Research on improved Unet in microscopic image segmentation of esophageal cancer

Abstract

Timely diagnosis can greatly reduce the mortality of esophageal cancer. However, the existing diagnostic methods can only rely on pathologists to observe and estimate through stained images of naked eye slices because of lacking advanced analysis methods. These methods will cause a long testing cycle and will have different degrees of error in the results of testing. Therefore, we put forward a cell segmentation model based on the basic framework of Unet network and introduced an attention mechanism to the model. First, the Unet with encoder-decoder structure is used for initial segmentation based on esophageal cancer cell images. Then, the attention mechanism is used to heavy the weight of the cell region to reduce the interference of the model caused by the low contrast between esophageal cancer living cells under an optical microscope image and the background. As a result, the recognition accuracy of the improved model for esophageal cancer cells can reach 96%. The experimental results show that the improved model reaches an accuracy of 96% in recognizing esophageal cancer cells. The proposed model can provide a real-time, objective, and accurate solution to the diagnosis of esophageal cancer.