An Efficient Deep Learning Framework using CapsNet and SOM for Multidrug-Resistant Tuberculosis Detection and Analysis in CXR Images

Early detection of Multidrug-resistant Tuberculosis (MDR-TB) is crucial for illness prevention, lowering mortality risk, and preventing the spread of infection to people. The chest X-ray images(CXRI), which is affordable and widely available data can be used for MDR-TB disease screening in clinics. However, radiologists must manually review CXRI, which places a significant strain on them and increases inter-observer variability. Therefore, it is difficult for researchers to suggest a reliable and affordable CAD method for detecting MDR-TB. In this study, Capsule Neural Networks and Self Organizing Maps, simple and effective deep learning networks were introduced to perform detection and analysis tasks respectively. CapsNet is used to reliably identify MDR-TB from the patient’s CXR images. SOM clusters and visualizes the data at each layering stage. This work proposes an innovative approach of fusing CapsNet and SOM to produce higher accuracy. Comparative analysis between three classification methods under CapsNet (Dynamic routing, Transformation matrix, Convolutional capsules) is also performed. ConvCaps outperformed than the other two methods with the accuracy of 89.47%. The model is assessed by few parameters such as Precision, Recall, Accuracy, and F1-score for each method. The accuracy of the proposed fused CapsNet and SOM model is 93.6%.