gavinlouuu/mib-cells-different-focus

MIB Cells in Different Focus

Private archival snapshot of the YOLO-guided MIB (microfluidic imaging biophysics) detection run over the "Cells in different focus" experiment. Four cell lines (C2C12, HEK293, HeLa, PANC-1) were recorded on the MIB instrument at seven focus-actuator voltages (40 V to 70 V in 5 V steps; 50 V is the nominal in-focus setting). Every row is one YOLO detection with its ROI crop, the MIB segmentation mask, and the MIB morphology metrics (area, area ratio, deformability, ring ratio, brightness quartiles) computed by the Biowork MIB processing core. These are the same quantities that mib-studio-qt produces during live processing.

Summary

• 399,287 detection rows across 23 recordings, 284,381 distinct frames
• 135,852 rows pass the MIB validity checks (is_valid)
• Source frames are 1184x240 8-bit grayscale; crops are the MIB ROI (YOLO box + 25 px padding)
• Pixel size 0.4886 um/px (pixel_to_micron)
• Detector: gavinlouuu/mib-notebook-yolo-v8n, confidence 0.05, IoU 0.7
• MIB processing core 0.2.0, profile yolo_guided, border check enabled, single inner contour required, all range checks disabled

Recordings

Cell line	Focus (V)	Rows	Valid rows	Frames	Source recording
C2C12	40	5,281	4,190	4,635	C2C12_in different focus/40v.h5
C2C12	45	8,998	2,529	7,735	C2C12_in different focus/45vh5.h5
C2C12	50	45,145	2,543	31,243	C2C12_in different focus/50v.h5
C2C12	55	15,324	324	11,812	C2C12_in different focus/55v.h5
C2C12	60	14,701	4,810	11,428	C2C12_in different focus/60v.h5
C2C12	65	12,197	7,663	9,784	C2C12_in different focus/65v.h5
C2C12	70	7,835	5,858	6,819	C2C12_in different focus/70v.h5
HEK293	50 (in focus)	16,498	146	11,026	HEK293_in different focus/50v_in focus.h5
HEK293	55	19,613	1,614	12,294	HEK293_in different focus/55v.h5
HeLa	40	11,867	7,447	9,081	Hela_in different focus/40v.h5
HeLa	45	15,219	11,543	11,745	Hela_in different focus/45v.h5
HeLa	50	26,446	7,076	16,612	Hela_in different focus/50v.h5
HeLa	55	22,695	1,485	14,997	Hela_in different focus/55v.h5
HeLa	60	20,322	3,281	13,839	Hela_in different focus/60v.h5
HeLa	65	21,453	11,002	14,406	Hela_in different focus/65v.h5
HeLa	70	17,506	12,453	12,821	Hela_in different focus/70v.h5
PANC-1	40	4,418	3,349	3,437	PANC-1/40v.h5
PANC-1	45	4,109	1,960	3,259	PANC-1/45v.h5
PANC-1	50 (in focus)	43,551	11,123	30,040	PANC-1/50_in focus.h5
PANC-1	55	18,197	1,521	12,582	PANC-1/55v.h5
PANC-1	60	4,968	1,653	3,642	PANC-1/60v.h5
PANC-1	65	38,468	28,727	27,600	PANC-1/65.h5
PANC-1	70	4,476	3,555	3,544	PANC-1/70v.h5

Six source files could not be read by the pipeline and are absent:

• C2C12_in different focus/40v_remasked.h5: Unable to synchronously open file (bad object header version number)
• HEK293_in different focus/40v.h5: Object visitation failed (addr overflow, addr = 2104, size = 328, eoa = 2048)
• HEK293_in different focus/45v.h5: Object visitation failed (addr overflow, addr = 2104, size = 328, eoa = 2048)
• HEK293_in different focus/60v.h5: Object visitation failed (addr overflow, addr = 2104, size = 328, eoa = 2048)
• HEK293_in different focus/65v.h5: Object visitation failed (addr overflow, addr = 2104, size = 328, eoa = 2048)
• HEK293_in different focus/70v.h5: Object visitation failed (addr overflow, addr = 2104, size = 328, eoa = 2048)

The five HEK293 files are 8 KB stubs on the share (truncated recordings), so HEK293 only has 50 V and 55 V.

