Deep Learning-based Framework for Mycobacterium tuberculosis Bacterial Growth Detection for Antimicrobial Susceptibility Testing

Tuberculosis (TB) kills more people annually than any other pathogen. Resistance is an ever-increasing global problem, not least because diagnostics remain challenging and access limited. 96-well broth microdilution plates offer one approach to high-throughput phenotypic testing, but they can be challenging to read. Automated Mycobacterial Growth Detection Algorithm (AMyGDA) is a software package that uses image processing techniques to read plates, but struggles with plates that exhibit low growth or images of low quality. We developed a new framework, TMAS (TB Microbial Analysis System), which leverages state-of-the-art deep learning models to detect M. tuberculosis growth in images of 96-well microtiter plates. TMAS is designed to measure Minimum Inhibitory Concentrations (MICs) rapidly and accurately while differentiating between true bacterial growth and artefacts. Using 4,018 plate images from the CRyPTIC (Comprehensive Resistance Prediction for Tuberculosis: An International Consortium) dataset to train models and refine the framework, TMAS achieved an essential agreement of 98.8%. TMAS offers a reliable, automated and complementary evaluation to support expert interpretation, potentially improving accuracy and efficiency in tuberculosis drug susceptibility testing (DST). Author summaryTuberculosis (TB) is one of the worlds leading causes of death from infectious diseases, with drug resistance making treatment increasingly difficult. Accurate and timely drug susceptibility testing (DST) is essential to determine which antibiotics remain effective against a given TB strain. A widely used DST method involves 96-well broth microdilution plates, where bacteria are exposed to different antibiotic concentrations to determine the minimum inhibitory concentration (MIC), the lowest concentration that fully prevents bacterial growth. However, manually interpreting these plates is time-consuming and prone to human error, while existing automated methods often struggle with imaging artifacts such as condensation, shadows, and contamination. In this study, we developed the Tuberculosis Microbial Analysis System (TMAS), a machine learning-based tool designed to automate MIC determination from plate images. Training on advanced deep learning models, TMAS can learn to detect bacterial growth with high precision, distinguishing true growth from artefacts such as shadows, bubbles, sediment, condensation and contamination. When tested on a large dataset, TMAS outperformed existing automated methods in accuracy, reliability and efficiency. By reducing the reliance on manual interpretation, TMAS has the potential to streamline TB diagnostics, improve efficiency in laboratories, and improve access to high-quality DST, particularly in resource-limited settings.