GMDM-MoE: A biologically-inspired growth-to-morphology and dual-magnification mixture-of-experts for bacterial detection

Abstract

Microscopic analysis of bacteria is crucial, thereby accurate and timely bacterial detection is essential, yet manual analysis is labor-intensive. Automated bacterial detection methods improve the efficiency. But they deviate from expert practice and neglect two essential aspects, i.e., the bacterial temporal growth dynamics and the complementary multi-scale features under different magnifications. As a result, they struggle with some clinical issues such as scale variability, morphological overlap with impurities, and dense clustering.

We propose GMDM-MoE: A Biologically-Inspired Growth-to-Morphology and Dual-Magnification Mixture-of-Experts, which emulates two strategies used by microbiologists. The GM-pipeline simulates recalling temporal growth history during single-frame observation: multi-frame pre-training with explicit temporal encoding captures growth dynamics that are transferred to a Morphological Characterization Phase for single-frame inference. The DM-MoE simulates switching between magnifications. The global context under the low-magnification and the detailed features under the high-magnification are simultaneously learned through the independent structures of two experts, which are the feature gate router and the specifically designed detection head.

Experiments on real bacterial datasets show that GMDM-MoE achieves state-of-the-art performance under challenging conditions, demonstrating that biologically inspired designs substantially enhance both accuracy and deployability in bacterial detection.