Deep learning speeds up TB drug testing

Tuberculosis (TB) remains a leading cause of death from infectious disease, with drug resistance making treatment harder and timely diagnosis more important than ever. Diagnosing which antibiotics will work requires drug susceptibility testing (DST), but many laboratories still rely on slow, manual methods that are difficult to scale. One widely used approach exposes bacteria to different drug concentrations on 96-well broth microdilution plates or 96-well microtiter plates and then reads whether bacterial growth is present to determine the Minimum Inhibitory Concentrations (MICs). Manual reading is time-consuming and prone to human error, and earlier automated tools have had limitations. For example, the Automated Mycobacterial Growth Detection Algorithm (AMyGDA) uses image processing to read plates but struggles with plates that show little growth or with low-quality images. To address these problems, Hai Thi Vo and collaborators developed a new framework called TMAS (TB Microbial Analysis System). TMAS applies recent advances in deep learning to identify growth of M. tuberculosis in plate images, aiming to speed up results while helping distinguish true bacterial growth from misleading visual artefacts.

TMAS is built around state-of-the-art deep learning models trained to detect M. tuberculosis growth on plate images and to produce MICs rapidly and accurately. The team trained their models and refined the TMAS framework using 4,018 plate images drawn from the CRyPTIC (Comprehensive Resistance Prediction for Tuberculosis: An International Consortium) dataset. A key goal for TMAS was not just to see growth, but to tell true growth apart from common artefacts that can confuse readers: shadows, bubbles, sediment, condensation and contamination. In testing on the CRyPTIC images, TMAS achieved an essential agreement of 98.8% when compared to reference readings. Compared with previous automated approaches such as AMyGDA, which can falter on low-growth plates or poor-quality images, TMAS showed improved accuracy, reliability and efficiency. The system is presented as an automated and complementary evaluation to support expert interpretation rather than a wholesale replacement of laboratory expertise.

The development of TMAS has several practical implications for TB diagnostics. By automating the visual reading of 96-well plates and providing rapid MIC estimates, TMAS can reduce the time technicians spend on repetitive image interpretation and decrease the chance of human error. Its ability to distinguish true bacterial growth from artefacts such as shadows or condensation addresses a common obstacle in both manual and earlier automated readings. Because the framework was trained on a large, real-world dataset (the CRyPTIC images) and reached high essential agreement, TMAS could serve as a reliable second opinion to support laboratory staff, improving consistency across readings. The authors suggest that such automation could make high-quality DST more efficient and accessible, especially in settings where expert reviewers are scarce, potentially helping clinicians get the drug-resistance information they need more quickly to guide patient treatment.