Ground Truth Protocol Guidelines
Tommy Macrina's Directives (August - September 2025)
Executive Summary
This document compiles protocol clarifications and best practices for ground truth segmentation tasks as directed by Tommy Macrina in the #putracers Slack channel between August 1 and September 3, 2025. All guidelines are referenced to specific Slack conversations for verification.
I. Mitochondria Segmentation Protocol
1.1 Individual Segmentation Required
Each mitochondrion must be assigned a separate segment ID and color. Do not paint all mitochondria with the same segment.
Reference: Message m10
1.2 Oversegmentation Directive
The original task description stating "do NOT oversegment mitochondria" was incorrect and should be disregarded. Mitochondria must be oversegmented individually.
Reference: Message m8
1.3 Membrane Handling
Do not paint mitochondrial membranes. Leave a visible border/gap around each mitochondrion to represent the membrane boundary.
References: Message m8, Message m15
II. Myelin Segmentation Protocol
2.1 Inner Boundary (Cell Interface)
No padding between myelin and the cell it ensheathes. The myelin segment should directly contact the inner cell.
Reference: Message m15
2.2 Myelin-to-Myelin Contact
No padding between adjacent myelin segments. Myelin segments may touch each other.
Reference: Message m15
2.3 Outer Boundary (Adjacent Cells)
Myelin should approach but not touch the membranes of neighboring cells. Maintain a small gap between myelin exterior and other cell membranes.
Reference: Message m15
III. Upsampled Image Best Practices
3.1 Two-Pass Labeling Approach
Recommended workflow:
- First pass: Paint annotations on the original image (thicker sections)
- Second pass: Copy labels onto the upsampled image and connect objects naturally
Reference: Message m2
3.2 Multi-Resolution Viewing
Working exclusively from upsampled images can be misleading. Always cross-reference with original image data.
References: Message m3, Message m4
Visual Reference: Side-by-side comparison
3.3 Handling Artifacts
"Hallucinations" (artificial structures in upsampled data) should be ignored. Focus on anatomically sensible structures.
References: Message m1, Message m11
IV. Ambiguous Cases & Edge Cases
4.1 Uncertain Organelles
When uncertain whether an object is a mitochondrion (vs. peroxisome, lysosome, etc.), segment it as a mitochondrion. Objects can be removed during review if needed.
Reference: Message m9
4.2 EM Damage Regions
Follow team consensus for handling EM damage. Specific protocols to be determined based on damage type.
Reference: Message m17
V. Workflow & Quality Assurance
5.1 Mandatory Peer Review
Before marking tasks as finished, at least one other team member must inspect your work. Prioritize sharing areas where you struggled.
Reference: Message m13
5.2 Sharing Annotations for Review
Correct procedure in WebKnossos:
- Navigate to Tasks
- Search for tasks under the relevant project
- Click the plus symbol (+) for the task
- Select "Actions" button (right side)
- Choose "View"
References: Message m13, Message m14
VI. WebKnossos Tools & Techniques
6.1 Interpolation Tool
The volume interpolation feature is particularly useful for objects perpendicular to the sectioning plane in higher z-resolution data. May require manual cleanup.
Reference: Message m7
Documentation: WebKnossos Interpolation Guide
6.2 3D Mesh Visualization
- Right-click on any segment to generate a 3D mesh
- Press
. (period) to toggle between all views and the view under cursor
- Useful for quality control and identifying errors
Reference: Message m12
Documentation: Keyboard Shortcuts
6.3 SAM2 AI Segmentation (Experimental)
New AI-assisted segmentation tools available on experimental server:
- Most effective for larger, convex cross-sections without many intercellular objects
- Adjust confidence threshold and erosion parameters for optimal results
- Can use existing segment annotations as hints for consistency
Reference: Message m18
VII. Additional Resources
Key Discussion Threads
- m8: Mitochondria oversegmentation clarification thread
- m15: Comprehensive myelin padding protocol discussion
- m18: SAM2 AI tool implementation and testing
External Links
Reference Messages
Below are the complete Slack message threads referenced throughout this document.
