1. Introduction

Pleural effusion is an increase in fluid between the two membranes that cover the lungs the chest wall, indicating an imbalance between pleural fluid formation and removal. The amount of fluid is regularly spread thinly over the visceral and parietal pleura that acts as a lubricant between the two membranes. Pleural effusion is a significant increase in the amount of pleural fluid. Determining the cause is important for treating to treat pleural effusion suitably. Some of the most ordinary symptoms of pleural effusion are chest pain and pleurisy. Pleural effusion has many causes, such as heart and kidney failure, low albumin levels in the blood, infections, pulmonary embolism, and malignancies. Many cases show no symptoms but are discovered during a physical examination or are detected by Chest X-ray (CXR) imaging, which is the most advantageous method for making a diagnosis.

Coronavirus disease (COVID-19) is an infectious lung disease that has serious consequences for human health globally. According to the WHO ^[1], more than 37 million coronavirus infections have been detected worldwide, with approximately one million deaths. This highlights the stage of emergency caused by the coronavirus and the impact on the body of the infected patients. The lungs can be damaged directly after a coronavirus infection. Radiologists can diagnose pleural effusion based on the CXR images. Pleural effusion is one of the most prominent manifestations of COVID-19. The Guardian reports that once the lungs become infected, the small air sacs can fill with cells and fluid, blocking oxygen flow.

The symptoms of pleural effusion manifest as shortness of breath while lying down. The degree of dyspnea depends on the amount of fluid in the pleural space. A rapid increase in the amount of pleural fluid causes acute difficulty in breathing. The clinical examination reveals syndromes, such as decreased vocal fibrillation, decreased alveolar fibrillation, and percussion. The manifestation of the disease will differ according to the cause of the disease. Pleural effusion has many causes and can be benign, but there are many cases of malignant pleural effusion that can cause death. The typical causes of hydrocephalus include pleural effusion, pleurisy, cancer, previous systemic history of heart failure, malnutrition, cirrhosis, and other chronic diseases.

Pleural effusion can often be diagnosed based on the typical symptoms in a physical examination and the patient's history. The doctor listens carefully to the lungs using a stethoscope and measures the pulse and blood pressure. An x-ray of the lungs can also be taken. An X-ray can identify the position and detect the quantity of pleural effusion because they normally occur as whitish areas at the lung base and appear unilaterally or bilaterally. If a patient lies on their side for a few minutes, most pleural effusions will move and layer out along that side of the chest space, which was positioned downward because of the effects of gravity. This movement of the pleural effusion can be seen on a CXR image taken with the patient lying on their side (a lateral decubitus X-ray).

Currently, radiologists perform diagnostic imaging in hundreds of cases. The diagnosis process takes significant time, but errors may also increase, which affects the diagnosis quality. In addition, some diseases, such as pleural effusion, pleural effusion, and lung cancer, have similar symptoms. Therefore, it is easy to confuse and difficult to diagnose accurately using traditional methods. As a result, there might be many more false negatives that may lead to a deterioration of the patient's health condition, possibly even death. Therefore, the deep learning model helps doctors make more accurate decisions quickly in the pleural effusion diagnosis process, which can allow the development of the earliest treatment plan for the patient and reduce the risk of death.

Deep learning is a promising approach in the medicine domain. Considerable research has been carried out to examine chest and lung diseases using machine learning. Deep learning innovations have brought promising improvements in medical technology, such as discovering antibiotics, analyzing electronic health records, or predicting medical events. Several deep learning approaches in medical imaging have been proposed for disease detection and diagnosis. On the other hand, this technology does not mean the ultimate replacement of physicians, particularly radiologists. Instead, it helps radiologists diagnose patients more accurately. Because belief is fundamental, people can decide to act on a prediction or whether to implement a new model. Providing more model details is important to turn an unreliable model or prediction into a reliable one. Although machine learning is at the core of many recent advances in science and technology, the important human role is an overlooked aspect of this field. Whether people are directly using machine learning classifiers as tools or deploying models into the products, an important concern remains if they believe in a model or prediction or whether they will use it with greater confidence.

