Morphological characterization of Mycobacterium tuberculosis in a MODS culture for an automatic diagnostics through pattern recognition

• Published in: PloS one 2013-12, Vol.8 (12), p.e82809-e82809
• Main Authors: Alva, Alicia, Aquino, Fredy, Gilman, Robert H, Olivares, Carlos, Requena, David, Gutiérrez, Andrés H, Caviedes, Luz, Coronel, Jorge, Larson, Sandra, Sheen, Patricia, Moore, David A J, Zimic, Mirko
• Format: Article
• Language: English
• Subjects: Algorithms; Analysis; Antitubercular Agents - pharmacology; Automation; Bacteriological Techniques - instrumentation; Bacteriological Techniques - methods; Cell culture; Control; Decision theory; Developing countries; Diagnostic systems; Digital imaging; Disease; Disease susceptibility; Drug resistance; Feature extraction; Feature recognition; Health aspects; Humans; Image classification; Image processing; Medical diagnosis; Methods; Microbial Sensitivity Tests - instrumentation; Microbial Sensitivity Tests - methods; Microscopy; Microscopy - methods; Morphology; Mycobacterium tuberculosis; Mycobacterium tuberculosis - cytology; Mycobacterium tuberculosis - drug effects; Mycobacterium tuberculosis - growth & development; Neural networks; Object recognition; Pattern recognition; Pattern Recognition, Automated - methods; Sensitivity; Sensitivity and Specificity; Sputum; Telephony; Tuberculosis; Tuberculosis - diagnosis; Tuberculosis - microbiology; Visual observation
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0082809

Tuberculosis control efforts are hampered by a mismatch in diagnostic technology: modern optimal diagnostic tests are least available in poor areas where they are needed most. Lack of adequate early diagnostics and MDR detection is a critical problem in control efforts. The Microscopic Observation Drug Susceptibility (MODS) assay uses visual recognition of cording patterns from Mycobacterium tuberculosis (MTB) to diagnose tuberculosis infection and drug susceptibility directly from a sputum sample in 7-10 days with a low cost. An important limitation that laboratories in the developing world face in MODS implementation is the presence of permanent technical staff with expertise in reading MODS. We developed a pattern recognition algorithm to automatically interpret MODS results from digital images. The algorithm using image processing, feature extraction and pattern recognition determined geometrical and illumination features used in an object-model and a photo-model to classify TB-positive images. 765 MODS digital photos were processed. The single-object model identified MTB (96.9% sensitivity and 96.3% specificity) and was able to discriminate non-tuberculous mycobacteria with a high specificity (97.1% M. avium, 99.1% M. chelonae, and 93.8% M. kansasii). The photo model identified TB-positive samples with 99.1% sensitivity and 99.7% specificity. This algorithm is a valuable tool that will enable automatic remote diagnosis using Internet or cellphone telephony. The use of this algorithm and its further implementation in a telediagnostics platform will contribute to both faster TB detection and MDR TB determination leading to an earlier initiation of appropriate treatment.