Federal learning-based a dual-branch deep learning model for colon polyp segmentation

The incidence of colon cancer occupies the top three places in gastrointestinal tumors, and colon polyps are an important causative factor in the development of colon cancer. Early screening for colon polyps and colon polypectomy can reduce the chances of colon cancer. The current means of colon polyp examination is through colonoscopy, taking images of the gastrointestinal tract, and then manually marking them manually, which is time-consuming and labor-intensive for doctors. Therefore, relying on advanced deep learning technology to automatically identify colon polyps in the gastrointestinal tract of the patient and segmenting the polyps is an important direction of research nowadays. Due to the privacy of medical data and the non-interoperability of disease information, this paper proposes a dual-branch colon polyp segmentation network based on federated learning, which makes it possible to achieve a better training effect under the guarantee of data independence, and secondly, the dual-branch colon polyp segmentation network proposed in this paper adopts the two different structures of convolutional neural network (CNN) and Transformer to form a dual-branch structure, and through layer-by-layer fusion embedding, the advantages between different structures are realized. In this paper, we also propose the Aggregated Attention Module (AAM) to preserve the high-dimensional semantic information and to complement the missing information in the lower layers. Ultimately our approach achieves state of the art in Kvasir-SEG and CVC-ClinicDB datasets.