Deep Learning Models for Radiography in Chest Disease: History
View Latest Version
Please note this is an old version of this entry, which may differ significantly from the current revision.
Contributor:
Adnane Ait Nasser
,
Moulay A. Akhloufi

Chest X-ray radiography (CXR) is among the most frequently used medical imaging modalities. It has a preeminent value in the detection of multiple life-threatening diseases. Radiologists can visually inspect CXR images for the presence of diseases. Most thoracic diseases have very similar patterns, which makes diagnosis prone to human error and leads to misdiagnosis. Machine learning (ML) and deep learning (DL) provided techniques to make this task more efficient and faster. Numerous experiments in the diagnosis of various diseases proved the potential of these techniques.

  • radiography
  • chest X-ray
  • computer-aided detection
  • machine learning
  • deep learning
  • deep convolutional neural networks

1. Introduction

Chest X-ray radiography (CXR) imaging is a fast and cost-effective technique widely used by radiologists to diagnose multiple parts of the human body such as heart, lungs, bones, blood vessels, and airways [1]. It plays a major role in detecting diseases and abnormalities. CXR images are typically generated by projecting X-ray radiation through the body positioned against the metallic plate of the X-ray machine. The organs appear differently on the CXR image because of the amount of radiation absorbed by each organ. The organs that absorb more radiation (e.g., bones) appear in white color, while the parts that absorb less radiation (e.g., heart) appear in different shades of gray. The airways and the organs containing air (e.g., lungs) appear in a black color [2]. CXR examinations are affordable, non-invasive and painless. They are considered as a valuable tool for the detection of many diseases and abnormalities, which helps in diagnosing diseases and monitoring therapy [3].
In order to diagnose the patients, radiologists inspect visually the CXR images. This process is time and resource intensive, especially in areas where there is a shortage of qualified clinicians. The lower resolution of CXR images, the similarities between the signs of diseases, and the lack of experience and focus while examining a CXR image can make the interpretation a challenging task for radiologists as it can lead to potentially life-threatening diagnostic errors. Therefore, computer-aided detection systems (CAD), including computer vision, machine learning (ML) and deep learning (DL) algorithms, were proposed to provide a good decision-making tool for radiologists to diagnose different diseases [4][5][6].
For nearly a decade, ML techniques became more popular for medical imaging-based anomaly detection and classification, especially with the release of several datasets. These techniques were applied for various purposes in medical image analysis such as organs segmentation, diseases detection and classification. They showed high performance through numerous studies developed to classify several diseases such as tuberculosis (TB), pneumonia, edema, cardiomegaly and COVID-19.
While ML models require users and data scientists to select features from the input data, DL models perform automatic features extraction. ML algorithms reveal less performance using large datasets, while DL algorithms perform better with the availability of large quantities of data and higher computational power [7]. Therefore, researchers are focusing more on DL techniques to increase the performance of medical applications and decrease the time and cost of the diagnostic process.

2. Datasets

In the medical area, there are several types of image screening technologies, including ultrasound imaging, CT (computed tomography), MRI (magnetic resonance imaging), and X-ray imaging. Radiologists use these images to diagnose organs for the detection of abnormalities [8]. Detecting diseases from CXR images is always a difficult task for radiologists and sometimes leads to misdiagnoses. To address this purpose using computer-aided detection (CAD) systems, a large amount of data is required for training and testing. CAD systems in medical analysis are usually trained and tested on an ensemble of data called a dataset, that are generally composed of images and other important information called metadata (e.g., age of patient, race, sex, Insurance type). Some hospitals, universities and laboratories in different countries used several approaches to collect data that belong to patients [9]. Datasets collection in medical area aims to advance research in detecting diseases. DL techniques proved their efficiency and ability to detect most dangerous diseases using different datasets [10][11]. These techniques achieved expert-level performance on clinical tasks in many studies [12][13].

