Single-cell AI for Acute Promyelocytic Leukemia Recognition: Multicenter Validation and a Randomized Clinician-AI Collaboration Study

This is the reproduction package for the LEAP (Leukemia End-to-end Analysis Platform) manuscript. Whole-slide images and patch-level embeddings cannot be distributed, so this package is built on a frozen intermediate: per-slide, per-Expert output scores in CSV form. Every published result is a function of those scores plus routine clinical covariates, so the full results section regenerates on a laptop with no GPU and no imaging data. A single command: python reproduce.py, refits the Expert aggregation out of fold, recomputes every statistic, and prints a comparison of each reproduced value against the published one with a PASS/FAIL verdict. All 116 checks pass:APL detection across three cohorts, six molecular alteration targets, the multimodal survival model, and the randomized crossover reader study (NCT07203885). Model and training code: https://github.com/hms-dbmi/LEAP