Inference-Time ROI Policy for Domain-Robust Deep Autofocus in Smear Microscopy: Code, Models, Results, and Supplementary Evidence

This record contains source code, frozen model checkpoints, calibration metadata, support indices, fixed-model predictions, policy-level metrics, statistical analyses, tables, figures, the main manuscript draft, and comprehensive supplementary evidence for a study of inference-time region-of-interest selection in smear-microscopy autofocus. Seventeen ROI policies were evaluated across white blood cell (WBC), peripheral blood smear (PBS), and bone marrow aspirate (BMA) domains while model weights, preprocessing, sign-confidence thresholds, test support, and FOV-level aggregation were held constant. Adaptive hybrid ROI selection reduced domain-gap MAE by 29.39% relative to center-top1 while weighted pooled MAE remained nearly unchanged under WBC-dominated support. Raw microscopy images and the approximately 198 GB Phase 3 tensor cache are not included; manifests, indices, frozen predictions, checkpoints, and calibration metadata are provided.