of 87.54%, which is also better than end-to-end deep learning models. This method achieves accurate automated cancer segmentation without relying on cancer labels, reduces the cost of acquiring labeled data, and has very high practical feasibility.