Skip to content

slides_4runs.md

Latest commit

 

History

History
321 lines (220 loc) · 10.4 KB

slides_4runs.md

File metadata and controls

321 lines (220 loc) · 10.4 KB
Raw

marp: true theme: default paginate: true

Cross-Task Validation

13iii26 — same fish, two tasks, four runs

Kiitan Ayandosu Phase 4 of the pipeline (p4_universal.py)

What changed since 7iii25

7iii25 (old) 13iii26 (new)
File format many tp-*.lux.h5, each (Z=7, 2048, 2048) many Cam_long_*.lux.h5, each (1, 2048, 2048)
Z-planes per timepoint 7 1
Tasks per fish 4, 5, 6 task1, task3, task6
Frames per task 42–56 50
What we can test within-task baselines cross-task generalization

We saw four flavors of imaging data this month

  • 7iii25 — multi-tp: many timepoint files, Z=7 → can do plane-split
  • 13iii26 — multi-cam: many timepoint files, Z=1 → cross-task only
  • 20iv26 — single-movie: one big file, T=4900 frames → too big for naive load
  • 12iii26 / 17iv26 — corrupt or NO_DATA_KEY headers → skipped (~257 GB cleaned)

The pipeline now auto-detects which format it is loading and adjusts accordingly.

Why we settled on 13iii26 for this round

  • Three usable task folders under one fish (task1stack0, task3stack0, task6stack0)
  • All three have the same imaging plane (Z=1) on the same animal
  • Recordings are short (50 frames each) but clean and consistent
  • No file corruption, no missing data keys
  • Lets us ask: do tuned CNMF params transfer between behavioral tasks?

13iii26 isn't multi-Z — but it gives us multi-task on the same plane, which is a different and equally useful generalization axis.

Re-framing "multi-plane" → "multi-task"

Originally the professor's split-by-plane test required Z>1.

Since 13iii26 is Z=1, we adapted by treating each task folder as the equivalent of a "plane":

  • Within-file (time-split) — tune on first half of frames, test on second
  • Cross-file (file-plane-split) — tune on task1, test on task3 (and reverse)

Same logic, different generalization axis.

The Pipeline — p4_universal.py

.lux.h5 directory  ──→  auto-detect format  ──→  load (T, H, W) movie
    ↓
512×512 downsample   ←── tradeoff: speed for proof of concept
    ↓
column-median stripe removal
    ↓
brain mask via Otsu  ←── kills components in dark periphery
    ↓
CNMF + Bayesian search (8 trials, gSig / gSig_filt / min_corr / min_pnr / rf / p)
    ↓
quality filters: circularity ≥ 0.5, max-area, in-mask centroid, SNR
    ↓
positive composite score = (1 − recon_error) + ½·compactness − 0.3·log(sparsity) + stability

Why we deferred Sweet3D / 3D segmentation

Bio team raised three concerns:

  1. "Catching multiple neurons — how to stop that"
  2. "Shouldn't capture above the planes"
  3. "Sweet3D rather than CNMF for boundaries — when neurons fire together it's hard to distinguish"

Our staged response:

Concern This phase (p4) Future phase (p5)
Over-merged blobs max-area quality filter Sweet3D segmentation
Out-of-brain detections Otsu brain mask + in-mask centroid filter Volumetric mask
Non-circular footprints circularity ≥ 0.5 filter Sweet3D circular priors

We get most of the benefit now without rewriting the pipeline from scratch.

Preprocessing — same fish, same stripe artifact

w:900

The light-sheet column-stripe pattern is identical to 7iii25 — same hardware, same fix.

Brain mask — task1 vs task3 (same fish, same plane)

Task 1 — 20.7% coverage

Task 3 — 20.2% coverage

Mask outlines (lime green) match between tasks → confirms the same imaging plane.

Run 1 — Task 1 time-split

w:700

  • 66 kept neurons (266 raw → 75% rejected by quality filters)
  • Composite score +0.28 | tune↔full stability 0.43
  • Best params: gSig=2, min_corr=0.47, min_pnr=3, rf=80, p=1

Run 1 — sample traces

w:700

Some clean transients, some flat — typical of permissive min_pnr=3 thresholds with 50 frames.

