Multiple Deep Learning Approaches to Detect Pneumonia with Chest X-Ray

Abstract

Pneumonia is one of the top 10 causes of death worldwide and the leading cause of death in young children. Chest X-ray radiographs are often examined by highly qualified experts to look for pneumonia. Radiologists commonly disagree because of this laborious process. Computer-aided diagnostics systems have shown the ability to improve diagnosis accuracy. In this paper, we have given a computational technique for finding pneumonia regions using single-shot detectors, squeeze, and extinction deep convolution neural network (CNN) augmentations, and multi-task learning. One of the challenge\'s best results was obtained by a modified CNN model with the recommended procedure and got an accuracy of 96%, which was evaluated as part of the radiological society of North America\'s pneumonia detection challenge.

Introduction

I. INTRODUCTION

According to UNICEF's data, a child dies of pneumonia every 43 seconds in the world. More children than any other infectious disease pass away from pneumonia each year, killing about 700,000 kids under the age of five, or over 2,000 people each day. More than 200,000 infants are included too. Pneumonia affects more than 1,400 children worldwide per 100,000, or 1 kid in 71, with the highest incidence rates seen in South Asia (2,500 cases per 100,000 children) and Africa's West and Central region (1,620 cases per 100,000 kids). Pneumonia is an interstitial lung disease bacteria, fungi, or viruses. It killed about 880,000 kids in 2016, making up about 16% of the 5.6 million fatalities of children under five. Each year, it results in more than 50,000 fatalities and one million hospital admissions. In South Asia and Sub-Saharan Africa, pneumonia affects children most often. Even Western countries are suffering from the problem of pneumonia too. In the US, around 1.5 million people (about the population of West Virginia) receive a pneumonia diagnosis each year. Pneumonia caused the deaths of 131 450 persons in the European Union (EU) in 2016. 3% of all fatalities in 2016 were caused by pneumonia. The number of deaths among the elderly escalated, even though the mortality rate for those over the age of 70 marginally decreased. This age group accounted for 1.13 million pneumonia-related deaths. The prognosis for pneumonia in the elderly is poor, and they are more likely to develop severe pneumonia. Up to 20% of those who have severe pneumonia die [1]. The study by Kitamoto et al. found that 83% of pneumonia-related deaths were caused by organic problems with the respiratory system and alveoli, 48% of deaths were caused by respiratory failure brought on by the progression of pneumonia, and 35% were caused by prolonged respiratory failure despite the suppression of the development of pneumonic lesions [2].

II. LITERATURE SURVEY

Rajpurkar et al. [11] proposed CheXNet based on DenseNet and Grad-CAM. Four experienced radiologists evaluated CheXNet's performance using the F1 score method, which produced an F1 score of 0.435, which was higher than the radiologist average of 0.387. Despite low contrast surrounding structures on computed tomography, Roth et al.'s [12] presentation of a deep convolutional neural networks (CNN) capability to recognize lymph nodes in a clinical diagnostic test generated impressive results. CNN was developed by Ronneburger et al. [13] via data augmentation. The model captured outstanding accuracy even when trained on small samples of picture data using transmitted light microscopy, according to the scientists. Histogram equalization and thresholding were employed by Sharma and Raju [14] to find the pneumonia haze in chest X-rays. We must pre-process the image and occasionally crop it by a specific percentage to figure out that the lungs are the biggest form. SymText, an NLP-based technique for predicting pneumonia from a description of an X-Ray picture, was introduced by Fiszman [15]. To make this work, however, the system needs access to image interpretation. A sample of the 1.4 million images in the collection was taken in 2017 by Hsu et al. using 5000 images from the CXR-14 dataset [16]. They employed the ResNet50 model with information to train their model. An image analysis probabilistic neural network (PNN) model was introduced by Wozniak et al. [17]. To find abnormalities in chest x-rays, a tiny network model is used.

For deep feature detection, Ho and Gwak [18] employed the DenseNet121 pre-trained deep learning model with four features. They classified 14 distinct chest disorders using local binary patterns (LBP), scale-invariant feature transform (SFT), histogram generated gradient (HoG), and GIST with CNN features. Gue et al.'s 3 D CNN and multi-scale forecasting technique with cube clustering detection and multiscale model for cube prediction as well as lung area detection and segmentation was proposed in Gue et al.'s [19] study.

III. METHODOLOGY

A. Dataset Description

A chest X-ray database was used to experiment with this study of Pneumonia discovery. This database contains around 4000 viral pneumonia images and around 10,000 healthy images. As a result, the dataset includes studies of Pneumonia and healthy individualities with a matrix resolution of 299 × 299. EnsNet can automatically remove all the text or comment from an image without any former knowledge. Data addition and image enhancement ways are performed to enhance the volume and variety of images given to the classifier for classification Image augmentations used to consist of vertical flip, spin, width shift, and peak shift on all the extracted records from the unique dataset. A perpendicular flip was not applied because chest X-ray images are not vertically leveled.

D. Modified CNN Model Architecture

The CNN that has been suggested is not just compact in size but also computationally effective, showing to improved performance on both large and small data sets.

First it applies a standard normalization to comfort zero mean and unit variance, and it performs a scaling by a term and shifting by another term. This result is again passed to the activation function, which provides the input for the coming convolutional layer. These terms are learned during network training via backpropagation to find the best values for the given task.

Pooling layers follow convolutional layers to reduce the number of parameters and range of the point maps. This is achieved so that the algorithm could learn quickly and become more productive by down-sample features it has detected. Overall, the main concept behind pooling layers is that they encapsulate the crucial points of the feature chart, by referring some kind of pooling to it. In our case, max pooling was applied, with both the strides and filter size equalling the output of the convolutional layers is data that represents the features learned from the layer. This data is stored in a matrix which at this point gets leveled out so neurons could have full connections to all activations in the former layer. This allows the network to learn the combinations of these features (in a non-linear way) to make the most stylish prediction.

