Interactive segmentation of biostructures through hyperspectral electron microscopy

Abstract

Microscopic imaging at the nanometer level has been automated resulting in Gigabyte images, similar to Google earth. However, the absence of a ground-truth database in electron microscopy (EM) poses a challenge for automated analysis of the grey scale images. Analytical analysis in EM using techniques such as energy-dispersive X-ray (EDX) imaging facilitates mapping nanometer-scale structures by capturing hyperspectral information for each pixel. Spectral signatures are linked to their elemental compositions, providing an objective, data-driven approach to automatic segmentation. Nevertheless, the absence of a ground-truth spectral database in EM remains a significant challenge for these automated methods. Here, we present a user-in-the-loop workflow combining pre-processing, filtering, dimensionality reduction, and interactive clustering to segment biostructures in large-scale EM datasets. We demonstrate that our recursive clustering tool can be used to segment and quantify biostructures effectively. This approach enhances the interpretation of relevant image regions and shows promise for downstream tasks such as automated segmentation and analysis in high-dimensional EM data, providing a scalable solution for complex biological systems.