Segmentation Methodologies for the Construction of Hyperspectral Cell Nuclei Databases in Histopathology

Area of Science:

• Biomedical Imaging
• Computational Pathology
• Spectroscopy

Background:

• Hyperspectral imaging (HSI) integrates spatial and spectral data for enhanced histopathology.
• Accurate cell nucleus segmentation is crucial for building spectral databases.
• Existing segmentation methods have limitations in handling hyperspectral data.

Purpose of the Study:

• To develop and evaluate a comprehensive methodology for generating hyperspectral nuclear databases.
• To compare the performance of spectral-only, spatial-only, and spatial-spectral nucleus segmentation approaches.
• To assess the suitability of different methods for instance-level separation in histopathological samples.

Main Methods:

• Hyperspectral data acquisition and preprocessing.
• Implementation and evaluation of three nucleus segmentation methods: spectral-only, spatial-only (RGB-derived), and spatial-spectral.
• Assessment using a proprietary dataset of 30 hyperspectral cubes of brain tissue annotated by pathologists.

Main Results:

• Spectral-only segmentation yielded a 61.89% Dice Similarity Coefficient (DSC) with significant over-segmentation.
• Spatial-only segmentation achieved the highest pixel accuracy (78.97% DSC) but caused underestimation of nucleus counts.
• The spatial-spectral method achieved a 73.13% DSC with only a 4% mean cell count deviation, demonstrating superior instance separation.

Conclusions:

• Pixel-wise accuracy is insufficient for generating reliable hyperspectral nuclear databases.
• Combined spatial-spectral approaches provide more accurate and reliable instance-level nucleus segmentation.
• This methodology advances quantitative analysis in hyperspectral histopathology.