Active Transfer Learning for 3D Hippocampus Segmentation
Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Standard
Active Transfer Learning for 3D Hippocampus Segmentation. / Wu, Ji; Kang, Zhongfeng; Llambias, Sebastian Nørgaard; Ghazi, Mostafa Mehdipour; Nielsen, Mads.
Medical Image Learning with Limited and Noisy Data - 2nd International Workshop, MILLanD 2023, Held in Conjunction with MICCAI 2023, Proceedings. ed. / Zhiyun Xue; Sameer Antani; Ghada Zamzmi; Feng Yang; Sivaramakrishnan Rajaraman; Zhaohui Liang; Sharon Xiaolei Huang; Marius George Linguraru. Springer, 2023. p. 224-234 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 14307 LNCS).Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
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 -
ID: 372613841