Discretely-constrained deep network for weakly supervised segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Weakly-supervised segmentation often relies on constraints or regularization priors for target regions.
• Current methods integrate these constraints within continuous optimization frameworks for convolutional neural networks (CNNs).
• Discrete formulations offer a more efficient approach for optimizing segmentation constraints and regularization priors.

Purpose of the Study:

• To propose a novel method for training CNNs using discrete constraints and regularization priors.
• To apply this method to medical image segmentation with weak annotations, incorporating size constraints and boundary length regularization.
• To demonstrate improved performance over existing approaches.

Main Methods:

• Development of a training method for CNNs utilizing discrete constraints and regularization priors.
• Application of the Alternating Direction Method of Multipliers (ADMM) algorithm for optimization.
• Implementation of size constraints and boundary length regularization for medical image segmentation.

Main Results:

• Significant improvements in segmentation accuracy compared to existing methods.
• Enhanced constraint satisfaction in segmented medical images.
• Faster convergence speeds during the training process.
• Validation on two benchmark medical image segmentation datasets.

Conclusions:

• The proposed ADMM-based method effectively trains CNNs with discrete constraints for medical image segmentation.
• This approach offers superior performance in accuracy, constraint adherence, and training efficiency for weakly-supervised tasks.
• The method shows promise for improving medical image analysis workflows.