Détection automatique de la tuberculose à partir de sons de toux

Abstract

Tuberculosis remains one of the leading global health threats, with over ten million new cases reported annually according to WHO. Despite their effectiveness, current diagnostic methods (molecular assays, cultures, imaging) are still limited by cost, infrastructure requirements, and turnaround time, which restricts their widespread use in low-resource settings. This thesis introduces an innovative automated tuberculosis detection system based on cough sounds, leveraging advanced signal processing and deep learning. Using the dataset provided by the CODA TB DREAM Challenge 2022, we designed a complete pipeline comprising audio preprocessing, feature extraction (MFCC, Mel spectrograms, and global statistics), data augmentation, hybrid deep neural architectures CNN–BiGRU–Attention, and hyperparameter optimization with Optuna. To ensure robust and generalizable results, strict subject-level cross-validation was applied throughout the experiments. The findings demonstrate the potential of cough sound analysis as a complementary, non-invasive, and cost-effective approach for early tuberculosis screening, with significant implications for reducing community transmission and improving global health outcomes.

Date	2 Mar 2026
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	David Labbé (Supervisor) & Neila Mezghani (Co-supervisor)