Two short case studies: low-risk win and high-risk caution

Case A — Lab imaging QC (low risk)

A microscopy core built a small CNN to flag out-of-focus images so technicians could re-acquire them. They used an internal labeled set (5,000 images), compared a simple threshold baseline, and deployed the model as a human-in-loop tool. Validation included an external holdout from a different machine and a weekly monitoring job that reports the proportion of flagged images. Result: throughput improved and human review caught rare false positives. Key lessons: simple baseline, human oversight, light monitoring.

Case B — Predicting patient risk for clinical trial enrollment (high risk)

A translational team developed a predictive model to prioritize patients for a trial. Data contained PHI and demographic imbalances. The team paused deployment after the risk assessment showed high consequence and regulatory obligation. They performed external validation at a partner site, added calibration and uncertainty reporting, obtained IRB and legal review, and delayed any automated triage until human clinicians reviewed recommendations. Key lessons: triage with conservative thresholds, external validation, documented approvals, and mandatory human review.