Learning with Limited Annotations: A Survey on Deep Semi-Supervised Learning for Medical Image Segmentation (2207.14191v3)

Abstract: Medical image segmentation is a fundamental and critical step in many image-guided clinical approaches. Recent success of deep learning-based segmentation methods usually relies on a large amount of labeled data, which is particularly difficult and costly to obtain especially in the medical imaging domain where only experts can provide reliable and accurate annotations. Semi-supervised learning has emerged as an appealing strategy and been widely applied to medical image segmentation tasks to train deep models with limited annotations. In this paper, we present a comprehensive review of recently proposed semi-supervised learning methods for medical image segmentation and summarized both the technical novelties and empirical results. Furthermore, we analyze and discuss the limitations and several unsolved problems of existing approaches. We hope this review could inspire the research community to explore solutions for this challenge and further promote the developments in medical image segmentation field.

Deep Semi-Supervised Learning for Medical Image Segmentation: An In-Depth Survey

This paper offers a comprehensive survey on the application of deep semi-supervised learning (SSL) techniques in medical image segmentation, a critical component of modern image-guided clinical procedures. Medical image segmentation is essential for delineating anatomical structures such as organs and tumors, and SSL emerges as a practical solution to the challenge of limited annotated data, common in medical imaging where expert annotations are costly and scarce.

Technical Overview

The authors categorize existing methods into three distinct strategies:

1. Pseudo Labels: Pseudo labeling is an intuitive approach wherein a model, initially trained on a limited labeled dataset, generates pseudo labels for unlabeled data. These pseudo-labeled images are iteratively used for further training. The paper outlines various techniques for refining pseudo labels, such as confidence-based selection and label propagation, to mitigate noise and improve label quality.
2. Unsupervised Regularization: This strategy involves integrating unlabeled data with labeled data during training through unsupervised loss functions. Key methodologies include:
   • Consistency Learning: Enforcing prediction consistency under various perturbations to leverage low-density separation.
   • Co-Training: Utilizing multiple models trained on different views of the data to provide diverse pseudo labels.
   • Entropy Minimization: Encouraging low-entropy predictions to push decision boundaries towards low-density regions.
3. Knowledge Priors: Incorporating anatomical knowledge, such as shape and positional information, to enhance model training. This includes the use of self-supervised tasks and anatomical constraints to strengthen representation abilities.

Empirical Analysis and Findings

The survey compares numerous approaches on benchmark datasets like LA, Pancreas CT, and BraTS, emphasizing the considerable improvements SSL methods achieve compared to fully supervised counterparts, especially when labeled data is scarce. The empirical results highlight the capability of SSL to attain performance levels close to fully supervised models using a fraction of annotated data.

Limitations and Open Challenges

While semi-supervised approaches have shown promise, the paper identifies several challenges:

• Distribution Alignment: Handling misaligned distributions between labeled and unlabeled data, which can adversely impact model performance.
• Model Robustness: Ensuring models focus on informative regions in data, especially when faced with noisy or conflicting pseudo labels.
• Integration with Other Methods: There is scope to synergize SSL with methods like transfer learning and few-shot learning to further reduce annotation burdens.

Future Directions

The paper suggests potential future developments:

• Developing frameworks that can adeptly handle data distribution challenges.
• Enhancing strategies to effectively focus model training on reliable pseudo labels.
• Exploring integration with foundation models like SAM for improved pseudo label generation.

Conclusion

The survey critically evaluates current semi-supervised learning techniques, outlining the progress and potential of SSL in advancing medical image segmentation. This comprehensive analysis not only marks the existing achievements but also delineates pathways for future research, aiming to inspire further innovation and development in the field. Through detailed categorization and empirical analysis, the paper serves as a valuable resource for researchers seeking to understand and contribute to this evolving area of paper.