3. Image Preprocessing Techniques

Preprocessing of X-ray images is the operation that consists of improving their quality by converting them from their original form into a much more usable and informative form. Most of CXR images are produced in DICOM (Digital Imaging and Communications in Medicine) format with large set of metadata, which makes it challenging to understand by experts outside the field of radiology [14]. In other areas such as computer vision, DICOM images are usually stored in PNG or JPG formats using specific algorithms. These algorithms allow the compression without losing important information in the images. This process has two main steps, first is to de-identify the information of patients (privacy protection). Second is to convert DICOM images into PNG, JPEG, or other formats. Normal X-ray images have dimensions of 3000 × 2000 pixels, which requires high computational resources if used in their original size. Therefore, radiological images must be resized without losing the essential information they contain. Most of the datasets have resized images, such as Indiana dataset, which has CXR images resized to 512 × 512 pixels [15] and ChestX-ray dataset that has resized images with a dimension of 1024 × 1024 pixels [16].
Datasets are most of the time imbalanced or contain low-quality images, which usually contain noise and unwanted parts. In the process of developing a CAD system, the image preprocessing techniques play a crucial rule in enhancing and improving the quality of images. They help to remove the irrelevant data, to extract the meaningful information, and to make the ROI clearer. These techniques improve the performance of CAD systems and reduce their error rate. Preprocessing techniques applied on CXR images, consist of several methods including augmentation, enhancement, segmentation, and bone suppression.

3.1. Augmentation

Training a Deep Convolutional Neural Network (DCNN) on an imbalanced dataset mostly leads to overfitting, makes the model unable to generalize to novel samples and does not provide the desired results. To cope with this situation, many transformations can be employed by position-based augmentation (cropping, rotating, scaling, flipping, padding, elastic deformations) and color-based augmentation techniques (hue, brightness, contrast) to increase the number of samples in the dataset by making slight adjustments to existing images.

3.2. Enhancement

Image enhancement techniques are generally used to improve the information interpretability in images. For CXR images, these techniques are used to provide a better image quality to human readers (radiologists) as well as to automated systems [17]. To improve the quality of a CXR image, multiple parameters can be considered (contrast, brightness, noise suppression, edge of features, and sharpness of edges) using different methods including histogram equalization (HE) [18], high and low pass filtering [19], and unsharp masking [20]. Figure 1 depicts an example of enhancement applied to a CXR image.
/media/item_content/202301/63c748c3060a3diagnostics-13-00159-g005.png
Figure 1. (a) Noisy CXR image from a low quality version of CheXpert dataset [21]; (b) Enhanced CXR image.

3.3. Segmentation

Image Segmentation has a critical rule in image preprocessing techniques. It is usually necessary to divide a visual image into fragments. For CXR images, this technique allows segmenting the thoracic image into areas in order to extract the ROI. Figure 2 depicts examples of ROIs overlaid on CXR images.
/media/item_content/202301/63c748e53e7aediagnostics-13-00159-g006.png
Figure 2. (a) Examples of CXR images from CheXpert dataset [21]; (b) Samples of ROIs overlaid on CXR images.

3.4. Bone Suppression

Bone suppression is a technique that can be applied on CXR images. It is an important step in the process of lung segmentation and extraction of features from thoracic images. Bone suppression technique is based on removing the bones from the CXR images, as depicted in Figure 3. It helps to increase the visibility of obscure zones and to prevent the overlap of signs of diseases with ribs and clavicles. 
/media/item_content/202301/63c7490a8cd95diagnostics-13-00159-g007.png
Figure 3. (a) CXR image from CheXpert dataset [21] before bone suppression; (b) CXR image after bone suppression.

3.5. Evaluation Metrics

Several metrics for evaluating the performance of CAD systems are available. The commonly used metrics in medical imaging analysis are: Accuracy (ACC), Precision (PRE), F1-score, Sensitivity (SEN), Specificity (SPE), Area under curve (AUC), Dice index, and Jaccard index.

