BioImage Hackathon — BEH Dataset

Briefing page: https://www.dcs.warwick.ac.uk/~u1898019/hackathon-beh/ (full pipeline + findings + caveats).

GPU-accelerated pipeline for nested cell segmentation and dynamics quantification on a 30-file epiflourescense timelapse dataset. Built for the Warwick BioImage Analysis Hackathon (April 2026).

Live briefing page: https://www.dcs.warwick.ac.uk/~u1898019/hackathon-beh/

Team

Name	Role
Haili Yuan (CS)	pipeline, dynamics, visualization
Edward Offord (CS)	segmentation, cell centering
Badeer Ummat (biology)	dataset, scientific question

Dataset at a glance

Property	Value
Files	30 (3 conditions × 10 positions)
Conditions	WT (Wild-Type), KO (Knock-Out), KI (Knock-In)
Modality	2D time-lapse epifluorescence, single channel
Acquisition	Zeiss + Micro-Manager 2.0, 60 s frame interval
Pixel size	~0.318 µm/px
Bit depth	16-bit unsigned
Frames per file	67 to 360 (varies by position)
Total dataset	1.3 GB, ~5,700 frames

Key finding: two intensity batches

While inspecting the data we discovered each condition cleanly splits into two intensity batches, with a 5-10x gap in pixel intensity. Almost certainly two separate imaging sessions, not biological variation. Implications:

• Cannot normalise globally across all 30 files
• Statistical comparison should use a mixed-effects model with batch as a random effect
• Findings should hold across both batches to be considered robust

Condition	Low batch (mean ~200-240)	High batch (mean ~500-700)
WT	wt_5, wt_9, wt_23, wt_31, wt_37	wt_wt4, wt_wt9, wt_wt13, wt_wt18, wt_wt25
KO	ko_3, ko_8, ko_17, ko_21, ko_25	ko_11, ko_1717, ko_2121, ko_28, ko_333
KI	ki_3, ki_11, ki_14, ki_23, ki_25	ki_1, ki_6, ki_13, ki_18, ki_28

Biology context (Day 2, confirmed by biology lead)

The fluorescence labels Arp2/3 (Actin Related Proteins 2/3 complex), which nucleates branched actin networks. Arp2/3 is recruited to the edge of cell protrusions; the inactive pool diffuses through the cytoplasm. So in our images:

• Brighter pixels = lamellipodia (Arp2/3 concentrated at the leading edge)
• Dimmer pixels = cytoplasmic Arp2/3 pool

The perturbed gene is important for both cell protrusion and cell adhesion to the substrate:

• WT: normal coordination between protrusion and adhesion
• KO (gene completely deleted): impaired cell adhesion
• KI (point mutation introduced): coordination between adhesion and protrusion is disrupted

Biological process being characterised: cell migration, following the canonical 4-step model:

1. Protrusion of the leading edge (Arp2/3-driven, lamellipodia)
2. Adhesion of the protrusion to the substrate (integrin-mediated)
3. Generation of traction forces (actomyosin contractility)
4. Release of older adhesions at the rear

This framing turns our segmentation/centering pipeline into the front-end of a canonical cell-migration analysis. The trajectory data we already saved (per-frame centroids in centering_manifest.json) contains the migration readout directly.

Pipeline

raw .tif (T, H, W)
  → adaptive intensity normalize (only for low-batch files, p99 < 1000)
  → Cellpose cyto3 segmentation (GPU)
  → keep largest mask, fall back to last centroid if empty
  → shift image so cell sits at image center
  → save *_centered.tif (uint16) + *_mask.tif (uint8 0/255)

The adaptive step is the key: low-batch files need their dynamic range stretched before Cellpose can see structures (raw signal sits in the bottom 1% of uint16), but high-batch files already span enough range that the stretch creates saturation artefacts. The decision threshold (p99 < 1000) cleanly separates the two batches without hardcoding filenames.

Run

On the GPU partition (recommended)

sbatch pipeline/run_centering.sbatch

The job uses 1 GPU on the falcon or gecko partition, 8 CPUs, 32 GB RAM, with a 2-hour wall-clock limit. On an NVIDIA RTX A5000 the full 30-file dataset processes in ~17 minutes.

