Navigating Tversky Loss Function Hyperparameter Spaces using Particle Swarm Optimization for Myocardial Scar Segmentation

Medical image segmentation faces challenges due to class imbalance, where the foreground often occupies a much smaller volume compared to background tissues. This imbalance significantly impacts deep learning model performance, as different loss functions exhibit varying levels of robustness to such disparities. The Tversky loss function was specifically designed to address this issue. However, dataset-specific characteristics can still affect network performance, necessitating hyperparameter fine-tuning. This study proposes using particle swarm optimization (PSO) to automatically search for optimal Tversky loss hyperparameter values for myocardial scar segmentation models. Moreover, the hyperparameter space was reduced by simplifying the hyperparameters. This approach was evaluated against both a baseline configuration and other state-of-the-art loss functions. The results outperform other loss functions with scar segmentation DSC of 71.81% and F2-score of 0.7870. This method efficiently finds optimal hyperparameter values, demonstrating its potential for robust and accurate medical image segmentation tasks. In conclusion, this work introduces a novel PSO-based approach for optimizing Tversky loss hyperparameters in DeepLabV3+ models for myocardial scar segmentation. This method achieves both better performance and efficient optimization, demonstrating its potential for robust and accurate medical image segmentation tasks.