Projected pooling loss for red nucleus segmentation with soft topology constraints

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Neuroscience

Background:

• Deep learning is standard for medical image segmentation but struggles with small datasets and anatomical accuracy.
• Incorporating anatomical knowledge as constraints can improve deep learning segmentation.
• Quantitative susceptibility mapping (QSM) is crucial for studying parkinsonian syndromes, requiring accurate red nucleus segmentation.

Purpose of the Study:

• To propose a novel deep learning loss function for medical image segmentation that incorporates soft topological constraints.
• To improve the segmentation accuracy and anatomical correctness of deep learning models, especially with limited training data.
• To apply and validate the proposed method for segmenting the red nucleus in QSM data.

Main Methods:

• Developed a new loss function using projected pooling to enforce soft topological constraints.
• Employed MaxPooling operations on projected structures (ground truth and prediction) to penalize topological errors.
• Evaluated the method on red nucleus segmentation from QSM and three tasks from the Medical Segmentation Decathlon (MSD).

Main Results:

• Achieved high accuracy (Dice 89.9%) and zero topological errors in red nucleus segmentation.
• Demonstrated improved Dice accuracy over baseline methods with small training sets.
• Reduced topological errors in MSD tasks (heart, spleen, hippocampus) with comparable Dice accuracies.

Conclusions:

• The proposed loss function effectively introduces topology constraints for deep learning segmentation.
• The method offers an efficient and implementable solution for accurate red nucleus segmentation.
• This approach enhances the reliability of deep learning models in medical image analysis, particularly in challenging scenarios.