Machine Learning-Assisted Morphological Assessment of Stem Cell Colony Quality

Abstract

Background: Automated morphological assessment of stem cell colonies is widely reported to exceed ninety per cent accuracy,
but published figures rest on incompatible validation designs.
Objective: To read the field through the validation regime attached to each accuracy figure and ask what the reported numbers
actually measure.
Methods: Narrative review of colony imaging studies, model benchmarks and inter-observer agreement data.
Findings: On 165,190 images across 32 cell lines and 93 batches, one model reached 99.9% accuracy under a batch-stratified split
and 89.1% under a batch-separated split. A model trained on single-cell bright-field images reached 0.923 accuracy on the held-out
20% but 71.89-82.73% on separate experiments using inducers absent from training. Two experts agreed on 48% of 449 colony
images.
Conclusions: The evidence supports reporting split construction with every accuracy and grouping splits by batch and donor; it
does not support ranking published accuracies against one another.