Introduction
Pneumonia is a symptom of inflammation of the lungs, and can be caused by bacteria, viruses, or fungi. It accounts for more than 15% of deaths in children under 5-year [
1,
2]. Especially in developing and underdeveloped countries, pneumonia are more likely to occur due to severe environmental pollution, unsanitary living conditions and inadequate medical infrastructure [
3]. And pneumonia could be extremely dangerous and life-threatening if not be detected in the early stages [
4]. Therefore, early diagnosis and interventional management are extremely important for patients with pneumonia. It can prevent the disease from becoming fatal.
And in clinical practice for screening pneumonia, chest X-ray imaging is the most commonly used method for diagnosing pneumonia. Because it is fast, low-invasive, low-cost, and simple to implement, X-ray imaging has become the standard method of screening pneumonia [
5,
6]. However, the chest X-ray examinations for pneumonia screening is challenging. Radiologists with different experience may miss and misdiagnose due to subjective variability [
7,
8]. Therefore, there is an urgent need for an accurate and automated computer-aided diagnosis for pneumonia detection.
Deep learning is a new technology used in image recognition, natural language processing, speech recognition and other fields. In recent years, deep learning algorithm, especially convolutional neural networks (CNNs) were used in medical image analysis. And achieved remarkable performance in different tasks, such as, medical image classification, lesion segmentation, lesion detection, etc. Specifically, deep learning has been used in brain tumor segmentation [
9], breast cancer diagnosis [
10], lung nodule detection [
11‐
13], abdominal disease diagnosis [
14,
15], and bone disease diagnosis and measurement [
16,
17]. Moreover, there has been some studies on pneumonia detection [
18‐
22]. Many studies are used to diagnose COVID-19 pneumonia, and some studies only classify different pneumonias, and cannot use limited information to complete the localization of pneumonia areas.
Therefore, in this study, we developed a DL model for automatic and accurate pneumonia detection using weak supervision information. The model can not only detect pneumonia cases from normal cases, but also localize areas of pneumonia. Furthermore, we evaluated model performance using independently public dataset and performed well.
Discussion
In this study, we proposed a deep learning weakly supervised algorithm for pneumonia detection. The model we developed can not only distinguish between normal cases and pneumonia cases, but also localize pneumonia regions using image-level information. The model achieved a high AUC of 0.9949, with a accuracy of 98.29%, sensitivity of 99.29% and specificity of 95.87%. And in the cohort 2, we also achieved competitive result with an AUC of 0.9835.
There are already studies on the diagnosis of pneumonia. Rohit Kundu et al. used an ensemble of deep learning methods on the two pneumonia datasets [
1]. Many deep learning models are combined for pneumonia detection, so the research method is more complicated and requires higher computing speed and capacity. In our study, we simplified the deep learning algorithm and achieved state-of-the-art results on external datasets. Compared with this study [
18], they conducted a differential diagnosis study of different pneumonias, including COVID-19, normal, Viral and Bacterial pneumonia. And many studies have designed different deep learning algorithms for the diagnosis of pneumonia, but using machine learning (ML) methods has several limitations, including complexity, overfitting and poor performance while training with small dataset size [
18,
19,
21,
22,
28‐
32]. Some studies in above studies, the models were not independently validated with external dataset, and some models cannot locate the pneumonia area. Therefore, we proposed a simple deep learning algorithm for pneumonia detection and used image-level weakly supervised information to localize the pneumonia region, and achieved remarkable performance.
Compared with some others studies [
33‐
35], the network framework used in this study is lighter and more concise, more convenient to apply in clinical practice, and the model performance is better, which can assist doctors in making clinical decisions. Specifically, we trained the model using a public database and conducted independent external testing using another public database. The results of external tests illustrate the good robustness and generalization ability of our model. In order to better reproduce the most advanced methods, we found methods with open source code for comparison. We fully trained the model using the same data and compared it on an external test set. We have reproduced the methods of these three studies, and the three models have achieved 87.82% [
36], 82.33% [
37] and 85.79% [
38] accuracy respectively in the external test data, and lower than our study. Our model’s excellent performance illustrates the potential of this study’s model to assist clinical decision-making.
From Table
3, we can see that the identification accuracy of the model for bacterial pneumonia is higher than that for viral pneumonia in the test dataset. The reason for this situation may be due to the fact that there are more bacterial pneumonia cases than viral pneumonia cases in the training data. Since there are more bacterial pneumonia cases than viral pneumonia cases, the deep learning model can better learn the features of bacterial pneumonia, resulting in better identification of bacterial pneumonia cases by the model. The same situation happens with external validation dataset. In the external validation dataset, the identification accuracy of the model for viral pneumonia is higher than that for COVID-19 pneumonia. Patients diagnosed with COVID-19 present symptoms similar to pneumonia. And, some of the findings frequently encountered in COVID-19 pneumonia are: ground glass opacities (GGO), consolidation, crazy paving and enlargement of subsegmental vessels (diameter greater than 3 mm) in areas of GGO [
39‐
43]. It is not completely consistent with other pneumonia manifestations [
44]. In the training set, because the model was not trained with COVID-19 data and could not learn the features of COVID-19, so the model had a low identification accuracy for COVID-19. Although the accuracy of the model for the identification of COVID-19 pneumonia is relatively low, it also achieves an accuracy of 83.54%, which has good clinical value for preliminary screening of COVID-19 pneumonia.
Besides, the CAM experiment was conducted to test whether pneumonia regions could be accurately distinguished from other normal regions. After testing, we found that the CAM are quite useful to precisely and accurately locate pneumonia regions in provided X-ray images. Since deep learning is a black box, it cannot be well explained. However, we found that the diagnosis of this study’s model is not incomprehensible and is trustworthy. The model pays more attention to the lesion area than normal area in the pneumonia X-ray images. It also indicates that the model can effectively distinguish the pneumonia area from the none-pneumonia area. It proves that our model is based on the pneumonia region to complete the diagnosis work. At the same time, we can also use image-level information to complete the detection of pneumonia lesions and locate it.
Our study had some limitations. First, the model was only trained and test on public datasets, and no further validation was performed on data from actual hospital institutions. Second, there is no bacterial pneumonia in the external validation dataset, so the model cannot be further validated for the diagnostic performance of bacterial pneumonia. For the above two limitations, in future study, we will collect data from our hospital to further validate the model’s performance. Third, we did not perform quantitative analysis. Only some images were randomly selected to show whether the model detected pneumonia is accurate. Therefore, in future research, we will compare whether the pneumonia area marked by the doctor is consistent with the pneumonia area identified by the model. Fourth, the method we proposed is to use the threshold set by the result of the CAM, which can obtain the specific location of pneumonia. However, this threshold is fixed, so there may be some deviations in locating pneumonia regions in some images. In our subsequent studies, we plan to make this threshold a learnable parameter for the model to increase the accuracy of localization.
In conclusion, we proposed a deep learning algorithm can accurately detect pneumonia and locate the pneumonia area based on weak supervision information, which can provide potential value for helping radiologists to improve their accuracy of detection pneumonia patients through X-ray images.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.