Run 2 — Task 3 time-split

w:700

  • 51 kept neurons (236 raw)
  • Composite score +0.35 | tune↔full stability 0.50
  • Best params: gSig=2, min_corr=0.51, min_pnr=3, rf=80, p=1

Almost identical params to Task 1 → pipeline is consistent across tasks.

Run 3 — Task 1 → Task 3 (cross-task transfer)

w:700

  • Tune on task1, apply best params to task3 unseen
  • 50 kept neurons in task3 with task1's params
  • Stability = 1.00 — every footprint from task1 found a match in task3

This is the headline result.

Run 4 — Task 3 → Task 1 (reverse direction)

w:700

  • Tune on task3, apply to task1
  • 56 kept neurons | stability = 0.96
  • Composite score +0.82

The transfer is bidirectional and near-perfect.

Bayesian search converged to the same answer in all four runs

Run gSig gSig_filt min_corr min_pnr rf
task1 time-split 2 3 0.47 3 80
task3 time-split 2 3 0.51 3 80
task1 → task3 2 3 0.52 3 80
task3 → task1 2 3 0.52 3 80

The optimizer finds the same regime regardless of which task we tune on.

w:700

Quality filter breakdown — what's killing components

w:900

Filter Avg rejection What it catches
circularity (≥ 0.5) ~6% Stringy / non-blob shapes
max-area ~36% Over-merged neuron blobs
in-mask ~32% Components outside the brain

≈75% of CNMF's raw output is filtered away → only ~25% survives. This is exactly addressing the bio team's "catching multiple neurons" concern.

Stability tells the cleanest story

Comparison Stability Interpretation
time-split: tune-half ↔ full movie 0.43 / 0.50 Half the tune neurons recovered when CNMF sees full movie
time-split: tune-half ↔ test-half 0.0 25 frames of test data → too sparse to recover
cross-task: task1 ↔ task3 0.96 / 1.00 Same fish, same plane → spatial footprints transfer perfectly

Cross-task transfer is dramatically more stable than within-task halves — because the limiting factor is frame count, not parameter quality.

Summary — four runs

Run Mode Raw → Kept Composite Stability
Task 1 time-split 266 → 66 +0.28 0.43
Task 3 time-split 236 → 51 +0.35 0.50
Task 1 → Task 3 file-plane-split 215 → 50 +0.86 1.00
Task 3 → Task 1 file-plane-split 234 → 56 +0.82 0.96

Verdict: parameters tuned on one task transfer essentially perfectly to another task on the same fish/plane.

What's working

  • Pipeline runs end-to-end on any of our 4 file formats (auto-detected)
  • Composite score is now positive and interpretable (was negative before)
  • Quality filters reject ~75% of CNMF noise → output is dominated by real-looking neurons
  • Cross-task footprint stability = 0.96–1.00 — the strongest result so far
  • Bayesian search converges in ~4 useful trials → cheap to repeat on new datasets

Current limitations

  • Time-split is data-starved at 50 frames per task (test-half stability = 0)
  • Brain mask too tight at 20% coverage — rejects ~32% of components, some likely real
  • Max-area filter too aggressive at default factor=4 — rejects ~36%, again partly real
  • 3 of 8 Bayesian trials wasted on unproductive min_pnr ≥ 5 regime
  • Running at 512×512 (downsampled from 2048×2048) — gives speed but loses detail

Next steps

  1. Loosen quality filters (--max-area-factor 8, dilate brain mask) → expect ~30% more kept neurons
  2. Tighten Bayesian prior to min_pnr ∈ [3, 5] → no more wasted trials
  3. Re-run at full 2048×2048 on best params from this phase (no info loss)
  4. Phase 5: Sweet3D segmentation for true single-cell boundaries when neurons co-fire
  5. Push to longer recordings (200+ frames) so time-split mode actually discriminates

Questions?

Repo: github.com/Mikito-Coder/Fish_Brain_Dynamics

All outputs (plots, traces, parameter logs, JSON summaries, master CSV) under results/13iii26_task{1,3}_*/

Script: p4_universal.py — universal loader, 4 cross-validation modes, quality filters, brain mask, Bayesian tuning.