4. Deep Learning for Chest Disease Detection Using CXR Images

Several CAD systems were developed to detect chest diseases using different techniques. Early diagnosis of thoracic conditions gives a chance to overcome the disease. Diseases such as TB, pneumonia, and COVID-19 become more serious and severe when they are at an advanced stage. In CXR images, three main types of abnormalities can be observed: (1) Texture abnormalities, which are distinguished by changes diffusing in the appearance and structure of the area. (2) Focal abnormalities, that occur as isolated density changes and (3) Abnormal form where the anomaly changes the outline of the normal morphology.

4.1. Pneumonia Detection

Ma and Lv [22] proposed a Swin transformer model for features extraction with a fully connected layer for classification of pneumonia in CXR images. The performance of the model was evaluated against DCNN models using two different datasets (Pediatric-CXR and ChestX-ray8). Image enhancement and data-augmentation techniques were applied, which improved the performance of the introduced model, achieving an ACC of 97.20% on Pediatric-CXR and 87.30% on ChestX-ray8. Singh et al. [23] proposed an attention mechanism-based DCNN model for the classification of CXR images into two classes (normal or pneumonia). ResNet50 with attention achieved the best results with an ACC of 95.73% using images from Pediatric-CXR dataset. Darapaneni et al. [24] implemented two DCNN model with transfer learning (ResNet-50 and Inception-V4) for a binary classification of pneumonia cases using CXR images from RSNA-Pneumonia-CXR dataset. The best performing model was Inception-V4 with a validation ACC of 94.00% overcoming ResNet-50 which achieved a validation ACC of 90.00%. Rajpurkar et al. [25] developed CheXNet model, which is composed of 121-layer convolutional network to detect and localize the lung areas that show the presence of pneumonia. ChestX-ray14 was used to train the model, which was fine-tuned by replacing the final fully connected layer with one that has a single output. A nonlinear sigmoid function was used as an activation function, and the weights were initialized with the weights from ImageNet. CheXNet model showed high performance, achieving an AUC of 76.80%.

4.2. Pulmonary Nodule Detection

According to the WHO, lung cancer is one of the most dangerous diseases. It is the most frequent cancer in men and the third in women [26]. Lung cancer manifests as lung nodules. Early diagnosis of these nodules is extremely effective in treating lung cancer before it becomes dangerous.
To demonstrate the usefulness of DL-based systems to assist radiologists in detecting pulmonary nodules on medical images, several studies were carried out. DL algorithms proved high performances on different modalities of medical screening and specifically on X-ray radiographs. For instance, to compare the performance of radiologists for the detection of malignant pulmonary nodules with and without assistance of a DL based CNN system, Sim et al. [27] proved in a multi-centre study that the performance of 12 radiologists was improved with more than 5% in terms of SEN after being assisted by a DL system. The performance of radiologists increased from 65.10% to 70.30% when they used the proposed DCNN software. Cha et al. [12] trained a model based on a ResNet architecture for detecting operable lung cancer. The schloars used a dataset composed of 17,211 CXR images, augmented to 600,000 using different techniques such as cropping, resize and rotation. The model achieved a SEN of 76.80% and an AUC of 73.20% outperforming six radiologists in the detection of lung cancer.

