Learning Rate Optimization of U-Net Architecture Using Grasshopper Optimization Algorithm to Enhance Accuracy in CT Image Segmentation of COVID-۱۹ Patients

Background and Aims: In light of its excellent learning accuracy and rate, rapid data processing, and independence from large databases for network training, the U-Net architecture is a well-known and popular deep learning architecture for image segmentation and feature extraction. Learning rate selection and updating are crucial in network training. As U-Net is a completely nonlinear network, classical mathematical optimization algorithms increase the probability of local optima. When initializing a neural network and selecting its associated hyper-parameters, there are typically two approaches that can be taken. In the first scenario, values are assigned based on empirical knowledge or prior experience with similar models. The second scenario involves using a metaheuristic algorithm to search for optimal solutions within a range of values that we know empirically has a higher probability of network convergence. Methods: This analytical research paper used the grasshopper optimization algorithm (GOA) as a metaheuristic approach to optimize the learning rate of U-Net. The network was trained using ۲۵۶*۲۵۶ CT images of the lungs of COVID-۱۹ infected and uninfected individuals. A total of ۴۰۰ CT images were employed as the training dataset, whereas ۸۰ CT images were used as the testing data. Coding was implemented in MATLAB. Results: The optimization of the learning rate enhanced image segmentation accuracy by ۲.۲۳%. Conclusion: Iterative metaheuristic algorithms would lead to longer network training times. However, the proposed network optimization method could be very useful when large databases are not available for network training and higher accuracy is preferred over time savings. Keywords: Covid-۱۹, GOA, Image segmentation, Metaheuristic, U-Net

کلیدواژه ها:

Covid-۱۹ ، GOA ، Image segmentation ، Metaheuristic ، U-Net