The performance of intelligent systems depends on the size of the training data. One of the greatest drawbacks is that deep learning models are not generalizing well, and the amount of public data for model training could be limited for some diseases. Recently, several approaches have been made to solve this problem. A previous study ^[2] proposed the deep k-nearest neighbor (kNN) ^[3]. The research improved the learning strategy to unify the kNN classification and feature extraction on multiple small-class and class-imbalanced medical image datasets. Therefore, this method has solved the non-parametric characteristics of kNN while associated with a feature extractor. In addition, most medical images lack annotated data public using the supervised methods in deep learning models. Pooch et al. ^[4] examined some novel semi-supervised classification methods using the ChestX-ray14 dataset. They were based on pseudo-labeling and consistency regularization to produce multi-label classification and use a modern model architecture in CXR image classification. In recent years, deep learning techniques have emerged as state-of-the-art methods to classify medical images. Tuning the hyperparameters of the networks trained and using deep learning libraries is difficult. Inés et al. ^[5] investigated an Automated Machine Learning (AutoML) approach to solve these issues. Their approach associated transfer learning with a new semi-supervised learning method when several annotated images were available.

This paper proposes a Convolutional Neural Network architecture that compares with some pre-trained models to assist in the control and early detection of pleural effusion from CXR images. Moreover, applying the model explanation algorithm to assist in medical diagnosis is a promising method. Because deep learning models still function as a black box, interpreting the generated output is a challenge.

· A CNN model that can distinguish between normal samples and pleural effusion patients efficiently was proposed.

· This study compared the pleural effusion classifi-cation performance of various scenarios, such as using the architecture with only a Fully Connected Layer (FC) and the changes in the quantity of convolutional layers that were explored from one to three convolutional layers with 32 to 64 filters for each convolutional layer.

· The use of two popular activate functions, such as Sigmoid and Softmax, was investigated to improve accuracy.

· Local Interpretable Model-Agnostic Explanations (LIME) ^[6] for visual explanations and interpretation purposes in the pleural effusion diagnosing process were examined.

The remainder of this paper is as follows. Section 2 presents related work. Section 3 reports methods to solve the problem, including to support pleural effusion with a combination of CNNs and algorithms to provide explanations from the output of the CNNs. Section 4 presents the experimental results of the proposed method. The conclusions are reported in Section 5.

2. Related Work

One of the most challenges for doctors is to make an exact decision and reduce the time to detect diseases. Therefore, the use of deep learning algorithms in the analysis and processing of biomedical images has provided more satisfactory results. Advances in deep learning and big data have enabled artificial intelligence to replace humans gradually. Several approaches have been introduced to support the detection of pleural effusion using CXR images. Some of these methods use feature extraction techniques along with a machine learning algorithm as a classification technique, whereas others use deep learning techniques for feature extraction and classification.

Maduskar et al. ^[7] presented pre-trained models, namely DenseNet121 and MobileNetV2, using the CheXpert ^[8] to support the successful diagnosis of pleural effusion. In tuberculous pleural effusion (TPE) diagnostic models, a previous study ^[9] proposed four machine learning algorithms (MLAs), such as logistic regression, k-nearest neighbors (KNN) ^[3], support vector machine (SVM) ^[10], and random forest (RF) ^[11]. The random forest reached the best result with a sensitivity and specificity of 89.1% and 93.6%, respectively. TPE using RF was produced improved the results, lowered the computation cost, and resulted in a diagnosis process. Shu et al. ^[12] constructed a logistic regression model for the diagnosis of TPE with a sensitivity and specificity of 82.9% and 86.7%, respectively. CheXNeXt is a complex neural network that detects 14 different diseases simultaneously, including pleural effusion ^[13]. Cicero et al. ^[14] examined the diagnosis of pleural effusion using two popular CNN architectures that have shown excellent results on the ImageNet dataset: AlexNet ^[15] and GoogLeNet ^[16]. Bustomi et al. ^[17] combined image processing analysis with a Na\"{i}ve Bayes Classifier (NBC) to overcome the subjective problems when radiologists analyzed the medical image. The five image features were used as predictors to determine the lung class, i.e., normal lungs class, pleural effusion class, and lung cancer class. The NBC method showed an accuracy of 70%.

Sunny et al. ^[18] used Chexpert datasets to compare the performance of VGG16 and InceptionV3 to predict pleural effusion. The best model was the InceptionV3 multi-class model, in which the additional labels were provided to help detect pleural effusion using CXR images. Behzadi-khormouji et al. ^[19] proposed two types of deep convolutional neural networks to detect lung abnormalities through CXR images with certain accuracy. In the first type, the state-of-the-art Deep Convolutional Neural Networks (DCNNs), such as VGG16 ^[20], DenseNet121 ^[21], and pyramid convolutional structure ^[22], were applied to the Pediatric Chest X-ray dataset. The VGG16 and DenseNet121 were pre-trained on the ImageNet dataset, including 1.2 million color images ^[23]. In the second type, an efficient architecture design of deep convolutional neural networks called ChestNet was proposed. This model was constructed based on the context in the datasets under study.

