Narrow-band loss - a novel loss function focused on target boundary

Area of Science:

• Medical Image Analysis
• Deep Learning for Medical Imaging
• Computational Anatomy

Background:

• Current region-based loss functions (e.g., Dice) in medical image segmentation treat all pixels equally, neglecting the critical boundary regions.
• Clinical applications often require precise segmentation of organ boundaries, an area underserved by existing loss functions.

Purpose of the Study:

• To introduce a novel loss function, narrow-band loss, designed to enhance focus on target organ boundaries in medical image segmentation.
• To evaluate the efficacy of narrow-band loss in improving segmentation accuracy, particularly around boundaries, compared to standard loss functions.

Main Methods:

• Designed narrow-band loss, a region-based loss function that integrates predicted probabilities within a narrow band around the target boundary.
• Generalized narrow-band loss for multi-target segmentation scenarios.
• Tested the proposed loss function on brain and abdominal imaging datasets.

Main Results:

• Narrow-band loss demonstrated significant improvements in both the Hausdorff 95 (hd95) and Dice similarity coefficient (DSC) metrics compared to the baseline Dice loss.
• The method effectively guides segmentation networks to better delineate target boundaries and their surrounding regions.

Conclusions:

• Narrow-band loss offers a superior approach for medical image segmentation by emphasizing boundary accuracy.
• This novel loss function shows promise for improving clinical applications requiring precise organ segmentation.