Output goes to ~/Desktop/BioImageHackathon_BEH_centered/ with:

BioImageHackathon_BEH_centered/
├── wt/  (10 paired _centered.tif + _mask.tif)
├── ko/
├── ki/
└── centering_manifest.json   (per-file metadata + per-frame centroids)

Locally on CPU (slow, for prototyping)

python pipeline/process_all_centering.py

CPU processing is ~12.6 s/frame versus ~0.18 s/frame on GPU.

Demos

Day 0 baseline 1: Multi-Otsu 4-layer segmentation

Static mid-frame segmentation per condition using only classical thresholding (no deep learning). Verifies the data IO and tooling chain runs end-to-end.

Day 0 baseline 2: Dynamics through time

Same approach extended to full timelapse with global thresholds. Bottom plot shows per-frame ring intensity for the entire time-lapse (745 frames across 3 conditions).

Cellpose + cell centering (Edward's pipeline)

Per-frame Cellpose segmentation followed by centroid-based image shift. Sanity check on wt_wt9.tif (3 frames, CPU).

Failure case: low-batch + Cellpose cyto

Running Cellpose with default settings on ko_8 (low batch) returns empty masks on 88% of frames. Same algorithm works fine on ki_18 (high batch). The intensity histogram at the bottom shows why: the low batch occupies less than 1% of the available 16-bit dynamic range.

Fix: per-frame normalization

Stretching each frame to its 1-99 percentile range before segmentation. Now Cellpose sees consistent contrast on both batches.

Pipeline iterations (v1 → v2 → v3)

We ran three GPU iterations to handle the low-vs-high batch issue cleanly:

Version	Strategy	Outcome
v1	cyto model on raw images	Works on high batch (0-4% empty). Fails on low batch (e.g. ko_8 88%, ko_21 87% empty).
v2	cyto3 + 1-99 percentile normalize on every frame	Fixes low batch (ko_21 87% → 0.6%) but breaks 3 high-batch files (wt_wt25 0% → 77%, ko_28 0% → 34%).
v3	adaptive: normalize only when p99 < 1000	Best of both. Final hybrid output is v2 low-batch results + v3 high-batch results.

Total empty mask rate across the 5,263 frames in the dataset:

Iteration	Empty rate	Files with >15% empty
v1	9.7% (511 / 5263)	5
v2	7.9% (416 / 5263)	4
v3 (final)	2.8% (148 / 5263)	2 (ko_8, ko_28)

Per-file comparison

Sorted by v1 difficulty (worst on the left). Files in [low] brackets are low-batch (intensity stretched in v3), [high] are high-batch (raw input in v3).

v3 sample outputs

Four representative files. Left: raw input. Middle: Cellpose mask overlaid in green. Right: image shifted so the cell sits at the centre. Both rescued low-batch files and stable high-batch files work cleanly through the same pipeline.

ko_8 (50% empty) and ko_28 (17% empty) remain difficult — ko_8's raw signal is the lowest in the dataset and ko_28 has the largest field of view, so cyto3 may benefit from manual diameter tuning. Both flagged for the biology lead to review.

Validation on the Categorised_Data dataset

Mid-hackathon, Badeer uploaded a new categorized dataset organized as {wt,ko,ki}{1,2}/X.tif. We cross-verified that his 1 / 2 folder split is identical to our intensity-based low / high batch detection (100% concordance on every sampled file). 22 of the 30 files are new positions extending the original dataset (5,263 → 6,221 frames, +18%).

Re-running the v3 pipeline on this expanded dataset:

Folder	Files	Empty / Total	%
ki1 (KI low)	5	0 / 900	0.0%
ki2 (KI high)	5	17 / 1,467	1.2%
ko1 (KO low)	5	133 / 630	21.1% ← problem cases here
ko2 (KO high)	5	2 / 1,434	0.1%
wt1 (WT low)	5	22 / 701	3.1%
wt2 (WT high)	5	2 / 1,089	0.2%
TOTAL	30	176 / 6,221	2.83%

The headline number is essentially unchanged from the original dataset (2.81% vs 2.83%), which means the pipeline generalises cleanly to new positions.

The ko1 folder concentrates the failures: ko1/4.tif (95% empty, only 60 frames so likely an aborted or very short recording) and ko1/8.tif (50% empty, identical to the original ko_8.tif — confirmed source-data limitation, not a pipeline issue). Together they account for 117 of the 176 empty masks. A plausible biological reading is that the KO knockout itself produces dimmer cells with weaker GFP signal, so segmentation difficulty in this folder may be part of the phenotype rather than an artefact. Worth confirming with Badeer.