3. The Proposed Scheme

This study examined the CheXpert dataset ^[24] (the smaller data set size was 11GB) to propose an automated CXR interpretation framework (Table 1). The dataset was collected from Stanford Hospital, which was performed between October 2002 and July 2017 in inpatient and outpatient centers, along with their associated radiology reports. Many board-certified radiologists annotated the dataset manually. Several samples from the original dataset were modified to improve the validation accuracy of the study. All images with the same resolution as the original image size were downsized to 64 x 64 for the prediction task.

4. Method

The CNN model was designed from the beginning. The Keras library with TensorFlow was applied to use the models. The LIME method was used to explain the predictions. The trust of the model was measured using AUC and accuracy. The proposed CNN model with various models reported elsewhere. The research methodology is presented in Fig. 1. Three Chest X-ray images were fetched into the proposed framework. In the training phase, the Chest X-ray image data will go through pre-processing. This paper proposes the CNN model and pre-trained models, such as ResNet, InceptionResnetV2, Xception. Finally, the CNN model, which was designed from scratch, was selected, and LIME was applied to visualize the explanations that helped highlight the effusion in the lungs. The output showed the results and probability of three Chest X-ray images after diagnosis.

Fig. 1. Proposed Architecture for the Pleural Effusion Diagnosis Support System.

4.1 Convolutional Neural Network Architectures

Several shallow CNN architectures were investigated, and the efficiency of each architecture on the CXR images for Pleural Effusion classification was evaluated. The number of convolutional layers increased from 1 to 3, while the number of filters per convolutional layer was 32 and 64. Two popular activation functions consisting of Softmax and Sigmoid were examined in such a configuration.

From the experimental results, the architecture shown in Fig. 2 was used to apply model explanation in the next steps for Pleural Effusion diagnosis. The input image size was 64 $\times ~ $64 with 32 filters, and ReLU activates the function for training. Three convolution layers with up to 256 features per class were added to extract the object better and improve the result. In addition, the max-pooling layer was used after the convolutional layer with 2 $\times $ 2 dimensions. The learning rate was 0.000001 to decrease the overfitting problem. The cost was compared using the Adam optimizer with Categorical Crossentropy. The softmax function with 20 epochs was suggested. The proposed algorithm focused primarily on the classification of images, which presents two outputs, including pleural effusion and normal.

Fig. 2. Architecture of shallow Convolutional Neural Network for classifying the Pleural Effusion.

4.2 Explainable Deep Learning using LIME

Ribeiro et al. proposed Local Interpretable Model-Agnostic Explanations (LIME) ^[6] in 2016. The primary idea was that it is easier to interpret a locally approximated black-box model using a simpler glass-box model. LIME is a popular approach to enhance the interpretability and visualize the explainability of black-box Deep Learning (DL) algorithms. LIME frequently produces an explanation for a single prediction by any model by training a simpler interpretable model, e.g., linear classifier around the prediction by producing simulated data around the instance by random perturbation and obtaining the feature importance by applying some form of feature selection ^[25]. Therefore, LIME provides a patient-specific explanation for a given classification that enhances the possibility of any classifier to supply efficient usage in clinical conditions. LIME is clinically relevant with a very high concordance with explanations provided by physicians, according to the evaluation by field experts.

LIME can be explained with a local model that shows the axes of the values of two continuous explanatory variables. The colored regions correspond to the decision areas for a binary classifier, i.e., combinations of the values of the two variables. The model classifies the observation to one of the two classes. The black crosses present the observation of interest. The dots show the artificial data generated to understand the local behavior of the complex model around the point of interest. The dashed line indicates a simple linear model fitted to construct the local approximation in Fig. 3. The local behavior of the black-box model can be explained by a simple model around the observation of interest. The simpler model serves as a local explainer for the more complex model.

The features and interpretable data representations, which use an understandable representation to humans, can be separated. For example, the text classifier is a binary vector corresponding to the presence or absence of a word, or image classification has an interpretable representation that can be a binary vector that can present the presence or absence of a contiguous patch of similar pixels.

(1)

$$ \xi(x)=\operatorname{argmin}_{g \in G} L\left(f, g, \pi_{x}\right)+\Omega(g) $$

Formula 1 shows the LIME method that can produce an explanation by the following:

· $x~ \in R$ is designated as a prior representation of an explained instance.

· $g~ \in G,$where $G$ is a class of capability interpretable models as an explanation. ^[26]

· $\Omega(g)$ is a measure of the complexity of the explanation $g~ \in G$.

· $f\left(x\right)$ is the probability that $x$ belongs to a certain class.

