Active Transfer Learning for 3D Hippocampus Segmentation

Ji Wu, Zhongfeng Kang, Sebastian Nørgaard Llambias, Mostafa Mehdipour Ghazi, Mads Nielsen*

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

Abstract

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.

Original languageEnglish
Title of host publicationMedical Image Learning with Limited and Noisy Data - 2nd International Workshop, MILLanD 2023, Held in Conjunction with MICCAI 2023, Proceedings
EditorsZhiyun Xue, Sameer Antani, Ghada Zamzmi, Feng Yang, Sivaramakrishnan Rajaraman, Zhaohui Liang, Sharon Xiaolei Huang, Marius George Linguraru
PublisherSpringer
Publication date2023
Pages224-234
ISBN (Print)9783031471964
DOIs
Publication statusPublished - 2023
Event2nd Workshop on Medical Image Learning with Noisy and Limited Data, MILLanD 2023 - Vancouver, Canada
Duration: 8 Oct 20238 Oct 2023

Conference

Conference2nd Workshop on Medical Image Learning with Noisy and Limited Data, MILLanD 2023
Country/TerritoryCanada
CityVancouver
Period08/10/202308/10/2023
SeriesLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume14307 LNCS
ISSN0302-9743

Bibliographical note

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

Keywords

  • Active learning
  • domain adaptation
  • entropy sampling
  • medical image segmentation

Cite this