Constrained-CNN losses for weakly supervised segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Weakly-supervised learning reduces annotation effort in CNN segmentation.
• Inequality constraints leverage unlabeled data and domain knowledge.
• Existing methods using Lagrangian dual optimization are computationally expensive.

Purpose of the Study:

• To develop a computationally efficient method for weakly-supervised segmentation using inequality constraints.
• To improve segmentation performance by directly enforcing constraints within the loss function.
• To enable comparable performance to fully supervised methods with minimal annotations.

Main Methods:

• Introduced a differentiable penalty to enforce inequality constraints directly in the loss function.
• Avoided complex Lagrangian dual iterates and proposal generation.
• Applied the method to CNN segmentation tasks with linear constraints like region size and image tags.

Main Results:

• The penalty-based approach significantly outperformed previous Lagrangian-based methods.
• Achieved segmentation performance comparable to full supervision with partial pixel annotations.
• Reduced computational demand during training.

Conclusions:

• The proposed differentiable penalty method offers a more efficient and effective approach to weakly-supervised segmentation.
• The framework is extensible to various non-linear constraints, showing potential for semantic medical image segmentation.
• Publicly available code facilitates further research and application.