BiFusionPathoNet: fusion network for drug-resistant bacteria identification via optical scattering patterns

Abstract

The presented research introduces a new method to identify drug-resistant bacteria rapidly with high accuracy using artificial intelligence combined with Multi-angle Dynamic Light Scattering (MDLS) signals and Raman scattering signals. The main research focus is to distinguish methicillin-resistant Staphylococcus aureus (MRSA) and methicillin-sensitive Staphylococcus aureus (MSSA). First, a microfluidic platform was developed embedded with optical fibers to acquire the MDLS signals of bacteria and Raman scattering signals obtained by using a Raman spectrometer. After that, for the detection of both scattering signals of MRSA and MSSA, three models were developed: (1) ResistNet, a hybrid model combining a Transformer Encoder with ResNet, with an accuracy of 83.8% on the MDLS dataset.; (2) SERB-CNN, which attained 91.84% accuracy on a Raman scattering public dataset and 93.5% on a custom-built dataset; and (3) BiFusionPathoNet, a multimodal fusion model that reached 96.8% accuracy, significantly outperforming single-modal approaches. The acquired results demonstrated the effectiveness of this multimodal strategy for the rapid detection of drug-resistant bacteria.