Non-destructive label-free automated identification of bacterial colonies at the species level directly on agar media using digital holography and convolutional neural network algorithms

Objectivesthis study aimed to develop a fully automated, non-destructive and label-free identification method of bacterial colonies, directly on agar plates, using a combination of digital holography and artificial intelligence and to evaluate its performances. Methodhigh-resolution holographic images of individual colonies on translucent brain-heart agar plates were taken every 30 minutes throughout an 18-hour incubation period (530 MPx for the full plate) using a large field 1x magnification system, a partially coherent LED light source and a high-resolution CMOS sensor. A database containing 49 490 digital holograms of individual colonies from 276 clinical strains belonging to ten of the most prevalent pathogenic bacterial species was used to train the convolutional neural network (CNN). Improvement in the accuracy of the prediction from the CNN algorithms was achieved using the information at different phylogenetic levels. Resultsthe performance of the BAIO-DX solution was assessed on 232 strains belonging to the 10 species used to train the algorithms but also on 64 strains from 8 species not included in the training database. For the species included in the training dataset, this new method identified 86.6% of the strains at the species level with a positive-percent agreement of 96.5%. An additional 48% of the strains not identified at the species level could be identified at the genus level thanks to the phylogenetic interpretation of the results. Conclusionsthese first results validate this approach as a candidate to obtain a fully automated non-destructive and label-free solution for bacterial identification in clinical microbiology laboratories. IMPORTANCE STATEMENTIdentification of pathogenic bacteria by culture-based methods are typically performed using MALDI-TOF mass spectrometry or biochemical systems. While automation and interpretive algorithms based on agar plate imaging and artificial intelligence (AI) has reduced manual steps, bacterial identification is still labor-intensive. Here we developed a fully automated, non-destructive and label-free identification method of bacterial colonies at the species level, directly on agar plates, using a combination of digital holography and convolutional neural network algorithms. After training the system with 276 strains belonging to ten of the most frequent pathogenic bacterial species, the BAIO-DX solution was able to identify 86.6% of new strains from these 10 species with a positive-percent agreement of 96.5%. These thorough proof of concept shows that imaging methods coupled to AI algorithms are promising to reach a fully automated identification of a significant proportion of pathogenic bacteria and has potential to enhance diagnostic workflows in clinical microbiology.