4.3. Tuberculosis Detection

According to the WHO, TB is ranked on the top 10 diseases leading to death. TB is ranking as the second infectious disease leading to death after COVID-19 and above HIV/AIDS. In 2020, around 10 million people suffered from TB (1.1 million children). It killed a total of 1.4 million people in 2019 and 1.5 million people in 2020. TB is caused by the bacillus mycobacterium TB, which spreads when people who are sick with TB expel bacteria into the air (by coughing or sneezing). The disease typically affects the lungs [28].
An early diagnosis of TB saved an estimation of 66 million lives between 2000 and 2020. The variety of manifestations of pulmonary TB on CXR images makes the detection a challenging task. DL proved its high efficiency in the detection and the classification of TB. Ahmed et al. [29] proposed an approach to overcome TB detection problem using an efficient DL network named TBXNet. TBXNet is implemented using five dual convolution blocks with different filter sizes of 32, 64, 128, 256 and 512, respectively. The dual convolution blocks are merged with a pre-trained layer in the fusion layer of the architecture. Moreover, the pre-trained layer is used to transfer pre-trained knowledge into the fusion layer. The proposed TBXNet obtained an ACC of 99.17%. All experiments are performed using image data acquired from different sources (Montgomery, a labeled dataset created by different institutes under the ministry of health of the Republic of Belarus and a labeled dataset, that was acquired by the kaggle public available repository).

4.4. COVID-19 Detection

Malathy et al. [30] presented a DL model called CovMnet to classify CXR images into normal and COVID-19. The layers in CovMnet include a convolution layer along with a ReLU activation function and a MaxPooling layer. The output of the last convolutional layer in the architecture is flattened and fit to fully connected neurons of four dense layers, activation layer and Dropout. Experiments are carried out for deep feature extraction, fine-tuning of CNNs (convolutional neural networks) hyperparameters, and end-to-end training of four variants of the proposed CovMnet model. The introduced CovMnet achieved a high ACC of 97.40%. All experiments were performed using CXR images from Pediatric-CXR dataset. 

4.5. Multiple Disease Detection

In some cases, a patient may suffer from more than one disease at the same time, which can put his life at higher risk. It may be difficult for radiologists to detect more than a pathology using CXR images due to the similarities between the signs of diseases. In such a situation, more details and more exams may be needed. To deal with this challenge, several DL systems were carried out using different algorithms. For instance, Majdi et al. [31] proposed a fine-tuned DenseNet-121 to classify CXR images into pulmonary nodules and cardiomegaly diseases. Images from CheXpert dataset were used for the experiment. The model obtained an AUC of 73.00% for pulmonary nodule detection and 92.00% for cardiomegaly detection. Bar et al. [32] employed a DL technique for the detection of pleural effusion, cardiomegaly, and normal versus abnormal disease by using a combination of features extracted by the DCNN model and the low-level features. Preprocessing techniques were applied on the used dataset that contains 93 CXR images collected from Sheba Medical Center. They attained an AUC of 93.00% for pleural effusion, 89.00% for cardiomegaly, and 79.00% for normal versus abnormal cases. Cicero et al. [33] used GoogleNet model to classify frontal chest radiograph images into normal, consolidation, cardiomegaly, pulmonary edema, pneumothorax, and pleural effusion. GoogleNet achieved an AUC score of 86.80% for edema, 96.20% for plural effusion, 86.10% for pneumothorax, 96.40% for normal, 87.50% for cardiomegaly and 85.00% for consolidation. The study proved that the DCNN model can achieve high performance even if trained with modest-sized medical dataset. Wang et al. [16] used a weak-supervised method for the classification and detection of eight chest diseases presented on ChestX-ray8 dataset.

