Validation of Machine Learning-Based Segmentation for Automated 3D Reconstruction in Electron Microscopy: Application in Life and Materials Science

Abstract

In recent years, automation of electron microscopy has enabled high-throughput acquisition of large, high-resolution serial-section image volumes. Three-dimensional reconstruction of these data has become essential for visualizing fine structural details in biological and material samples [1]. However, despite rapid advances in automated data acquisition, image analysis remains a major bottleneck. Conventional image processing methods, such as gray-level thresholding and manual annotation, require extensive labor and suffer from reduced reproducibility due to operator subjectivity [2]. To establish a highly efficient three-dimensional measurement workflow from imaging to quantitative analysis, we applied machine-learning (AI) to the most challenging image-analysis step and benchmarked its effectiveness against conventional methods. We prepared serial sections of Chlamydomonas for this evaluation. Continuous serial-section SEM images were acquired using a Hitachi High-Tech scanning electron microscope equipped with Auto Capture for Array Tomography (ACAT) and a focused ion beam scanning electron microscope (FIB-SEM) [3,4]. We performed appropriate sample pretreatment and optimized imaging conditions to clearly visualize the target chloroplast structures, followed by the automatic acquisition of continuous serial-section SEM image stacks. The obtained images were processed by cropping regions of interest, aligning images, adjusting contrast, and applying filters to facilitate structural identification. For segmentation and three-dimensional reconstruction, both a conventional method combining thresholding and manual correction [2] and a machine-learning-based approach (AIVIA, Leica Microsystems) trained on annotated data were employed [5]. In Figure 1(b), the region selected by the conventional method is shown in blue. Regions with contrast resembling that of the U-shaped chloroplast in Chlamydomonas were also selected. In contrast, the deep-learning-based method automatically extracted multi-scale features, such as intensity (gray-level), edges, and curvature, from the annotated regions and classified pixels individually. This facilitated the extraction of chloroplast regions, even in images containing structures with similar contrast (Figure 1(c)). The three-dimensional images reconstructed from the automatically segmented regions (Figures 2(a) and 2(b)) confirmed the presence of large openings and multiple micropores in the chloroplasts. Only 10 out of 60 annotation slices were required, significantly reducing manual annotation time compared to the conventional method. High reproducibility was also achieved in three-dimensional measurements. Furthermore, we acquired continuous serial-section SEM images of HIPS resin and an aluminum alloy using FIB-SEM and performed three-dimensional reconstruction combined with machine-learning-based segmentation. This presentation shows that integrating automated image acquisition with deep-learning-based segmentation streamlines the workflow from acquisition through three-dimensional reconstruction. It presents quantitative evaluation results and demonstrates the method’s utility for high-throughput three-dimensional analysis [7]. Comparison of chloroplast segmentation in Chlamydomonas. (a)An SEM image acquired using an FE-SEM equipped with ACAT, (b) the regions segmented using a threshold-based method, (c) the regions segmented using AIVIA. Three-Dimensional reconstruction of a Chlamydomonas chloroplast. (a) Three-dimensional reconstruction of the chloroplast obtained through automatic segmentation, (b) The same chloroplast viewed from a different orientation. The large opening and multiple micropores are found at the positions indicated by the white arrows.

Plant phenotyping relevance

Chlamydomonasの chloroplast 構造を対象に、機械学習セグメンテーションと3D再構築を従来法と比較・評価しており、植物構造の定量的取得ワークフローが研究の中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.