Abstract

Traditional data augmentation methods, which employ static hyperparameters, often lead to model overfitting. To address this limitation, a novel hyperparameter-driven data augmentation approach (HDAug) is introduced in this study. Training images that simulate a plethora of lighting and exposure conditions are synthesized by HDAug through the stochastic sampling of augmentation hyperparameters, within predefined ranges. Additionally, HDAug does not rely on prior knowledge of specific datasets, endowing it with superior generalization capabilities. The Dice coefficient was utilized as the primary evaluation metric. Experimental results demonstrate that HDAug achieves significant performance improvements in two challenging cross-modality medical image segmentation datasets, with average Dice coefficients of 86.77%, 88.08%, and 84.11%, respectively. The superiority of HDAug lies in its ability to substantially enhance model robustness across diverse imaging conditions while circumventing the overfitting issues inherent in conventional methods. Furthermore, HDAug is computationally efficient and is integrated into existing medical image segmentation workflows.OPEN ACCESS Received: 30/08/2024 Accepted: 22/10/2024 Published: 07/04/2025