Owner-provided biology: approximately 1 µm width and 5–10 µm rod length, with mixed sizes. All wells received WGA, but not all cells have detectable labeling. No known biologically unlabeled wells are currently identified. At the current calibration, these priors correspond to about 9 pixels wide and 46–92 pixels long; they are starting expectations, not hard exclusion rules.
The owner reports possible focus loss partway through the first acquisition round. Audit by actual acquisition sequence and available time metadata, with z-stack and biological-feature review. Fixed camera dust can remain sharp when bacteria are out of focus, so whole-image sharpness alone is insufficient. Distinguish usable sample focus, suspected focus failure, empty with independent focus evidence, and unresolved. Check whether a blank-looking interval is followed by a return to good biological focus. A cell surface outside all five planes may be unrecoverable.
Previously selected background-only fields are suspended as confirmed empty controls pending this audit. Do not convert failed focus to zero counts; track usable fields and valid area separately. Inspect original slices and sample regions before selecting correction templates.
Prepare numbered cell and cluster crops with brightfield, FITC, and original z slices, retaining experiment/well/field/pixel coordinates. Suggested labels are provisional until reviewed. Show clear and difficult examples of round versus long cells, low versus clear WGA, localized versus broad WGA, rosette candidates, camera artifacts, and uncertain attachment. Do not claim that every requested subtype is present or observable.
| Stage | Work | Evidence required to proceed |
|---|---|---|
| 1. Reference gallery | Review representative cells and clusters together; record corrections and a short labeling guide. | Shared definitions, accepted example IDs, and explicit uncertain/unobservable categories. |
| 2. Background and dust correction | Use float64. Compare channel-specific background scale/offset and spatial patterns. Mask persistent severe artifacts where correction is unreliable. Preserve negative values. | Before/after crops; held-out candidate-background residuals; visible weak cells preserved. Report excluded image area rather than inventing cells behind artifacts. |
| 3. Cell detection and orientation | Detect brightfield candidates; measure area, width, length, aspect ratio and roundness in physical units. Inspect z behavior. Test whether the median projection suppresses moving objects without losing desired cells. | Numbered overlays showing missed, false, merged and duplicate detections. Roundness is an orientation proxy, not proof of attachment. |
| 4. Per-cell WGA | Measure corrected FITC in/around each cell against local background. Use maximum projections for screening and original z planes for position. Record continuous signal measures as well as tentative labels. | Reviewed clear-positive, weak and absent examples. Without unlabeled cells, label thresholds are operational and require review; empty fields alone do not measure cellular autofluorescence. |
| 5. Rosettes | Flag candidate groups and distinguish central from broad fluorescence only after reviewing actual examples. Preserve cluster IDs and cell membership where resolvable. | No double-counting cells; unresolved cluster cell numbers remain unresolved rather than being counted as one cell. |
| 6. Held-out pilot | Freeze parameters on tuning fields; evaluate separate wells/fields that were not used for tuning. | Detection errors, count errors, orientation and WGA agreement, and uncertainty rates reported separately. Revise if a failure mode is material. |
| 7. Full run | Process remaining data by experiment, with QC flags and example images per well. | Proceed after the pilot meets agreed criteria and its overlays have been reviewed. |
Proposed starting fields: first folder D10 F1/F3, C10 F2 and A11 F2; second folder A02 F4 and A05 F3. Include second-folder C09 F1/F3 as provisional background checks. D10 fields belong in the same tuning or validation partition to avoid splitting one well across both. Add a suitable third-folder field after the gallery review. These choices cover varied morphology, labeling, crowding and cluster candidates; they are not biological control assignments.
Five planes at 0.3 µm spacing cover only 1.2 µm from first to last plane. An upright 5–10 µm cell may extend beyond the stack. Fluorescence in the highest acquired plane must not automatically be called the top pole; bottom/top/all-over labels require evidence of the cell’s observed extent and the biological surface direction. Keep “unobservable” distinct from WGA absent.
Start with a small numbered gallery, then assemble approximately 50–100 independently reviewed cell labels spanning the classes and difficult cases. Include full-field count checks on held-out fields. Suggested discussion targets are ≤10% relative count error and ≥90% agreement for confidently observable orientation and WGA presence; these are proposals, not accepted or achieved thresholds. Report precision/recall, merges/splits and the uncertain fraction so good total counts cannot hide wrong labels.
Cell-level records: experiment, well, field, cell ID, coordinates, measurements, attachment status, roundness/orientation, WGA signal/category, observed z extent, cluster ID and QC reason. Well-level tables: total confidently attached cells, polar/upright cells, polar cells with detectable WGA, flat cells, flat cells with detectable WGA, plus uncertain counts, rosette summaries, fields and valid image area. Attached cells with uncertain orientation can make the total exceed polar + flat; show that explicitly.
The selected publishing target is WeissLab/AdhesionScreening96. The owner has said the anonymized report may be public. Publish a portable static report with HTML, previews and tables; keep raw ND2 data and local environments on D:. The owner has created the repository privately. The current GitHub connection cannot yet access it. Password-protected Render hosting is prepared locally, with proposed address adhesion96.weisslab.ca; DNS is managed in Cloudflare. No remote upload or deployment has succeeded.