AOA-guided hyperparameter refinement for precise medical image segmentation
Hossam Magdy Balaha, Waleed M. Bahgat, Mansourah Aljohani, Amna Bamaqa, El-Sayed Atlam, Mahmoud Badawy, Mostafa A. Elhosseini
Abstract
Medical image segmentation faces significant challenges, including the need for extensive annotated data, the impact of hyperparameters, and the limitations of traditional CNN models. Breast cancer (BC) and COVID-19 imaging, in particular, require precise segmentation for accurate diagnosis and treatment planning. This study introduces a novel framework that utilizes the Archimedes Optimization Algorithm (AOA) to optimize hyperparameters for medical image segmentation, aiming to enhance accuracy and efficiency. We propose a four-stage framework for hyperparameter optimization using AOA. The framework consists of: (1) Population Initialization, (2) Fitness Function Evaluation, (3) Population Updating, and (4) Results Logging. The framework optimizes key hyperparameters, including activation functions, loss functions, optimizers, and batch sizes, using a hybrid loss function combining Focal Tversky and IoU. The proposed framework was rigorously evaluated on two medical datasets: the BUSI dataset for BC and the COVID-19 CT scan lesion segmentation dataset. The R2 U-Net 2D model achieved an accuracy of 95.7% on the BUSI dataset, while the V-Net model achieved 99.2% accuracy on the COVID-19 dataset. The AOA-guided framework demonstrated superior performance compared to existing methods, with Dice coefficients of 0.675 and 0.723 for the BUSI and COVID-19 datasets, respectively. Convergence curves and performance metrics validated the stability and efficiency of AOA in optimizing hyperparameters. The AOA-guided framework significantly improves medical image segmentation by automating hyperparameter optimization. The results highlight the potential of AOA as a powerful tool for enhancing segmentation accuracy and robustness. • Using Archimedes Optimization Algorithm (AOA) for hyperparameter tuning in medical image segmentation. • Integrating AOA with V-Net, U-Net 2D, and R2 U-Net 2D to enhance segmentation accuracy. • Evaluating the performance on the BUSI dataset (breast cancer) and the COVID-19 CT scan lesion dataset. • Comparative analysis shows AOA outperforms existing methods in segmentation accuracy.