The suggested network model's structure:

 Trainable params: 9,014,595

Non-trainable params: 0

1. Result Analysis

During the evaluation procedure, predictions are framed for the testing data, and the conclusions are collected in the confusion matrix, which reflects the number of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). These are stated below:

True-positive (TP): refers to instances of pneumonia that have been correctly classified.

False-positive (FP): refers to instances of pneumonia that were erroneously classed as healthy.

True-negative (TN): refers to instances that are correctly classified as healthy.

False-negative (FN): refers to instances of pneumonia that were erroneously categorized as healthy.

Using all elements we can classify model accuracy, precision, recall, F1 score, negative predicted values, and specificity.

Conclusion

The chest X-ray images (Pneumonia, healthy) were applied to dissect lung problems. This study was done to understand the specific strengths and weaknesses of deep-learning models to identify Pneumonia with high accuracy. This is critical for a doctor’s decision-making since each has benefits and downsides. Similarly, a physician can be obliged to select only one modality when time, resources, and the patient\'s health are constrained. In this work, deep learning techniques were used for automatic Pneumonia detection from chest X-ray images. For this, four different models were implemented. Among them, three models are existing CNN models while the last one is modified CNN, a novel approach suggested in the study for more accurate classification with the least compilation time. The classification accuracy of the modified CNN model is 96% in 1 h 45 min 25 s. The precision, recall, sensitivity, and F1 score of the model are 95%, 94%, 96% and 96%, respectively. This is the best accuracy among these four models. This will be more efficient for doctors to detect disease in a short span of time. In future work, the proposed method could be implemented on a dataset with more classes of diseases such as asthma, and lung cancer, and with more images on the dataset.

References

[1] Mandell, L. A. (2007). Infectious diseases society of America; American thoracic society. Infectious diseases society of America/American thoracic society consensus guidelines on the management of community-acquired pneumonia in adults. Clin Infect Dis, 44(2), S27-S72. [2] Kiyokawa, K., & Kawai, S. (2010). Efficacy of steroid therapy on adult patients with severe pneumonia. Journal of infection and chemotherapy, 16(4), 266-271. [3] McIntosh, K. (2002). Community-acquired pneumonia in children. New England Journal of Medicine, 346(6), 429-437. [4] Mccracken Jr, G. H. (2000). Diagnosis and management of pneumonia in children. The Pediatric infectious disease journal, 19(9), 924-928. [5] Yamaya, M., Yanai, M., Ohrui, T., Arai, H., & Sasaki, H. (2001). Interventions to prevent pneumonia among older adults. Journal of the American Geriatrics Society, 49(1), 85-90. [6] Tilve, A., Nayak, S., Vernekar, S., Turi, D., Shetgaonkar, P. R., & Aswale, S. (2020, February). Pneumonia detection using deep learning approaches. In 2020 international conference on emerging trends in information technology and engineering (ic-ETITE) (pp. 1-8). IEEE. [7] Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., ... & Ng, A. Y. (2017). Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225. [8] Ra?i?, L., Popovi?, T., & Šandi, S. (2021, February). Pneumonia detection using deep learning based on convolutional neural network. In 2021 25th International Conference on Information Technology (IT) (pp. 1-4). IEEE. [9] Hashmi, M. F., Katiyar, S., Keskar, A. G., Bokde, N. D., & Geem, Z. W. (2020). Efficient pneumonia detection in chest X-ray images using deep transfer learning. Diagnostics, 10(6), 417. [10] Sirish Kaushik, V., Nayyar, A., Kataria, G., & Jain, R. (2020). Pneumonia detection using convolutional neural networks (CNNs). In Proceedings of First International Conference on Computing, Communications, and Cyber-Security (IC4S 2019) (pp. 471-483). Springer Singapore. [11] Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., ... & Ng, A. Y. (2017). Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225. [12] Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18 (pp. 234-241). Springer International Publishing. [13] Jamaludin, A., Kadir, T., & Zisserman, A. (2016). SpineNet: automatically pinpointing classification evidence in spinal MRIs. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19 (pp. 166-175). Springer International Publishing. [14] Sharma, A., Raju, D., & Ranjan, S. (2017, November). Detection of pneumonia clouds in chest X-ray using image processing approach. In 2017 Nirma University International Conference on Engineering (NUiCONE) (pp. 1-4). IEEE. [15] Fiszman, M., Chapman, W. W., Aronsky, D., Evans, R. S., & Haug, P. J. (2000). Automatic detection of acute bacterial pneumonia from chest X-ray reports. Journal of the American Medical Informatics Association, 7(6), 593-604. [16] Hsu, J., Khosla, K., & Lu, P. (2007). Predicting Thorax Diseases with NIH Chest X-Rays. [17] Wo?niak, M., Po?ap, D., Capizzi, G., Sciuto, G. L., Ko?mider, L., & Frankiewicz, K. (2018). Small lung nodules detection based on local variance analysis and probabilistic neural network. Computer methods and programs in biomedicine, 161, 173-180. [18] Khanh Ho, T. K., & Gwak, J. (2019). Multiple feature integration for classification of thoracic disease in chest radiography. Applied Sciences, 9(19), 4130. [19] Gu, Y., Lu, X., Yang, L., Zhang, B., Yu, D., Zhao, Y., ... & Zhou, T. (2018). Automatic lung nodule detection using a 3D deep convolutional neural network combined with a multi-scale prediction strategy in chest CTs. Computers in biology and medicine, 103, 220-231.

Copyright

Copyright © 2023 Dhvani Ramdevputra, Dhruv Ramdevputra, Abhishek Shekhada, Ansh Shah. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.