Motivation
The "early" stage spans 350 minutes — by far the longest bin in the classification. Sub-staging it (2-cell, 4-cell, 8-cell, gastrulation, etc.) would dramatically improve time-to-hatching predictions and give biologists much richer developmental tracking.
Two levels of ambition
Lightweight: nuclear counting
CellPose (or a lighter segmentation model) segments nuclei in the volume. The count maps directly to sub-stage:
- 2 nuclei → 2-cell (~40-55 min post-fertilization)
- 4 nuclei → 4-cell (~55-80 min)
- 8 nuclei → 8-cell (~80-100 min)
- 50+ → gastrulation
- 200+ → late gastrulation / morphogenesis
This could be a perception tool: when the VLM classifies "early", a CellPose subagent segments and counts.
Full: lineage tracing
StarryNite/AceTree-style tracking through cell divisions. C. elegans has invariant lineage — every cell follows the same division pattern. This is still a perception problem (a different kind of perception) but operates as a separate analysis pipeline.
Key challenges
Speed
CellPose is slow. For real-time perception during acquisition, the nuclear counter needs to run within the acquisition interval (seconds, not minutes). Options:
- Lighter segmentation models (thresholding + connected components for early stages where nuclei are well-separated)
- GPU acceleration
- Run asynchronously (don't block acquisition)
Resolution is tunable
Z-step and magnification are acquisition parameters, not fixed constraints. For lineage tracing during the early cell-division period, finer z-steps and higher magnification may be worth the cost (tradeoff: temporal resolution vs spatial resolution vs phototoxicity). The perception system could even recommend acquisition parameter changes for the early period.
What does lineage tracing ADD?
Beyond stage classification: individual cell identity enables detecting division timing anomalies, asymmetric divisions, cell death events. These are scientifically interesting observations that the VLM cannot make from projections alone.
Motivation
The "early" stage spans 350 minutes — by far the longest bin in the classification. Sub-staging it (2-cell, 4-cell, 8-cell, gastrulation, etc.) would dramatically improve time-to-hatching predictions and give biologists much richer developmental tracking.
Two levels of ambition
Lightweight: nuclear counting
CellPose (or a lighter segmentation model) segments nuclei in the volume. The count maps directly to sub-stage:
This could be a perception tool: when the VLM classifies "early", a CellPose subagent segments and counts.
Full: lineage tracing
StarryNite/AceTree-style tracking through cell divisions. C. elegans has invariant lineage — every cell follows the same division pattern. This is still a perception problem (a different kind of perception) but operates as a separate analysis pipeline.
Key challenges
Speed
CellPose is slow. For real-time perception during acquisition, the nuclear counter needs to run within the acquisition interval (seconds, not minutes). Options:
Resolution is tunable
Z-step and magnification are acquisition parameters, not fixed constraints. For lineage tracing during the early cell-division period, finer z-steps and higher magnification may be worth the cost (tradeoff: temporal resolution vs spatial resolution vs phototoxicity). The perception system could even recommend acquisition parameter changes for the early period.
What does lineage tracing ADD?
Beyond stage classification: individual cell identity enables detecting division timing anomalies, asymmetric divisions, cell death events. These are scientifically interesting observations that the VLM cannot make from projections alone.