Optimizing stroke lesion segmentation: A dual-approach using Gaussian mixture models and nnU-Net

Machine learning-based stroke lesion segmentation models are widely used in biomedical imaging, but their ability to detect treatment effects remains largely unexplored. Gaussian Mixture Models (GMM) and nnU-Net are among the most prominent and well-established segmentation workflows. GMM has been widely used for probabilistic tissue classification for decades, while nnU-Net has established itself as a leading deep learning framework for biomedical image segmentation, with hundreds of applications in preclinical and clinical research. Despite their widespread adoption, these methods are typically evaluated using segmentation metrics alone, without assessing their reliability in detecting therapy-induced changes - a critical factor for translational research and clinical decision-making. In this study, we systematically evaluate GMM and nnU-Net to determine their effectiveness in identifying therapy-related changes in stroke volume. Both methods demonstrate strong segmentation performance; however, nnU-Net trained solely on manual segmentations fails to detect significant therapy-induced stroke volume reductions, leading to false negative study outcomes despite achieving excellent segmentation metrics. This limitation is particularly relevant given the increasing integration of nnU-Net into biomedical research, multi-center trials and clinical workflows. To further investigate this issue, we evaluated nnU-Net trained with GMM-derived ground truth (GT) labels and observed that it more accurately detected therapy response compared to training with Manual-GT. These results illustrate how different GT definitions can influence model performance in therapy assessment. While the integration of probabilistic methods with deep learning has been previously described, our results demonstrate its practical impact in a controlled experimental setting. By systematically evaluating two widely used segmentation methods under therapy conditions, this study highlights the importance of considering therapy detection as a key evaluation criterion, rather than relying solely on segmentation accuracy.