This entry is adapted from the peer-reviewed paper 10.3390/diagnostics13010159

References

  1. Abiyev, R.; Ma’aitah, M.K.S. Deep Convolutional Neural Networks for Chest Diseases Detection. J. Healthc. Eng. 2018, 2018, 4168538.
  2. Radiological Society of North America. X-ray Radiography-Chest. Available online: https://www.radiologyinfo.org/en/info.cfm?pg=chestrad (accessed on 1 November 2022).
  3. US Food and Drugs Administration. Medical X-ray Imaging. Available online: https://www.fda.gov/radiation-emitting-products/medical-imaging/medical-x-ray-imaging (accessed on 1 November 2022).
  4. Avni, U.; Greenspan, H.; Konen, E.; Sharon, M.; Goldberger, J. X-ray Categorization and Retrieval on the Organ and Pathology Level, Using Patch-Based Visual Words. IEEE Trans. Med. Imaging 2011, 30, 733–746.
  5. Jaeger, S.; Karargyris, A.; Candemir, S.; Folio, L.; Siegelman, J.; Callaghan, F.; Xue, Z.; Palaniappan, K.; Singh, R.; Antani, S.; et al. Automatic Tuberculosis Screening Using Chest Radiographs. IEEE Trans. Med. Imaging 2014, 33, 233–245.
  6. Pattrapisetwong, P.; Chiracharit, W. Automatic lung segmentation in chest radiographs using shadow filter and multilevel thresholding. In Proceedings of the International Computer Science and Engineering Conference (ICSEC), Chiang Mai, Thailand, 14–17 December 2016; pp. 1–6.
  7. Piccialli, F.; Di Somma, V.; Giampaolo, F.; Cuomo, S.; Fortino, G. A survey on deep learning in medicine: Why, how and when? Inf. Fusion 2021, 66, 111–137.
  8. Elangovan, A.; Jeyaseelan, T. Medical imaging modalities: A survey. In Proceedings of the International Conference on Emerging Trends in Engineering, Technology and Science (ICETETS), Pudukkottai, India, 24–26 February 2016; pp. 1–4.
  9. Saczynski, J.; McManus, D.; Goldberg, R. Commonly Used Data-collection Approaches in Clinical Research. Am. J. Med. 2013, 126, 946–950.
  10. Horng, S.; Liao, R.; Wang, X.; Dalal, S.; Golland, P.; Berkowitz, S. Deep learning to quantify pulmonary edema in chest radiographs. Radiol. Artif. Intell. 2021, 3, e190228.
  11. Tolkachev, A.; Sirazitdinov, I.; Kholiavchenko, M.; Mustafaev, T.; Ibragimov, B. Deep learning for diagnosis and segmentation of pneumothorax: The results on the kaggle competition and validation against radiologists. J. Biomed. Health Inform. 2021, 25, 1660–1672.
  12. Cha, M.J.; Chung, M.J.; Lee, J.H.; Lee, K.S. Performance of deep learning model in detecting operable lung cancer with chest radiographs. J. Thorac. Imaging 2019, 34, 86–91.
  13. Schultheiss, M.; Schober, S.; Lodde, M.; Bodden, J.; Aichele, J.; Müller-Leisse, C.; Renger, B.; Pfeiffer, F.; Pfeiffer, D. A robust convolutional neural network for lung nodule detection in the presence of foreign bodies. Sci. Rep. 2020, 10, 12987.
  14. Johnson, A.; Pollard, T.; Berkowitz, S.; Greenbaum, N.; Lungren, M.; Deng, C.y.; Mark, R.; Horng, S. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Sci. Data 2019, 6, 317.
  15. Demner-Fushman, D.; Kohli, M.; Rosenman, M.; Shooshan, S.; Rodriguez, L.; Antani, S.; Thoma, G.; McDonald, C. Preparing a collection of radiology examinations for distribution and retrieval. Am. Med. Inform. Assoc. 2016, 23, 304–310.
  16. Wang, X.; Peng, Y.; Lu, L.; Lu, Z.; Bagheri, M.; Summers, R.M. ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017; pp. 2097–2106.
  17. SK, S.; Naveen, N. Algorithm for pre-processing chest-x-ray using multi-level enhancement operation. In Proceedings of the International Conference on Wireless Communications, Signal Processing and Networking (WiSPNET), Chennai, India, 23–25 March 2016; pp. 2182–2186.
