Quality-Guided Subset Selection for Bacterial Microscopy Classification: A Compact Multi-Seed Benchmark on DIBaS

This preprint presents a reproducible bacterial microscopy classification benchmark built on the Digital Images of Bacterial Species (DIBaS) dataset using a ResNet18 backbone, explicit image-quality review, and multi-seed evaluation. The prepared dataset contains 629 images across 30 classes. The study compares a baseline training pipeline against a quality-guided subset-selection arm derived from blur, contrast, and brightness cues. Across three seeds, the baseline achieved 0.9719±0.0122 accuracy and 0.9701±0.0127 macro-F1, while the quality-guided arm achieved 0.9298±0.0161 accuracy and 0.9275±0.0147 macro-F1. The main conclusion is methodological: strict threshold-based quality control proved too restrictive in this compact setting, so the final subset-selection mechanism relied on ranked class-preserving rescue. This record contains the manuscript PDF for citation and public access.