· $\pi _{x}\left(z\right)$ is a proximity measure to identify locality around x between $z$ to $x$.

· $L\left(f,g,\pi _{x}\right)$ measure $g$ approximating $f$ in the locality defined by $\pi _{x}$. $L\left(f,g,\pi _{x}\right)$ is needed to enhance the interpretability and local faith, and $\Omega(g)$ should be low to be explainable by humans.

Fig. 4 gives an example of the explanations of an image by LIME. The original image is shown in Fig. 4(a), and 100 \textit{superpixels} are identified in Fig. 4(b). Figs. 4(c) and (d) present an explanation for the poodle class, goose class, respectively.

In medical systems, the basis for providing analytical results also needs to be provided to implement processes correctly based on AI analysis. For example, an AI system can analyze whether a patient has cancer and show why it came to such conclusions. Furthermore, it is almost impossible to design a safety model without understanding how it works. Increasing the interpretability of the AI systems helps analyze their logical vulnerabilities or data blind spots, thereby improving the security of the AI systems. Malhi et al. ^[28] proposed an explainable machine-learning tool that can be used potentially for decision support in medical image analysis cases, particularly on in-vivo gastral images obtained from capsule endoscopy. They applied LIME to visualize the explanations to help health experts trust the black box predictions. The method presented elsewhere ^[29] highlights the need for explainable AI in the medical field to help medical experts make interpretable, transparent, and easy to understand decisions that are easy to deploy, have a fast computation time, and provide short, selective explanations that suit busy physicians.

LIME is widespread across different fields, especially the medical domain. The interpretation is generated by initiating the original model (black box) closer to the explained model. The explainable model was produced based on the perturbation of the original example from which the chosen components were excluded, e.g., removing words or hiding a part of an image dividing the image into interpretable parts. The image was altered in several ways to enable or disable some of the interpretable components. For each of the generated images, the probability on it was obtained. To learn a linear model weighed locally on the dataset, it is important to understand how it conducts. The original model, in which the components with the highest weights are presented, are explained. Because the true positive predictions showed the visual explanations with the largest weights, the LIME identifies the relevant superpixels, and more true positive explanations are produced. LIME provides visual explanations that can interpret examples with fewer essential features. The selected feature in a good explanation meets the need because people prefer sparse explanations. LIME can predict the change in probability of a label based on the change in an input property, which provides a good contrast explanation.

Fig. 3. Idea behind the LIME approximation with a local glass-box model.

Fig. 4. Example of the Explanations of an Image Obtained by the VGG16 network with ImageNet weights for the half-duck and half-horse image.

5. Results

5.4 Performance Analysis

These measurements require True Positive (TP), False Positive (FP), True Negative (TN), and False Negative (FN) values. A TP presents the positive results of pleural effusion that were classified correctly by the system. TN is the amount of negative or normal data that were classified correctly by the system. FN is the amount of negative data that were classified incorrectly by the system. FP is the amount of positive data that were classified incorrectly by the system. Fig. 6 presents a confusion matrix of the TP, FP, TN, and FN measurements.

Fig. 7 presents the statistics of the training process with different metrics, such as Accuracy and AUC after 20 epochs. Both line graphs provide information on the proposed method throughout two metrics: Accuracy and AUC. The Training and Validation Accuracy showed a rapid increase in the performance until 20 epochs. On the other hand, at some epoch, test accuracy reached the saturated value. Similarly, the other chart also shows the proportion of the AUC that rises significantly at the initial stage and gains a stability value of more than 93%.

Fig. 5. Explaining the prediction of pleural effusion in positive patients.

Fig. 6. Confusion matrix for our proposed method.

Fig. 7. Experimental results of the proposed architecture in various evaluation metrics, including Accuracy, AUC.

6. Conclusion

Pleural effusion is generally diagnosed with CXR images. A CXR image contains abnormalities as consolidation resulting from pleural effusion. In this study, a method was proposed to detect imaging features of pleural effusion efficiently using convolutional neural networks. The proposed approach is expected to support doctors for pleural effusion diagnosis.

A CheXpert dataset of chest radiographs, which have uncertainty labels and radiologist-labeled reference standard evaluation sets, was evaluated. A variety of CNN architectures were examined to discriminate between normal samples and pleural effusion images.

LIME was leveraged to provide explanations from the trained model on the CXR images. This promising technology can help explain many tasks related to assessing the trust for effective human interactions with machine learning systems and obtaining more informative predictions. The proposed method will provide growth and acknowledgment of CXR interpretation models to improve healthcare in the future.