A deep learning-based framework for automatic detection of drug resistance in tuberculosis patients

Tuberculosis (TB), caused by mycobacterium tuberculosis, is one of the most severe respiratory diseases that kill thousands of people annually around the world. Late diagnosis and drug resistance adversely affect the treatment plan. This paper introduces several deep learning models for TB diagnosis from CT scans. Specifically, for the detection of multi-drug resistance and for the detection of the TB types. These models are multi-channel models as they take, as input, the image frame, the mask frame and the gender/age data to make the final classification. Transfer learning based on VGG19 and ResNet neural networks were used for feature extraction from CT scans. The best performing model, for the classification of multi-drug resistance, was a three-channel model which used VGG19 for feature extraction and a cascade of convolutional and dense layers for classification with accuracy equals to 74.13% and AUC equals to 64.2%. By comparison, the best performing model for the TB type classification, employed the ResNet for feature extraction and a cascade of convolutional and dense layers for classification with accuracy equals to 53% and Kappa index equals to 34.3%. These results outperform the reported results in the literature for the same type of tasks. The dataset was obtained from the Image CLEFF 2018 data on TB.