Sample outputs across all 6 folders

One representative file from each folder. Same pipeline (cyto3 + adaptive normalize) handles low-batch and high-batch input cleanly, and the centering step is robust across the wide size variation (smallest FOV 165×165 in ki1/7, largest 594×456 in ko2/ko11 and 558×507 in ko2/ko5).

Lamellipodia separation, direction sweep

Edward added a percentile-based separator (pipeline/otsu_threshold.py) that splits each cell mask into "cell body" (brighter pixels) and "lamellipodia" (dimmer pixels) using a tunable percentile threshold. To help the biology team pick the right convention before we run it on the full dataset, we ran a parameter sweep on wt2/wt4.tif covering both directions and three percentile values, plus a Multi-Otsu 3-class variant and a distance-transform-constrained variant.

• Red = cell body, Blue = lamellipodia in every overlay.
• Top row is Edward's original convention (brighter pixels = body).
• Bottom row swaps it (brighter pixels = lamellipodia).
• Last column on top is Multi-Otsu 3-class (body / transition / lamellipodia all separable in one shot).
• Last column on bottom adds a spatial constraint: a pixel is only labelled lamellipodia if it is also within ~12 px of the cell boundary.

The biology lead can pick whichever overlay best matches the expected morphology, which fixes the convention and the percentile value in one decision. Once chosen, the separator runs over the whole 30-file dataset in a few seconds per file.

The biology lead has now confirmed the direction: brighter pixels = lamellipodia (Arp2/3 concentrates at the leading edge), so the bottom row of the prototype is the correct convention. Edward's otsu_threshold.py defaults still need to be flipped accordingly.

Cell migration analysis (Day 2)

Since the v3 pipeline already saved per-frame centroids during the centering step, we can compute canonical cell-migration metrics directly from centering_manifest.json, with no additional GPU work required.

Metrics extracted (pixel size 0.318 μm/px, frame interval 60 s):

• Mean / median speed (μm/min)
• Total path length (cumulative travel)
• Net displacement (start to end straight line)
• Persistence index (= net / total, 1 = directed motion, 0 = random walk)

Mask-stability filter (Day 2 afternoon)

Edward noticed that wt1/33 had a suspiciously high speed (5.55 μm/min). On inspection we found Cellpose was fragmenting the cell mask in some frames — the centroid would jump from a 10,000-pixel mask to a 200-pixel fragment and back, producing fake 30 μm/min "displacements". See demos/wt1_33_diagnostic.png and demos/wt1_33_mask_diagnostic.png.

We added a per-pair mask-stability filter to extract_migration.py:

• Frame valid iff mask area >= 30% of per-cell median
• Pair valid iff both frames valid AND consecutive masks overlap with IoU >= 0.3

After filtering, mean IoU during accepted pairs is 0.80, and the worst single-frame jumps (>15 μm/min) drop out of the metric — wt1/33 now reports 1.88 μm/min instead of 5.55, in line with the rest of the WT cohort.

Per-condition summary (N=10 each, mask-stability filter applied)

Condition	Mean speed	Median speed	Total path	Net disp.	Persistence
WT	1.42 ± 0.47 μm/min	1.00 ± 0.30 μm/min	228 ± 141 μm	43 ± 20 μm	0.26 ± 0.15
KO	1.15 ± 0.47 μm/min	0.89 ± 0.33 μm/min	237 ± 152 μm	68 ± 53 μm	0.35 ± 0.29
KI	1.06 ± 0.60 μm/min	0.73 ± 0.14 μm/min	242 ± 144 μm	50 ± 27 μm	0.25 ± 0.16

Three coherent patterns:

1. WT migrates fastest (1.42 μm/min vs KO 1.15 vs KI 1.06), consistent with intact protrusion / adhesion coordination supporting effective traction.
2. KO is slower but most persistent (persistence 0.35 vs WT 0.26), consistent with impaired adhesion compressing path length but trajectories remaining the most directional. KO cells slip in straighter lines, perhaps because adhesion failures prevent the complex curving moves WT cells do.
3. KI is slowest with the lowest persistence (1.06 μm/min, persistence 0.25), the "worst of both worlds": broken protrusion-adhesion coordination costs both speed and directionality simultaneously, more functionally damaging than complete loss of the gene.

