U-Net Architecture and Hyperparameter Optimization via Human Learning Optimization Algorithm for Medical Image Segmentation

The health status of the placenta significantly influences fetal development, with abnormalities in its function potentially compromising normal fetal growth and posing serious threats to life in severe cases. Thus, precise quantitative assessment of the placenta is of high clinical importance. To overcome the difficulties associated with manual hyperparameter tuning in traditional U-Net models and insufficient segmentation accuracy on complex placental boundaries in medical imaging, this study proposes an automatically optimized U-Net model (HLO-UNet) leveraging the Human Learning Optimization (HLO) algorithm. This method establishes a synergistic framework that integrates HLO with neural architecture search, enabling autonomous co-evolution of hyperparameters and network architectures. Experimental results demonstrate that HLO-UNet significantly outperforms the original U-Net model in placental segmentation tasks, achieving a Dice coefficient of 0.892\(\pm\)0.062 (2.7% improvement over U-Net) and an IoU of 0.805\(\pm\)0.077 (4.3% improvement over U-Net), with smaller standard deviations indicating enhanced model robustness. This research introduces a novel adaptive network optimization approach tailored specifically for medical image segmentation, exhibiting distinct advantages in accurately segmenting complete placenta previa cases. It provides a valuable methodological reference, paving the way for further research and practical applications in the domain of placental medical imaging.