A comparison of U-net backbone architectures for the automatic white blood cells segmentation
Mohammed Hakim Bendiabdallah, Nesma Settouti · 2021
Reliable recognition of white blood cells is an essential step in the diagnosis of several types of cancer. Therefore, the segmentation of white blood cells plays an essential role and is an important part of the medical diagnostic system. Manual cell diagnosis involves doctors visually examining microscopic images to detect any cellular abnormalities. This step is costly and time-consuming. An automated system based on white blood cell identification provides a more accurate result than the manual method. Image segmentation is one of the crucial contributions of a deep learning community to the medical field. In this paper, we demonstrate how the U-Net type architecture can be improved by the use of the pre-trained encoder, a comparison of several efficient methods for automatic recognition of white blood cells using the original U-NET, different pre-trained classification networks are used as the backbone to obtain better performance. The architecture of RESNET-50 obtains the best segmentation results on testing data for automatic recognition in cytological images with a less amount of training epochs.