TY - GEN

T1 - Active Transfer Learning for 3D Hippocampus Segmentation

AU - Wu, Ji

AU - Kang, Zhongfeng

AU - Llambias, Sebastian Nørgaard

AU - Ghazi, Mostafa Mehdipour

AU - Nielsen, Mads

N1 - Publisher Copyright: © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

PY - 2023

Y1 - 2023

N2 - 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.

AB - 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.

KW - Active learning

KW - domain adaptation

KW - entropy sampling

KW - medical image segmentation

UR - http://www.scopus.com/inward/record.url?scp=85174736778&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-44917-8_22

DO - 10.1007/978-3-031-44917-8_22

M3 - Article in proceedings

AN - SCOPUS:85174736778

SN - 9783031471964

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 224

EP - 234

BT - Medical Image Learning with Limited and Noisy Data - 2nd International Workshop, MILLanD 2023, Held in Conjunction with MICCAI 2023, Proceedings

A2 - Xue, Zhiyun

A2 - Antani, Sameer

A2 - Zamzmi, Ghada

A2 - Yang, Feng

A2 - Rajaraman, Sivaramakrishnan

A2 - Liang, Zhaohui

A2 - Huang, Sharon Xiaolei

A2 - Linguraru, Marius George

PB - Springer

T2 - 2nd Workshop on Medical Image Learning with Noisy and Limited Data, MILLanD 2023

Y2 - 8 October 2023 through 8 October 2023

ER -