Trajectories per folder

Each subplot shows 5 cells. Coloured solid lines: pairs that passed the mask-stability filter. Light grey dashes: rejected segments (Cellpose mask fragmentation, real motion uncertain). Black dot: start of valid trajectory. Coloured star: end. Equal aspect ratio across panels so paths are visually comparable.

Caveats

• N=10 per condition (5 cells per session × 2 imaging sessions). Variability is large; statistical tests are pending.
• ko1/4.tif has 95% empty masks, so it has zero valid pairs after filtering and contributes nothing to the KO summary.
• Imaging session is a confound; the next analysis pass should fit a mixed-effects model with batch as a random effect and condition as a fixed effect.
• The mask-stability filter currently rejects any pair with IoU < 0.3 — this is a conservative threshold. A handful of cells (notably wt1/33, wt1/12, wt2/wt10, ko1/8) have many frames excluded; their reported metrics are based on the surviving valid pairs.
• No photobleaching correction yet, so intensity-based metrics may carry some imaging artefact even after the centering / size normalisation.

Cellpose self-training (Day 2 afternoon)

Edward suggested fine-tuning Cellpose on our own data using v3 outputs as pseudo-ground-truth, instead of relying on hand-drawn annotations we don't have. We built a 3-stage pipeline and ran it.

Training setup

1. Prepare pseudo-GT: take v3 Cellpose masks, exclude ko1/4 + ko1/8
   (known stubborn cases), filter out frames with mask-fragmentation
   (area < 30% of per-cell median, or IoU < 0.5 with previous frame),
   subsample every 10th frame to keep training set manageable.
   → 586 (image, mask) pairs across 28 cells
2. Fine-tune cyto3 with --min_train_masks 1 (single-cell masks),
   100 epochs, SGD lr=0.05, weight_decay=1e-4, batch_size=8.
   → 12 minutes on RTX A5000, train loss 0.23 → 0.016 (14x reduction)
3. Re-run the v3 pipeline with the fine-tuned model on all 30 cells.
   → Output to Categorised_Data_finetuned/, parallel to the baseline

Results: 22x reduction in empty masks

	Baseline (cyto3 zero-shot)	Fine-tuned cyto3	Improvement
Total empty rate	176 / 6221 (2.83%)	8 / 6221 (0.13%)	22×
WT empty	1.34%	0.06%	20×
KO empty	6.54%	0.24%	27×
KI empty	0.72%	0.08%	9×

The two stubborn cases that resisted every previous fix are largely solved:

Cell	Baseline empty	Fine-tuned empty
ko1/4	57/60 (95%)	5/60 (8%)
ko1/8	60/120 (50%)	0/120 (0%)
ko1/16	15/180 (8%)	0/180 (0%)
wt1/12	13/180 (7%)	0/180 (0%)
wt1/33	8/81 (10%)	0/81 (0%)
ki2/4	13/240 (5%)	0/240 (0%)

Mean IoU between baseline and fine-tuned masks per cell averages 0.83-0.90 across conditions — i.e. the model still detects the same cell, but adjusts mask boundaries to be more reliably populated. It is not a different segmentation, it is a better-behaved one.

Six representative cells, mid-frame each. Green = baseline mask, purple = fine-tuned mask, difference column with green/red/blue showing agreement / baseline-only / finetuned-only pixels.

Why this works

The pseudo-GT is filtered v3 output. The fine-tuned model is essentially learning "produce the kinds of masks v3 produces when v3 produces good masks." That is enough to:

• Make boundaries smoother and more consistent across consecutive frames (the mask-stability filter we added earlier rejected ~127 frames; the fine-tuned model produces fewer such frames in the first place)
• Recognise dim cells that the off-the-shelf cyto3 missed (ko1/8 with weak signal, ko1/4 with truncated recording)
• Specialise to the imaging conditions of this dataset (epifluorescence, this objective, this exposure)

Honest caveats

• Self-training can amplify systematic biases in the pseudo-GT. We did not fix those biases by training on the masks they produced.
• The 22× empty-mask reduction is on the same dataset the training data was drawn from. Generalisation to a held-out experiment is untested.
• For genuine model improvement assessment, hand-drawn ground truth would still be the right reference. Self-training is a useful intermediate step, not a substitute.

