| Mycobacterium tuberculosis,also known as the Tubercle Bacillus,is the cause of Pulmonary Tuberculosis,which kills 1.8 million people worldwide.It is mainly damage or affect the lungs,but it can also affect other body parts.Tuberculosis is one of the top 10 deadly diseases.If it is not diagnosed at the beginning,it can be life-threatening.Chest X-ray imaging is the most common and effective method of diagnosing Pulmonary Tuberculosis.Pulmonary Tuberculosis is a more complicated disease than other lung diseases.It shows a variety of manifestations on chest X-rays,such as consolidation,effusion,fibrosis,infiltration,mass,nodule,and pleural thickening.As a result of these differences in pathology,it becomes more difficult for doctors to detect pulmonary tuberculosis,affecting the accuracy of their judgments.Additionally,without professional training and long-term experience,it can be difficult to distinguish between lung abnormalities and soft tissues with similar textures.The lack of adequate funding and inadequate medical infrastructure in resource-poor and marginalized areas leads to a limited number of radiologists who receive professional training and lowquality chest X-rays,which affects the detection time of Pulmonary Tuberculosis and the quality of diagnostic evaluations.Furthermore,fatigue resulting from the large volume of workload while reading chest X-rays makes it even more difficult for human experts to complete the task efficiently and with a high level of accuracy.The automatic detection of Pulmonary Tuberculosis from chest X-rays can aid radiologists in reducing workloads and improving diagnoses quality.In the past decade,automated medical image analysis has dramatically changed,largely due to neural networks success in classification,segmentation,and quantification tasks.Convolutional neural networks achieve superhuman performance in many of these tasks.Chest X-ray images are being analyzed with convolutional neural networks,but the high spatial resolution,the lack of large datasets with reliable ground truth,and the wide variety of diseases are significant challenges associated with developing a clinical application for these networks.Furthermore,these challenges are the driving force behind the contributions made in this thesis.A public dataset of 7000 chest X-ray images(3500 tuberculosis-infected and 3500 normal)is used in this study to identify Tuberculosis from chest X-ray images using image preprocessing and deep learning approaches.To extract deep features from chest X-ray images,we applied three pre-trained networks(ResNet101,VGG19,and DenseNet201).The Tuberculosis and normal cases are classified using the eXtreme Gradient Boosting(XGBoost)model.The proposed model operates in a sequential manner in which it first assesses the resolution of input images.The 224 x 224 is found to be the optimal resolution for the proposed approach.Secondly,various image preprocessing techniques are compared to find the best preprocessing technique at the optimal resolution.The third step is to evaluate the techniques for identifying an optimal layer for deep features extraction of the deep pre-trained convolutional neural networks.The last Convolutional blocks are found to be the best layer for extracting deep features.The deep features are extracted after the last convolutional block.The fourth step is finding the best classification method to classify the chest X-ray images with high sensitivities and specificities.The XGBoost classifier is used to improve the performance further to overcome the problem of high False Positive and False Negative rates.To tune the model hyperparameters,we use the Bayesian Optimization technique with HYPEROPT.The combined hybrid methods with hyperparameter adjustment are then evaluated.The highest performance(accuracy 99.92±0.14%,sensitivity 100±0.10%,specificity 99.85±0.20%,precision 99.85±0.20%,F1-score 99.92±0.14%and Area Under Curve of 99.93 ±0.13%,)obtained with DenseNet201-XGBoost model,for classification of tuberculosis chest X-ray images,comparing to the ResNet101-XGBoost and the VGG19XGBoost methods.This novel technique provides hope to medical institutions and radiologists in developing countries to solve the early detection problem of tuberculosis. |
