HS-GC-IMS couples with convolutional neural network for Burkholderia gladioli pv. Cocovenenans detection in Auricularia Auricula.

The shortage in early detection methods for the pathogen Burkholderia gladioli pv. cocovenenans (BGC) and its toxin bongkrekic acid rises the risk for food poisoning. Combining Headspace-Gas Chromatography-Ion Mobility Spectrometry (HS-GC-IMS) with convolutional neural network, we established a HS-GC-IMS-VGGNet architecture for the detection of BGC in edible fungus Auricularia auricula (AA) contaminated by six microorganisms species and the overall accuracy was 93.8 %. Meanwhile, the network achieved a limit of detection (LOD) of 80 CFU/mL and limit of quantification (LOQ) of 241 CFU/mL for BGC biomass. An LOD of 0.25 mg/L for bongkrekic acid detection covering the range of 0-1.54 mg/kg was achieved directly in the AA matrix. Furthermore, 28 microbial organic volatiles were extracted by Gradient-weighted Class Activation Mapping (Grad-CAM) and identified as conducive to the BKA detection. In all, the detection system established for BGC and its bongkrekic acid toxin is of good accuracy, precision and possesses greener mode.