smear-microscopy-autofocus-benchmark

This record provides version 1.0.0 of the software and reproducibility outputs supporting the study “Weakly Supervised Cross-Dataset Benchmarking and Interpretable Symbolic Fusion of Focus Measures for Smear Microscopy Autofocus.”The archive implements a weakly supervised cross-dataset autofocus benchmark across five smear-microscopy datasets: WBC, TBI, PBS, BMA, and TBF. It includes 32 implemented handcrafted focus measures, ten autofocus evaluation metrics, leave-one-dataset-out genetic programming with ten seeds per held-out dataset, a ten-seed final all-dataset refit, and evaluation of fourteen retained composite focus measures.Included materials comprise the complete Python source code, experiment configuration, source/surrogate/leave-one-out label arrays, raw and normalized focus curves, timing measurements, single-measure results, 50 LODO GP seed runs, 10 final-refit runs, composite evaluations, statistical analyses, publication tables and figures, and reviewer-response sensitivity computations.Raw microscopy images and the approximately 46 GB reconstructed stack cache are not redistributed because their access, licensing, privacy, and storage requirements must be handled separately. Authorized users can rebuild the stack cache using the supplied pipeline and dataset configuration instructions.Surrogate labels are consensus labels and must not be interpreted as optical ground truth. The internal “Curvelet Transform Sharpness Index” name maps to Wavelet Detail Energy (db1). The optional downstream analysis is a focus-quality proxy and not diagnostic validation.Production GP experiments used a CPU backend. The archived environment and staged reproduction instructions are included.