Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Deep learning is effective for medical image segmentation (2D and 3D).
• Segmentation performance degrades with class imbalance, common in rare observations.
• Existing loss functions (weighted cross-entropy, sensitivity, Dice loss) have limitations.

Purpose of the Study:

• Investigate deep learning loss function behavior with varying label imbalance rates.
• Assess sensitivity of loss functions to learning rate tuning in 2D and 3D segmentation.
• Propose an improved loss function for unbalanced medical image segmentation tasks.

Main Methods:

• Comparative analysis of weighted cross-entropy, sensitivity, and Dice loss functions.
• Evaluation across diverse 2D and 3D medical image segmentation tasks.
• Testing under various label imbalance scenarios and learning rate adjustments.

Main Results:

• Loss functions exhibit varying sensitivities to learning rate tuning under class imbalance.
• Generalized Dice overlap demonstrates robust performance in unbalanced segmentation.
• The proposed approach mitigates performance degradation caused by label imbalance.

Conclusions:

• Loss function choice is critical for deep learning segmentation, especially with imbalanced data.
• Generalized Dice overlap offers a promising, accurate, and robust loss function for rare medical image segmentation.
• Further research can refine deep learning strategies for challenging segmentation tasks.