Neural Architecture Search Using Metaheuristics for Automated Cell Segmentation.

Deep neural networks give successful results for segmentation of medical images. The need for optimizing many hyper-parameters presents itself as a significant limitation hampering the effectiveness of deep neural network based segmentation task. Manual selection of these hyper-parameters is not feasible as the search space increases. At the same time, these generated networks are problem-specific. Recently, studies that perform segmentation of medical images using Neural Architecture Search (NAS) have been proposed. However, these studies significantly limit the possible network structures and search space. In this study, we proposed a structure called UNAS-Net that brings together the advantages of successful NAS studies and is more flexible in terms of the networks that can be created. The UNAS-Net structure has been optimized using metaheuristics including Differential Evolution (DE) and Local Search (LS), and the generated networks have been tested on Optofil and Cell Nuclei data sets. When the results are examined, it is seen that the networks produced by the heuristic methods improve the performance of the U-Net structure in terms of both segmentation performance and computational complexity. As a result, the proposed structure can be used when the automatic generation of neural networks that provide fast inference as well as successful segmentation performance is desired.