Boundary loss for highly unbalanced segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Machine Learning

Background:

• Standard loss functions like Dice and cross-entropy in Convolutional Neural Network (CNN) segmentation struggle with highly unbalanced datasets.
• Regional integrals in these functions lead to performance and stability issues due to orders-of-magnitude differences in class values.

Purpose of the Study:

• To introduce a novel boundary loss function for CNN segmentation that addresses challenges posed by highly unbalanced data.
• To improve the training stability and performance of segmentation models on datasets with significant class imbalance.

Main Methods:

• Proposed a boundary loss function operating on contour spaces, utilizing integrals over region interfaces instead of unbalanced regional integrals.
• Formulated a non-symmetric L2 distance on contour spaces as a regional integral, inspired by graph-based active contour flow optimization.
• Integrated the boundary loss with regional softmax probability outputs for easy implementation within existing N-D segmentation architectures.

Main Results:

• Demonstrated significant performance increases on various unbalanced segmentation problems.
• Showcased improved training stability compared to traditional regional loss functions.
• Validated the effectiveness and general applicability across different network architectures and dimensions.

Conclusions:

• The proposed boundary loss effectively mitigates issues associated with highly unbalanced segmentation problems.
• This approach offers a complementary strategy to regional losses, enhancing overall segmentation accuracy and robustness.
• The boundary loss is a versatile tool for improving deep learning-based segmentation tasks, with publicly available code for implementation.