Migration re-analysis with fine-tuned masks

We re-ran the migration pipeline on the fine-tuned outputs (centroids from Categorised_Data_finetuned/). The clean-up of mask-fragmentation events drops noise dramatically, and the biological story sharpens:

Condition	Baseline (filtered)	Fine-tuned	Change
WT mean speed	1.42 ± 0.47 μm/min	1.03 ± 0.33 μm/min	std down
KO mean speed	1.15 ± 0.47 μm/min	1.06 ± 0.29 μm/min	std down
KI mean speed	1.06 ± 0.60 μm/min	0.68 ± 0.08 μm/min	std 7.5× smaller
WT persistence	0.26 ± 0.15	0.31 ± 0.16	up
KO persistence	0.35 ± 0.29	0.38 ± 0.17	up, std halved
KI persistence	0.25 ± 0.16	0.31 ± 0.14	up

Updated reading of the three patterns:

1. WT and KO migrate at similar speeds (~1.0 μm/min). The original "WT fastest" interpretation was inflated by a few mask-fragmentation outliers in WT cells (wt1/12, wt1/33 had artificial bursts). Once the mask noise is gone, KO cells move just as fast as WT.
2. KO retains the highest persistence (0.38), now with half the variance (0.17 vs 0.29). Impaired adhesion still results in straighter trajectories.
3. KI is dramatically slower than both WT and KO (0.68 vs 1.03/1.06 μm/min, Cohen's d > 1.4 against either). With only 0.08 μm/min std, KI cells are remarkably consistent in being slow. This is the cleanest single signal in the dataset.

Biology reading: broken protrusion-adhesion coordination (KI) is much more functionally costly than complete loss of the gene (KO). KO cells can still migrate at WT speed, just less manoeuvrable; KI cells barely migrate at all. Consistent with the idea that the KI point mutation produces a non-functional product that interferes with normal migration machinery, rather than simply being absent.

Mask comparison utility (Day 2 afternoon)

Edward asked for an automated way to compare Cellpose mask coverage against a reference segmentation. We wrote a generic utility (pipeline/compare_masks.py) that takes any two parallel mask directories and outputs per-frame metrics:

• IoU (Jaccard index)
• Dice coefficient
• Centroid distance (px)
• Area difference and area ratio
• Boundary IoU (intersection over union on 1-pixel-thick boundaries only)
• Hausdorff 95 (95th-percentile boundary distance, robust to outliers)

Demo: Cellpose v3 vs whole-image Otsu baseline

To prove the pipeline works while waiting for hand-drawn reference masks, we ran it on Cellpose v3 vs an Otsu-threshold baseline (whole-image threshold, morphological cleanup, keep largest connected component).

Condition	Mean IoU	Mean Dice	Mean centroid distance	Mean HD95
WT (N=10)	0.798 ± 0.070	0.886	6.9 px	19.0 px
KO (N=10)	0.789 ± 0.069	0.875	8.5 px	23.7 px
KI (N=10)	0.850 ± 0.078	0.917	4.0 px	14.2 px

Cellpose and Otsu agree on roughly 80% of the cell area on average across the dataset — Otsu is a respectable baseline for these images. KI cells have the highest agreement (0.85), KO the lowest (0.79). Worst single cell: ko1/16 at IoU 0.63 with HD95 80 px (mask boundaries strongly disagree).

Six representative cells, mid-frame each. Green = both masks agree, red = Cellpose only, blue = Otsu only. Per-cell IoU / Dice / HD95 in the rightmost panel.

How to swap in a reference

When real ground-truth masks become available, point the same utility at the new directory:

python pipeline/compare_masks.py \
    --a /path/to/cellpose_masks/ \
    --b /path/to/reference_masks/ \
    --out output_dir/

The utility will pair files by matching relative paths and emit comparison_per_frame.csv and comparison_summary.csv with the same metrics. Drop-in replacement.

Generalisation test on a fresh batch (Day 2 EOD)

A second upload of 30 fresh cells (2ndUpload/, same 6 categories, 5,862 frames) arrived after the main analysis. We re-ran the full pipeline on it using the same fine-tuned model with no retraining or parameter changes. This is a strong test of generalisation: the model never saw these cells.

