GT Protocol Guidelines

Protocol clarifications and best practices for Ground Truth segmentation, consolidated from the team's directives.

Mitochondria segmentation

• Individual segmentation: each mitochondrion gets its own segment ID and color — don't paint all mitochondria as one segment.
• Oversegment: mitochondria must be oversegmented individually. (An earlier task description saying "do not oversegment mitochondria" was a mistake and should be disregarded.)
• Membranes: don't paint mitochondrial membranes; leave a visible border/gap around each mitochondrion.

Myelin segmentation

• Inner boundary: no padding between myelin and the cell it ensheathes — the myelin segment should directly contact the inner cell.
• Myelin-to-myelin: no padding between adjacent myelin segments; they may touch.
• Outer boundary: myelin should approach but not touch neighboring cell membranes — keep a small gap.

Upsampled images

• Two-pass labeling: first paint the original (thicker) sections, then copy labels onto the upsampled (thinner) sections and connect objects naturally.
• Cross-reference resolutions: working only from the upsampled image is misleading — always check against the original image data.
• Artifacts: ignore "hallucinations" (artificial structures in upsampled data); focus on anatomically sensible structures.

Ambiguous & edge cases

• Uncertain organelles: if unsure whether an object is a mitochondrion (vs. peroxisome, lysosome, etc.), segment it as a mitochondrion — it can be removed during review.
• EM damage: follow team consensus for handling damaged regions.

Workflow & quality assurance

• Mandatory peer review: before marking a task finished, at least one other team member must inspect your work — prioritize sharing the areas where you struggled.
• Sharing for review (WebKnossos): Tasks → search for tasks under the relevant project → click the + for the task → Actions → View.

WebKnossos tools

• Interpolation: volume interpolation helps for objects perpendicular to the sectioning plane in higher z-resolution data; may need manual cleanup.
• 3D mesh: right-click a segment to generate a 3D mesh; press . to toggle between all views and the view under the cursor — useful for QC.
• SAM2 (experimental): AI-assisted segmentation on the experimental server; most effective for larger, convex cross-sections; adjust confidence and erosion parameters, and use existing annotations as hints.