Message & Thread Index
•
m1 — putracers — 2025-08-05 19:38 UTC — Hallucination handling directive
•
m2 — putracers — 2025-08-05 19:41 UTC — Two-pass labeling approach
•
m3 — putracers — 2025-08-05 19:41 UTC — Upsampled image caution
•
m4 — putracers — 2025-08-05 20:18 UTC — Side-by-side comparison link
•
m5 — putracers — 2025-08-05 20:41 UTC — Volume visibility troubleshooting
•
m6 — putracers — 2025-08-05 20:41 UTC — Segmentation visibility confirmation
•
m7 — Thread in putracers — 2025-08-01 13:34 UTC — Interpolation tool discussion
•
m8 — Thread in putracers — 2025-08-01 15:25 UTC — Mitochondria oversegmentation clarification
•
m9 — Thread in putracers — 2025-08-04 18:54 UTC — Ambiguous organelle decision
•
m10 — Thread in putracers — 2025-08-05 14:31 UTC — Individual mitochondria segmentation
•
m11 — Thread in putracers — 2025-08-05 19:38 UTC — Difficult hallucinations discussion
•
m12 — Thread in putracers — 2025-08-05 20:03 UTC — 3D meshing usage
•
m13 — Thread in putracers — 2025-08-05 20:20 UTC — Peer review requirement
•
m14 — Thread in putracers — 2025-08-06 14:19 UTC — Task sharing troubleshooting
•
m15 — Thread in putracers — 2025-08-06 15:17 UTC — Comprehensive protocol best practices
•
m16 — Thread in putracers — 2025-08-14 17:58 UTC — Task transition and IP updates
•
m17 — Thread in putracers — 2025-08-18 19:37 UTC — EM damage handling
•
m18 — Thread in putracers — 2025-08-31 19:07 UTC — SAM2 AI tool deployment
Standalone Messages by Tommy Macrina
From my spot checks, you've all done well: just ignore them.
Context: Response regarding how to handle "hallucinations" in upsampled segmentation data.
I realize now that we should probably handle labeling this data in two passes:
1. paint the original image (thicker sections)
2. copy those labels onto the upsampled image (computationally thinner sections),
and just connect the objects as "naturally" as possible
I've found only working from the upsampled image to be misleading.
Here's a side-by-side of the original image and the upsampled image:
https://spelunker.cave-explorer.org/#!middleauth+https://global.daf-apis.com/nglstate/api/v1/5712377138905088
Open left sidebar and see if "volume" is flipped on?
I see your segmentation still.
Threaded Discussions with Tommy Macrina
Thread Summary: Discussion of WebKnossos interpolation tool effectiveness
Ryan (2025-08-01 13:34:20 UTC):
Oh that's what that does. That would be useful with the larger, simpler segments
we typically start off with, but the smaller neurites need more guidance. I agree
with Jay that it might require clean up but I'll give it a try
Tommy Macrina (2025-08-01 15:30:16 UTC):
Yeah, this has been our general experience, too.
Tommy Macrina (2025-08-01 15:31:38 UTC):
Because this image has a higher z-resolution, more objects that are perpendicular
to the sectioning plane don't change that much. So you might find more objects
where interpolation is useful.
Ryan (2025-08-01 15:59:35 UTC):
Yea I tried it on a simple round segment with minimal movement between slices with
a 2 slice gap and it was pretty perfect. But I tried 5-6 slices in between and it
got a bit confused by the size change in the segment and under-colored. But I'm
gonna keep playing with it and find the sweet spot
Thread Summary: Critical correction regarding mitochondria segmentation protocol
Jay (2025-08-01 15:25:29 UTC):
@Tommy Macrina @Eric Are we definitely oversegmenting mitochondria?
The task directions explicitly say not to.
Attached: image.png
Tommy Macrina (2025-08-01 15:26:57 UTC):
Yes, it was a mistake to assign that task description.
Jay (2025-08-01 15:27:16 UTC):
Ok, thanks for the confirmation, we'll oversegment!
Eric (2025-08-01 15:27:41 UTC):
Sorry for the confusion, just was getting to this
bun (2025-08-01 15:27:46 UTC):
do we leave a border between mitos?
bun (2025-08-01 15:27:57 UTC):
for membrane
Tommy Macrina (2025-08-01 15:28:20 UTC):
Yes, try not to paint the mitochondria's membrane.
bun (2025-08-01 15:30:37 UTC):
is there a way we can get Kisuk's tool that expands or reduces the segments
so we can make the membrane unpainted while still using the fill tools?
Tommy Macrina (2025-08-05 19:43:23 UTC):
> Kisuk's tool that expands or reduces the segments
This is some post-processing he does. Handling it manually for now is the
right call, until we have more experience with this data.
Tommy Macrina (2025-08-05 19:43:50 UTC):
It'd be nice if webknossos made it easy to extend their toolkit...