Segmentation quality

Batch	Cells	Frames	Empty	Empty rate
Batch 1 (original 30)	30	6,221	8	0.13%
Batch 2 (fresh 30)	30	5,862	75	1.28%

Empty rate is 10× higher on fresh cells but still under 2%. 25 of 30 fresh cells had zero empty masks. Five had non-zero rates, with ko1/7 the worst (28%). Expected pattern: fine-tuning generalises broadly but does not solve every imaging artefact.

Findings reproduce

Migration speed (μm/min):

	Batch 1	Batch 2
WT	1.03 ± 0.33	0.95 ± 0.24
KO	1.06 ± 0.29	1.19 ± 0.43
KI	0.68 ± 0.08	0.72 ± 0.10

Lamellipodia / cytoplasm intensity ratio:

	Batch 1	Batch 2
WT	1.53 ± 0.29	1.60 ± 0.37
KO	1.47 ± 0.45	1.64 ± 0.56
KI	1.83 ± 0.33	1.83 ± 0.27

Cohen's d (effect size):

	Batch 1	Batch 2
Migration KI vs WT	-1.40	-1.22
Migration KI vs KO	-1.68	-1.44
Lam ratio KI vs WT	+0.90	+0.67
Lam ratio KI vs KO	+0.85	+0.41

The KI mean lam/cyto ratio is 1.827 in batch 1 and 1.830 in batch 2: the two means agree within 0.003 on independent groups of 10 cells. With samples this small this is not formal proof, but agreement on independent data is a strong signal that the result is not specific to the cells used during self-training. Both findings (KI is slowest, KI is most polarised) reproduce.

Per-cell trajectories on the fresh batch (KI panels in the bottom row cover noticeably shorter ground than WT and KO):

Reproduce

# Segment + center on the new batch
sbatch pipeline/run_2nd_with_finetuned.sbatch
# Downstream analyses
python pipeline/extract_migration_2nd.py
python pipeline/extract_lamellipodia_2nd.py
# Side-by-side comparison
python pipeline/compare_batch1_vs_batch2.py

Validation under a biologist-trained model (Day 3 morning)

The biology team supplied 14 cells with Quimp active-contour reference masks (Fiji + biologist supervision, file label seg4CellPose_Retrain). We trained a parallel Cellpose model on these biologist-validated masks (no overlap with our self-training pipeline) and ran the entire downstream analysis again on both batches. The original self-trained model and all earlier results are untouched.

Setup

• 14 cells, ~2,500 frame-mask pairs after stride 5 sampling, un-shifted from Quimp's centered frame back to raw frame.
• Same hyperparameters as self-training: cyto3 base, 100 epochs, SGD lr=0.05.
• Train loss 0.335 → 0.014 in 9 minutes on RTX A5000.
• Inference on all 60 cells (batch 1 + batch 2) takes ~17 minutes.

Empty-mask rate (segmentation quality)

Model	Batch 1	Batch 2
Self-trained	0.13%	1.28%
Quimp-trained	2.12%	1.40%

Quimp model is slightly less aggressive on Batch 1 (only 14 of 30 cells were in its training set; the rest are unseen). Both stay under 2.5% — production-grade quality.

Lamellipodia / cytoplasm intensity ratio

Condition	self_b1	quimp_b1	self_b2	quimp_b2
WT	1.53 ± 0.29	1.56 ± 0.34	1.60 ± 0.37	1.63 ± 0.40
KO	1.47 ± 0.45	1.50 ± 0.50	1.64 ± 0.56	1.70 ± 0.59
KI	1.83 ± 0.33	1.88 ± 0.35	1.83 ± 0.27	1.86 ± 0.29

KI lam/cyto ratio reproduces with essentially identical magnitude across both models on both batches (1.83-1.88). Cohen's d for KI vs WT is +0.90 (self_b1) versus +0.87 (quimp_b1), and +0.67 (self_b2) versus +0.64 (quimp_b2). The two segmentation philosophies (Cellpose deep-learning vs Quimp active contour) converge on the same number.

