Optimizing Hyperparameters of Resnet50 for 3t2fts-Based Tumor Grading

Abstract

Three-dimensional (3D) MR imaging allows for a comprehensive assessment of crucial factors like brain tumor volume and location. Also, two-dimensional (2D) analysis comprises more methods than 3D-based analysis to be applied for tumor classification, e.g. high-grade glioma (HGG) vs low-grade glioma (LGG). Regarding this, to accurately classify between HGG and LGG glioma types, each slice of 3D MR images can be analyzed using 3D to 2D feature transformation methods. In this paper, we utilize optimized-ResNet50 models to accurately classify 2D identity information (2D-ID) of HGG-and LGG-typed tumors that is the output of a 3D to 2D feature transformation strategy (3t2FTS). At this point, state-of-the-art algorithms including crystal structure algorithm (CSA), chaotic dynamic weight-particle swarm optimization (CDW-PSO), and modified social group optimization (MSGO) are compared to tune the hyperparameters of ResNet50, i.e. continuous & discrete optimization problem. The experimental results indicate that the ResNet50-CSA model achieves the highest accuracy (91.23%) on the 3t2FTS-based data. Furthermore, ResNet50-MSGO provides 87.72 % accuracy, and its operation time is nearly half of the time belonging to ResNet50-CSA. As a result, MSGO and especially CSA, demonstrate significant performance to catch a tradeoff between searching capacity (convergence capability), accuracy and operation time. © 2024 IEEE.