High-frequency energy fusion (HFEF) for nuclei segmentation with boundary-aware loss

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Deep Learning

Background:

• Accurate nuclei segmentation is crucial for quantitative pathology but challenged by clustered nuclei, heterogeneous morphology, and unclear boundaries.
• Existing transformer models struggle with explicit boundary feature enhancement and minimizing false negatives at nuclei edges.

Purpose of the Study:

• To develop a boundary-enhanced nuclei segmentation framework using High-Frequency Energy Fusion (HFEF) and a sensitivity-aware loss.
• To improve the accuracy and boundary precision of nuclei segmentation in histopathological images.

Main Methods:

• Proposed an HFEF framework integrating high-frequency components (from Stationary Wavelet Transform and Laplacian filtering) as an additional input channel.
• Introduced a sensitivity-tuned loss (ST-Loss) to penalize low-confidence false negatives, especially near boundaries.
• Implemented and evaluated the framework with two Swin Transformer-based models on three public datasets using region and boundary metrics.

Main Results:

• HFEF consistently improved segmentation performance, increasing Dice Similarity Coefficient (DSC) by 1.07% and Intersection over Union (IoU) by 2.20%.
• Boundary delineation significantly improved, with a 27% reduction in Hausdorff Distance (HDF) and a 0.89% increase in Boundary F1 score (BF1).
• ST-Loss further enhanced recall by 2.48%, effectively reducing boundary-related false negatives and improving visual results for clustered nuclei.

Conclusions:

• The HFEF framework with sensitivity-aware loss significantly enhances nuclei boundary representation and reduces segmentation errors.
• This model-agnostic approach improves segmentation accuracy and boundary precision, applicable to medical imaging tasks requiring precise delineation.
• The framework offers a promising solution for challenging nuclei segmentation problems in digital pathology.