BENDING LOSS REGULARIZED NETWORK FOR NUCLEI SEGMENTATION IN HISTOPATHOLOGY IMAGES

Area of Science:

• Digital pathology
• Computational imaging
• Biomedical image analysis

Background:

• Accurate nuclei segmentation is crucial for histopathology image analysis.
• Existing methods struggle with segmenting overlapped nuclei, limiting their clinical utility.
• Overlapped nuclei present a significant challenge in automated analysis of tissue samples.

Purpose of the Study:

• To develop an improved nuclei segmentation method for histopathology images.
• To address the limitations of current approaches in segmenting overlapped nuclei.
• To introduce a novel bending loss function for enhanced contour accuracy in nuclei segmentation.

Main Methods:

• A novel bending loss regularized network was developed for nuclei segmentation.
• The bending loss function penalizes contour points with high curvature, preventing contours from encompassing multiple nuclei.
• The proposed method was validated on the MoNuSeg dataset using standard quantitative metrics.

Main Results:

• The proposed bending loss network demonstrated superior performance in nuclei segmentation compared to six state-of-the-art methods.
• Significant improvements were observed in Aggregate Jaccard Index, Dice, Recognition Quality, and Panoptic Quality metrics.
• The method effectively reduces errors caused by overlapping nuclei.

Conclusions:

• The bending loss regularized network offers a promising solution for accurate nuclei segmentation in histopathology.
• This approach enhances the separation of overlapped nuclei, advancing digital pathology analysis.
• The method provides a more reliable tool for quantitative analysis of tissue microenvironments.