Migration speed (μm/min)

Condition	self_b1	quimp_b1	self_b2	quimp_b2
WT	1.03 ± 0.33	1.04 ± 0.45	0.95 ± 0.24	1.03 ± 0.34
KO	1.06 ± 0.29	1.36 ± 1.01	1.19 ± 0.43	1.20 ± 0.48
KI	0.68 ± 0.08	0.79 ± 0.14	0.72 ± 0.10	0.91 ± 0.21

KI is the slowest condition under all four model × batch combinations. Cohen's d shifts from "very large" (-1.40 self_b1) to "medium-large" (-0.71 quimp_b1) because Quimp's larger boundary makes the centroid more sensitive to halo shape changes, which inflates within-condition variance. The direction of the finding is preserved.

Visual side-by-side

Six representative cells, mid-frame each, raw-frame overlay. Blue = self-trained mask, red = Quimp-trained mask, green = both. The pattern is consistent: the self-trained model produces a tighter cell-body mask, the Quimp-trained model adds a halo ring around it (q/s area ratio ~1.17 across the 30 cells of batch 1). Despite the boundary-philosophy difference, both models land on the same cell and the downstream lam/cyto ratio is essentially unchanged (1.83 vs 1.88 for KI), because the percentile-50 split inside the mask separates the bright leading-edge signal from dim interior signal regardless of where the boundary is drawn.

Conclusion

The KI dominant-negative interpretation is not an artefact of the self-training pseudo-ground-truth. The lamellipodia finding survives a switch to biologist-validated training data with essentially the same Cohen's d. The migration finding survives in direction and ranking; effect size is somewhat smaller because of boundary-philosophy differences.

Reproduce

# Build (image, mask) training pairs (un-shifts Quimp masks back to raw frame)
python pipeline/prepare_training_data_quimp.py
# Fine-tune cyto3 on Quimp masks (~9 min on A5000)
sbatch pipeline/run_finetune_quimp.sbatch
# Run inference on both batches with the new model (~17 min)
sbatch pipeline/run_quimp_inference.sbatch
# Downstream analyses
python pipeline/extract_migration_quimp.py
python pipeline/extract_lamellipodia_quimp.py
# Cross-model comparison + Cohen's d
python pipeline/compare_self_vs_quimp.py
# Optional: agreement metrics (IoU/Dice/HD95) between Cellpose and Quimp masks
python pipeline/compare_quimp_vs_cellpose.py finetuned
python pipeline/compare_quimp_vs_cellpose.py v3

Files

Path	Description
pipeline/process_all_centering.py	Main pipeline: Cellpose segment + centroid + shift
pipeline/run_centering.sbatch	SLURM submission script for falcon/gecko GPU partitions
pipeline/otsu_threshold.py	Lamellipodia separator (Edward) — Otsu / percentile threshold inside cell mask
pipeline/extract_trajectories.py	Per-cell lamellipodia / cytoplasm metrics from v3 outputs
pipeline/extract_migration.py	Cell-migration metrics (speed, path, displacement, persistence) from saved centroids
pipeline/compare_masks.py	Generic mask-comparison utility (IoU, Dice, Hausdorff, etc.)
pipeline/demo_cellpose_vs_otsu.py	Demo runner: Cellpose vs Otsu baseline
pipeline/process_2nd_with_finetuned.py	Run fine-tuned pipeline on the 2nd-upload batch (generalisation test)
pipeline/run_2nd_with_finetuned.sbatch	SLURM submit for the generalisation test
pipeline/extract_migration_2nd.py	Migration metrics, 2nd batch
pipeline/extract_lamellipodia_2nd.py	Lamellipodia metrics, 2nd batch
pipeline/compare_batch1_vs_batch2.py	Side-by-side comparison: per-condition stats + figure
pipeline/plot_trajectories.py	Generic 2x3 trajectory plot (one panel per folder) from any trajectories.json
pipeline/prepare_training_data_quimp.py	Build training pairs from Quimp reference masks (un-shifts to raw frame).
pipeline/run_finetune_quimp.sbatch	SLURM script: fine-tune cyto3 on Quimp masks.
pipeline/process_with_quimp.py	Run the Quimp-trained model on both batches.
pipeline/run_quimp_inference.sbatch	SLURM script for the above.
pipeline/extract_migration_quimp.py	Migration metrics under the Quimp model.
pipeline/extract_lamellipodia_quimp.py	Lamellipodia metrics under the Quimp model.
pipeline/compare_self_vs_quimp.py	Side-by-side comparison: self-trained vs Quimp-trained, both batches.
pipeline/compare_quimp_vs_cellpose.py	Per-frame IoU/Dice/HD95 between Cellpose mask and Quimp reference.
pipeline/inspect_quimp_mismatch.py	Visual inspection figure for low-IoU cells (frame-level overlay).
pipeline/visualize_self_vs_quimp_masks.py	Side-by-side figure: 6 cells, raw-frame overlay of self-trained (blue) and Quimp-trained (red) masks.
models/cellpose_1777550975.949237	Self-trained cyto3 weights (26 MB). See models/README.md.
models/cellpose_1777590610.8133328	Quimp-trained cyto3 weights (26 MB). 14 cells × 502 pairs from biologist-validated reference.
data/categorised_data_manifest.json	First 30-file dataset manifest
data/categorised_data_2nd_manifest.json	Second 30-file dataset manifest (generalisation batch)
data/per_cell_summary.csv	Lamellipodia metrics, batch 1 (30 cells)
data/lamellipodia_per_cell_summary_2nd_finetuned.csv	Lamellipodia metrics, batch 2 (30 cells)
data/migration_metrics.csv	Migration metrics, batch 1 (30 cells)
data/migration_metrics_2nd_finetuned.csv	Migration metrics, batch 2 (30 cells)
data/batch1_vs_batch2_summary.csv	Per-condition summary across both batches
data/cohens_d_batch1_vs_batch2.csv	Effect-size comparison across both batches
data/empty_rate_batch1_vs_batch2.csv	Per-cell segmentation quality across both batches
data/cellpose_vs_otsu_summary.csv	Per-cell mask agreement metrics, Cellpose vs Otsu
demos/*.png	All demo screenshots embedded in this README

