Active Transfer Learning for 3D Hippocampus Segmentation

Insufficient data is always a big challenge for medical imaging that is limited by the expensive labeling cost, time-consuming and intensive labor. Active learning aims to reduce the annotation effort by training a model on actively selected samples, most of them adopt uncertainty measures as instance selection criteria. However, uncertainty strategies underperform in most active learning studies. In addition, inaccurate selections worse than random sampling in initial stage referred to as “cold start” problem is also a huge challenge for active learning. Domain adaptation aims at alleviating the cold start problem and also reducing the annotation effort by adapting the model from a pre-trained model trained on another domain. Our work focuses on whether active learning can benefit from domain adaptation and the performance of uncertainty strategy compared to random selection. We studied 3D hippocampus images segmentation based on 3D UX-Net and four MRI datasets Hammers, HarP, LPBA40, and OASIS. Our experiments reveal that active learning with domain adaptation is more efficient and robust than without domain adaptation at a low labeling budget. The performance gap between them diminishes as we approach to that half of the dataset is labeled. In addition, entropy sampling also converges faster than random sampling, with slightly better performance.