Automatic segmentation of esophageal cancer pathological sections based on semantic segmentation

The cancer area segmentation of esophageal histopathology images is a crucial step in determining the stage of esophageal cancer. This task is very important. However, manual segmentation will cost a lot of time. The rise of computational pathology has led to the development of automatic methods for cancer area detection. In the automatic segmentation problem, a well-labeled dataset is the most important part. One of the main contributions of this paper is to establish a dataset contains 1388 patches (958 Normal and 430 Abnormal containing tumor cells), marked with cancer, all of which are manually labeled and supervised by professional pathologists. We test the currently popular networks on our dataset, such as DeeplabV3, FCN+ResNet, Unet and so on. And FCN+ResNet achieves the best performance on our dataset with the highest Mean IoU (85.06%) and Pixel Acc (92.63%).