Status (Day 2 EOD)

• Per-file manifests built for both datasets (original BioImageHackathon_BEH and Categorised_Data).
• Three baseline demos running end-to-end (Multi-Otsu static, Multi-Otsu dynamics, Cellpose centering).
• Cellpose pipeline iterated v1 → v2 → v3 with adaptive normalize, then self-trained on filtered v3 outputs to produce a fine-tuned cyto3.
• Empty-mask rate cut from 9.7% (v1) to 0.13% (fine-tuned) on the original 30 cells: a 75-fold reduction.
• GPU pipeline runs the full dataset in ~10 minutes on an A5000 with the fine-tuned model.
• Lamellipodia direction confirmed by biology lead: brighter pixels = lamellipodia (Arp2/3 concentrated at the leading edge).
• Biology context aligned: GFP labels Arp2/3, gene important for protrusion + adhesion coordination, process is cell migration.
• Migration finding on 30 cells: WT 1.03, KO 1.06, KI 0.68 μm/min (Cohen's d ~1.4 to 1.7).
• Lamellipodia finding on 30 cells: WT 1.53, KO 1.47, KI 1.83 lam/cyto ratio (Cohen's d ~0.9).
• Combined biology reading: KI cells over-polarise Arp2/3 yet migrate ~35% slower. Broken protrusion-adhesion coordination is more functionally costly than complete deletion. Dominant-negative interpretation.
• Generalisation test: same fine-tuned model on a fresh batch of 30 cells (2ndUpload/). Empty rate 1.28% (vs 0.13% on the original cells, 9.95x in relative terms but still under 2% in absolute terms), both findings reproduced. KI mean lam/cyto ratio is 1.827 (batch 1) vs 1.830 (batch 2): means agree within 0.003 on independent groups of 10 cells.
• Generic mask-comparison utility ready for any future ground-truth annotation.

Open questions

Most of the original day-1 questions have been answered by the biology lead. Still open:

1. What gene specifically is perturbed in KO and KI? Function is described (protrusion + adhesion coordination) but the gene is not named.
2. Are ko1/4 and ko1/8 technical or biological? Both are very dim. ko1/4 was a 60-frame aborted recording, ko1/8 stays consistently dim.
3. Why does ko1/7 still fail on the fresh batch (28% empty) when 25 of 30 fresh cells worked perfectly?
4. Should we refine to 3 layers (body / transition zone / lamellipodia) instead of 2?

License

Internal hackathon code. Dataset belongs to the Warwick CAMDU group and is not redistributed via this repository.