Class-conditional conformal coverage of a cellfinder classifier fails to transfer from serial two-photon training data to public light-sheet datasets

Whole-brain cell-detection pipelines such as cellfinder classify each candidate cell with a network that is often trained on another laboratory's data. Conformal prediction could flag unreliable candidates, but its coverage guarantee assumes that new data are exchangeable with the calibration data. We asked whether class-conditional conformal sets, calibrated on serial two-photon training data to contain the true label for 90% of cells and 90% of non-cells, keep that coverage on two public light-sheet datasets. Six pre-registered studies found that they did not. On CellSeg3D the classifier itself failed. On DELiVR, a classifier trained on per-cube standardised inputs was 72.9% accurate on 2,988 test candidates, yet its sets covered 52.5% of non-cells and 34.8% of cells. The failure held under two label rules, in each of four test patch groups and, in a smaller earlier test, after intensity matching. Class-conditional coverage is unaffected by changes in class proportions, so this failure means that scores shifted within each class. Controlled perturbations of a held-out source brain showed that blur, downsampling, noise and object size cause such shifts in the standardised model; noise cut cell coverage by 27.5 percentage points while accuracy fell by 1.5. No tested remedy restored reliable coverage on held-out target groups. A label-shift correction cannot repair within-class shift, and covariate-shift weighting failed. With 160 target labels from two other patch groups, recalibrated sets covered both classes at 85% or more in 56.5% of resampled label sets, against 86.0% expected without shift. Coverage should be checked on labelled target data from each acquisition group.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24369430
Primary Topic
Cell Image Analysis Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Class-conditional conformal coverage of a cellfinder classifier fails to transfer from serial two-photon training data to public light-sheet datasets

S. M. Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
Cell Image Analysis Techniques
preprint

Class-conditional conformal coverage of a cellfinder classifier fails to transfer from serial two-photon training data to public light-sheet datasets

S. M. Abdulla Al Mamun
preprint en

Abstract

Whole-brain cell-detection pipelines such as cellfinder classify each candidate cell with a network that is often trained on another laboratory's data. Conformal prediction could flag unreliable candidates, but its coverage guarantee assumes that new data are exchangeable with the calibration data. We asked whether class-conditional conformal sets, calibrated on serial two-photon training data to contain the true label for 90% of cells and 90% of non-cells, keep that coverage on two public light-sheet datasets. Six pre-registered studies found that they did not. On CellSeg3D the classifier itself failed. On DELiVR, a classifier trained on per-cube standardised inputs was 72.9% accurate on 2,988 test candidates, yet its sets covered 52.5% of non-cells and 34.8% of cells. The failure held under two label rules, in each of four test patch groups and, in a smaller earlier test, after intensity matching. Class-conditional coverage is unaffected by changes in class proportions, so this failure means that scores shifted within each class. Controlled perturbations of a held-out source brain showed that blur, downsampling, noise and object size cause such shifts in the standardised model; noise cut cell coverage by 27.5 percentage points while accuracy fell by 1.5. No tested remedy restored reliable coverage on held-out target groups. A label-shift correction cannot repair within-class shift, and covariate-shift weighting failed. With 160 target labels from two other patch groups, recalibrated sets covered both classes at 85% or more in 56.5% of resampled label sets, against 86.0% expected without shift. Coverage should be checked on labelled target data from each acquisition group.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
Cell Image Analysis Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon