Abstract
This protocol, developed by the CellMap Project Team at HHMI Janelia Research Campus, describes a standardized workflow for creating precise 3D digital boundaries (masks) for cell bodies in complex microscopy data. Specifically optimized for Expansion-Assisted Iterative Fluorescence In Situ Hybridization (EASI-FISH) samples, the method utilizes Amira 3D 2021.1 software to generate high-fidelity masks. These masks are essential for downstream quantitative analysis, such as single-molecule spot detection using Radial Symmetry-FISH (RS-FISH), ensuring that RNA transcripts are accurately localized to individual cells within a 3D tissue volume, such as the Drosophila brain.
Introduction
In advanced imaging techniques like EASI-FISH, researchers can visualize individual RNA molecules as tiny fluorescent spots within a 3D tissue sample. However, detecting these spots is only the first step; to understand gene expression at a cellular level, each spot must be assigned to a specific cell body. Automated segmentation of cell bodies in these datasets is notoriously difficult because fluorescent signals can be dim, irregular, or crowded. This protocol addresses the need for a "gold standard" semi-manual segmentation method that ensures biological accuracy for quantitative research.
The Challenge of 3D Segmentation
Fluorescent light-sheet microscopy of expanded samples provides incredible detail but presents significant challenges for automated algorithms. If a mask is too large, it captures "noise" from neighboring cells; if it contains holes or gaps, critical data is lost. The goal of this protocol is to provide a manual workflow that overcomes these hurdles by leveraging human anatomical knowledge and sophisticated software tools.
Methodology: The Amira Workflow
The workflow centers on the Amira 3D 2021.1 Classic Segmentation Workroom. The key technical steps include:
- Data Handling: Strategies for loading massive
.czi(Zeiss) files by reading complete volumes into memory and isolating specific fluorescent channels (typically Channel 0) dedicated to cell body morphology. - Threshold-Assisted Painting: Using the "Paint" tool combined with a "Masking" feature. This allows the user to set a signal threshold so that the brush only selects voxels (3D pixels) within a certain intensity range, preventing the mask from "bleeding" into the background or adjacent structures.
- 3D Interpolation: To optimize efficiency, users annotate every fifth slice of the image. The software then "interpolates" the frames in between to create a continuous 3D volume, significantly reducing the manual labor required for large datasets.
- Smoothing and Refinement: Applying kernel-based smoothing (typically kernel sizes 2–5) to even out jagged edges caused by slice-by-slice annotation, resulting in a more biologically realistic cell boundary.
- Object Individualization: Assigning each cell its own unique "material" ID, which is essential for per-cell statistical reporting and comparative analysis.
Key Anatomical Identifiers
The protocol identifies two "critical classifiers" that help manual segmentors distinguish true cell bodies from artifacts:
- The Nucleus: A dark, round region typically found in the center of the cell body.
- The Axonal "Tail": The point where the cell body narrows into an axon. The protocol explicitly instructs users to cut off the mask at this junction to ensure RNA counts are localized to the soma (cell body) rather than being diluted by axonal projections.
Additionally, users are cautioned against including the fluorescent "halo"—the diffuse light often seen around the membrane. Staying strictly inside the plasma membrane ensures the mask remains conservative and accurate.
Results and Applications
The primary output of this protocol is a 3D TIFF mask file. This spatial map can be imported into other software environments, such as Fiji, for use with the RS-FISH plugin. By combining these high-precision masks with RS-FISH's 2D and 3D spot detection capabilities, researchers can achieve single-molecule resolution that is correctly localized to individual cells in complex 3D environments.
Limitations and Future Directions
While this workflow provides a "gold standard" for accuracy, it remains labor-intensive due to the manual intervention required. It is currently best suited for high-quality ground-truth generation or smaller-scale studies. Furthermore, the reliance on proprietary software (Amira) may be a barrier for some labs. However, the manual techniques described here are instrumental in creating training data for machine learning models, which will eventually enable fully automated, high-throughput segmentation of whole-brain imaging data.
Summary updated