- Title
- Black lung detection on chest X-ray radiographs using deep learning
- Creator
- Devnath, Liton
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2021
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- Coal Worker Pneumoconiosis (CWP), commonly known as Black Lung (BL), is an incurable respiratory disease caused by long-term inhalation of respirable dust, such as coal, asbestos, and silica. If parenchymal and pleural abnormalities in the lung indicate the presence of BL, stages of the disease are classified as either simple BL or complex BL, depending on the size of the opacities, localized/diffuse pleural thickening and costophrenic angle obliteration. As an initial diagnostic tool for a variety of clinical conditions, chest X-ray radiography (CXR) is the most commonly used radiological examination. Sometimes it is difficult even for experts to diagnose BL from CXRs, and disagreement on diagnosis among medical doctors happens often. Education about the diagnosis for inexperienced doctors is based on the experience of experts. For these reasons, it would be desirable to develop CAD (Computer-Aided Diagnosis) systems to provide an automatic quantitative evaluation of BL, and also to serve as an initial screening process, as well as a second opinion, for medical doctors. In this thesis, deep learning (DL) based methods were used for CWP detection in CXR images. Because of the small BL incidence and restrictions on the sharing of patient data, the number of available CWP X-rays is insufficient, which introduces significant challenges for training deep convolutional neural networks (CNNs) models. A training image set of real, segmented chest X-ray images, with and without BL, was used as a benchmark for this study. To detect BL in CXRs, we have pursued six research objectives (sections 1.2.1-1.2.6), including with and without transfer learning, comparison transfer learning using seven DL models, ensemble of DL models predictions, an application of the leave-one-out cross-validation (LOOCV) method using DL models, multilevel CNN features learning with Support Vector Machine (SVM) classifier, and high-level features learning based on the ensemble learning of multiple classifiers. Finally, Grad-CAM was applied to produce a coarse localized map highlighting the most important regions of interest (ROIs) in the black lung radiographs. The first research objective started with CNNs, or in other words, using the DL method without applying transfer learning. This study applied two CNN architectures, one smaller and the other larger, to non-transferable learning for BL detection. In the next objective, we iterated the same, but with the transfer learning of seven pre-formed DL models, VGG16, VGG19, InceptionV3, Xception, ResNet50, DenseNet121, and CheXNet. In both objectives, the size of our original dataset was 153 images, of which 71 BL diseases were used as experiment 1. We also added synthetic radiographs from the original dataset using two data generators, Cycle-Consistent Adversarial Networks (CycleGAN) and Keras, in experiments 2 and 3. In order to cross-validate all methods, the training and testing images were randomly split into three folds before each experiment. Based on the results of these two objectives, it was found that the method with transfer learning performed better, where the CheXNet achieved the highest accuracy (about 85.37%) on experiment 1. The VGG16 and VGG19 were excluded from the next objectives, as they exhibited the lowest overall accuracy in all experimental datasets. Therefore, it was decided to extend the analysis further using the remaining five DL methods with transfer learning on the original dataset (experiment 1). In the third objective, three ensemble learning strategies, i.e., majority voting, weights averaging and simple averaging of predictions probabilities, were investigated using five selected DL models. It was initially expected that these approaches would provide better performance with regards to accuracy. However, the performance was unsatisfactory in the simple averaging strategy, although the detection accuracy of BL improved from 85.37% to 87.80% in the multi-weighted ensemble with CheXNet. In the fourth objective, in order to further investigate the individual performance of these five models in the detection of BL disease, the LOOCV method was implemented on the original dataset. Using the LOOCV method, we were able to find a robust DL model for all of our data. It was found that the performance was improved for the CheXNet model (achieved about 90.20%). That led us to select CheXNet for analysis in the following fifth objective. Afterwards, multilevel deep features from the CheXNet architecture were extracted to automatically detect CWP in CXRs. The deep features were mapped to three different kernel patterns of Support Vector Machine (SVM) classifiers and CNN-based aggregation methods. The experimental results showed that the DL and SVM with high-level feature sets outperformed when the proposed framework reached 92.68%, 87.80% and 85.37% accuracy in the three-fold of original datasets, respectively. The performance of the high-level features of CheXNet led us to the ensemble learning in the final objective. In the final objective, nine machine learning classifiers’ performances on the same high-level CheXNet features set only were evaluated. The efficiency of each classifier and their ensemble was compared on the basis of the AUC-PR (area under the precision-recall curve) and AP (average-precision) values. The proposed learning outperformed other objectives of this research with an accuracy of 92.68%, 90.24%, and 87.80% in three-fold datasets, when the best AP values reached 0.9490, 0.9234, and 0.9005, respectively. Additionally, Grad-CAM was applied to produce an approximate localized map, highlighting that the most important ROIs were created by averaging multiple CAMs generated from four convolutional blocks of the CheXNet model. The accuracy of the average-IOU (intersection over union) between the predicted and true ROIs assesses the positivity of the BL compared to the AP values. These experiments on BL detection have shown that the proposed system has the potential to be used in clinical practice.
- Subject
- pneumoconiosis; coal worker pneumoconiosis; VGG19; inceptionV3; Xception; ResNet50; DenseNet121; CheXNet; cycle-consistent adversarial networks; grad-CAM; LOOCV; ensemble learning; black lung; chest x-ray radiography; computer-aided diagnosis; deep learning algorithm; convolutional neural networks; transfer learning; machine learning algorithm; VGG16
- Identifier
- http://hdl.handle.net/1959.13/1479344
- Identifier
- uon:50291
- Rights
- Copyright 2021 Liton Devnath
- Language
- eng
- Full Text
- Hits: 686
- Visitors: 842
- Downloads: 271
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | ATTACHMENT01 | Thesis | 4 MB | Adobe Acrobat PDF | View Details Download | ||
| View Details Download | ATTACHMENT02 | Abstract | 484 KB | Adobe Acrobat PDF | View Details Download |