smith6jt-cop Tile SNAKE Pattern Stitching - Research Notes
Experiment Overview
| Item | Details |
|---|---|
| Date | 2024-12-15 |
| Goal | Correctly stitch multiplex microscopy tiles acquired in SNAKE pattern |
| Environment | KINTSUGI pipeline, Python 3.10+, CuPy GPU |
| Status | Success |
Context
CODEX and similar multiplex imaging systems acquire tiles in a SNAKE pattern (also called serpentine or boustrophedon). This pattern alternates direction on each row to minimize stage movement. Incorrect tile ordering causes stitching failures with misaligned or scrambled images.
SNAKE vs RASTER Pattern
SNAKE Pattern (Default for CODEX)
Even rows scan left-to-right, odd rows scan right-to-left:
Row 0: → → → (tiles 1, 2, 3)
Row 1: ← ← ← (tiles 6, 5, 4)
Row 2: → → → (tiles 7, 8, 9)
RASTER Pattern
All rows scan left-to-right:
Row 0: → → → (tiles 1, 2, 3)
Row 1: → → → (tiles 4, 5, 6)
Row 2: → → → (tiles 7, 8, 9)
Verified Workflow
Building Tile Coordinates for SNAKE Pattern
from itertools import chain, repeat
n_rows = 2 # Grid rows
n_cols = 2 # Grid columns
# Row indices (same for both patterns)
rows = list(chain.from_iterable(repeat(row, n_cols) for row in range(n_rows)))
# Column indices - SNAKE PATTERN
# Even rows: 0,1,2... (L→R)
# Odd rows: ...2,1,0 (R→L)
cols = list(chain.from_iterable(
range(n_cols) if row % 2 == 0 else range(n_cols - 1, -1, -1)
for row in range(n_rows)
))
# For 2x2 grid:
# rows = [0, 0, 1, 1]
# cols = [0, 1, 1, 0] <- SNAKE: second row reversed
Test Data Tile Reordering
When test data tiles are renamed by POSITION (1,2,3,4 for grid positions) but were acquired in SNAKE order, reorder before stitching:
# Files load as position order: [tile1, tile2, tile3, tile4]
# Grid positions:
# [1] [2] row 0 (L→R)
# [3] [4] row 1 (should be R→L for snake)
#
# For snake acquisition, row 1 was acquired as [4, 3] not [3, 4]
# So actual acquisition order was: 1, 2, 4, 3
#
# Reorder to match acquisition order:
if n_rows == 2 and n_cols == 2:
corrected = corrected[[0, 1, 3, 2]] # [1,2,3,4] → [1,2,4,3]
Failed Attempts (Critical)
| Attempt | Why it Failed | Lesson Learned |
|---|---|---|
Using RASTER pattern cols = [0,1,0,1] |
Tiles misaligned at row boundaries | CODEX uses SNAKE pattern - odd rows are reversed |
| Not reordering position-named test tiles | Stitching scrambled because tile order didn't match stitch model | Test data tiles named by position need reordering to match snake acquisition |
| Assuming tile filenames reflect acquisition order | Test data was renamed by grid position, not acquisition order | Always verify whether tile names are position-based or acquisition-order-based |
Final Parameters
SNAKE Pattern Coordinates (2x2 grid)
rows = [0, 0, 1, 1]
cols = [0, 1, 1, 0] # NOT [0, 1, 0, 1] (that's RASTER)
SNAKE Pattern Coordinates (3x3 grid)
rows = [0, 0, 0, 1, 1, 1, 2, 2, 2]
cols = [0, 1, 2, 2, 1, 0, 0, 1, 2] # Row 1 reversed
General Formula
cols = list(chain.from_iterable(
range(n_cols) if row % 2 == 0 else range(n_cols - 1, -1, -1)
for row in range(n_rows)
))
Key Insights
- SNAKE pattern is standard for CODEX and similar multiplex imaging systems
- The pattern minimizes stage travel time between tiles
- Even rows: left-to-right (columns 0,1,2...)
- Odd rows: right-to-left (columns ...2,1,0)
- When test data tiles are renamed by position, they must be reordered to match acquisition order before stitching
- The stitch model is computed from tile positions, so tile array order must match the coordinate lists
References
- CODEX multiplex imaging documentation
- KINTSUGI Notebook 1 (Single Channel Eval) and Notebook 2 (Cycle Processing)