Medium-aware multi-task learning for robust bacterial identification using Raman spectroscopy

Raman spectroscopy has demonstrated significant potential for rapid bacterial identification; however, bacterial spectral signatures are frequently confounded by complex culture medium background signals, leading to limited generalization under varying cultivation conditions. Conventional single-task deep learning models struggle to effectively decouple environmental noise, while training paradigms that neglect environmental information inherently compromise recognition robustness. To address these challenges, we propose a Hybrid Medium-Aware Network (HMA-Net) that transforms culture medium backgrounds from interfering noise into informative contextual representations to guide precise bacterial feature extraction. Specifically, we develop a Residual Cross-Task Attention mechanism that dynamically recalibrates bacterial spectral features using environmental context, and construct a Mixture-of-Experts (MoE)-based classifier that adaptively adjusts classification logic according to medium conditions. Furthermore, a hybrid loss function combining contrastive learning and Focal Loss is introduced to enhance intra-class compactness and mitigate severe class imbalance. Experimental results on a complex mixed dataset comprising 10 bacterial strains across 4 culture media demonstrate that our method achieves 97.45% strain identification accuracy and a Macro F1-score of 0.9729. Compared to single-task baseline (1D-CNN) and standard multi-task learning (MTL) models, the accuracy improvements reach 6.39% and 2.90%, respectively. These substantial gains confirm that by synergistically learning environmental context, the proposed model effectively mitigates culture medium interference, providing a robust and practically meaningful solution for robust bacterial identification in complex environments.