A Survey on Shape-Constraint Deep Learning for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Anatomy

Background:

• Fully convolutional deep neural networks (DNNs) revolutionized medical image segmentation but often produce artifacts due to pixel-level classification.
• Sparse annotations in medical datasets exacerbate issues like fragmented structures and topological inconsistencies.
• Segmentation artifacts hinder downstream clinical tasks such as surgical planning and prognosis.

Approach:

• This review surveys recent literature integrating explicit anatomical constraints into medical image segmentation.
• Methods discussed include Markov/Conditional Random Fields, Statistical Shape Models, and Active Contours.
• The review examines shortcomings, opportunities, and the emerging trend of implicit shape modeling.

Key Points:

• Explicit anatomical constraints are crucial for robust and clinically relevant medical image segmentation.
• Traditional DNNs struggle with anatomical consistency, leading to segmentation artifacts.
• Emerging implicit shape modeling offers a promising direction for future research.

Conclusions:

• Integrating anatomical knowledge into deep learning models is essential for reliable medical image segmentation.
• Future work should focus on implicit shape modeling to overcome limitations of current methods.
• This review provides a comprehensive overview and tabulated details of relevant segmentation techniques.