This study introduces a new method that enhances training for deep learning models in auto-segmentation by combining density and diversity criteria to select the most informative samples, thereby reducing manual annotation effort.
Experiments on MRI and CT images of lower extremities show that this hybrid approach outperforms existing methods, resulting in statistically significant improvements in metrics like dice score and reduced annotation cost (RAC).
The findings suggest that integrating density and diversity criteria in Bayesian active learning increases efficiency in annotating medical images, ultimately streamlining the segmentation process in medical settings.