Columns

Column	Meaning
image	PNG grayscale ROI crop from the raw h5 frame (roi_xywh region)
mask	PNG binary mask (0/255) of the MIB contour, same shape as image
cell_line, focus_voltage_v, nominal_in_focus	Experimental condition parsed from the folder and file name
recording, recording_stem	Source h5 path relative to the experiment folder
record_index	Row index within the original mib_processing_manifest.jsonl
frame_index, frame_timestamp_ns	Frame position in recorded_frames/images and the camera timestamp from recorded_frames/metadata
yolo_detection_index, yolo_confidence, yolo_bbox_xyxy	YOLO detection within the frame (pixel coordinates in the full frame)
roi_xywh	Crop origin and size in full-frame pixels
object_id, object_count	MIB object id inside the ROI (-1 means MIB found no valid contour) and contour count
frame_type, is_valid, touches_border, in_range, is_target_group, has_single_inner_contour, inner_contour_count	MIB validity flags
area_px, area_um2, area_ratio, deformability, ring_ratio, youngs_modulus	MIB morphology metrics (youngs_modulus is 0 because the E-modulus model was disabled)
brightness_q1 to brightness_q4	Brightness quartiles inside the contour
laplacian_variance	Contract-2 focus metric computed on the Contract-1 object: variance of the 3x3 Laplacian of the raw ROI crop over the filled object contour (science::calculateLaplacianVariance, mib-studio-qt). NaN when no object
laplacian_contour_area_px	Area of the contour the Laplacian was evaluated on (equals area_px / area_ratio; validation column)
laplacian_core	Identity of the Laplacian core that produced the two columns above
track_*	Tracking fields, unused in this run (track_id is -1)
pixel_to_micron	Calibration used for area_um2

Rows with object_id == -1 are YOLO detections where MIB found no usable contour; their metrics are zero and the mask is empty. Keep them if you want to study detector versus segmenter disagreement, filter on is_valid for morphology analysis.

Configs

• default (data/*/*.parquet): Contract 1 detections (saturating background subtraction, ring ratio) with the Laplacian focus metric added on top of the Contract-1 objects (2026-09-24).
• contract2 (data_contract2/*/*.parquet): the same YOLO ROIs re-processed under Processing Contract 2 (mib-processing 0.3.1: cv::absdiff background comparison, objects are top-level contours, ring width abolished, per-object Laplacian variance). One row per Contract-2 object; contract1_object_count / contract1_valid_count link back to the default config. Extra columns: centroid_xy and bbox_xywh in full-frame pixels, channel_band_rows (the channel detected on the recording's background with scripts/channel_roi.py, a port of mib-studio-qt detectChannelRoi, issue #295) and in_channel (object centroid inside that band; false marks channel-wall objects to reject). The core's own gates are unchanged from the Contract-1 run (area gate off), so filter on is_valid & in_channel and an area band for cell-level analysis. Background per recording is the mean of 128 YOLO-negative frames (indices in provenance/recordings/*/contract2_reprocess.json).

Analysis

analysis/ holds the ring-ratio versus Laplacian-variance comparison across focus voltage (figure, hexbin, summary CSV, notes).

Usage

from datasets import load_dataset
ds = load_dataset("gavinlouuu/mib-cells-different-focus", split="train")
valid = ds.filter(lambda r: r["is_valid"])

import pyarrow.parquet as pq
t = pq.read_table("data/HeLa/HeLa_50v.parquet", columns=["focus_voltage_v", "area_um2", "deformability", "is_valid"])

Structure

data/<cell_line>/<cell_line>_<voltage>v[_infocus].parquet
provenance/run/EXPORT_MANIFEST.json            per-recording counts from the export
provenance/run/run_manifest.json               pipeline run log including skipped inputs
provenance/run/sparse-mib-contours-v3-status.json
provenance/recordings/<cell_line>/<stem>/mib_processing_results.json   MIB config, background, throughput per recording
provenance/recordings/<cell_line>/<stem>/mib_processing_metrics.csv    original metrics table per recording
dataset_metadata.json
checksums.sha256

Provenance

Raw recordings live on the lab NAS at DC/RW experiment/Cells in different focus. The YOLO inference run inference-dataset-20260818T074408Z and the MIB processing run of 2026-08-19 were exported by Biowork as mib-contours-inference-dataset-20260818T074408Z. This dataset was assembled on 2026-09-21 by cropping the ROI of each manifest record from the raw h5 frames and decoding the packed masks from mib_artifacts.tar. Full frames are not included; they total roughly 190 GB and remain on the NAS.

Limitations

• Detections come from a YOLOv8n trained on 24 positive frames, run at confidence 0.05, so many rows are false positives or fragments. is_valid is the MIB filter, not a human label.
• Focus voltage is a nominal actuator setting, not a measured defocus distance.
• HEK293 is only available at 50 V and 55 V.
• No public-data licensing review has been done; keep the repository private until ownership and release terms are confirmed.

Intended use

Internal research on focus sensitivity of MIB morphology metrics, segmentation model training and evaluation, and comparison with mib-studio-qt live processing. Not for clinical or diagnostic use.