Interpretable Multi-Feature Optical Analysis for Stage- and Batch-Aware Quality Assessment of Brain-Organoid Cultures

Reliable quality assessment of brain organoids is important for reproducible culture and downstream experimentation, yet routine evaluation relies largely on qualitative brightfield inspection. We investigated whether visual QA criteria could be represented by interpretable image descriptors while accounting for developmental stage and culture batch. We analyzed 692 brightfield image–ROI pairs from six culture batches. Binary QA labels were assigned by one expert. We compared 115 handcrafted features with simple morphology and frozen pretrained DINOv2 and ResNet-50 representations using logistic-regression and random-forest classifiers. Leave-one-batch-out (LOBO) testing was the primary exploratory evaluation; repeated image-level cross-validation and maturation-only analyses provided complementary assessments. Handcrafted logistic regression achieved a pooled LOBO ROC AUC of 0.836 (conditional 95% batch bootstrap interval 0.625–0.960), with substantial variation among held-out batches. Mean repeated-CV AUC was 0.933 across all stages and 0.895 within maturation. At a fixed 0.5 score cutoff, all-stage LOBO sensitivity was 0.656 and specificity was 0.869. The comparison concerns the evaluated frozen-feature pipelines and does not establish superiority over fine-tuned deep learning. These results provide an exploratory image-analysis baseline for brightfield QA; label reproducibility and performance beyond the observed laboratory workflow require further validation.