  18. Reza, A. Realization of the contrast limited adaptive histogram equalization (CLAHE) for real-time image enhancement. J. VLSI Signal Process. Syst. Signal Image Video Technol. 2004, 38, 35–44.
  19. Agaian, S.; Panetta, K.; Grigoryan, A. Transform-based image enhancement algorithms with performance measure. IEEE Trans. Image Process. 2001, 10, 367–382.
  20. Chen, S.; Cai, Y. Enhancement of chest radiograph in emergency intensive care unit by means of reverse anisotropic diffusion-based unsharp masking model. Diagnostics 2019, 9, 45.
  21. Stanford ML Group. ChexPert a Large Chest X-ray Dataset and Competition. Available online: https://stanfordmlgroup.github.io/competitions/chexpert/ (accessed on 1 November 2022).
  22. Ma, Y.; Lv, W. Identification of Pneumonia in Chest X-ray Image Based on Transformer. Int. J. Antennas Propag. 2022, 2022, 5072666.
  23. Singh, S.; Rawat, S.; Gupta, M.; Tripathi, B.; Alanzi, F.; Majumdar, A.; Khuwuthyakorn, P.; Thinnukool, O. Deep Attention Network for Pneumonia Detection Using Chest X-ray Images. Comput. Mater. Contin. 2022, 74, 1673–1691.
  24. Darapaneni, N.; Ranjan, A.; Bright, D.; Trivedi, D.; Kumar, K.; Kumar, V.; Paduri, A.R. Pneumonia Detection in Chest X-rays using Neural Networks. arXiv 2022, arXiv:2204.03618.
  25. Rajpurkar, P.; Irvin, J.; Zhu, K.; Yang, B.; Mehta, H.; Duan, T.; Ding, D.; Bagul, A.; Langlotz, C.; Shpanskaya, K. Chexnet: Radiologist-level pneumonia detection on chest X-rays with deep learning. arXiv 2017, arXiv:1711.05225.
  26. World Health Organization. World Cancer Report. Available online: https://www.who.int/cancer/publications/WRC_2014/en/ (accessed on 1 November 2022).
  27. Sim, Y.; Chung, M.J.; Kotter, E.; Yune, S.; Kim, M.; Do, S.; Han, K.; Kim, H.; Yang, S.; Lee, D.J.; et al. Deep Convolutional Neural Network–based Software Improves Radiologist Detection of Malignant Lung Nodules on Chest Radiographs. Radiology 2020, 294, 199–209.
  28. World Health Organization. Tuberculosis. Available online: https://www.who.int/news-room/fact-sheets/detail/tuberculosis (accessed on 1 November 2022).
  29. Iqbal, A.; Usman, M.; Ahmed, Z. An efficient deep learning-based framework for tuberculosis detection using chest X-ray images. Tuberculosis 2022, 136, 102234.
  30. Jawahar, M.; Anbarasi, J.; Ravi, V.; Jayachandran, P.; Jasmine, G.; Manikandan, R.; Sekaran, R.; Kannan, S. CovMnet-Deep Learning Model for classifying Coronavirus (COVID-19). Health Technol. 2022, 12, 1009–1024.
  31. Majdi, M.; Salman, K.; Morris, M.; Merchant, N.; Rodriguez, J. Deep learning classification of chest X-ray images. In Proceedings of the Southwest Symposium on Image Analysis and Interpretation (SSIAI), Albuquerque, NM, USA, 29–31 March 2020; pp. 116–119.
  32. Bar, Y.; Diamant, I.; Wolf, L.; Greenspan, H. Deep learning with non-medical training used for chest pathology identification. In Medical Imaging 2015: Computer-Aided Diagnosis; SPIE: Bellingham, WA, USA, 2015; pp. 215–221.
  33. Cicero, M.; Bilbily, A.; Colak, E.; Dowdell, T.; Gray, B.; Perampaladas, K.; Barfett, J. Training and validating a deep convolutional neural network for computer-aided detection and classification of abnormalities on frontal chest radiographs. Investig. Radiol. 2017, 52, 281–287.
More
© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license; additional terms may apply. By using this site, you agree to the Terms and Conditions and Privacy Policy.
This entry is offline, you can click here to edit this entry!
Video Production Service
Feedback