[Thread continues with discussion of shape deforming tools and interpolation - see full message archive for complete context]
Thread Summary: Guidance on handling uncertain organelle identification
bun (2025-08-04 18:54:04 UTC):
should i over segment this object in the cell body? i'm not 100% sure its a
mitochondria. it might be a peroxisome or lysosome or something else
@Tommy Macrina @Eric
Attached: image.png
Tommy Macrina (2025-08-05 19:33:04 UTC):
Link or location?
Tommy Macrina (2025-08-05 19:34:15 UTC):
It's not jumping out at me in your volume. I think your volume looks pretty good.
bun (2025-08-05 19:34:48 UTC):
295127, 258550, 3573
bun (2025-08-05 19:35:24 UTC):
i went ahead and segmented it. figured we could always get rid of it if we need to.
Tommy Macrina (2025-08-05 19:35:29 UTC):
Yeah, I think you make the right call to segment it.
Thread Summary: Clarification on segment ID assignment for mitochondria
Jay (2025-08-05 14:31:09 UTC):
@Eric @Tommy Macrina Point of clarification: should all mitochondria be segmented
with the same segment/color, or should they each be given an individual segment/color?
Sorry, it's been so many years since I've done one of these tasks that I don't
remember what the correct protocol is
Tommy Macrina (2025-08-05 19:27:41 UTC):
Different segment for each mitochondria.
Thread Summary: Discussion of challenging artifacts in upsampled data
Tommy Macrina (2025-08-05 19:38:27 UTC):
Does anyone have a "hallucination" that they found particularly difficult?
bun (2025-08-05 19:52:18 UTC):
294885, 258672, 3576,
294908, 258553, 3589,
295009, 258619, 3565
bun (2025-08-05 19:52:41 UTC):
just off the top of my head. these areas were a bit tricky
Ryan (2025-08-06 15:09:27 UTC):
So far, because these slices are artificial, I've just done my best to annotate
something that makes sense with the flow/movement and what makes sense with known
morphology, even if it slightly goes against the shape/borders the up-sampling
came up with
Thread Summary: Introduction to 3D visualization features in WebKnossos
Tommy Macrina (2025-08-05 20:03:05 UTC):
How much have you all used the ad hoc meshing in webknossos?
Jay (2025-08-05 20:14:40 UTC):
The interpolation? I tried it out for a while, but I found it produced enough
errors that I was spending just as much time cleaning it up as I would be
tracing the slices myself so I stopped using it
Tommy Macrina (2025-08-05 20:17:36 UTC):
No. If you right-click on a segment, you can generate a 3D mesh.
Jay (2025-08-05 20:24:52 UTC):
I didn't even realize there was a 3D view. How do you access it? I don't see
any menu options, and the "keyboard shortcuts" help option is just a 404 error
Tommy Macrina (2025-08-05 20:46:19 UTC):
Hit "." to toggle between all views and the view you hover over.
Tommy Macrina (2025-08-05 20:46:28 UTC):
Docs for webknossos are online, too
Tommy Macrina (2025-08-05 20:46:29 UTC):
https://docs.webknossos.org/webknossos/ui/keyboard_shortcuts.html
Thread Summary: Mandatory quality control process
Tommy Macrina (2025-08-05 20:20:14 UTC):
Before calling these tasks finished, would you mind having at least one other
team member inspect your work? You might share the places where you struggled
the most.
bun (2025-08-05 20:39:13 UTC):
can you remind us how to do that in webknosses. i can't seem to "View" other
peoples tasks
Tommy Macrina (2025-08-05 20:45:02 UTC):
Go to tasks, search for tasks under project `mec_cutout_0_upsampled`, click on
plus symbol for any task, go to `actions` button on the right side, select `View`.
Attached: image.png
bun (2025-08-05 20:48:52 UTC):
thanks!
Thread Summary: Technical issues with viewing shared annotations
[Extended thread discussing WebKnossos UI challenges and annotation sharing - abbreviated for brevity]
Tommy Macrina (2025-08-06 15:22:36 UTC):
No idea how that link doesn't want to load the image data.
Tommy Macrina (2025-08-06 15:22:57 UTC):
If you work through tasks > actions > view, it shows up:
http://wk.zetta.ai:9000/annotations/688a4519010000db0079097b#295030,259728,3539,0,0.35
Tommy Macrina (2025-08-06 15:40:04 UTC):
Hit `.`
Tommy Macrina (2025-08-06 15:40:59 UTC):
Ohhh... maybe Austin copied the task to his own account. That could explain
the missing image.
