Neural Architecture Search for Microscopy Cell Segmentation

摘要

Live microscopy cell segmentation is a crucial step and challenging task in biological research. In recent years, numerous deep learning based techniques have been proposed to tackle this task and obtained promising results. However, designing a network with excellent performance is time-consuming and labor-intensive, which limits the progress of biological research. In this paper, we propose a neural architecture search (NAS) based solution for cell segmentation in microscopy images. Different from most of the current NAS-based solutions that search the network using basic operations, we restrict the search space by exploring sophisticated network blocks. In this way, both expert knowledge and NAS are considered to facilitate the network searching. We attempt NAS with two prevailing backbone networks of U-net and Unet++. The experimental results on seven cell tracking challenge (CTC) microscopy cell data sets demonstrate that the searched networks achieve better performances with much fewer parameters than the baseline method. Thanks to its simplicity and transportability, the proposed method is applicable to many deep learning based cell segmentation methods.