Optimization of deep learning based segmentation method

Abstract

The use of deep learning models has become widespread in different computer vision problems such as classification, detection, and segmentation. Many deep learning models have been developed in the segmentation of medical images. Although segmentation accuracy has been increased, segmentation performance needs to be improved due to the variability of tissue, cell and image acquisition methods. In the deep-learning-based segmentation and classification methods, the parameters of the method should be optimized in order to obtain more successful results for segmentation. In this study, the optimization of the parameters has been performed with five optimization algorithms according to segmentation loss. These algorithms are Grey Wolf Optimizer, Artificial Bee Colony (ABC), Genetic Algorithm, Particle Swarm Optimization (PSO), and Black Widow Optimization (BWO). In the experimental studies, each algorithm was run independently ten times and ABC obtained the lowest average segmentation loss with a value of 0.135. However, ABC achieved this performance about seven hours longer than PSO and about 5 h longer than BWO. Since the parameter optimization of CNN-based models takes much more time than other benchmarks, the convergence speed of algorithms is very important. For this reason, it has been observed that PSO is much more successful than other algorithms with an average run time of 9.438 h. As a result, considering the Jaccard similarity coefficient, it was seen that the model performance increased by 8.1% with the optimization compared to manual parameter selection.