Domain Adaptation and zero shot learning for medical image segmentation

Introduction

Precise segmentation of affected region in medical imaging is key to appropriate diagnosis. But there is a number of problems need to be resolved before the actual issue got addressed. Firstly, availability of data as medical images is confidential, secondly, they need laborious effort of labelling by specific domain experts, thirdly the variation in not only the anatomical structure but also the modality and finally the cost of training the model for automated segmentation. In recent years the research has opened many new horizons to overcome above mentioned issues. One of previously known methods is domain adaption. Domain adaptation is a method of transferring knowledge from source domain with sufficient labeled data to target domain containing insufficient labeled data[5]. Figure:1 and Figure:2 are showing an overview of Domain Adaptation and its different types respectively.

Figure 1

Figure 2

While one of very recent approach discussed is zero shot learning methods for image segmentation. Zero shot learning is a method in which it is assumed that the training and testing classes are disjoint [6]. But it is not possible to implement the same definition on medical image segmentation. The labeled source/training data can be categorized as the “seen” classes, while the unlabeled target data as the “unseen” classes.

Method:

Unsupervised domain adaptation needs fully annotated training data while ZeroShot Learning method needs some auxiliary information describing the domain. In our study, we propose an approach that will combine the strengths of two approaches i.e. domain adaptation and zero shot learning to overcome their weaknesses. An ablation study at [2] shows an Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation that successfully implemented the approach on cardiac and abdominal datasets and gives reasonable good results.

References:

[1] M. Bateson, H. Kervadec, J. Dolz, H. Lombaert, and, I. Ayed, ”SOURCE-FREE DOMAIN ADAPTATION FOR IMAGE SEGMENTATION” Medical Image Analysis 82 (2022) 102617

[2] C. Bian , C. Yuan, K. Ma , S. Yu , D. Wei , and Y. Zheng “Domain Adaptation Meets Zero-Shot Learning:An Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation” IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 41, NO. 5, MAY 2022

[3] E. Ahn, A. Kumar, M. Fulham, D. Feng and J. Kim ,” Unsupervised Domain Adaptation to Classify Medical Images Using Zero-Bias Convolutional Auto-Encoders and Context-Based Feature Augmentation” , IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 39, NO. 7, JULY 2020

[4] S. Ma, , X. Li, J. Tang and F. Guo, “A ZERO-SHOT METHOD FOR 3D MEDICAL IMAGE SEGMENTATION”, 2021 IEEE International Conference on Multimedia and Expo (ICME) | 978-1-6654-3864-3/20/$31.00 ©2021 IEEE | DOI: 10.1109/ICME51207.2021.9428261

[5] https://www.v7labs.com/blog/domain-adaptation-guide#h1

[6] https://www.v7labs.com/blog/zero-shot-learning-guide#h1

[7] R. Feng , X. Zheng, T. Gao, J. Chen ,W. Wang, D. Z. Chen and J. Wu, “Better Medical Image Segmentation”, IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 40, NO. 10, OCTOBER 2021

[8] R. Zoetmulder, E. Gavves, M. Caana, H. Marqueringa, “Domain- and task-specific transfer learning for medical segmentation tasks”, Computer Methods and Programs in Biomedicine 214 (2022) 106539

Creator:

Najam-us-Sahar

PhD Scholar Computer Science

Department of Computer Science

COMSATS University Islamabad, Lahore campus

Email: fa22-pcs-007@cuilahore.edu.pk

Page updated

Google Sites

Report abuse