Ensemble deep learning for tuberculosis detection

Abstract

Tuberculosis (TB) is one of the deadliest infectious disease in the world. TB is caused by a type of tubercle bacillus called Mycobacterium Tuberculosis. Early detection of TB is pivotal to decrease the morbidity and mortality. TB is diagnosed by using the chest x-ray and a sputum test. Challenges for radiologists are to avoid confused and misdiagnose TB and lung cancer because they mimic each other. Semi-automated TB detection using machine learning found in the literature requires identification of objects of interest. The similarity of tissues, veins and small nodules presenting the image at the initial stage may hamper the detection. In this paper, an approach to detect TB, that does not require segmentation of objects of interest, based on ensemble deep learning, is presented. Evaluation on publicly available datasets show that the proposed approach produced a model that recorded the best accuracy, sensitivity and specificity of 91.0%, 89.6% and 90.7% respectively.

Machine and Deep Learning for Tuberculosis Detection on Chest X-Rays: Systematic Literature Review

“…Various ML and DL methods have been applied in the included studies: 7/47 (15%) studies [13,14,37,40,43,63,67] focused on using ML approaches, while 34/47 (72%) studies [17][18][19][20][34][35][36]39,41,[44][45][46][47][48][49][50][51][52][53][54][55][56][58][59][60][61][62]65,66,[69][70][71][72]74] used DL approaches; 4/47 (9%) studies [38,64,68,73] used both ML and DL approaches, while 2/47 (4%) [42,57] focused on industrial-grade DL image analysis software and various deep AI models without further information on the types of AI techniques used.…”

“…The most popular DL architectures used in the included studies were ResNet-50 (n=11), followed by VGG-16 (n=8), VGG-19 (n=7), and AlexNet (n=6). However, it is noteworthy that various DL ensemble (n=9) [17,18,36,48,54,[60][61][62]64] and custom (n=9) [20,35,[45][46][47]50,53,58,59] methods were also introduced by authors in this field. For the ML approaches, SVM (n=5) was the most applied method, followed by KNN (n=3) and RF (n=2).…”

“…Twenty-five of the included studies [14,[17][18][19]34,[36][37][38][47][48][49]52,53,56,59,[62][63][64][65][66]68,70,[72][73][74] reported comparison results with several other previous studies, while the remaining 22 did not. Twenty-six of 47 included studies [19,20,35,[38][39][40]42,43,45,[47][48][49][50]53,55,56,[58][59][60][61][62]65,[69][70][71]74] were funded or sponsored by companies, private...…”

Automated Tuberculosis Classification with Chest X-Rays Using Deep Neural Networks -Case Study: Nigerian Public Health

“…M. H. A. Hijazi et al [28] also used the two publicly available datasets MC and SZ, in which they first preprocessed the CXRs images to retain only the Region of Interest (ROI). And also used two different pre-trained architectures; VGG 16 and InceptionV3, and further developed a custom CNN architecture with 15 layers to detect TB from the preprocessed CXR images, where the process took less time to complete training.…”

“…K. T. Hwa, M. H. A. Hijazi et al [29] also used the publicly available datasets MC and SZ similar to other literature, in which they first pre-processed the CXRs images to obtain edge feature using Canny Edge detector, which they believe CXR images with more unusual edges could increase the detection rate. They perform an ensemble of two pre-trained CNN architectures as in [25] and [28]. They believe sensitivity is considered more in medical image analysis; hence their work was to get a high sensitivity score.…”