New preprint: a deep learning model that can read 96-well broth micro dilution plates

The CRyPTIC project used bespoke 96-well broth microdilution plates to measure the minimum inhibitory concentrations (MICs) of 13 different antibiotics; to reduce the error in the measurements, photographs of each plate were taken after two weeks incubation and stored. Hence we have available over 20,000 images of M. tuberculosis growing on these plates along with MICs measured using a variety of approaches.

In this preprint, a group of enterprising students at RMIT University in Ho Chi Minh, Vietnam, retrained a deep learning image classification algorithm (YOLOv8) on a subset of this dataset and achieved very good performance on a hold-out validation dataset. The call this model the TB Microbial Analysis System, TMAS for short.

This is another step towards developing machine learning models that can be used to support the reading of such plates in low resource settings where genetics is currently too complex and expensive for tuberculosis drug susceptibility testing (DST). Even then, we will never be able to stop phenotypic DST since the catalogues of mutations a genetics-based microbiology rests on will need continuous validation and improvement.