Thread Summary: Comprehensive discussion of segmentation best practices
bun (2025-08-06 15:17:12 UTC):
As for the GT best practices I know Tommy mentioned the cutouts shouldn't have
had that description applied: "Segment all cells. Segment myelin as a separate
object, without padding between the myelin and neighboring objects. Pay special
attention to segment any invaginations, or invagination-like structures. Do NOT
oversegment mitochondria. Leave extracellular space unsegmented."
I think the most up to date ground truth best practice would be this: "Segment
all cells. Segment myelin as a separate object, *without padding between* the
myelin and neighboring objects. Pay special attention to segment any invaginations,
or invagination-like structures. Segment mitochondria *without membrane painted*.
Segment extracellular space."
@Tommy Macrina Is that fair to say?
Ryan (2025-08-06 15:29:49 UTC):
I think we're supposed to avoid the mitochondria's border/membrane:
[link to earlier message]
Jay (2025-08-06 15:30:26 UTC):
Yeah to clarify, I think the following schema should be correct for myelin:
1. No padding between myelin and the cell it surrounds
2. No padding between myelin and other myelin
3. Myelin should butt up against the cell membranes of other cells, but not
touch the actual segment
Attached: image.png
bun (2025-08-06 15:35:06 UTC):
Yes I think this is mostly correct. The outside of myelin shouldn't touch other
cells/membranes, but it might be the case that it can touch other myelin. Not
100% about that though. The inside should touch the cell it is myelinating like
shown. @Tommy Macrina
bun (2025-08-06 15:35:41 UTC):
@Ryan yes, "Segment mitochondria *without membrane painted*."
Ryan (2025-08-06 15:36:34 UTC):
Oh yes I definitely misread that!
Tommy Macrina (2025-08-06 15:39:49 UTC):
Yeah, that's right, @bun
Thread Summary: Project management and technical access
Jay (2025-08-14 17:58:06 UTC):
@Austin.Burke @bun @twister @wileykit @Ryan This round of GTs are up; if y'all haven't been invited to the project yet LMK & I'll get that taken care of
[Thread primarily regarding task scheduling and IP whitelist updates]
Tommy Macrina (2025-08-14 21:30:22 UTC):
@bun yeah, if your IP address changes, it needs to be updated.
Thread Summary: Handling electron microscopy damage in segmentation
Austin.Burke (2025-08-18 19:37:34 UTC):
@Tommy Macrina What would be the protocol for EM damage like this? Should it be
left alone as is or segmented as the same segID or a separate one?
Attached: image.png
Tommy Macrina (2025-08-20 13:16:58 UTC):
Agree with @bun below
[Note: bun's recommendation not included in Tommy's message; refer to full thread for context]
Thread Summary: New AI-assisted segmentation capabilities
ranl (2025-08-31 19:07:47 UTC):
@channel, I deployed a experimental webknossos server at http://ranl.zetta.ai:9000/,
upgraded to the new 25.09 release and include new features like SAM2. The new
server does not contain any old data, so you have to register again. Migrate the
database might be difficult so we probably switch after the current tracing tasks
are finished, meanwhile you can give it a try and report if there are any issues.
Tommy Macrina (2025-09-02 10:38:04 UTC):
@ranl Can you give me access?
ranl (2025-09-02 12:49:03 UTC):
done
Tommy Macrina (2025-09-02 13:56:33 UTC):
Thanks!
Tommy Macrina (2025-09-02 14:06:09 UTC):
Pretty good at bigger convex cross sections without many intercellular objects. Cool!
Tommy Macrina (2025-09-02 14:06:29 UTC):
Got it working on some objects for 20 sections just pointing & clicking.
[Thread continues with technical discussion of SAM2 parameters and performance - see full archive for complete details]
Document Information
Source: Slack workspace (#putracers channel)
Date Range: August 1 - September 3, 2025
Primary Subject Matter Expert: Tommy Macrina
Compilation Date: November 6, 2025
Total Messages Referenced: 18 (6 standalone + 12 threaded discussions)
Images: Referenced images are stored in tommy_macrina_images/ directory
This document was automatically generated from Slack message exports and formatted for professional distribution. This Google Docs-ready version has been optimized for copy/paste into Google Docs with simplified styling and no interactive JavaScript features.
Document Control
Created: November 6, 2025
Last Updated: November 6, 2025
Owner: Benjamin Silverman
Review Frequency: As needed when processes change